Nonlinear

Finite Element Analysis

in Structural Mechanics
Proceedings of the Europe-U.S. Workshop
Ruhr-Universitiit Bochum, Germany,
July 28-31,1980

Editors
W Wunderlich, E. Stein, and K.-J. Bathe

With 272 Figures

Springer-Verlag Berlin Heidelberg GmbH

Prof. Dr.-Ing. W WUNDERLICH
Institut fUr Konstruktiven Ingenieurbau
Ruhr-Universitiit Bochum
UniversitiitsstraBe 150,4630 Bochum, Germany

Prof. Dr.-Ing. E. STElN

Fachbereich Bauingenieur- und Vermessungswesen
Universitiit Hannover
CallinstraBe 32, 3000 Hannover, Germany

Prof. K.-J. BATHE, Ph. D.

Department of Mechanical Engineering, MIT
Cambridge, MA 02139, USA

ISBN 978-3-642-81591-1 ISBN 978-3-642-81589-8 (eBook)

DOI 10.1007/978-3-642-81589-8

This work is subject to copyright. All rights are reserved, whether the whole Of part of the material is concemed,,
specifically those of translation, reprinting, re-use of iIIustrations, broadcasting, reproduction by photocopying
machine or similar means, and storage in data banks. U nder § 54 of the German Copyright Law where copies are
made forotherthan private use a fee is payable to 'Verwertungsgesellschaft Wort',Munich.
© Springer-Verlag Berlin, Heidelberg 1981
Originally published by Springer-Verlag Berlin Heidelberg New York in 1981
Softcover reprint ofthe hardcover Ist edition 1981

The use of registered names, trademarks, etc. in this publication does not imply. even in the absence of a specific
statement, that such names are exempt from the relevant protective laws and regulations and therefore free for
general use.

2061/3020 - 543210
Preface

With the rap1d development of computational capab1lities,

nonl1near f1nite element analys1s 1n structural mechan1CS has
become an 1mportant field of research. Its objective is the
real1stic assessment of the actual behaV10r of structures by
numerical methods. Th1S requires that all nonlinear effects,
such as the nonl1near character1stics of the mater1al and
large deformations be taken 1nto account. The act1vities in
th1S f1eld be1ng worldw1de, d1rect 1nteraction between the
various research groups 1S necessary to coordinate future
research and to overcome the time gap between the generat10n
of new results and the1r appearance 1n the 11terature.

The f1rst U.S.-Germany Sympos1um was held 1n 1976 at the

Massachusetts Inst1tute of Technology. Under the general to-
P1C "Formulat1ons and Computat1onal Algorithms in Fin1te Ele-
ment Analysis" 1t prov1ded an opportun1ty for about 20 re-
searchers from each country to present lectures, hold discus-
sions, and establ1sh mutual contacts. The success of th1S first
sympos1um was so encourag1ng that 1t seemed natural to organ-
1ze a second bilateral meet1ng, this time 1n Germany, and to
1nv1te researchers from other European countr1es as well.

The purpose of the Europe-U.S. Workshop held 1n July 1980

1n Bochum was to exchange new results of research 1n the f1eld
of nonl1near f1n1te element analysis w1th emphas1s on struc-
tural problems. It was felt that glv1ng the meeting a workshop
character would encourage thorough and act1ve d1Scussions and
establ1sh new or tlghten old contacts between 1ndividuals and
research organ1zat1ons. The organ1zers also hoped that the ad-
vantages of a workshop with a relatively small group of re-
searchers - representing most un1vers1t1es and organizations
work1ng in the f1eld - would outwe1gh some shortcom1ngs IY1ng
1n the lim1tat1on of the number of part1c1pants. On the
VI

European side the organizing committee consisted of W. Wun-

derl~ch of the Ruhr-Un~vers~ty, Bochum, Chairman, E. Stein
of the Technical University, Hannover and O.C. Zienk~ew~cz

of Un~versity College of Wales, Swansea. On the U.S. side the

organ~zing committee consisted of K.J. Bathe of MIT, Cambrid-
ge, Chairman, J.T. Oden of the Un~vers~ty of Texas, Austin and
S. Key of Sandia Laboratories, Albuquerque. Financial support
was provided by the Deutsche Forschungsgemeinschaft and the
U.S. National Science Foundation. The workshop was held at the
Ruhr-University, Bochum dur~ng the week of July 28 - 31, 1980
with 15 participants from the U.S. and 30 participants from
var~ous European countries. In all, 38 lectures were given and
supplemented by numerous d~scussions. This volume contains
the papers presented at the sympos~um.

The f~nite element method has been created as a numer~cal

tool for the analysis of engineering problems. For its full

development and applicat~on to nonl~near problems knowledge
from a number of d~scipl~nes is ~mportant: continuum mechan-
~cs, mater~als science, numerical analysis, functional anal-
ys~s and computer sc~ence. Use of th~s knowledge and an under-
standing of the various, somet~mes d~verse, v~ewpo~nts pre-
valent ~n these f~elds is ~ndispens~ble if cont~nued success
in the development of nonl~near f~nite element methods ~s to
be ach~eved. It is ~n this sp~r~t that the partic~pants gath-
ered at the workshop to exchange their ~deas and exper~ences

and to report on the~r most recent research results. We be-

l~eve that the papers presented there, and compiled in this
book, give a true representation of the state-of-the-art of
nonlinear finite element analys~s in structural mechan~cs.

The editors wish to thank all participants for the~r con-

tr~butions and Dr. Obrecht for his help in organizing the
 VII

meetlng. The flnanclal asslstance of the Deutsche Forschungs-

gemelnschaft and the U.S. Natlonal SClence Foundatlon lS most
gratefully acknowledged, and thanks are extended to Springer-
Verlag for a speedy publicatlon of these proceedlngs.

October 1980 W. Wunderllch

E. Steln
K.J. Bathe
Contributors
S.N. Atluri, H. Murakawa, School of Civll Engineerlng,
Georgla Instltute of Technology, Atlanta, Georgia 30332, USA.

I. Babuska, Institute for Physical Science and Technology,

Universlty of Maryland, College Park, MD 20742, USA.

J. Banovec, University Edvard Kardelj of Ljubljana, Faculty

for Architecture, Clvil Engineerlng and Survey, Yugoslavla.

K.J. Bathe, L.W. Ho, Department of Mechanical Engineering,

Massachusetts Institute of Technology,Cambrldge, Massachu-
setts 02139, USA.

T. Belytschko, Department of Civil Engineering, Northwestern

University,Evanston, Illinois 60201, USA.

P.G. Bergan, Dlvision of StructuraL Mechanlcs, The Norweglan

Institute of Technology, Trondheim, Norway.

J.F. Bessellng, Laboratory for Engineerlng Mechanics, Univer-

Slty of Technology Delft, The Netherlands.

F. Brezzl, IStltuto dl Analisi Numerlca del C.N.R. and Istl-

tuto di Matematica Appllcata dell'Universita dl Pavia
27100-Pavia, Italy.

K. Brink, W.B. Kratzlg, Ruhr-Unlversltat Bochum, Institut fUr

Konstruktiven Ingenieurbau III, 4630 Bochum, West-Germany.

G.F. Carey, Texas Instltute for Computational Mechanlcs, Uni-

verslty of Texas, Austin, Texas 78712, USA.

L. Corradi, G. Maier, Department of Structural Engineerlng,

Technlcal University (Politecnico), Mllano, Italy.

M. A. Crlsfleld, Transport and Road Research Laboratory,

Crowthorne, Berkshire, U.K ..

O.M. Eldsheim, P.K. Larsen, Department of Civll Englneerlng,

The Norweglan Institute of Technology, Trondhelm, Norway.

R. Glowinski, P. Le Tallec, V. Ruas de Barros, Paris VI Univer-

sity, L.A. 189, Tour 55.65, 75230 Paris, France.

T.J.R. Hughes, W.K. Liu, I. Levit, Division of Englneering

and Applied SClence, California Institute of Technology,
Pasadena, California 91125, USA.

C. Johnson, R. Scott, Chalmers Unlverslty of Technoloqy,

Department of Computer Sciences, 402 20 Goteborg 5, Sweden.

J.J. Kalker, H.J.C. Allaert, J. de Mul, Delft Universlty of

Technology, Delft, The Netherlands.
 IX

H.A. Kamel, J.M. Tan, Aerospace and Mechanlcal Englneerlng

Department, Unlverslty of Arlzona, Tucson, Arlzona 85721, USA.

D. Karamanlldls, A. Honecker, K. Knothe, Technlsche Unlver-

sltat Berlln, Instltut fur Mechanlk, StraBe des 17. Juni 135 ,
1000 Berlln 12, West Germany.

S.W. Key, C.M. Stone, R.D. Krleg, Applied Mechanlcs Dlvlslon,

Sandla Laboratorles, Albuquerque, N.M. 87115, USA.

B.H. Kroplin, H. Duddeck, Technlsche Unlversltat Braunschwelg,

Institut fur Statlk, Beethovenstr. 51, 3300 Braunschwelg,
West Germany.

D.S. Malkus, Department of Mathematlcs, Illinols Instltute

of Technology,Chlcago, Illlnois 60616, USA.

G. Mehlhorn, D. Klein, Technische Hochschule Darmstadt, Instl-

tut fur Massivbau, Alexanderstr. 5, 6100 Darmstadt, West Germany

M.J. Mlkkola, H.S. Slnisalo, The Academy of Flnland, Depart-

ment of C1Vll Engineerlng, Helsinkl Unlverslty of Technology,
02150 Espoo 15, Finland.

H.D. Mlttelrnann, Unlversltat Dortmund, Abtellung Mathematlk,

August-Schmldt-Str. 1, 4600 Dortmund, West Germany.

A.K. Noor, J.M. Peters, George Washlngton UnlverSlty Center

at NASA Langley Research Center, Hampton, Virginia 23665, USA.

J.T. Oden, Texas Instltute for Computatlonal Mechanics, The

Unlversity of Texas, Austln, Texas 78712, USA.

P.J. Pahl, Technlsche Unlversitat Berlln, Instltut fur Allge-

meine Bauingenleurmethoden, StraBe des 17. Juni 135,
1000 Berlln 12, West Germany.

J. Paulun, E. Steln, Technlsche Unlversltat Hannover, Lehr-

stuhl fur Baurnechanlk, Call1nstr. 15, 3000 Hannover, West
Germany.

T.H.H. Plan, K. Kubomura, Massachusetts Instltute of Techno-

logy, Cambridge, Massachusetts 02139.

E. Rarnrn, Universltat Stuttgart, Institut fur Baustatik, Pfaffen-

waldring 7, 7000 Stuttgart, West Germany.

F.G. Rarnrnerstorfer, D.F. Flscher, Voest-Alplne AG, Hauptver-

waltung Llnz, A-4010 Linz, Austrla.

A. Samuelsson, M. Froler, Department of Structural Mechanics,

Chalmers Unlversity of Technology, 41296 Goteborg, Sweden.

R.L. Taylor, E.L. Wllson, S. J. Sackett, Department of Civil

Englneerlng, Unlversity of Call£ornla, Berkeley, CA 94720, USA.
x

J.C.F. Telles, C.A. Brebbia, Civil Engineering Department,

University of Californ1a, Irvine, CA 92717, USA.
L.C. Wellford, Jr., S.M. Hamdan, Department of Civil Eng1-
neering, Un1versity of Southern California, Los Angeles,
California, USA.
W. Wunderlich, H. Obrecht, Ruhr-Universitat Bochum, Institut
fUr Konstruktiven Ingenieurbau IV, 4630 Bochum West Germany.
O.C. Zienkiewicz, Institute of Numerical Methods in Enginee-
ring, University College of Swansea, U.K .•
Contents
Part I: General Formulatlons

A-posterlorl error estlmatlon for the flnlte

element method
I. Babuska . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Another look at the applicatlon of the prlnciple

of virtual work with partlcular reference to
flnlte plate and shell elements
J.F. Bessellng . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

New general and complementary energy theorems,

finite strain, rate sensltlve lnelastiClty and
flnite elements: some computatlonal studies
S.N. Atlurl and H. Murakawa . . . . . . . . . . . . . . . . . . . . . . . . . 28

Formulation of contact problems by assumed stress

hybrld elements
T.H.H. Pian and K. Kubomura . . . . . . . . . . . . . . . . . . . . . . . . . 49

Part II: Geometrically Nonlinear Problems -

Rods and Shells . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Strategies for tracing the nonlinear response

near limit points
E. Ramm ••••••••••••.•••••••.••••••.•••••••••••.••••• 63
Parameter sensitivlty of nonllnear structures
concernlng stablllty limlt
F.G. Rammerstorfer and D.F. Flscher . . . . . . . . . . . . . . . . . 90
Behaviour of finl te element solutions near a
blfurcatlon pOlnt
F. BreZZl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Some results ln the analysls of thin shell structures
K.J. Bathe and L.W. Ho . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Nonllnear dynamlc flnlte element analysls of shells
T.J.R. Hughes, W.K. L1U and I. Levlt . . . . . . . . . . . . . . . . 151
Elastlc collapse analysis of shells via global-
local approach
A.K. Noor and J .M. Peters . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Large spatlal deformatlons of rods uSlng generali-
zed variational principles
W. Wunderlich and H. Obrecht . . . . . . . . . . . . . . . . . . . . . . . . 185
Large deflection finite element analysis of pre-
and postcritlcal response of thin elastlc frames
D. Karamanlldis, A. Honecker and K. Knothe ........... 217
Geometrlcally correct formulatlons for curved
finite bar elements under large deformations
K. Brlnk and W.B. Kr!tzig . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
XII

Part III: Physically Nonllnear Problems ••••..•••••• 257

Some nonllnear problems of sOlI statlcs and dynamlcs

O.C. Zlenkiewicz . . . . . . . . . . . . . . . . . • . . . . . . . . . . . . . . . . . . 259

Numerlcal methods ln elasto-plasticlty -

A comparative study
A. Samuelsson and M. FreSler .......•..•.............. 274

Finlte element elastoplastlc and limlt analysls:

Some consistency criterla and their lmplications
L. Corradl and G. Maler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290

A flnite element method for problems ln perfect

plastlclty usingdlscontlnuous trial functions
C. Johnson and R. Scott . . . . . . . . . . . . . . • . . . . . . . . . . . . . . 307

Finite element analysis for combined material

and geometric nonlinearities
M.A. Crlsfield ........•.....•.............•...•..... 325

Incremental elastlc-plastlc deformatlons of

stlffened plates in compresslon and bending
J. Paulun and E. Steln • . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339

A study of some generalized constitutive models

for elasto-plastic shells
O.M. Eidsheim and P.K. Larsen . . . . . . . . . . . . . . . . . . . . . . . 364

An efficient finlte element method for elastic

plastlc analysls of plane frames
J. Banovec . . . . . . . . . . . • . . . . . . . . . . . . . . . . . . . . . . . . . . . • . . 385

Elasto-plastic boundary element analysls

J .C.F. Telles and C.A. Brebbla .........•..........•. 403

Slmpllfied calculatlon models applled to

postbuckling analysis of thln plates
B.H. Kr6plln and H. Duddeck . . . . . . . . . . . . . . . . . . . . . . . . . 435

Finite element analysis of relnforced concrete

slabs and panels
G. Mehlhorn and D. Klein . . . . . . . . • . . . . . • . . . . . . . . . . . . . 452

Part IV: Nonlinear Dynamics .•••••.•..••••••.••.••. 479

Finite element analysis of transient nonlinear

response of relnforced concrete structures
M.J. Mikkola and H.S. Sinisalo ...........•.•........ 481

An analysls of the stabllity and convergence pro-

perties of a Crank-Nicholson algorithm for
nonlinear elasto-dynamlcs problems
L.C. Wellford,Jr. and S.M. Hamdan . . . . . . . . . . . . . . . . . . . 502
 XIII

Part V: Solutlon Methods . . . . . . . . . . . . . . . . . . . . . . . . . 519

Dlrect solution of equatlons by frontal and

varlable band, actlve column methods
R.L. Taylor, E.L. Wllson and S.J. Sackett 521

Solutlon by lteratlon In dlsplacement and

load spaces
P.G. Bergan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553

Partltloned and adaptlve algorithms for

expllcit tlme integratlon
T. Belytschko . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 572

Dynamlc relaxation applled to the quasi-static,

large deformatlon, lnelastlc response of
aXlsymmetric solids
S.W. Key, C.M. Stone and R.D. Krleg .................. 585

On the efflclent solutlon of nonllnear flnite

element systems
H. D. Ml ttelmann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621

The numerlcal calculatlon of the contact problem

In the theory of elastlclty
J.J. Kalker, H.J.C. Allaert and J. de Mul . . . . . . . . . . . . 637

Exterior penalty methods for contact problems

in elastlclty
J .T. Oden . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655

ApproXlmate solutlon of nonlinear problems in

lncompresslble flnite elasticlty
R. Glowlnski, P. Le Tallec and V. Ruas de Barros ..... 666

Incompresslble flnlte elements: The LBB

condltion and the dlscrete elgenstructure
D.S. Malkus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696

Part VI: Computational Algorithms 713

Data management In finite element analysis

P.J. Pahl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715

Performance of flnite element algorithms on an

array processor-minlclomputer based system
H.A. Kamel and J .M. Tan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742

High-speed processors and implication for

algorithms and methods
G.F. Carey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 758
Part I:
General Formulations
A-posteriori Error Estimation for the Finite Element Method

I. BABUSKA
UlliverSlty of Maryland, College Park, USA

Summary

The paper shows how a flnite element solver for linear ellip-
tic equations with an a-posteriori estimator can be extended
to the case of eigenvalue problems and nonlinear elliptic
problems.

Introduction

Recently in the project FEARS we developed a finite element

solver for linear elliptic equations WhlCh has adaptive
features and provides a-posteriori error estimation of the
error. In this presentation we show how this kind of solver
wlth a-posteriori estimator can be extended to the eigenvalue
problems and nonllnear elllptlc problems. Flnally tYPlcal
results of FEARS are illustrated in two examples.

The Linear Problem

Let us consider an elliptic linear boundary value problem on

r1CR n :

Lu = f (la)

with boundary conditions

u = 0 on 3r1 (lb)

We will assume that there exists 0 < c l ,c 2 < 00 such that

4

for any u,v E'Ho

1
un
2
clll u II 1 < (Lu,u), I(Lu,v)l.::. c 2 1I u ll l Ilvll l (2)
H (n) H (n) H (n)

We denoted by Hl(n) (respective to H~(n)) the usual

Sobolev space of functions with square integrable first de-
rivatives (respective with zero traces on an). By L (n) ,
p
1 .::. p < 00 we denote the usual space of functions with pth
power integrable. The respective norms are denoted by
II • II 1 , I I • IlL (n) , etc.
H (n) p

Further we denote I lui IE = (LU,u)1/2 Because of (2)

I I· I IE and I I· I I 1 are equivalent.
H (n)

In addition we assume that f E' R, where R is a cone of

admissible right hand sides and we assume that for every
1
f E'R there exists exactly one (weak) solution uE'HO(n) of
the problem (1) .

Let S be a family of finite element meshes II and by S*

we denote the totality of finite element functions associated
to the meshes II E'S For a fixed II E' S we define S* (ll)
as the linear space of all finite element functions (in H~)
associated with the mesh ll. Further we associate to every
llE'S a cone R*(ll) c R . For given llE'S fE'R(ll) we de-
note U(ll) E'S*(ll) the finite element solution of (1) (usinq
the same test and trials functions) associated to the mesh II
and the right hand side f. We shall assume that U(ll)
exists and is uniquely determined.

We will assume also that there is an error estimator E(U(ll))

which is computable only by use of the finite element solu-
tion U(ll) and the input data such that

1) There exist constants Di' i = 1,2 dependent only on

S, R*, L but are independent of particular mesh II , f,u
such that

( 3)
 5

2) For any 1 < P < 00

(4)

where K(E,p) is independent of u,~ etc. and K(E,p) ~ 0

as E ~ 0

3) There exists a sequence of meshes ~i ES i 1,2, ...

such that for i ~ 00

Ilu(~i) - uil E ~ 0 (Sa)

E(U(lli»
9 (Sb)

The Eigenvalue Problem

Let us assume in this section that in addition the problem

(1) is selfadjoint, j.e.

(Lu,v) = (u,LV) (6)

for any

Consider now the eigenvalue problem

Lu = AU (7a)

u = 0 on an (7b)

and assume that for any v(~)ES*(~),v(~)ER*(~)CR. By

u(~) E S*(~), Ilu(~) IlL (n) = 1 we denote the finite element
eigenfunction associate~ to the approximate eigenfunction
A(~) • Assume further that there exists sequence ~i such
that A(~i)~ A and u(~i) ~ u where u and A are eigen-
function and eigenvalue of (7). Obviously u(~) is the
finite element solution of (1) with f(~) = A(~)U(~) •

Denote by V(~) = u the exact solution of this problem. The

error estimator E (u(~» estimates the norm IIV-u(~) liE as
explained in the previous section. Now we have

THEOREM 1. There exists eigenvalue Aj such that

6

I A(]li)-A j I < A1/2 (]li)E(]li)D 2 (8)

and in addition

(9 )

Proof. Denote the complete system of normalized

eigenfunctions.

Then
LC. 2 1 ,
1.

Denoting C; V-u(]l) we have

(LV,v) = (Lu(]l) ,v) + (LC;,v)

and of course

Denoting e. =U(A) - <p. we get

] ]

(Le j , v) - Aj (e j , v) = (A (11) - AJ) (u ( 11) , v) - (L C; , u)

Set now Then we get

and therefore we get

minIA(Il)-A·1 < Ilu(Il)II E IIC;II E ~ Al / 2 (1l).E(Il)D 2 (10)

j ] -

where D2 is given in (3). When Ili is used for 11 then

the right side of (8) can be replaced by Al / 2 (Il J )E(llj)
(1+0(1)). Let us remark that (8) is independent of the spec-
trum, but the 0(1) term in (8) is dependent on it. The theo-
rem 1 can be generalized in various ways. Let us mention one.

THEOREM 2. Let (5b) hold. Then E2(llj) (1+0(1)) < IA(llj)-AI

<
2
E (11.) (1+0 (1) ) The proof goes by analogous way of [lJ.
- ]

The Nonlinear Problem

Let us be interested in the nonlinear problem
 7

Lu + g(u) = f (lla)

u = 0 on 3>1 (llb)

About g we will assume the following. For any

1
u,vE'HO(>I)
2m
and p ~ m+2 we have

Ilg(u) IlL (>I) 2 cllull E ( 12)

P
(g(u+v) - g(u) ,v) ~ 0 (13)

Ilg(u+v) - g(u) IlL 2 cllvll E (14)

P
In addition we will assume that for any f E' R there exists
unique solution (in H~) of (11) and if uE'S*(lJ) then,
g(u) €R*(lJ) I and fE'R*(lJ) for any lJ E'S We assume also
that the finite element solution u(lJ) exists and is unique.

Let now V be the exact solution of (1) with the right hand
side f - g(u(lJ» and v(lJ) be the finite element solution.
Obviously u(lJ) = v(lJ) • Denote by E(u(lJ» the estimator
for Ilv-u(lJ)II E • Nowwehave

THEOREM 3. be a sequence of meshes such that

E (lJi) + 0 as and (5) is satisfied. Then

1
Proof. Write V = ~+u(lJ) . Then for "Iv E' HO (>I)

(LV,v)+(g(V) ,v) = (f,v) + ((g(V)-g(V-~» ,v)

(f,v) + (R,v) (16)

We have also

(Lu,v) + (g(u) ,v) = (f,u) ( 17)

and therefore denoting z = V-u we get (Lz,v) + (g(u+z)-

g(u) ,v) = (R,v) which yields
 8

and therefore

I I z I IE < E (u ( )l) ) cr ( 1) as E -* 0 .
Now

~ (l+cr(l))E(u()l))

and because

we get the other side of (15).

This theorem can be generalized in various ways, but we will

not elaborate more.

The estimators for the elliptic problem can be also used in

other finite element problems as in linear and nonlinear
evolution equations.

The FEARS
During recent years the solver FEARS (finite element adaptive
research solver) was developed which has various special fea-
tures which will not be discussed here. For more information,
see [ 2 1, [3 1, [4 1, [5 1, [ 6 J •

Here we will illustrate the behaviour of the effectivity in-

dex 9 = Ell lu-u()l) I IE introduced in (5b) on two examples.

Example 1. Elasticity problem for homogeneous isotropic me-

dium (v .25) on the ring

with displacement ul,u 2 prescribed on internal boundary

8 1 = [(xl'x 2)I (Xi+x ;) 1/2 = .25J namely u l = 1 , u 2 = o. On
the external boundary 8 2 = [xl,x21 (Xf+x~)1/2 = lJ we take
u l = u 2 = o. The exact solution is known. The computation
was performed on a quarter ring in the adaptive mode by (map-
 9

ped) piecewise bilinear elements. The (energy) error is as-

ymptotically of order N- l / 2 where N is the number of de-
grees of freedom. Table 1 shows essential results.

DOF 100
*+t E
neTT
20 14.697 1. 015
40 11. 664 1. 095
52 10.168 .981
76 8.136 1. 0002
130 6.332 .9895
173 5.744 .97107

Table 1. Elasticity problem on

the ring Q.

We see higher effectivity of the estimator when I lei I + 0 .

Example 2. Elasticity problem for homogeneous isotropic me-

dium with v = 0.0 on the rectangle

R = [(xl,x2l1'O < xl < 1, -4 < x2 < 0]

On the sides

R2 = [(x l ,x 2 )! 1/2 < xl < 1, x2 0]

we assume zero traction. On

R3 = [(xl'x 2)! 0 < xl < 1/2, x2 0]

we prescribe ul = 0 , u2 = 1 On

we assume u2 = 0 and zero traction in xl direction. On

RJ = [e x l ,x 2)! xl = 0, -4 < x2 < 0] we assume u l = 0 and
zero traction in x 2 direction.

computation has been made in the adaptive mode. The essen-

10

tial data are presented in Table 2.

DOF 100 *++ E

nerT
32 41. 76 .672
50 35.04 .672
80 25.25 .764
106 21. 82 .764
168 17.29 .773
229 14.10 .827

Table 2. Elasticity problem on the

rectangle 2.

The solution has oscillatory singularity in the point xl = 0,

x 2 = 1/2 of the type rl/2. For a uniform mesh the assymp-
totical order of convergence is N- l / 4 , for adaptive meshes
N- l / 2 The table shows that very clearly the effectivity
E
index e = nerT converges to 1 and experience shows that
about for 5% accuracy .85 < e < 1.15. In most cases much
better effectlvltv lndex occurs.

References
[lJ Babuska, I.i Rheinboldt, W.C.: A-posteriori error esti-
mates for the finite element method. Int. J. Num. Mech.
Eng. 12 (1978) 1597-1615.

[2 J Babuska, I.i Rheinboldt, W.C.: An~lysis of optimal fi-

nite element meshes in Rl. Math. of Compo 33 (1979)
431-463.

[3J Babu~ka, I.i Rheinboldt, W.C.: Error estimates for adap-

tive finite element computations. SIAM J. Numer. Anal.
15 (1978) 736-719.

[4J Babu~ka, I.i Miller, A.: A-posteriori estimates and

adaptive techniques for a finite element method. to
appear.

[5 J Babu~ka, I.i Rheinboldt, W.C.: Reliable error estima-

tion and mesh adaptation for the finite element method.
Compo Meth. Nonlinear Mechanica, J.T. Oden, edltor,
North Holland Publ. Co. (1980) 61-108.

[6J Rheinboldt, W.C.: On a data structure for adaptive fi-

element mesh refinements. ACM transaction on Math. Soft-
ware 6 (1980) 166-187.
Another Look at the Application of the Principle of Virtual Work
with Particular Reference to Finite Plate and Shell Elements

J F BESSELING
UOlverslty of Technology, Delft, The Netherlands

Summary

In the finite element model each element may be looked upon as a

deformable body, however with a limited number of deformation modes,
determined by the number of nodal displacement and rotation compo-
nents on its boundary. In terms of a corresponding number of suitably
defined strain parameters finite element properties are derived for
plate and shell elements, valid for arbitrarily large displacements
and rotations. The membrane properties are derived from a strain- or
a elastic potential distribution in the domain of the element. The
bending properties follow from an equilLbrium field of internal
moments and are restricted to small deformations of the element. It
is shown that for triangular elements all properties can be given in
closed form, but for large rotations the equations of equilibrium
can only be solved as rate equations if bending is taken into account.

1. Introduction

The variational statement of the dynamics of solid bodies in terms

of the princLple of virtual work is a suitable starting point for
the discrete modelling of the mechanical behaviour of such bodies,
in particular if they represent complex structures. We shall employ
this principle in the version which in [IJ is attributed to Piola.
The principle of virtual work according to Piola requires the virtual
work of the external forces and of the inertial forces to be zero
for all infinitesimal rigid body motions. Infinitesimal rigid body
displacements and rotations can be specified as subsidiary condi-
tions. Introduced into the virtual work condition with the aid of
mUltipliers the resulting contribution may be interpreted as the
virtual work of deformation, where the multipliers are components of
a stress field.
We may conclude that for a dLscrete model it is sufficient to define
12

for each finite element the necessary and sufficient conditions for
rigid body motion in terms of the nodal velocities, or equivalently
to define the deformation modes by properly chosen generalized
strain measures. If these strains by constitutive equations are
related to the dual stress multipliers, we arrive at the system of
equations by which we can solve a stress- and deformation problem.
This approach, that was indicated in [2J, will be discussed here in
some more detail.

2. Principle of virtual work

Let us consider a body B of massdensity p, subjected to surface

tractions !, a volume force density p~ and an acceleration field u.
Written with rectangular cartesian components of the vector quantities
involved, the principle of virtual work states as a necessary condi-
tion for any motion of the body B

JdB t.ou.dA +
~ ~
JB pf.ou.dV
~ ~
= JB pu.ou.dV
~ ~
(2.1)

for all rigid virtual displacement fields ou .• A rigid virtual dis-

~

placement field is a field of infinitesimal, kinematically admissible

displacements satisfying the conditions

doui dOU.)
[ - - + --1. = oe: .. D. (2.2)
dX. dX. ~J
J ~

We observe that Oe: .. is a symmetric tensor and hence the conditions

~J
(2.2) can be introduced into the variational statement of the prin-
ciple of virtual work by means of a symmetric tensor field of multi-
pliers t ..• Thus (2.1) is transformed into
~J

JdB t.ou.dA JB pf.ou.dV = fB (pu.ou.

~ ~
I-
~ ~ ~ ~
+ t .. oe: .. ) dV,
~J ~J
(2.3)

which now must hold for all kinematically admissible virtual displace-
ments ou .•
~

By applying Green's transformation we derive from (2.3)

t.
1.
t .. n. (2.4)
1.J J

and
dt ..
--2:.J. + p f. (2.5)
dX. ~
J
 13

which identifies t .. as the stress tensor of Cauchy. Hence t . . 0£. in

~J ~J ~

the right-hand side of (2.3) may rightly be called the virtual work
of deformation per unit volume.
In the case of elastic material behaviour, or if the inelastic
deformations remain sufficiently small, deformations with respect to
an initial geometry of the body are physically relevant variables.
The deformation tensor may then be considered as the sum of an elastic
and an inelastic part, such that for the elastic components hold
unaltered the elastic~ty relations that are valid in the ~nitial

geometry. It should be stressed however that this is no longer true

after large inelastic deformations [3J.
If we restrict ourselves to elastic and small inelastic deformations
it is appropriate to express the volume integrals in (2.3) in terms
of the initial geometry Bo of the body. We define a i as the cartesian
coordinates of the material points in this initial geometry. They
identify as material coordinates the material points during the
deformation process.
The difference of the square of a line-element in the deformed and in
the undeformed geometry is defined by

(2.6)

where EaB is the lagrangian strain tensor

(2.7)

A rigid virtual displacement field can now be defined by the condi-

tion that the first variation of the strain tensor (2.7) vanishes in
all points of the body

(2.8)

Using the equality pdV = podVo and again with the aid of a symmetric
tensor field of multipliers, craB' the principle of virtual work is
expressed with volume integrals in terms of the initial geometry B
o
by

J
dB
t.ou.dA +
~ ~
J
B
p f.ou.dV
0 ~ 1 0
JB(pOu.ou.
ll
+ cr BoE B) dV,
a a 0
(2.9)
o o
14

which must hold for all kinematically admissible virtual displacement

fields.
The tensor 0aB is the so-called pseudo stress tensor of Kirchhoff,
related to the stress tensor of Cauchy by

(2. 10)

This relation follows from the identities

J
B
t .. OE. ,dV
~J ~J
J
B
o

(2. II)

The principal difference between O£ij and oE aB lies in the fact that
O£ij cannot, and OE aB can be derived from a measure of finite strain
as a first variation. The strains EaB are by their elastic components
, 0
EaB = EaB - EaS related to the stresses GaB' that are the proper duals
of the virtual deformations oE aS '
To allow for a maximum freedom in approximation methods we now write
the virtual work condition (2.9) for an assembly of N finite elements
, h anumb er 0 f mu It'~p I'~er f unc t'~ons, 0aB'
w~t * * and t*,~ *.
ti'

N
L J
e=1 Be
f-p 0
(f,-u:)ou: +
~ ~ ~
o

+ 00*
aB

N
+ L J(lBe [ ( 0u ~ ~
- 0u: ) t ~ + (u ~ - u:) 0 t ~] dA
~ ~ ~ ~ ~ 0
+
e=1
o

+ JdB u [-ou~t~* ~ ~
+ (~, -u~) ot~*]
~ ~ ~
dA= O. (2.12)

The displacement fields u~ are only defined on the external surface

~

and on the element interfaces, while G~S is related to E~B' for

instance through the linear isotropic constitutive relations
 15

O~Q = e -a(T-T )} 0 +2G(E e -E" - -31 EkekoNS)'

c {E kk (2.13)
~" 0 as as as ~

Here E~S represents the small inelastic stralns and aCT-To) is the
specific thermal expansion. The prescribed quantities ~. and
1
r. 1
will
usually be given on the external boundary in the deformed state.
e ' ou e ' and
By considering the variations oE aS ~ui
* we derive the
k
following equations

0* (2.14a)
as

3u e
_3_{ 0* (0 + ~)} + pofk (2.14b)
3a S as ka 3aa

e
0*
as
[0 ka
3Uk )
+ 3a nS t*
k
on 3B e
o '
(2.14c)
a

* - t*
t i (!!) =
on element interfaces, (2.14d)
i(-!!)

t~* dA = t*1 dA0 on 3B u , (2.14e)

1

t~ dA = t. dA on 3B t . (2.14f)
1 0 1

From the variation of the multipliers we have an additional set of

equations

e e e e
3u 3u S 3uk 3uk I
Ee
as 2
1 a
-- +
3a s 3a + 3aa
a
aa; J in Be0 ' (2.ISa)

e
u. u~ on 3B e (2.ISb)
1 1 o '

u~ u. on 3B u . (2.ISc)
1 1

The most common type of finite elements is the so-called kinematically

conforming type. The equations (2.15) are satisfied apriori. With
the aid of proper interpolation polynomials in (2.9) from the prin-
ciple of virtual work a set of equations is obtained, which after
differentiation of the appropriate constitutive equations are of the
form

(2.16)

The matrices are symmetric and the geometrical non-linearities

16

present themselves in the matrices ~u, ~u and ~.

For the analysis of shells and of large bending deformations of plates
this approach meets with fundamental difficulties. For geometrically
nonlinear problems of initially flat plates, in addition to the rigid
body motions, simple independent states of either constant membrane
deformation or bending deformation must be represented in each finite
element. For shells even in the linear case it is difficult to ensure
that bending deformation of a finite element is not accompanied by
spurious membrane deformations. If inextensional bending is permitted
by the shell geometry, the finite shell element should reproduce
exactly simple independent states of membrane and bending deformation.
Otherwise, the relatively large stiffness against membrane deformation
will obscure the inextensional bending modes of deformation in the
finite element model of the shell.
It is not possible to meet the above requirements by a straightforward
displacement approach to the finite element modelling on the basis
of the variational statement (2.9). In the next paragraph we shall
have another look at the application of the principle of virtual
work.

3. The consequences of the representation of the displacement field

by a finite dimensional vector space

In the finite element method at interelement boundary points, -lines,

and -surfaces the adjoining elements must have common displacements
in order to ensure continuity of the structure. If these displace-
ments are to be approximated by a finite set of functions, the dis-
placements at the interelement boundaries and at the natural bound-
ary of the structure can be taken as interpolations between nodal
displacements at the interface between any two elements or between
an element and the surroundings of the structure. Continuity of the
structure is then ensured for all values of the nodal displacements
and the displacement field is represented by a finite dimensional
vectorspace Ue with elements in Rn.
We recall that in the formulation of the principle of virtual work
rigid virtual displacements had to be characterized. The inter-
polations between nodal displacements must be such that all possible
rigid motions of the finite element are represented. Even for
elements with curved boundaries this requirement is met by the
 17

isoparametric concept [4J. For the transformation to natural coordi-

nates ~a of the finite element and for the interpolation between
nodal displacements the same interpolation functions are used. In
the case of material finite elements the coordinates of the material
points of a finite element are then given by

0.1)

Infinitesimal rigid body motions of a finite element are now defined

by a subspace of Ue , which will be called the vectorspace of degrees
of freedom, Ve • It is appropriate to call the complementary subspace,
Ee , the deformationspace of the finite element.
Ue '" Ve $ Ee • 0.2)

The conditions for a rigid body motion are expressed by oEaS '" O.
However, in the derivation of the finite element equations by appli-
cation of the principle of virtual work it suffices to require
locally oEaB '" 0, such that for each finite element a number of
linearly independent conditions is obtained that is equal to the
dimension of its deformation space. In other words, for the finite
element model we may derive the contribution of the subsidiary condi-
tions in the variational statement (2.9),

0.3)

from a strain distribution with parameters obtained by local appli-

cation of (2.7). The number of these strain parameters is determined
by the number of nodal displacements U:i minus the number of degrees
of freedom of the finite element as a rigid body. The strain para-
meters are by application of (2.7) expressed as nonlinear functions
of the nodal displacements with the aid of the interpolations (3.1):

E.=D.(u.). (3.4)
L L Ie

The strain parameters thus defined are elements of a vector of

generalized strains of the finite element, Ee ' for which holds
oEeE Ee •
Instead of (3.3) we now may add 0ioEi to the virtual work condition
in order to characterize the rigid body motions. However, in order to
end up with a physically defLned set of equations we must be able to
18

derive constitutive equations for the generalized stresses, ~, which

are the duals of the virtual strains, D£. It is then in the first
place necessary to describe the deformed state of each finite element
over its whole domain. For this we can introduce a set of functions
of the natural coordinates, such that

(3.5)

For the linear constitutive equations (2.13), provided also the

inelastic and thermal strains are given as functions of the natural
coordinates with £~ as parameters, the virtual work condition (2.9)
1 m.
can be evaluated and we obtain (D
1,
k = ~)
Q~

DU k [ -f k + ~R-iiR- + C\Di,k] = 0 V D~, (3.6)

where

0. S .. (£.-£~), S .. = S ... (3.7)

1 1J J J 1J J1

In the case of nonl1near elasticity we have

(3.8)

The elastic potential per unit undeformed volume, poe, is a compli-

cated function of the components of the strain tensor, and hence for
the finite element a complicated function of the generalized strains
£i' But rather than resort to numerical integration for the evalu-
ation of (3.3) in each step of the deformation proces&, we can deter-
mine first the elastic potential of the finite element as a whole,
as a function of the generalized strains £i' With the aid of inter-
polation functions we obtain by integration over the domain of the
finite element from local functional relationships,

p e
o
= p0 e(E a 13) = p 0 e(£.,
1
~
a
), (3.9)

the finite element expression for the elastic potential.

E(£.) =
1
JBe p e(£.,~ ) dV •
0 1 a 0
(3.10)
o

The virtual work condition again assumes the form (3.6), but the
generalized stresses 0i are now determined by
 19

dE
a. = (3. II)
l. Cle:i'

The approach outlined above proves to be particularly fruitful in

dealing with the geometrical nonlinearities of the membrane defor-
mations of plates and shells. It avoids the appearance of spurious
membrane deformations, which are most detrimental to the accuracy of
the solution, in particular if the plate or shell would permit
inextensional deformation. The success of reduced integration tech-
niques for the improvement of the conditioning of the matrix and of
the accuracy of the solution can also be understood in terms of the
local enforcement of the conditions oE a8 , in number not exceeding
the number of linearly independent virtual deformation parameters,
as determined by the number of nodal displacements.
Since local application of the expressions (2.7) give the components
of strain as simple quadratic functions of the nodal displacements
all properties of triangular (and rectangular) elements can be
determined in closed form.

4. Bending of plates and shells

The description of the bending of plates and shells by finite elements

meets with the wellknown difficulty, that not only the deflectl.ons but
also the rotations must be continuous at element interfaces. For a
kinematically conforming approach this implies contl.nuity conditions
for the normal derivatives of the deflections, which are not easily
satisfied. Here a derivation of the element properties on the basis
of a bending and twisting moment distribution provides a simpler
formulation than a derivation on the basis of displacement inter-
polation functions.
We observe that for the static case (u:l. = 0) in the absence of volume
forces (f i = 0) the displacement fields u~ disappear from expression
(2.12), if the equations (2.14b) and (2.14c) are satisfied. By a
rigid rotation of the finite element, provided the element is chosen
sufficiently small in relation with its deformations, these equations
can always be reduced to the linear form with any desired degree of
accuracy.

0, a*
a8 8
n = t*a . (4. I)

When in addition we satisfy the relations (2.14a), a~8 a: 8 ,

20

expression (2.12) reduces to

t.ou~dA+ e=1~ [-JBe ocr*SE*SdV

~ ~ a a
+JaBe (ou~t~+u.ot~)dA
0 ~ ~ ~ ~ 0
lJ+
o 0

(4.2)

We should realize that t~ in (4.2) and t* in (4.1) differ by the

~ a
rigid rotation, that had to be applied to each of the finite elements
in order to free the equilibrium equations from the displacements.
Since the orthogonal transformation, that represents this rigid
rotation, has to be expressed in terms of displacements, the
geometrical nonlinearity enters condition (4.2) through this trans-
formation.
We shall now discuss the implications of the above reasoning for the
analysis of the bending of plates by triangular finite elements.
The deformation of each single element is assumed to remain small,
such that with respect to a plane through the vertices of a triangular
element we may apply linear bending theory. Let we denote the dis-
placement component normal to the basic triangle. With moments MaS
and curvatures kaS we obtain for each finite element, analogues to
the linearized contributions of the elements in (2.12),

(w*_w e ) oV* + r0<1>* + (lOwe) M* +

n l an n

+ rl <1>* + aw*) oM* ] ds +

an n i=1
~{( ow*-owe ) H* . + (w*-w e ) oH*. }
c~ c~

(4.3)

If we let M~S = M:~, V~, M~, and H~i satisfy the equations

(4.4a)

(4.4b)
 21

(4. 4c)

then the displacement we disappears from (4.3) and we have for each
element

3
oM* k* dA
as as 0
+ L J [ow*v* + o~*M* +w*ov* + ~*oM* ]ds
n n n n
i=1 s.
~

3
+ L (ow*H*
c~
+ w*oH*.).
c~
(4.5)
i=1

On the basis of linear moment distributions, with the values of M~S

at the vertices as generalized stresses, in [5J element propert~es
were derived. By a quadratic interpolation for w*, and a linear inter-
polation for ~* along each of the three sides of the basic triangle,
nine linearly independent deformation modes were defined, correspon-
ding to the nine generalized stresses. With the aid of the linear
elasticity relations expression (4.5) can then be evaluated and we
find:

e e- 1 e w * ~ * e w ~
-~ui (Sij OJ + Dikwk + Dii~i) + 0i (Dikow~ + Diio~~). (4.6)

The matrices lS~jJ, [D7kJ , and ID!i J have been given in [5J 1)
The displacements w~ and the rotations ~~ at the boundary are defined
with respect to the basis triangle of the element under consideration.
Continuity of the structure requires that they are expressed in terms
of the displacement and rotation functions defined on the interfaces
of the finite elements. Apart from a rigid translation, which has no
consequences, this implies for the displacements an orthogonal
transformation. The elements of the complete transformation matrix,
as determined by the displacements of the vertices of the basic
triangle in the case of small membrane deformations, are given in the
appendix. However, for the rotations it is a different matter.
Infinitesimal rotations can be decomposed into orthogonal components.
The infinitesimal rotation o~* for a particular element is then
equal to the difference between the tangential component of the

1) The submatrix C33 ' as given in [5J, must be multiplied by a

factor 2.
22

infinitesimal rotation of the plate surface at the interelement

boundary and the tangential component of the infinitesimal rotation
of the basic triangle. The latter is determined by the displacements
of the vertices and their variat~ons, while the tangential component
of the infinitesimal rotation of the plate surface at the inter-
element boundary must be the same for adjacent elements in order to
ensure continuity of the structure. Thus the infinitesimal rotations
in (4.6) can, just as the infinitesimal and finite displacements, be
expressed in terms of functions defined on the interfaces of the
finite elements, such that continuity between elements is ensured.
This is not possible however for the finite rotations $*, if these
rotations as well as the rotations of the basic triangle do not
remain sufficiently small.
If the rotations do remain small, we may follow the approach indi-
cated in [6J. Otherwise we meet with the difficulty that finite
rotations can only be characterized by orthogonal transformations,
defined in terms of angular coordinates, which obey complicated non-
linear transformation rules. Then we are compelled to define $* for
each finite element separately as the normal slope of the plate at
the element boundary with respect to the basic triangle of the
element under consideration, $* = - a~e • Continuity of the structure
can now be ensured by deriving the rate of change of $* at any moment
from the difference between the tangential component of the angular
velocity of the plate surface at the interelement boundary and the
tangential component of the rate of rotation of the respective
triangle. As a consequence for arbitrar~ly large rotations for the
finite element model under consideration only rate equations can be
formulated. In view of the numerical procedures, required to solve
the highly nonlinear equations in the case of large rotations, this
is not felt as a serious disadvantage of the model.
The fact however that $* cannot be expressed in terms of nodal
coordinates implies that the generalized strains from (4.6),

E.
~
(4.7)

cannot be written as functions of the type (3.4). The rate

equations on the basis of (3.6) are of the form

(4.8)
 23

The symmetry of the geometrical stiffness matrix, [o.D. k O]' is

~ ~, "
clearly due to the analytic relation between generalized stra~ns and
nodal coord~nates. It turns out that this symmetry is lost if bending
~s taken into account by (4.7), because the generalized bending
strains (4.7) are determined by the nodal coordinates as follows

$*m (4.9)

Here ~ is the angular veloc~ty at the interelement boundary. The

transformation coefficients ~R, and Bmk are functions of the nodal
displacements at the element vertices. Some details are given in the
appendix.
The equilibrium equations according to the principle of virtual work
now read

o.~ (D~~, k -

(4.10)

from which the following rate equations are derived

[(D~~, k - D~~mBmk) S~J

.. (D~ ° -D~ B 0)
J," J n n"
+ o.(D': kO
~ ~,,,
-D~~m
u ~ ~"u
+ (D. k - D. Bmk ) S .. D. ° ~ ° fk '
~, ~m ~J J" "

D~oS •.
~"~J
(D':
J ,m
-D~ B
In nm
)~
m
+ D~ S.
~R, ~J
D~
Jm m
$ = f~
R, .
(4.11 )

Since Bmk,R, will in general be unsymmetric in the indices k and R"

the symmetry of the geometrical stiffness matrix is lost. It should

be observed however that the geometrical nonlinearity of the bending
deformations, embodied in the transformation ~, need only be taken
into account in the case of rotations for which no longer holds
~2 « 1. This in contrast with the geometrical nonlinearity of the
membrane deformations, which for instance determines the buckling
behaviour [7].
It has been shown in [8J that for doubly curved shell elements,
provided they are taken suff~c~ently small, the bend~ng behaviour
can be approximated to any desired degree of accuracy by the bending
behaviour of the flat triangular element. For the membrane behaviour
a more accurate description for equal size of elements is obtained
24

by going from the flat plate approximation to the flat plate with
initial curvature, determined by the midside initial displacements
w*o
Finally it can be observed that ~n case of distributed loading
normal to the surface no improvement of accuracy may be expected
going from equivalent nodal forces corresponding to a linear deflec-
tion interpolation to equivalent nodal forces found by the quadratic
interpolation which is possible for the TRIM-6 element. This is due
to the fact that the bending properties of the element are derived
from a moment distribution, that satisfies the homogeneous equilib-
rium equations (4.4a). However, as it was mentioned in [6J, for
pressure loaded shells, that permit a membrane solution of the equi-
librium equations, we can enter the finite element into the structure
in a state of initial membrane deformation with the corresponding
membrane forces acting on the surrounding structure. Thus spurious
bending stresses are avoided.
 25

References

1. C. Truesdell and R. Toupin, 'The classical field theories',

p. 596, Encyclopedia of Physics, Vol. III/I, Springer Verlag,
Berlin, Gottingen, Heidelberg (1960).
2. J.F. Besseling, 'Finite element methods', Trends in Solid
Mechanics, Delft University Press, Sijthoff and Noordhoff Int.
Publ. (1979), pp. 53-78.
3. J.F. Besseling, 'A thermodynamic approach to rheology', Proc.
IUTAM Symp. on irreversible aspects of continuum mechanics,
Springer Verlag, Wien (1968) pp. 16-53.
4. o.c. Zienkiewicz, The finite element method, third edition,
McGraw -Hill Book Co., London (1977).
5. J.F. Besseling, Postbuckling and nonlinear analysis by the finite
element method as a supplement to a linear analysis, ZAMM 55,
(1975), pp. T3-TI6.
6. J.F. Besseling, L.J. Ernst, A.U. de Koning, E. Riks, K. van der
Werff, 'Geometrical and physical nonlinearities, some developments
in the Netherlands', Proc. Fenomech 1978, North Holland Publ. Co.,
Amsterdam; Compo Meths. Appl. Mech. Eng. 17/18 (1979), pp. 131-
157.
7. L.J. Ernst, A geometrically nonlinear finite element shell theory;
applications to the postbuckling behaviour of shells, Dept. Mech.
Eng. T.H. Delft, WTHD-126 (1980).
8. L.J. Ernst, A f~nite element approach to shell problems, Dept.
Mech. Eng. T.H. Delft, WTHD-114 (1979).
26

Appendix

The coordinates in the x-y plane of the three cornerpoints of a

triangular element in the undeformed state define the following
geometrical quantities

al x3 - x 2 bl Y2 - Y3 22 a2 + bL
1 1 1

a2 xI - x3 b2 Y3 - YI 22 a2 + b2
2 2 2
a3 x 2 - XI b3 YI - Y2 22 a2 + bL
3 3 3

and

For each element the following vectors are introduced where u, v, w

represent displacement components of a cornerpoint.

I
bl
I
al
e lUI e hi e
WI

1::\'
d l = 2A b 2 ' d 2 = 2A a2 u U2 v W2
b3 a3 IU 3 = w W3

If the deformations of the basic triangle may be neglected its

rotation can be characterized by the orthogonal transformation

_ bTw e + veT(baT_ab T) we
_aTw e _ ueT(baT_ab T) we
I + bTu e + a Tv e + u eT (baT-ab T) v e

Now the transformation coefficients Ak2 Ln (4.9) follow from

The angular velocity of the basic triangle is determined by the

spinmatrix

-w

~z
z
=[ o
-w w
x
y

The tangential unLt vector along the side pq of the basic triangle
is given by
 27

x -x +u -u y -y +v -v W -w
q P q P e + qpqpe~e.
L -x \ -y R,i -z
~

The tangential component of the angular velocity of the basic

triangle about the side pq, expressed in terms of uk and ~, follows
as the scalar product of the vectors ~ and !i' Thus the transform-
ation coefficients BkR, in (4.9) are determined.
New General and Complementary Energy Theorems,
Finite Strain, Rate Sensitive Inelasticity and Finite Elements:
Some Computational Studies

S.N.ATLURI, H.MURAKAWA
Georgia Institute of Technology, Atlanta, USA

Abstract:

General varlational theorems for the rate problems of rate-de-

pendent finlte strain lnelasticity, in terms of the appropri-
ate rates of the flrst and second Piola-Kirchhoff stress ten-
sors, the symmetrized Biot-Lure' stress tensor, and their con-
jugate measures of straln-rate, are dlscussed. Certaln new
rate-complementary-energy principles, involving the rate of
spin and the rate of the symmetrized Biot-Lure' stress tensor
as variables, are stated for finlte straln analysis of rate-
sensitlve materials, such as those exhiblting elasto-vlSCO-
plastlc and creep behavior. Unlqueness and stability criterla
for those lnelastlc SOllds, uSlng the flnite element counter-
parts of the new complementary energy rate principles, are
dlscussed. Computational studies, uSlng the complementary ener-
gy methods, discussed herein include: (i) blfurcation neck-
ing and post-buckling analyses of inltlally perfect elasto-
plastlc bars, and (ii) post-buckllng and large deformation
analyses of thin elastic plates under inplane compresslon and
transverse bending loads.

Introduction:
The topic of rate (incremental), multi-fleld, varlational
principles, in general, and the rate complementary energy
principles, in particular, and the corresponding finite ele-
ment methods, for finlte strain analysls of compressible non-
linear-elastic solids were discussed in detail by Atluri and
Murakawa [1]. Also dlscussed in [1] were the contributions of
Koiter, Zubov, and Fraeijs de Veubeke, dealing with the sub-
 29

ject of complementary energy principles, governing the total

deformations of semi-linear and/or nonlinear compressible iso-
tropic elastic materials. It was shown in [1], that the con-
cept of treating the angular momentum balance condition as an
a posteriori constraint through a complementary energy princi-
ple involv~ng the symmetrized Biot-stress (or what is also re-
ferred to as the symmetrized Lure'-stress or the Jaumann-
stress) as well as the orthogonal tensor of rigid rotation, as
variables, as first introduced by F. de Veubeke, has certain
fundamentally novel features that makes it attractive for
practical application.
The ideas of discretizing the angular momentum balance condi-
tions through a complementary energy principle has been ex-
tended by the authors in (i) the incremental (rate) analysis
of finite strains in compressible as well as incompressible
nonlinear elastic materials [2-4], (ii) the rate problems of
classical (rate-independent) finite strain, elasto-plasticity
[5-9], and (iii) nonlinear stability and post-bifurcation
analysis of semilinear isotropic elastic beams [10].
In the present paper the authors' earlier work, [2-10], is ex-
tended to the cases of finite strain analyses of materials
with rate-sensit~ve behavior such as elasto-viscoplasticity
and creep, and post-buckling and large-deformation behavior of
structural members such as plates and shells, undergoing large
rotations and large stretches.
The summary of the topics presented in the following is: (i)
a discussion of general (multi-field) variational principles,
with emphasis on complementary energy, in terms of alternate
stress-rates and conjugate measures of strain-rate, for rate-
sensitive inelastic materials, (ii) rate complementary energy
potentials, for the chosen stress-rates, for rate-dependent as
well as rate-independent materials, (iii) criteria for unique-
ness and stability of solutions, (iv) numerical study of neck-
ing of an initially perfect elasto-plastic bar, and (v) numer-
ical study of post-buckling of an axially compressed plate of
semilinear isotropic elastic material, undergoing large rota-
tions, as well that of a thin plate undergoing large displace-
ments due to transverse loading.
30

PRELIMINARIES:
We use a fixed rectangular cartesian coordinate system, and
employ the notation: (_) denotes a second-order tensor; (_)
denotes a fourth-order tensor; (-) 1mplies a vector; ~ = A.b
implies a. = A .. b.; A.b implies a product such that (A.B) ..
1 1)) - - - - 1)
= A .. B. k ; A:B = trace (AT.B) = A .. B .. ; and u.t u.t .. A parti-
1J J - - - - 1J 1J - - 1 1
cle in the undeformed body has a position vector ~ = Xa~a(a=1 .. 3)
where e are unit cartesian bases. The gradient operator Vo
-a -
in the undeformed configuration C is VO = (e a/ax ). The po-
o - -a a
sition vector of the particle in the deformed configuration,
N
1-1.. The gradient operator in CN is -v =
say CN' is Lv = y.e
(~ia/aYi). The deformation gradient tensor is ~ = (ZO~)T,
such that F.1a y.1, a = (ay./ax
1 a ). The nonsingular F has the
polar-decomposition, ~ = 0. (!+9) where the rotation a is or-
thogonal and the stretch h is symmetric and +ve definite. The
Green-Lagrange strain is g T T T
1/2(~ .~-!) 1/2(~+~ +~.~)
o T -
where e = (Z ~) , ~ = ~-~.

For the present purposes, we introduce the stress measures (i)

the "true" Cauchy stress !; (ii) a weighted tensor, the Kirch-
hoff stress tensor 0
_ = JT where J is determinant of matrix
N

[yo ]; (iii) the first Piola-Kirchhoff stress tensor t; (iv)

1,a -
the second Piola-Kirchhoff stress tensor ~, and (v) the sym-
metrized Biot stress tensor (or what is also often referred to
as the symmetrized Lure', or the Jaumann stress tensor)~. As
discussed in [1,5], and elsewhere, the above stress measures
are related as:
1 T 1
J~.~. ~ = J:: (1)
-1 -l-T
t J(~ .!); s = J(~ .!.~ ) ( 2)

r (3)

The tensors !, ::' ~, and r are symmetric, wh11e t is unsymme-

.
tr1c. In the a b o vF-eT ,_(F'-l)T an d th e superscr1pt
. T d e-
notes a transpose.

RATE FORMULATIONS:
Now, we consider the (incremental) rate analysis of finite
strain problems of an inelastic solid with a rate-sensitive
constitutive law. In doing so, one can choose an arbitrary
reference frame. In practice, however, two choices, one the
so-called total-Lagrangean (TL) and the other, the so-called
 31

updated-Lagrange an (UL) reference frames are appealing. Even-

though the choice of a reference frame does not, per se, af-
fect the theoretical or computational approaches, we discuss
the details of a UL formulation, since the rate consitutive
relations of an inelastic solid depend, naturally, on the cur-
rent state of true stress.

In the UL formulation, the solution variables in the generic

state CN+ l are referred to the configuration of the body in
the immediately preceding state, CN' which is known. In the
UL formulation, one is essentially concerned with an initial
stress problem: the initial "true" stress in CN is the Cauchy
stress !N, while the initial displacements in CN as referred
N
to CN are, obviously, zero. Let Yi be the current spatial co-
ordinates of a particle in CN. Let yN be the gradient operator
in CN (ie., vN=e.a/dy~) and let Ii be rate of deformation (ve-
- -1 1 -
locities) from CN. We define the rate of displacement gradi-
V u ) T an d wr1. t
en t e· = (U. e ·e = s. + w. were
h s• [ s• .. = -2 au. I'ay.
1 ('" N +
- - - - - - - -lJ 1 J
. I ay.N)].1S t h e symme t r1C
au. . .
UL straln-rate an d w
• [w
• .. = -2 • I ay.N
1 ( au.
J IN - 1J 1 J
-au./ay.) is the skew-symmetric spin-rate. Let i, and
J 1 -
£(=IN[!+(~:~)~N]) (where IN=polpN, Po and PN being the mass-
densities in Co and CN' respectively), be the substantial de-
rivatives of the Cauchy and Kirchhoff stresses respectively.
As is well-known, these stress-rates are not objective. Let t,
~, and! represent the appropriate stress rates referred to CN;
N+l N N+l .
ie., for instance, s6t = ~N -~ where ~N 1S the second
Piola-Kirchhoff stress in CN+ l referred to (and measured per
unit area in) CN. It is shown in [5] that:

S (£_~.7N_qN.~T)/JN; t (~_~.7N)/JN (4a,b)

t (5a,b)

Unless large elastic deformations of a dilatational nature

have preceded the inelastic straining, one may, without sig-
N
nificant error, assume that J ~l.O.

The equations of linear momentum balance (LMB), angular momen-

tum balance (AMB), compatibility, and traction and displace-
ment boundary conditions (TBC and DBC) in the UL rate formu-
lation can be written as:

o (6a,b)
32

AMB: (7a,b)
or, equivalently,
. N • N . N .T
~.:t: + ~.:t: +~ T • W (7c)
compatlbility:

(Ba,b)

TBC: (9)

OBC: u =u at SUN (10)

Let us suppose, for the moment, that the consitutive law for a
rate-dependent material can be expressed (as shown later in
this paper) in terms of certain rate-potentials, as:

(lla-c)
We consider the Legendre (contact) transformations of the type:

S:E: ~(P s*(§); ~T:~_O(~)

r:E: - Q(P R* (!') (12a-c)
such that
as*/a§

As discussed in [1-6], the AMB conditions are embedded in the

structure of Wand O. As shown in [5], the complementary en-
ergy principles, and the Hellinger-Reissner type principles,
involving the stress-rates §,~, and!, are as below. In each
case the functional whose stationary condition is the prin-
ciple in question is given. The respective functionals are,
denoted by TI with the subscript C and HR denoting complemen-
tary and Hellinger-Reissner type functionals, respectively.

TIc(~'~) = JC {-s*(§) - !!N:[ (~N~). (~N~)T]}dv + JC t.~ds

VN SUN (l3)

TIHR(~'§) = J: VN
{-s*(:;;) + pN~·~+hN:[(2N~).(~N~)T]

+ !§:[(~N~) + (~N~)T]}dv - JC E.uds - ) ( t. (~-~)dS (14)

SaN SuN

TIC(~) f VN
-E* (t) dv + [
SUN
!.~ ds ( 15)

TIHR(~'~) = f VN
{-E*(t)dv p
N·
~.~
.
+ e: [(~N~) T] }dv + contd.
 33

JC t.uds - JC ~. (~-~)dS (16 )

ScrN SuN

ITc(~,0) = JC {-R*(~) + ~!N: (0 T .0) - t T :0}dv + JC t.~ds (17)

~ S~
N· ·T N.]T.T.
p B.u + t :[~ ~ -t :~}dv

(18)

In the above VN , ScrN' and SUN are the volume, prescribed-trac-

tion boundary, and prescribed displacement boundary, respec-
tively, of the solid in CN; and pN and ~, are respectively,
the mass-density and rate-of-body-force in CN. The above
functionals are valid, in general, for non-conservative (de-
format~on-dependent) surface tractions. If A is the set of
conditions that are satisf~ed a priori, and B is the set of
those conditions that are satisfied a posteriori in the vari-
ational principle, for each of the above functionals the sets
A and B are as follows: (i) Eq. (13): Set A(Eqs. 6a, 7a, and
9), set B(Eqs. 8b, and 10) (ii) Eq. (14): Set A (the exist-
ence of S* such that dS*/d~ = ~, and Eq. 7a), set B(Eqs. 6a,
8b, 9, and 10). (iii) Eq. (17): Set A(Eqs. 6b, 9, the de-
finition of t as in Eq. 5a and that w is skew-symmetric), set
B(Eqs. 7c, 8a, and 10); (iv) Eq. 18: Set A(the definition of
~ as in Eq. (5a), and that w is skew symmetric), set B(6b, 7c,
8a, 9, and 10).

The complementary, and Hellinger-Reissner type principles as

through Eqs. (17) and (18) were first stated by Atlur~ [5].
The general ~nvalidity of the principles through Eq. (15)
which was alluded to by Hill [11], and Eq. (16), were discus-
sed in [5]. Eventhough, Eqs. (13) and (14), and the attendant
variational principles, may be viewed as being consistent, the
limitations of practical applicability of these are discussed
in [5]. Especially, in the application of Eq. (13), the need
to select a symmetric §, such that Eq. (6a) (which involved
coupling with ~) is satisfied, a priori, is not an altogether
easy proposition. Several interesting ways of satisfying Eq.
(6a), and of applicat~on of Eq. (13), were d~scussed by
At 1 ur ~ [6, 7 , 12 ] .
34

However, the complementary principle of Eq. (17), ~ntroduced

in [5], has several attractive features for practical appli-
cation: (i) the LMB, Eq. (6b) can be easily satisfied by set-
• N .p
ting: ~ = ~ x~+~ where ~ are first-order (once differenti-
able) stress functions (ii) @can be selected to be skew sym-
metric, by setting w ij = ekijQk where e kij is an alternating
tensor). In general, even in a TL formulation, the constraint
a.a T = I is easily met by taking ~ to be a function of the 3
Euler-angles of rigid rotation [4]. In the case of plates and
shells, the concepts of a finite-rotation vector, as discussed
later, may be employed.
The application of the complementary energy principle as
through Eq. (17), and its TL rate counterpart, is illustrated
later in this paper.
RATE POTENTIALS FOR RATE-SENSITIVE MATERIALS:
As discussed in [5], and elsewhere, the principle of object-
ivity is met, in writing the rate constitutive law of the ma-
terial, by postulating the constitutive relation between the
objective strain-rate ~ and the objective co-rotational (or
also at times referred to as the Zaremba, or the rigid-body or
the Jaumann) rate of Kirchhoff stress, denoted here by 0*. It
is well-known that,
0* = 0 - 0.~N _ ~N.0T (19)
Thus, in view of Eqs. (4-5),
. • N N .)/ N •
( ~*-:.~ ( ~.*. N N .)/ N
-~.~ J: t -:.~ -~.r:: J
[ • 1(. N N .)]/N
~*-2 ~.~ +~.~ J (21)
Thus, if V is the postulated rate potential for 0* such that
3V/a~ = £*,
we can define:
JNW V N
0 :(~.~): aW!aE § (22)
1 N (.T .)
JNiJ W+ -0
2- : ~ ~ : .
au/a~T t (23)
JN Q V - -01 N
2-
: (~. p; a6/3 ~ = r (24)

From -chese, one can establish 5* (§) , E* (!.) and R*(t) as de-
fined ~n Eqs. (12a-c). Thus, we focus attent~on on the poten-
tial V.
It is worth noting that for materials with rate-independent
constitutive laws, such as classical elastic-plastic materials
 35

the der1vative (.) 1S considered to be with respect to a fic-

tition time. However, for rate-sensitive materials, such as
elasto-viscoplastic and creeping materials, the derivative (.)
is w.r.t. to natural time.

For rate-independent classical elastic-plastic mater1als, Hill

[13] presented the postulat1on:
·
V lL .. a ( • )2
= 2 ijklEijEkl - g AklEkl ( 25)

where L ijkl is a +ve definite symmetric (under ij ~~ kl 1nter-

change) tensor of instantaneous elastic modulii, a = 1 or 0 ac-
cording to whether Akl Ekl is positive or negative, g is a
scalar related to the measure of hardening, and A.. is a tensor
1J
normal to the interface between elastic and plastic domain in
the skl space. Prandtl-Reuss type rate equations of type (25)
for classical 1sotropically hardening mater1als can easily be
derived, formally, to be [5]:

2 (~:S!')S!'
q* 2\1~ + A(~:!)! - l2a\1 (26)
(a' :a') [6\1+2(aF /aW P )]
- - 0
where, A and \1 are Lame' constants, a ' - S!-1/3(S!:~)~ is devi-
ator1c Kirchhoff stress, and the Y1eld-surface is represented
h
by F [3J 2 (a')]2 - F O.
- 0

Here F =F (WP ) where

00
wP = Ja:sPdt; and J
-- 2-
(a') = (1/2)(a':a').
--
A rate-sensitive constitutive law of a considerable generality,
as glven by Perzyna [14], can be eas1ly written for finite
strains, when an associative flow-rule is used, as:
.a
E = y< ¢ F
( ) >'3aF
CJ (27)

where < ¢> denotes a specific function, such that < ¢> ¢ (F)
for F>O, and ¢=O for F~O. The parameter y is called the
fluidity parameter and ~a is the general anelast1c strain. It
has been shown by Zienkiewicz and Coworkers [15], and Argyr1s
and Coworkers [16] that classical, rate-1ndependent elasto-
plastic solutions can be obtained from the above theory, when
(i) either y~oo or (ii) a stationary solution of the visco-
plastic flow 1S sought. Various forms of ¢ were reviewed
by Perzyna [14]. For the Hencky-Mises-Huber Y1eld cr1ter1on,
one can define F to be:

F = [3J 2 ( l! ' ) ] 1:; FO - aeq F (28)

o
36

where 0 is the equivalent Kirchhoff stress. A simple choice

eq
for ~(F) can be:
~(F) = F n (29)
With Eq. (29), the viscoplastic strain rate as in Eq. (27) can
be seen to correspond to the well-known Norton's power law for
steady-state creep when Fo+O. We now derive rate potentials
6V(=V6t),6W, etc. for the viscoplastic constitutive laws given
by Eqs. (27-29).

Let at times tN and t N+6t, the Kirchhoff stresses be oN+06t -

~N+6~, respectively, where ~ 1S the substantial derivative.
The inelastic strain-rates corresponding to Eqs. (27-29) at
times tN and t N+6t, are given, respectively, by:
.a (~a) N
~N -
yF (aF/aq) = y(3/20 ) (0 -F )no' (30)
- N eq eq 0 -

In the above 0' - (qN) , (ie. , superscript N dropped for con-

2
ven1ence) and 0 = (3/2)0':0'. Likew1se at a time 6t later,
eq
.a -1
~N+l
Y (3/2) (0 Ho ) (0 Ho -F )n(O'+60') (31)
eq eq eq eq 0
By stra1ght forward algebra, it can be shown that,

~Na+l = £Na +y (3/20 ){-(60 /0 ) (0 -F )no'+(o -F )n 60 ,

- eq eq eq eq 0 - eq 0 -

+ n{o -F) n-l 0'60 } (32)

eq 0 - eq
However, it can be easily shown that
60 (3/2)[{o':60')/0] (33)
eq - - eq
Using (33) in (32) we obtain:
.a .a + V A
(34)
~N+l = ~N :LlO
wherein the definition of Y is apparent. Since 6q can 1n-
clude the effects of pure spin between tN and t N+ l , one can
replace Eq. (34) by:
.a = £a + V:60* (35)
~N+l -N
where 60* = cr*6t is the corotational increment of Kirchhoff
stress. Now, the corotational rate g* can be written as:
0* = L : (s_sa) (36)
;:e --
Where L is the tensor of instantaneous elastic moduli.
;:e
t N+6t
Thus, 60* = L :6e: - L:
- :::e - e t
f Eadt
-
(37)
N
one may use the approximation,
 37

.a (1 ).a .a t (38)
E = -6 ~N + B~N+l N~t~tN+~t

when Egs. (35) and (38) are used, Eg. (37) becomes,

~a*
~
= L :~E - ~tL : (Ea+Bv~a*)
~e ~ ~e ~N ~ ~
(39)
From which, upon rearranging terms,
~ a * = M:
~ ::; (~E -~ Ea ) ~ ~
( 40)
wherein the definition of ~ is apparent, and ~~a is known and
is given by: ~Ea ~tE~. From Eg. (40) one can immediately
write

~v = ~ijk~~Eij~Ek~ - Mijk~~Eaij~Ek~ (41)

From Eg. (41), the potentials ~W, ~U, and ~Q can easily be ob-
tained through Egs. (22-24).

It is noted that Wang [17] attempted to derive a relation bet-

ween ~s, ~E, ~; a , and hence ~W directly. However, this deriv-
~ ~ ~

at ion appears to be in error, since, among other reasons, the

transformation between the deviatoric part of ~ and the devi-
atoric part of T was assumed to be the same as that between ~

and I themselves.

All the above developments for the UL rate formulation can be

converted to a TL rate formulation by noting the relations [5]:

E' (fN)T.~.fN;~, = ~.fN; ~' = IN(fN)-l.~. (fN)-T

t' IN(FN)-l.~i~' = (1/2) [tN.~,+~,T.tNT+t,.~N+~NT.t,T] (42)

where E' is TL rate of Gree-strain, and ~', ~', t', and~' are
TL rates of e, s, t, and r respectively. Now a' is subject to
~ ~ NT ~
the constralnt that a .a' is skew-symmetric.

UNIQUENESS & STABILITY CRITERION:

In the present paper, the applicatlon of the rate complementary
energy principle as embodied in Eg. (17), or its TL counter-
part, will be used in some computational studies. In the di-
rect application of Eg. (17), the assumed stress field ~ must
not only satisfy the LMB condition within each element, but
also the traction reciprocity condition at the interelement
. + . -
boundary, viz., 0 at PmN (where + and -,
(~*.~) +(~*.~) =
respectively, indicate the two sides of PmtJ ' the interface
between mth and (m+l)th elements in CN). In the present work
this interelement condition is introduced as a posteriori con-
straint, through a Lagrange multiplier ~P at P mN , in the
38

functional of Eq. (17), thus leading to a 'hybrid' finite ele-

ment method. The thus modified functional is:

TrHS(0,t'~p)=ln{1 [-R*(~)+(1/2)~N: (0 T .0)-tT .0]dv

VmN

+ JC (~*.~) .~dS + JC (~*.~) '~pdS (43)

SumN 0 mN

At the point of bifurcatlon, or instability, from the concept

of adjacent compatible states, the following criterion can be
shown to hold:
Tr HS = 0 and CTr HS = 0 (44)
with the constraints:
~
N'
.t =
O.
; IE
.T
= -w n*.t
and u
-
= 0 at S N
um
( 4 5)
In the case of linear pre-buckllng states, the above criterion
reduces to an eigen-value problem, wlth the eigen-value depen-
N
ding on :E •
EXAMPLE PROBLEMS:
Necking of an Initlally Perfect, Plane-Strain, Elastic-
Plastlc Bar.
Cowper and onat [18] examined the above bifurcation necking
problem, for a bar of rigid-plastic work-hardening material,
under uniform tension applied at the ends of the bar. Mises'
Yleld, and isotropic hardening crlteria were used [18]. In
[18] only the eigen-value problem for the applied tension at
which necking would initiate in the bar was treated, but the
phenomenon of post-bifurcation necking was not treated in [18].
L (L) and B (B) are the initial (current) length and width of
o 0 N N
the bar respectively. If Yl and y? are the current carteslan
coordinates of a material particle, the boundary conditions
are: (i) at yN=O: ul=-V; E2=o; (ii) at yNl=L: ul=+V; t 2=O;
N 1 .!..!.
(iii) at Y2=~(B/2):tl=t2=O.
In the present analysis, the complementary energy formulation
based on the TL rate equivalent of Eq. (17) was used. The
problem parameters used are: (B /L ) = (1/3); T (yield
4 0 0 y
stress) = 4xlO psi; the true stress versus logarithmic strain
(In(l/l » curve was assumed to be bilinear, with the two
o 7 4
slopes, E=lO psi, and h = 5xlO psi. The notation
 39

n =
(L-L )L ~ is used. The eigen-value solution for n at bi-
o 0
furcation-necking for the perfect bar, as obtained in [18]
for the presen~ linear-hardening but rigid-plast1c material,
. C
1S n 0.4B.
TOTAL DEGREE OF
FREEDOM 3/2

E =10
1/=03
O"y=40000
12
Mlx BI~TlON H = 50000
LOAD POINT
11
--- FUNDAMENTAL SOLUTION
10 - NECKING SOLUTION

O~~_~ __~__~~~
o 01 02 03 04 05 06
7]-(L-L ol/L o 02 04 067]=LL~o

F1gure 1 Figure 2

The variation of total applied load with n is shown in Fig.

1, from which it is seen that necking starts at nC = 0.4B2
(which is 1n excellent agreement with the value of n C = 0.4B
of [18]). It is also seen that necking starts after the
maximum load is attained. The convergence of nC with the
finite element mesh has been reported elsewhere [B,9], with
the mesh as shown in Fig. 2, which is used to obtain the
remainder of the reported results, being the finest mesh
reported in [B]. The variation of the width reduction ra-
tio, (B-B )/B is shown in Fig. 2, from which it is seen that
CO 0
at n = n the width reduction becomes much more pronounced
as compared to the rigid-plastic fundamental solution.
The variation of o/L (with 0 being defined as the difference
o
of widths at loading edge and the necking sections, respec-
tively) with n is shown in Fig. 3. The slope of this curve
at the beginning of necking, viz, n = nC = 0.4B, was obtained
in an asymptotic analysis in [lB]. The present result for
this initial slope is in excellent agreement with [18]. How-
ever as necking develops, the slope (an/ao) decreases from
the initial value at n = nC , which appears to be in contra-
40

diction with the result of McMeeking and Rice [19]. The

necked profile of the bar for n > nC , are shown in Fig. 4.

z
:;{
1/=0.3 TJ
0:
1-.6 Lo =12# 5931%
If) /;57.61%
Bo =4' 05
-l4 E =10
7
pSI
Ilr5536%
<1:. 5286%
z CTy=40000 pSI 04
~.2 h = 50000 PSI
z
O,~--~--~~-------- or 48~%J I

o <l
I

02 .04 .06 o/Lo 02 04 06

F~gure 3 Figure 4
The ~ecked profiles of the bar, and the progressive develop-
ment of unloaded regions (shaded) are shown in Fig. 5 at
various values of n. Note that unloading begins at the cen-
ter of the loaded face of the bar at n = nC = 0.482. Final-
ly the distribution of Cauchy stresses, '11 (in the direc-
tion of loading) '22 (in the width loading), and '33 (in
the thickness direction of this plane-strain specimen), at
the neck (y~ = L/2) are shown in Fig. 6. These results are
in excellent qualitative agreement with those of Needleman
[20] who also analyses the necking and post-necking problem
of an initially perfect cylindr~cal bar. It ~s noted that

A \ =59.31%
8: \=56.86%
A
~111
2.0 8
..!<
0 ..!< ;i2 ..!< ;i2 0
I.D
0
I.D 0 0-
eo ::: rI') ~ I.D ~
N
If)
""
If)
If)
If)
I.D
If)
I'-
If)
eo
If)

" "
t=:" "
t=:" "
t=:" "
t=:" "
t=:"
A

Figure 5 Figure 6
 41

the problem of necking of an elastic-plastic bar, with

1nitial imperfect1ons, was also analysed by Osias [21),
McMeeking and R1ce [19), and Nemat-Nasser and Taya [21).
It is observed that the finite element meshes used in [19-
21) are very much finer than the one used presently. Even-
though a precise mathematical statement as to this appears
impossible, the above comparison of the mesh-sizes appears
to indicate the relat1ve advantages of the present comple-
mentary energy method.

F1nite Deformation and Post-Buckling Analyses of a Thin

Plate:
We consider large deformations (large rotations and large
stretches) of a thin plate made of a semi-linear isotropic
mater1al (ie., a material exhibiting a linear relation be-
tween the stretch tensor h and its conjugate stress-measure,
the symmetrized Biot-Lure' or Jaurnann tensor r). We invoke
the well-known Kirchhoff-Love type plausible deformation
hypotheses that a normal to the m1dplane of the initially
flat undeformed plate remains normal to the deformed mid-
plane and that there is no thickness stretch. In order to
derive a consistent complementary energy princ1ple for
this constrained deformation problem, we start with the
general (Hu-Washizu type) variational principle involving
~, g, ~, and ~ as variables.

By introducing the appropriate approximations to all these

variables, we derive a general variational principle for
the plate problem; from this we proceed to construct a con-
sistent complementary energy principle for the plate prob-
lem. It 1S shown in [1) that a general functional, for
the three-dimensional finite elasticity, whose stationary
conditions lead to all the appropriate field equations, is
given by:

-Is GO
t.uds (46)
42

where W(h) is the strain energy density per unit of initial

volume V , as a function of pure stretch hiP is mass density/
0 - 0
unit initial volume, and ~ are body forces/unit mass.

Let xl' x 2 be cartesian coordinates in the mid plane, and x3

normal to the midplane, of the plate. The position vector of
an arbitrary material point in the undeformed plate is x=x.e.
- 1-1
(i=i, ... 3). Under the present deformation hypotheses, the
posit1on vector of the same particle after deformation is,
y = (xl+ui)~l + (x2+u2)~2 + u~~3 + x3~ where ~ is a unit nor-
mal to the deformed midplane. Further, the displacement u~
1
(i=1,2,3) are functions of xl and x 2 only. Thus the displace-
ment of an arbitrary particle in the plate, is y-~ = ui~l +
u2~2 + u~~3 + x3(~-~3)· The base vector at an arbitrary point
in the deformed plate are given by:

d via xa
...
(cS
la
+u~ )e.+N, x3=G
1,a-1-a a
(a=1,2i i=1,2,3) (47)

The deformation gradient is

T
F = (yy) = ~a~a + ~~3 (a=1,2) (48a)

Further, for the present kinematic hypotheses, we assume thp

stretch tensor to be:

11.: h e e .
af)-a- 13' ha13 = ha ~x.)
1 [i=1, ••• 3i a, 13=1,2] (48b)
From Eqs. (48a,b) it is seen:

N = ~·~3 = [cz. (:~+Q) ] . ~3 ~·~3 (49)

Thus the displacement vector can be written as:

~(x.) = u~(x)e.+(a-I).e3x3 [i=1,2,3i a=1,2] (50)

1 1 a -1 - - -

Further, we assume that h Q(x,) can be approximated as:

afJ 1

h a13 (x i ) = h~13(xa) + x 3 XaS (x cS ) [a,S,y=1,2] (51)

ie. _ + x 3-X
h_ = h*

For the semilinear isotropic material we assume the constitu-

tive law:

r 21l~ + \(h: I) I (52)

 43

Since, for isotropy, !1, g, and E' are coaxial, Eg;. (3) becomes,
r = t.a (53)
The tensors t and a are assumed to be:
t = t .. e . e .; a =a .. e . e . [ i, J' =1,2,3; 1 (54)
lJ-l-J lJ-l-J
where, t .. t .. (x l ,x 2 ,x 3 ); a .. = a .. (x l ,x 2 ) (55)
lJ lJ lJ lJ
Finally, the external forces distributed on the plate are as-
sumed to be specified per unit area on the mid plane of the
plate to be g.1 = g.1 (x a ) (i=1, .. 3, a=1,2). When the assumptions
in Eqs. (50-55) are substituted in Eq. (4-6), we find through
atraight-forward algebra, that

G [ui, h~S' XaS ' a ij , Tal' Mail

Is
1T

{"1'1*(!1*,c:) - 9:.~* + ~T:[~a~a e

+ -u*,a-a
o

+ (<;<·~3)~3 - Q:.(;r:+b-)l + ~T:[(<;<·~3),a~a - <;<·X]}ds

-f Cu
[!'. (~*-~*) + ~.< (<;<-~) ·~3> ldc -1Co
[,!:.~*+~. (~-;r:) ·~3ldc
(56)
where ~ = M .e e
al-a-l
(a=1,2; i=1, ... 3), and,

M .
al

and (57)

In Eq. (56), So is the area of the undeformed mldplane, and

C and C are the displacement and traction prescribed boun-
u 0
daries of s. It is seen that only tai enter into the above
energy expression due to the presently invoked deformation as-
sumptions. The constitutive equations and LMB conditions ob-
tainable from Eq. (57) are:

ClW*/Cl!1* = ~.a -
Band ClW*/ClX = M.a - N
-- (58)
and T. + g. = 0 (a=1,2; i=l, ... 3) (59)
al,a 1

When Eqs. (58, 59) and the appropriate traction boundary

conditions on T . are satisfied a priori, one can eliminate
al
from Eq. (56), (i) ~* and c: through the usual contact trans-
formations and by establishing a complementary energy density
R* such that ClR*/d~ = ~* and dR*/d~ = 0' (ii) u* through
satlsfYlng (53) a priori. When this is done, we obtain a
44

complementary energy functlonal:

rrc(~, ~, ~) = JC-R*(~,~) + ~T:[~a~a+(~·~3)~3-~]

s

+ ~T:[(g.~3) , a~a])ds- .I rc ('!>~*+~.< (C:-C:-).~3»dc

u

-i Co
8· (g-p ·~3dc. (60)

In Eq. (60), g is required to be orthogonal and further g is, as

assumed in Eq.(55), a function only of xl and x 2 • Also the
variables Rand N in Eq. (60) are assumed to be deflned ln
terms of ~: ~, a~d a as in Eq. (58). To assume an orthogonal
a(x.,x 2 ), the concept of a flnite rotatlon vector [23] is use-
- 1
ful. Let w be the finite angle of rotation around an arbitra-
rily orlented unit vector e in the midplane of the plate. The
finite rotation vector is defined to be:
n = (sinw)~ (61)
The action of finite rotation n on a vector V can be written
[23] as the transformation of V to Y* as,

V* = V + n x V + [nx(nxV]/2cos 2 (w/2) = g.V (62)

- - - - -- 2-
where g = ~ + ~ x ~ + [(~x~).(~x~)]/2cos (w/2) (63)
It can be shown that g of Eq. (63) is orthogonal, ie.,
Cb<J.T = I.

The vector ~ in Eq. (61) can be written as:

e = ~lcOS8 + ~2 sin8 (64)

Thus, the rotation tensor g of Eq. (63) is a function of two

parameters: 8(x l ,x 2 ) and w(x l ,x 2 ). The explicit expreSSlon
for g can be shown to be:
g=<1-(1-COsw)sin28>~1~1+< (1-cosw)sin8cos8>~1~2
+ (sinwsin8)~1~3 + (1-cosw)sin8cos8~2~1
2
+ <l-(l-cosw)cos 8>~2~2 - sinwcos8~2~3 - sinwsin8~3~1

+ sinwcos8~3~2 + cosw~3~3·
In a Von-Karman type plate theory w is assumed to be moderate-
ly large, such that cosw ~ 1-(w 2 /2) and sinw = Wi while the
angle 8 can be assumed to be arbitrary.

Further details of the analysis of large rotations and

stretches of thin plates uSlng the complementary energy method
sketched above, which are omitted here for space reasons, will
 45

be reported elsewhere. In the following we present some re-

sults using a hybrid-stress finite element method based on Eq. (60) .

Post-Buckling of a Thin Plate:

The problem is that of a simply supported square plate [-a/2 ~

xI,x2~+a/2] that is subject to uniform uniaxial compression at

edges xl = ±a/2. Only a (2x2) mesh in a quarter of the plate
was used. The pertinent geometrical and material-property
data are indicated in Fig. 7.

In the lower half of Fig. 7, the maximum transverse displace-

- 2 2
ment (vJ It) is shown magnified for lower values of (p a lEt )
max x
where P is the aXlal compressive stress. In the present analy-
x 2 2 -5
SlS a small uniform lateral load of (Pa lEt ) = 2xl0 was ap-
plied along with the axial Px . The linear buckling load pred-
icted by Levy [24] is (P x a 2 /Et 2 ) = 3.6. The post-buckling dis-
placement W as obtained by Levy [24] is also shown in both
max
the parts of Fig. 7. It is seen that the present results a-
gree excellently wlth those in [24] upto displacement levels of
(1;vmax /t) e> 3.5 .

......
---ro 'X 2

Efj_
X
1 5 .5
l DIRECTION OF
4 L COMPRESSION
14-.5=.,j -XI .
3 8/2 BOUNDARY CONDITIONS
PRESENT FEM .....,
2 ~
'-..
UI =U2"'W=0. xl=~a/2or x2=±a/2
-- LEVY(NACA REP. 737) X Wx -0 x 2 " ±a/2
ro
0 Px i/Et2 -l Wy=O XI = ~a/2

E =ld
...... 0 8 12 16 20 24
----ro
X
E =107
V-0316
a = 1.0
i ~ PRESENT
• LEVY V=0.316 20
t =0.01
06 a=1.0
16
0.4 t =001
12
P=0.022 • LEVY( NACA REP.737)
02 08
~ PRESENT FEM.
0 Pxa2/Et2 O~ Pa4/Et
0 2 3 4 5 0 80 160 240 320 400

Flgure 7 Flgure 8
46

Large Deformation of a Transversley Loaded Plate:

The problem is that of a semilinear, isotropic thin elastic
simply supported plate that is subject to uniform transverse
pressure P. The pertinent geometric and material property
data is shown in Fig. 8. The presently computed variation of
the maximum transverse displacement Wmax with P is seen to a-
gree excellently with that of Levy [24).

The results presented in this paper, and further comFleted

work to be reported elsewhere, appear encouraging to warrant
us to pursue this line of thought in other areas of nonlinear
continuum mechanics. Our current studies in this area will be
reported on shortly.

Acknowledgements:
The research reported herein has been supported in parts by:
NASA under grant NASA-NAG3-38; by ONR under contract number
N00014-78-7636; and by AFOSR under contract number F49620-78-
C-0085. The authors gratefully acknowledge these supports.
The authors express thelr sincere thanks to Ms. Margarete
Eiteman for her untiring efforts in the preparation of this
typescript.

References:
[1) Atluri, S.N., and Murakawa, H., "On Hybrid Finlte Ele-
ment Models in Nonllnear Solid Mechanics", in Finite
Elements in Nonlinear Mechanics, Vol 1 (Ed. P.G. Bergan,
et all Tapir Press, Norway, 1977, pp 3-41.
[2) Murakawa, H., "Incremental Hybrid Finite Element Meth-
ods for Finite Deformation Problems (With Special Em-
phasis on the Complementary Energy Principle) Ph.D.
Thesis, Georgia Institute of Technology, Aug. 1978.
[3) Murakawa, H., and Atluri, S.N., "Finite Elasticity
Solutions Using Hybrid Finite Elements Based on a
Complementary Energy Principle" Journal of Applied
Mechanics Trans. ASME, Vol 45, 1978, pp 539-548.
[4) Murakawa, H., and Atl'Uri, S.N., "Finite Elasticity
Solutions Using Hybrid Finite Elements Based on a
Complementary Energy Principle - II. Incompressible
Materials", Journal of Applied Mechanics, Vol 46, 1979,
pp 71-78.
[5] Atluri, S.N., "On Rate Principles For Finite Strain
Analysls of Elastic and Inelastic Nonlinear SOllds",
ln Recent Research on Mechanlcal BehaVlor of SOllds
(Professor H. Miyamoto's 60th Anniversary Volume) Univ.
of Tokyo Press, Tokyo, Japan 1979, pp 79-107.
[6) Atluri, S.N., "On Some New General and Complementary
Energy ~heorems for the Rate Problems of Finite Strain,
Classical Elastoplasticity" Journal of Structural
Mechanics, Vol. 8, No.1, 1980, pp 61-92.
 47

[7] Atluri, S.N., "Rate Complementary Energy principles,

Fin~te Strain Plasticity Problems, and Finite Elements"
in Proc. IUTAM S osium on Variational Methods North-
western Univ., Sept. 1978 S. Nemat-Nasser and K.
Washizu, Eds.} Pergamon Press, 1980 (In Press).
[8] Murakawa, H., and Atluri, S.N., "Finite Element Solu-
tions of Finite Strain Elastic-Plastic Problems, Based
on a New Complementary Energy Rate Principle" in Ad-
vances in Computer Methods for Partial Differentia!
Equations (R. Vishnevetsky and B. Stepleman, Eds.) IMAS,
Rutgers Univ., 1979, pp 53-61.
[9] Atluri, S.N., "Finite-Strain Plast~c~ty Computations
Based on a New Complementary Rate Principle" Presented
at Int. Conf. on Finite Element Methods, Chalmers Inst.
of Tech. Goteborg, Sweden, Aug. 1979.
[10] Murakawa, H., Reed, K.W., Atluri, S.N., and Rubenste~n,
R., "Stability Analysis of Structures Via a New Comple-
Mentary Energy Method" Proc. of Symp. on Computational
Meth. in Nonlinear Structural and Solid Mechanics, (A.K.
Noor, and H.C. McComb, Jr., Eds.) Washington, D.C.,
Oct. 6-8, 1980 (In Press) •
[11] Hill, R., "Some Basic Principles in the Hechanics of
Solids Without a Natural Time", J. Mech. & Phy. Sol~ds,
Vol 7, 1959, pp 209-225.
[12] Atluri, S.N., "On the Hybrid Stress Finite Element
Model in Incremental Analysis of Large Deflection
Problems", Int. Jnl. Solids and Structures, Vol 9, 1973,
pp 1177-119l.
[13] Hill, R., "Eigenmodal Deformations in Elastic/Plastic
Continua" Jnl. Mech. and Phys. Solids, Vol 15, 1967,
pp 371-386.
[14] Perzyna, P., "Fundamental Problems ~n Visco-Plast~city"
Jnl. Mech.and Phys. Solids, Vol 9, 1966, pp 243-377.
[15] Zienkiewicz, O.C., and Cormeau, I.C., "Viscoplasticity-
Plasticity and Creep in Elastic Solids - A Unified
Numerical Solut~on Approach" Int. J. Num. Methods ~n
Engg., Vol 8, 1974, pp 821-845.
[16] Argyris, J.H., and Kleiber, M., "Incremental Formu-
lation in Nonlinear Mechanica and Large Strain Elasto-
Plasticity" Compo Meth. in Appl. Mech. Engg., Vol 11,
1977, pp 215-247.
[17] Wang, H.C., "Finite Deformation Elastic-Viscoplastic
Analysis By Finite Element Method" Ph.D. Thesis, Mas-
sachusetts Institute of Technology, Feb. 1980.
[18] Cowper, G.R., and Onat, E.T., "The Initiation of Neck-
ing and Buckling in Plane Plastic Flow", in Proc. 4th
U.S. Congress of Applied Mech., ASME, N.Y., 1962,
P 1023.
[19] McMeeking, R.M., and Rice, J.R., "Finite Element Formu-
lations for Problems of Large Elastic-Plastic Deforma-
tions" Int. Jnl. Solids and Structures, Vol 11, 1975,
pp 601-.
[20] Needleman, A., "A Numerical Study of Necking in Circular
Cylindrical Bars", Jnl. Mech. Phys. Solids, Vol 20,
1972, p Ill.
[21] osias, J.R., "Finite Deformation of Elastic-Plastic
Sol~ds: The Example of Neck~ng ~n Flat Tens~le Bars",
Ph.D. Thesis, Carnegie Mellon Univ. 1972.
48

[22] Nemat-Nasser, A., and Taya, M., "Model Studies of

Ductlle Fracture-Part II. Further Numerical Formu-
lations", in Finite Elements in Nonlinear Mechanics,
Vol 1 (P.G. Bergan et al Eds) Tapir, Norway, 1977,
p 21l.
[23] Lure', A.I., "Analytical Mechanics" (In Russian)
Moscow, 1961.
[24] Levy, S., "Bending of Rectangular Plates with Large
Deflections", NACA Tech. Note 846 and NACA Tech
Report 737, 1942, P 139.
Formulation of Contact Problems by Assumed Stress Hybrid
Elements

'IH.H. PIAN, K. KUBOMURA

Massachusetts Instttute of Technology, Cambndge, USA

Summary

A modi fled complementary energy principle has been derlved

by introduclng the contlnulty at contact surface as the con-
dltlon of constralnt and the reactlons at contact surface as
additional fleld varlables. In the Solutlon procedure an
assumed contact surface is introduced and lS divided lnto
elements. An lterative procedure lS used to obtaln the cor-
rect location and extent of the contact surface. Example
problems are limlted only to elastlc bodles under small de-
formatlons but are lncludlng extenslve slldlng between sur-
faces WhlCh may be elther frictlonless or frlctional.

Introductlon

Th~ assumed stress hybrld method has been demonstrated to be

very competltive to the conventlonal assumed dlsplacement
method ln finite element analyses. One advantage of the ap-
proach lS on its more accurate stress evaluatlon. Such ad-
vantage would be ampllfled for problems for which the solu-
tions are dependent on accurate evaluation of stresses. The
contact problems for WhlCh the behavlor of frictlonal slid-
lng depends on the contact pressure also belong to thlS cate-
gory.

This paper presents an lncremental flnite element formulation

of contact problems WhlCh may include extensive frictional
Slldlng such that node-to-node contact lS no longer malntain-
ed. The formulation is based on the introduction of contact
reactions at the contact surface as independent variables ln
a Lagrange multiplier method.
50

INCREMENTAL FORMULATION
Varlatl0nal principle for incremental finite element
methods are described in Ref. 1. Modifications of variation-
al principles for relaxing continuity conditions along
interelement boundaries have been used for the derivation
of many hybrid finite element models [2].
For th8 contact problem, consider two bodles A and
B which are divided into finite elements Vn where n refers
to the nth element. The boundary surface aV n of a typical
element may be composed of So n , Su n ' Sand
n Sc n ' which
are respectively, the surface with prescribed tractions,
the surface with prescribed displacements and the inter-
element surface and contact surface between A and B. The
entire contact surface is discretized into contact element
SCm" A modified variational principle which includes the
conditions of continuity and equilibrium on the contact
surface and for which the stresses satisfy the equilibrium
equation is

..
11
mc [[1
n V
flu
"..
dS +
n

stationary (1)

in which
£. = stresses
T = element boundary tractions
 51

...u element boundary displacements

A reactions at contact surface
~ = coordinate of material point prior to
deformation
The superscripts A and B refer to bodies A and B respec-
tively. With 6~, 62 and A as variables, and with
a satisfying the equilibrium conditions, the Euler's
6 ...,
equations of this variational principle provide, in addi-
tion to the compatibility conditions in Vn , the following
equations along the element boundaries:

on interelement
boundary between
Va and Vb

T + 6 ..,
..- T = if + 6 if on Sa
n

J
(T + 6 T)A + A 0

(T + 6 T)B - A 0

-A A
and u + 6 -A
u + x -B
u + 6 ~B + x (2)

These are, of course, the equlllbrlum condltlons

along the element boundaries and the condition of continu-
ity along the contact surface. For a contact problem with
frictional sliding, the following condition should also be
satisfied:
(3)

where the reactions at the contact surface are resolved

into the normal and tangential components, A and A ,
n s
and I.L is the coefficient of friction.
In the finite element implementation, the stress
increments are expressed as
(4 )

from which
R 6 e (5)
52

The boundary displacements are interpolated as

u = L q (6)

and
(7)

In the finite element implementation, the total contact re-

actions are actually resolved into their normal and tan-
gential components and are interpolated in terms of their
nodal values, !,
A= M t (8 )

· 1 acenents ~A-B
Th e d ~sp ~ and Q over each contact
element Sc are then interpolated in terms of the nearby
m
nodal displacements qA and qB of the neighboring elements
of the contacting bodies.
The funct~onal 1Tmc ~s in the form of

1T
mc L[ ~ 8 ~ T ~n 8! -
n

where

(lO)
 53

In the above functlonal ~mc the lncremental stress

parameters 8 ....,B are independent of those of other elements •
Thus, by setting 0 ~ mc = 0 one can obtain a set of equa-
tions which express 8! in terms of 8 q. Then, by elimina-
ting 8 B and by summing up matrices over all elements one
obtains

AT JA _
- ( 8q (11)

where
K = [ G T H -1 G
_n _n _n
n

8Q = '[ 8~n
n

R = Ln ~n Re Lm R
e
-m

~ LJ
_m
A JB
l J B
_m
(12)
m m

Setting o ~ me = 0 yields the matrix equation for the

finite element solution

,... + .....
K 8 q 8 Q R

(13 )

o
"'"
-
t
54

Iterative Procedure

Once the contact bodies are adequately constrained such that

the inverse of the global stiffness matrix, K- l in Eq. (13)
can be evaluated, it is used throughout the solut10n process.
In locating the contact surface only ~A and JB need to be re-
computed in each iteration. Even in the case of material
and/or geometrical nonlinearities, it is poss1ble to use a
modified Newton-Raphson method, hence to keep the global
st1ffness K constant during the iterative process for each
increment.

For a two-dimens1onal problem, before each iteration, the

contact surface is a line fixed in the coordinate system,
but not to the contacting bodies. Such a line is assumed
known hence integrals over Scm can be evaluated. The dis-
placement increments are then solved using Eq. (13) and new
range and location of the contact surface can be determined
for further 1teration. It has been found that, for best re-
sults, the length of the contact element should be the same
as the length of the contacting elements.

The procedure for solving the contact problem is as follows:

First an increment in the external load or prescribed dis-
placement is applied. Second, a contact surface is assumed
together w1th the points at which the nodes of the bod1es
are in contact. Also, the types of contact (s11ding or non-
sliding) at each contact p01nt are assumed. For the init1al
calculation of the first load1ng 1ncrement the above choice
is made simply by inspection, and for the first iteration
after each new loading increment, the converged solution of
the previous loading step is used. Th1rd, at the 1-th itera-
t10n of the N-th loading step, incremental displacements
~u. and contact reactions A. are determined.
_1
Fourth, know-
~1

ing the total d1splacement ~N-l at the end of the previous

loading step, the total displacements uN_l+~u.
_ ~1 on the boun-
A are checked to determLne
dary and the contact react10ns _1
1f the cond1tions of contact are satisfied. If these con-
ditions are not satisfied, the extent of the assumed contact
surface is modified for further iteration.
 55

To determine if the solution satisfies the conditions of con-

tact, the followlng assurances are made:

(1) That nodes of either body, beyond the last contacting

nodes from the previous iteration have not penetrated
the other body.
(2) That tractlons at the contacting nodes are compressive.
These normal tractlons can be calculated by three dlf-
ferent methods; (a) from the stress coefflcients,
(b) from the equlvalent nodal forces, and (c) from the
contact reactlons,~. Here the last method was used.
(3) That the relationship between normal and tangentlal
contact reactions,

is satisfied.

Dependlng on which of the above checks, if any, lS violated,

one of the following procedures is employed to modify the
assumed location of the contact surface.

(a) If (1) is violated, the contact surface may be extended

to include the points at which penetration has occurred.
(b) If (2) is violated, the contact surface is reduced by
excludlng nodes at which the tractions are tensile.
(c) If (3) lS violated, sllding is allowed to occur.

After the conditlons of contact are satisfied, a test for

convergence can be made by calculating the following quantity
from a representing nodal displacement ~u,

If R is less than a prescribed quantity, say 0.01, the solu-

tlon lS consldered as converged.
56

EXAMPLE SOLUTIONS

Example solutions have been carried out for problems of con-

tact between a desk and a semi-infinite half-plane. The
relevant dimensions and the fin~te element pattern are shown
in Fig. la, with the area

R 15 mm
x IT
I
J-V V -
~
.- 1
I-
1 I
I
I

128 mm

- - - ~

I""" -
.....- ~
(~)
153.6 mm
Overall Mesh Pattern ----- (b) Mesh Near Contact
Surface
F~gure 1 Elast~c Half-D~sk and Half-Plane Contact
(E = 21,000 Kg/mm2, v = 0.3)

immediately surround~ng the contact surface shown in great

detail in Fig. lb. The semi-infin~te half-plane has been
modeled by a finite one w~th overall dimens~ons much larger
than those of the disk. The basic element used are four-node
quadrllateral element derived by assuming seven S-parameters
and linear displacement distrlbution along each edge. Five-
node and slx-node elements are also introduced in transition
regions between coarser and finer meshes. Contact tractions
along each contact element are approximated by linear inter-
polations.

Non-Sl~ding Contact

Problems are solved for the case wlth both applled loads and
prescrlbed dlsplacements at the top of the disk. In the
problems, the ratio of Young's moduli are varied over a
 57

-4
range from 1 to 10 and Sllghtly different mesh patterns near
the contac~ surface are used to accommodate for node-to-node
and node-to-internode contacts. Loads or displacements are
appl1ed by three increments until the length of contact sur-
face becomes about 2.4 mm. For each 1ncrement, the converged
solutions are reached with three or four iterations. For
these solutions, the best results for contact tractions are
obtained when calculated from equivalent nodal forces and
are compared excellently with the Hertz solution in all cases.

Frictionless Contact With Extensive Slid1ng

The half disk and the semi-infinite half-plane are also used
to demonstrate the capab1lity of this formulation to solve
extensive sliding contact problems. Since the contact be-
tween the two bodies is frictionless, the solution is inde-
pendent of the path, thus a Hertz solution 1S again available
for comparison. Solutions are obtained for prescribed dis-
placements at the top of the disk with both vertical and hori-
zontal components. Stress distributions on the plane of con-
tact, for two prescribed displacements, are plotted 1n Fig.2,
where zero position represents the point of initial contact.

Displacement
at D1Sk Top
Traction
u -0.4 mm Kg/mm <)-_ Present Solut1on
v -0.3 mm (on plane, ~ = 0.2)
Present Solution
(on disk, ~ = 0.2)

Present Solution
(on plane, ~ = 0)

Hertz Solut1on
(~ = 0)

u = -0.2 mm
v -0.15 mm

x (mm)
-2.0 -1. 0 o 1.0
F1gure 2 Normal Traction on Contact Surface
58

Four increments were used to obtain each solution. It is

seen that the solution agrees almost exactly with that of
the Hertz solution. It is noted that the center of symmetry
of the stress distributlon moves to the left as the half-
dlsk slides in that direction.

Frictional Contact with Extensive Sliding

The same problem is again solved here with frlction (~ = 0.2)

between the disk and the half-plane as an added consideration.
The normal tractions at the contacting nodes between the disk
and the half-plane for every other displacement lncrement are
also shown in Fig. 2 and are compared with those of the fric-
tionless case. Because of friction, it can be seen that the
displacement of the contact surface is retarded. That the
normal tractions of the plane and disk are equal in magnitude
and Opposlte In sign is also eVldent in the figure. The con-
tact surface of the two bodies and the locatlons of contact-
ing nodes are shown in Fig. 3. It is seen that through
averaging over the entire contact surface, the contact con-
dition of no separation or penetration is satisfied.

-0.6 -0.5 -0.4 -0.3 -0.2 -0.1 (mm)x

o Disk (with friction)

o DlSk (no frlctlon)

~ Plane (wlth frictlon)

Figure 3 Locations of Contact Surface and

Nodes of Contactlng Elements
(Frictional Sliding)
 59

Conclusions

An incremental variatlonal princlple and a corresponding

finlte method have been formulated for contact problems
based on the assumed stress hybrid model. The method is
convenient for handling problems with node-to-internode
contact, and has been verified by example solutions which
involves only small deformation elasticity problems.

References

1. Pian, T.H.H., "Variatlonal Principles for Incremental

Finite Element Methods," J. Franklin Institute, Vol.302,
1976, pp 473-488.

2. Pian, T.H.H. and Tong, P., "Finite Element Methods in

Continuum Mechanics" Advances in Applied Mechanics,
Vol 12, Edited by C.S. Yih, Academic Press, 1972, pp.
1-58.
Partll:
(}eonnetrically~orllUnear
Problenns-
Rods and Shells
Strategies for Tracing the Nonlinear Response Near Limit
Points

E.RAMM
UOlversltat Stuttgart, Germany

Abstract
For the prebucklmg range an extenslVe llterature of effectIve
solutlOn techmq ues eXIsts for the numerIcal solutlOn of structural
problems but only a few algorIthms have been proposed to trace
nonllnearresponsefrom the pre-llmlt mto the post-limIt range.
Among these are the sImple method of suppressmg eqUlllbrlUm
IteratlOns, the mtroductlOn of artIflClal sprIngs, the dIsplace-
ment control method and the "constant-arc-Iength method" of
Rlks /Wempner. It IS the purpose of thIs paper to reVIew these
methods and to dISCUSS the modIfIcatIons to a program that are
necessary for theIr ImplementatIon. Selected numerIcal exam-
ples show that a modlfled Rlks /Wempner method can be espe-
cIally recommended.

1. IntroductIon

Usually postcrIhcal states are not tolerated In the desIgn of a

structure. However, the predlctlOn of response In thIS range
may stIll be of great value. A tYPIcal example IS the Imperfechon
senSItIvIty of certam structures whlCh m general IS dIrectly
related to the postcrIilcal response. In partIcular this IS true
for structures exhlbltmg a decreasIng post-lImIt characterIstic.
ThIS may res ult III a dynamIC snap- through or snap- back phenom-
enon dependlllg on whether the load or the dis placement controls
the system. However, a statIc analysIs traces the wQole post-
crItIcal range allowmg for a better Judgement of the overall
structural response.

It IS well known that the us ually applIed N ewton-Raphson iteratlOn

methods are not very effICIent and often fall m the neighborhood
64

of critical points. The stiffness matrix approaches smgularity result-

mg m an mcreasing number of iterations and smaller and smaller
load steps. Fmally the solution diverges. In recent years several
strategIes have been proposed to overcome these problems and to
trace the response beyond the critical point.

It is the purpose of this paper to describe some of the most commonly

used techmques. These are the method of suppressing the eqUlllbrlUm
IteratIOns in the neighborhood of the critical point, the method of
artlficial springs, the displacement control techmque and the "con-
stant - arc - length method" of Riks [1], [2J and Wempner [ 3J. In
partlcular an attempt IS made to show the correlation of the latter
procedures. Special emphasis IS given to some modifIcations of the
Rlks /Wempner method leadmg to an efflclent iterative technique
throughout the entire range of loadmg and not only near the crItical
pomt. Other methods for solvmg the same type of problem, e. g. the
perturbation method or dynamIc relaxation, are not studied.

The dIscussIOn refers to lImIt points only. BifurcatIOn problems may

be mcluded either by introducing a small perturbation m geometry
or load (Imperfect approach) or by superImposing on the displacement
field of the critlcal load a part of the eigenmode (perfect approach).
The procedures are descrIbed in conJunction WIth the Newton-Raphson
method in its standard or modifled versions. A combination WIth
accelerated quasi Newton methods IS possible. Proportional loading
IS assumed but few changes are necessary for non proportIOnal load-
ing.

2. Starting Pomt and Notation

The study IS based on the incremental/IteratIve solutIOn procedure

m a nonlinear fmite element analysIs, 1. e. the nonlmear problem is
stepWIse linearIzed and the lInearizatIOn error is corrected by addi-
tIOnal eqUIlibrIum iteratIOns, see for instance [4J. A left supers cript
mdlcates the current confIguration of the total dlsplacementsmu , the
 65

load vector mp , the mternal forces mF and the out-of- balance

forces ~ . For proportional loadmg the loads may be expressed
by one load factor r:n.".
(1)

where P IS a vector of reference loads. WIthin one increment from

confIguratIOn m to m + 1, the posItIons 1 and J = 1 + 1, before and
after an arbItrary Iteratlon cycle, are dIstmgUlshed (figure 1).

load
1

1=1,2/3···

'U Ju displacement
FIgure 1: NotatIOn

The total Increments between positions m and 1 are denoted by u(i),

P (1) and A(1) whereas the changes in increments from 1 to J are
denoted by AU(J), AP(J) and t:.A (J), respectively:

Jp = mp + P(I)+AP(J) and JA =m A + A(,) + .dA(J}

p(J) A(J)

Ju =mU + U(I) +AU(J)

U'{J) (2)
66

In VIew of the fact that Iteration takes place m the displacement and
load s pace the load level may change from one Iterate to the other.
J' 1
In thIS caseanmtermediateposltlon J' forthe same load level A = A
IS mtroduced before the final state J IS reached (flgure 1).

Supposedly conflguratlon 1 has already been determmed and the in-

cremental eqUllIbrIUm equatlOns may be expressed by the lInearIzed
stlffness expresslOn.

(3 a)

If the out-of-balance forces lR Ip _ IF are mserted

The tangent stlffness matrIX lK at posltlon 1 may mclude all possIble

nonlmear effects. It may be kept unchanged through severallteratlOn
cycles followmg the modIfied Newton-Raphson technique. Eq. (3) IS
the basIc relation used as the startmg pomt for the dIfferent Iterative
techmques described below.

The statIc stabilIty CrIterlOn indIcates a lImIt or blfurcatlon pomt by

(4)

where LI U C IS the eigenmode of the crItlcal pomt. The smgularIty is

usually checked by the determmant

det cK = 0 (5 )

The determmant can easIly be calculated as the product of all diagonal

terms m the trIangularIzed matrIX durmg GaussIan elimmation. Note
that a posltlve determmant IS not a sufficIent CrIterlOn for stable
equIlIbrIum. Rather, the sIgns of the dIagonal terms should be rnoni-

to red to detect negatIve eigenvalues. ThIS IS the pomt when the lImIt
load IS passed and unloadmg should start.
 67

3. DescnptlOn of Some Iterative Techmques

3.1 Suppressing EquilIbrlUm IteratlOns

As mentlOned the eqUllIbrlUm 1terations usually break down near the

hm1t pomt even 1f the load mcrement 1S small. The slmplest way of
avo1dmg th1S d1fficulty 1S to suppress the iterations m the cr1tical
zone. Th1S procedure 1S used w1th great success by Bergan [5J who
mtroduced the "current stiffness parameter" to gU1de the algonthm
(figure 2).

load
B C ___ ---OC I

Increm. +
deration

F1gure 2: Suppressmg 1terations due to Bergan [5J

At a prescnbed value of the stiffness parameter the 1teratIon proce-

dure 1S d1scontinued (pomt A). Then pure mcrementatlOn is used. If
the EuclIdean norm of the d1S placement mcrements exceeds a certam
prescr1bed lIm1t (pomt C') load and d1splacements are linearly
scaled back (pomt C). Here negative d1agonal elements may be de-
tected m Wh1Ch case negative load mcrements are applied (point D).
The 1teratlOn procedure is resumed when the stiffness parameter
agamreaches its prescr1bed value (pomt E). The lIm1t point is located
by a zero value of the stiffness parameter. The techmque requ1res
very small load mcrements to aVOld dnftmg away from the equll1b-
rlUm path.

3.2 Artlficial- Sprmg - Method

ThiS method was developed for frames by Wright and Gaylord [ 6J

and has been apphed to arch systems by Shanf1 and Popov [7J and
68

to shell structures by the author [ 8J. The techmque is based on the

observatlOn that a snap-through problem may be transformed mto one
w1th a pos1hve defmite character1stIc 1f linear artificIal sprmgs are
added to the system (f1gure 3).

load

IG

If

Ip

F1gure 3: ArtlfiClal s prmg method

The method IS descr1bed m detaIl m append1x I. It IS an essentlal

requ1rement that a separatlOn of the real problem must be possIble
after the analys1s of the stiffened system 1S obtamed, i. e. for each
stage only one load-reduchon factor 1S defmed. Furthermore the
symmetry of the augmented stiffness matrix should be preserved.
These requirements lead to sprmgs at all loaded degrees of freedom,
WhICh are coupled, and depend on one smgle reference shffness.
Th1S parameter has to be found by trial. The coupling of all arhflclal
stlffnesses may destroy the banded nature of the stiffness matrix.
In [8J the elements outside the band were omItted from the stiffness
matrix but were retamed on the right hand sIde to find the proper
mternal forces. Augmentmg the sprmg stiffnesses on the band by a
factor of three to five accelerates the convergence.

Because the "nonlinearity" of the system is d1mmished by the art1fi-

c1al sprmgs the total number of 1terations can nevertheless be re-
duced compared to the analys1s without springs. Numerical experI-
ence shows that the method is successful only in real snap-through
 69

problems where the sprlllgs can keep the destabllizlllg structure alIve.
The method cannot be recommended for structures wIth local buckllllg
or when a tendency to bIfurcation IS present.

3. 3 DIsplacement - Control

The most often used method to avoId the slllgularIty at the crItIcal
POlllt IS the mterchange of dependent and mdependent varIables. Here
a smgle dIsplacement component selected as a controllIng parameter
IS prescrIbed and the correspondmg load level IS taken as unknown.
The procedure was llltroduced fIrst by ArgyrIs [9J but III the meantIme
has been modIfIed by several authors.

For SImplICIty let us assume that the stIffness expreSSIOn, eq. (3),
IS reordered so that the prescrIbed component f::, u~J) = U2 IS the last
one m the dIsplacement vector AU(J). Then equatIOn (3) may be de-

composed llltO two parts

(6 )

Interchanglllg the varIables

(7)

It IS ObVIOUS that the loss of the symmetrIcal and banded structure of

the shffness matrIX IS a severe handIcap. Later It was recogmzed
that the solutIon of eq. (7) could be formed m two parts. The fIrst
lme of eq. (7)

+ IR _ IK A
(8)
1 12' U2

IS lInear m the unknown mcrement at the load parameter 6. A (J)

70

Therefore its solution may be decomposed mto (figure 4)

(9)

correspondmgtothe two parts of the right hand side of eq. (8). That
is, both solutiOns are obtamed simultaneously using two different
"load" vectors

'K " , - P,
AU W - (10 a)

'K 11 . AU(J)n-
, - 'R , - 'K12 · u"2 (lOb)

2' 2'
Q---------~ O------~~

IU, IU
2

Figure 4: Displacement - Control Method

i
The displacement mcrement ;:1"1 J ), eq. (9), is mtroduced mto the
second part of eq. (7). This allows the determination of the load
parameter t::. A(J):

(11 )
 71

Thus lllstead of solvmg an unsymmetrIcal equation the modifled shff-

ness express lOn, eq. (8), lS analysed for two rIght hand sldes pro-
vlded that
1
K 11 lS not slllgular. Smce the dlsplacement 112 lS held
flxeddurlllgthelterahonthe underhned terms in equations (10 b) and
(11) are oml tted III all further itera tlOn cycles.

ThlS modlfled dlsplacement control method was descrIbed flrst by

Plan and Tong [10J wlthout menhomng the out-of-balance terms.
Zlenkiewlcz [l1J refers to the standard programmlllg techmque and
gl ves a physlcallllterpretatlOn of the two step method. Sablr and Lock
[12J exphcltly mtroduced the out-of-balance terms mto the formula-
tIon. The method was also descrlbed III detall by Stncklm et al. [13J.
A slmllar procedure has been applied by Nemat-Nasser and Shatoff
[14J who used a dlreCt subshtutlOn method instead of the Newton-
Raphson techmque.

A valuable slmphflcatlOn was uhhzed by Batoz and Dhatt [15J. Slllce

the techmque above descnbed reqUlres a modiflcatlOn of the shffness
matrIX (1 K -7 lK 11) the authors point out that it lS not very likely
to obtam exactly the singular pomt. Hence the original matrIx 1 K may
shll be used and equatlOns (10) are replaced by

(12 a)

(12 b)

where the underhned term III eq. (10 b) lS not required to be formed.
Agalll both solutlOns are added:

(13 a)

The vector lllcludes also the prescrIbed component

(13 b)

ThlS constramt equatlOn used m the first iterahon cycle (m -7 J = 1)

72

allows the determmation of the mcremental load parameter

(14)

Supposedly the structure 1S m an eqUllibriUm state at the begmmng

ofastepsotheout-of-balanceforcesvamshand so does u~J)I1. Then
t:,

t:, X. (1) 1S slmply a scalmg factor prov1dmg the constramt t:, u~l) = u2 .
Batoz and Dhatt [15J even drop th1S first cycle. They update the dis-
placement field only by 1tS component t:, u~l) and start to iterate.

For all further cycles u (J) does not change 1. e. t:, u (J) 1S zero and
2 2
X. (J) 1S
t:,

J =2,3 . (15)

Applymg the mod1fied Newton-Raphson techmque eq. (12 a) needs to

be solved only when the stiffness matrix 1S updated. Then no additional
computer time 1S required and the only add1tional vector stored 1S
~ U ( 1 ) I. The 1teration is contmued until all other dis placement com-
ponents are adJusted and the new equilibriUm pos1tion 1S found (fig. 4).

The d1splacement control method 1S usually used only m the ne1ghbor-

hood of the critical pomt although it may be applied throughout the
entire load range. Obv10usly the method falls whenever the structure
snaps back from one load level to a lower one (see example 5.2).
,
Some knowledge of the fa1lure mode is requ1red for a proper choice
of the controlling dIsplacement. It m1ght even be necessary to change
the prescribed parameter. Therefore an ObViOUS mod1fication IS to
relate the procedure to a measure mcludmg all dIsplacements rather
than to one smgle component. Th1S 1S dIscussed m the next section.

3.4 ModifIed Constant - Arc - Length - Method of R1ks (Wempner

Th1S 1terative techmque has been mdependently mtroduced by R1ks

[ IJ. [2J and Wempner [3J. Both authors limit the load step t:, X. (1)
 73

by the constramt equatIOn

(16)

That IS, the generahzed "arc length" of the tangent at m IS fIxed to a

prescrIbed value ds. Then the IteratIOn path follows a "plane" normal
to the tangent (fIgure 5), so the scalar product of the tangent 1 (1) and
the vector t,-+u(J) contammg the unknown load and dIsplacement mcre-
ments must vanIsh:

(17 a)

or In matrIX notahon

(17 b)

J =2,3

'A
new to ngent
'normal plane'

tan gent

'u
FIgure 5: Constant - Arc - Length Method
74

The constramt equations orIgmally were added to the incremental

stIffness expressIOn destroymg symmetry and the banded structure of
the matrIx. It was realIzed by Wessels [16J based on geometrIcal
consIderatIOns that these dIffICultIes could be removed by a two step
techmque SImIlar to that descrIbed m the previous sectIon. It IS thIS
Idea followed m thIS study. i~)

Agam the unknown vector D. Q{J) IS formed in two parts

(18 a)

or m matrIX notatIOn equIvalent to eq. (13 a).

(18 b)

Also here AU{J)I and AU{J)II are obtamed by equatIons (12) usmg
eIther the reference load vector P (D. A = 1) or the out-of-balance forces
IR as rIght hand SIdes. Then eq. (18) lS mserted mto the constramt
eq. (17) and solved for the unknown load mcrement D. A(J)

(19)

GeometrIcally thls lS the mtersectIOn J of the new tangent t{J) wlth the
"normal plane" (flgure 5). Eq. (19) lS equivalent to eq. (15) but con-
tams the mfluence of alldlsplacement components man mtegral sense.
The load mcrement D.A (1) m the denommator, which ObVIOusly has
another dlmensIOn, expresses the different scalIng of the load aXlS
Wlth respect to the dlsplacement space. It may be seen for the one
degree-of-freedom system mflgure 6 a that a low value D. A(I) tends to
a dlsplacement control and a large value to a load control of the ltera-
tion. In many degree-of-freedom systems the value D. A(I) m eq. (19)
does not play an lmPOrtant role and may be suppressed.

i~) Durmg the preparatIOn of thls study the author became aware
of the valuable paper by Crlsfleld [ 17J devoted to the same
subJect.
 75

Agam the modIfIed Newton-Raphson techmque slmphfles the method

because eq. (12 a) IS solved only once at the beglllmng of the step and
may even be replaced by the flrst solutlOn LI U (1):

(20)

Instead of Iteratlllg III the " plane " normal to the tangent ~(1)
t It mIght
be useful to deflne a "sphere" wIth a center at m and a radius ds [ 17J
(see appendIx II). Alternatlvely the "normal plane" may be updated
llleverYlteratlOncycle (fIgure 6 b). That IS, III eq. (19) ~u(l) IS re-
(1)
placed by the total lllcrement U . It was found that except for very
large load steps the dIfferences resultlllg from these formulations are
mInor.

IU
oj bJ
FIgure 6: ModlflCatlOn of constant - arc - length method

NumerIcal experIence has shown that thIS IteratIve techmque IS very

effIcIent in the entIre load range partlCularly when automatic load lll-
crementatlOn based on eq. (16) IS used. The only addItIonal storage
reqUlred 1S the vector ilU(1). The extra computer tlme ~s neghglb1e.
76

In addItion to the "constant-arc-length" the step SIze may be scaled

by relatmgthe number of iterations, n, used in the previous step to a
1
desIred value, II. • It was found that a factor
1
n.1 In 1 res ults in oscillatlOns
m the number of iterations requIred from step to step so that JIl]"n'
1 1
IS recommended. If materIal nonlmeantIes are involved smaller load
steps should be defmed to avoId drIfting. Whenever a negative element
m the trlangularlzed matrIx IS encountered unloadmg IS ImtIated. The
convergence may be eIther monotomc or alternating and may m some
cases be slow. Then relaxation factors may accelerate the IteratlOn
process. For mstance, m the alternatmg case a cut-back of the next
load change to 50 % resulted m a conSIderable Improvement.

4. Summary of the DIsplacement Control and ModIfied

Rlks IWempner Method

The algonthms for the displacement control method and the modified
Rlks IWempner method dIffer only m the equation used for the evalua-
tion of D. A(!) The algorIthm IS summarized as follows:

1. Select a basIc load mcrement as the reference load P , thus

defmmg the length ds m the first step (eq. 16).

2. In any step:
a) Solve the equihbrlUm equations for P and linearly scale
the load and dIsplacements to produce the length ds. ThIS
determmes D.A(I), ilU(l).
b) AdJust the step slze to the desired number of iteratlOns n.
l'
e. g. Jfl{rl
1 1
.
c) Check the trlangulanzed matrlx for unloadmg.

3. a)* Update the stiffness matnx lK

b) and, sImultaneously, determine the out-of- balance forces1R
c)* Solve for P to determine ,dU (J)I.
d) and, slmultaneously, solve for the out-of- balance forces lR
to determme iI U (J) II.

Note: * mdicates a step which is omItted m the modIfIed Newton-

Raphson procedure.
 77

4. Use constralllt eq. (15) or (19) to determllle the load lllcrement

/::, A(J) and eq. (13 a) '" eq. (18 b) to determllle dIsplacement m-
crements ~U (J). (If needed use acceleratlon factors. )

5. Update the load level and the dIS placement fIeld.

6. Repeat steps 3 - 5 untll the desIred accuracy IS achIeved.

7. Reformulate the stlffness matnx and start a new step by re-

turmng to 2.

5. Numencal Examples

The examples have been analysed on CDC 6600/Cyber 174 computers

USlllg the nonlmear fmlte element code NISA [18J. The geometncal
nonlmeanty IS based on the total LagrangIan formulatlon. For the
arch example, an 8 node Isoparametnc plane stress element IS used
[4J. The plate and shell structures are IdealIzed by degenerated ISO-
parametnc elements developed m [8J, [19J. The modIfIed Rlks /
Wempner method, m comblllatlOn wIth the modlfled Newton-Raphson
techmque, has been applIed exclusIvely. The ratlo of the change of
the mcremental dIsplacements to the total dIsplacement mcrements,
uSIng EuclIdean norms, IS used for the convergence cnterIOn.

5. 1 Shallow Arch

The shallow cIrcular arch under umform pressure (flgure 7) has al-
ready been analysed III [8J applying the artlficlal sprlllg method
(c 11 = 28 lb/m), see also [7J. Ten 8 node Isoparametnc plane stress
elements were used for one half of the arch. The analysIs wIth a basIc
loadolp= 0.3 and usmg the constant-arc-length constraint shows the
tYPIcal step SIze reductIOn m the neIghborhood of the lImIt pOlllt.
ThIrty steps wIth 1 to 2 IteratIOns per step were needed. The analySIS
has been repeated for a basIc load step of p = 1. O. The step SIze has
been adJusted by the factor rn;;;-
I 1
wIth a desired number of IteratIOns
~ = 5. In addItion, the load lllcrement was reduced to 50 % whenever
1
It alternated and the absolute value decreased. Now only 9 steps are
78

suffICIent. The number of iteratlOns required are indicated III the

flgure. The diagram also shows the starting point III each step after
the flrst Newton-Raphson Iterate. Compared to the artifIcial sprlllg
technique consIderable savlllgs are achIeved.

14------------~----------------------
~ --03}0
•••• ~ - basIc load step
000 p=1.
12

a.

10.

06
R=100 In I h=2b= 2 In J

02=008 I E=10 7psl , fJ =025

0.4

02
R

001~----~----~--~----~----~----~
0.00 0.02 004 0.06
w/R
FIgure 7: Shallow cIrcular arch

5.2 Shallow CylIndrIcal Shell

The shallow cylIndrical shell under one concentrated load (figure 8)

IS hinged at the longitudlllal edges and free at the curved boundaries.
The structure exhibIts snap-through as well as snap-back phenomena
WIth horizontal and vertIcal tangents. The shell has been analysed by
 79

Sablr and Lock [20J who used a comblllahon of the dIsplacement and
load control techmques. In the present study one quarter of the shell
has been IdealIzed by four 16 node blcublC degenerated shell elements.
As the basIc load step, P = 0.4 kN was chosen. Agalll the load steps
were adJusted wIth ~ and the acceleratIOn scheme descrIbed for
1 1
the arch was applIed. The entlre load deflectIOn dIagram IS obtallled
III one solutIOn wIth 15 steps and 3 to 9 Iteratlons per step as llldlcated
III the fIgure. If the acceleratIOn techmque was not used the number
of IteratIOns Increased consIderably especIally at the mlmmum load.

•• Sablr an d Lock
[201
0.6
fI'----n.....
P I "-
[KNl 1 ' 4
I '\ ",W,
0.4
,,
'\ 5 •

\
"
, 5
0.;:
•

E=3103 KN/mm2,J..l=03

2L
7 I

-02- L I
I
\.
hi n ged

I
R=10 L= 2540 mm
h=635mm
\A
\~
-0.4
0 10 20 WC,W, [m m 1 30
FIgure 8: Shallow cyhndrIcal shell

ThIS part of the load-deflechon curve IS numerically d~fflcult because

of the abrupt changes of the response at, for lllstance, the POlllt i at
80

the free edge. The structure has also been analysed using 36 bIlinear
4 node degenerated elements in combmation with an umform 1 x 1 re-
duced mtegratIOnscheme. ApproxImately the same results have been
obtamed but at about 20 % of the CP-tlme.

5.3 Elastlc - Plastlc Bucklmg of a Plate

-p.. -P
Pcr
06

04

I~ b .~1
----'l .

0.2 T
a
..i
LLJLLL-4-LLl...L.LJ- ~
b
~b/2~ P=6h 2
b =4a =1680 mm I h= 6mm
E =210 kN/mm2 JJ=03
OO~----~----~------~----~----~----~
o 10 20 We [mml 30

FIgure 9: Buckhng of a long plate

The sImply supported plate shown m fIgure 9 has an aspect ratIO of

a. = 1/4 and IS loaded only on ItS mIddle part. The plate has an mltlal

geometrIcal ImperfectIOn, defmed by a double sm-functIOn, wIth a

 81

maXImum amplItude of 0.294 mm. The YIeld lImIt (J of the elastIc-

y
Ideally plastlc steel IS 240 N /mm2. EIghteen blcublC degenerated
elements unevenly spaced were used for one quarter of the plate. The
thIckness was dIvIded llltO seven layers. The total load P IS non-
dlmenslOnalIzed WIth the lInear elastIc buckling load P of the plate
cr
wIth umform load on the entIre boundary:

n 2 ·E·h 3
Pcr = k b
12 (1-fJ 2 ) 0 2
(21 )

The basIc load step chosen was p= 0.25. In fIgure 9 the normalIzed
load IS plotted versus the center lateral dIsplacement. The plate falls
under combllled geometncal and matenal faIlure. The mltlal YIeld
POlllt at a deflectlOn of about 6 mm IS ImmedIately followed by the
lImIt POlllt at about 8. 3 mm. ThIrty steps with 1 or 2 lteratlons per
step were used. The elasto-plastlc analysIs was supplemented by a
purely elastlc solutlOn also shown III the flgure. Here the typlcallll-
creaslllg postbucklIng response of plates IS recogmzed.

5.4 CylIndncal Shell under Wllld Load

The buckllllg analysIs of the closed cylmdncal shell under wmd load
(flgure 10) studIed III [21] has been extended to the postbucklIng range.

~
p
10

TL
1350 18
1t

1 -10

R= L/2 = 220 mm I h =0.105 mm

E=6 87 10 4 N/mm2 I }J =0.3

FIgure 10. Geometry and load functIon of a cyhndrlcal shell

82

The extremely thlll structure with a radIus to thickness ratIO of over

2000 IS sImply supported at both ends. The variation of the wind load
defmed III fIgure 10 IS taken as constant over the length of the cylinder.
The maxImal load p at the stagnatIon point is normalized to the lInear
buckling load of the shell under uniform pressure

= 0918 E(1/ is =
p
(22)
.h.- f1[ - 0 657
RVh
One quarter of the sheills Idealized by 2 x 18 bicubiC 16 node elements.
Two elements of unequal length are used in the aXIal dIrectIOn, whIle
the 18 elements in the circumferential direction are concentrated near
the stagnation zone. The first load lllcrement defllled the basic step
SIze as p = 0.25. Both the perfect and an imperfect shell have been
analysed. FIgure 11 shows the dIsplacement pattern of one quarter of

FIgure 11: DIS placement pattern

the shell near the lImIt pomt. A faIlure mode wIth one half a wave in
the aXial direction and a few bucklIng waves III the circumferentIal
direction, located III the compression zone, is indicated. The post-
bucklIng mimmum of the load-deflection diagram (figure 12) IS about
60 % of the limIt pOlllt. The ImperfectIOn assumed for the second
analysIs corresponds to the failure mode of the perfect structure. The
maXImum Imperfection amplitude is 2. 5 times the wall thickness. The
 83

load deflectlOn path (flgure 12) llldlcates a reductlOn of the lImIt load
to 68 %of that for the perfect shell. The postbuckllllg mlllima nearly
cOlllclde. It should be noted that the example IS numerically very
sensitIve because of the extreme slenderness ratIo and the local nature
of the failure mechamsm. In both cases over 60 steps were necessary.

15.-------------------------------------~

P
~I
perfect shell

10
"..--0--__ _.....,.,
/ " ~~
---0------ ~~~--~o_~
~ I mperfect shell
/
/
05 /
I
I
I
I
I
oo+-----------~----------~--------~----~
o 5 10 w/h 15
FIgure 12: Load - deflectIon - dIagram of a wllld loaded shell

6. Conclus lOns

ThIS study on IteratIve techmques for passlllg lImIt pOlllts allows the
followlllg concluslOns:

SuppresslOn of eqUllIbrIum IteratlOns near the lImIt pomt may

be a useful procedure but reqmres very small load steps.

if The method of artlflclal s prlllgs IS based on numerIcal experi-

ence and trIal solutIOns. For local fallure It may not be suc-

cessful.
84

* The displacement control method requires a proper selectIOn of

the controlling parameter. It fails m snap-back situatIOns.

* The constant - arc - length method of Riks /Wempner seems to

be the most versatlle techmq ue, being advantageous m the entire
load range.

Due to modificatIOns of the origmal method the constramt equa-

tIOn does not need to be solved simultaneously with the equihb-
num equatIOns.

Automatic adJustment of the load step and acceleration schemes

may further Improve the performance. Only mmor changes m
coding are necessary. Applymg the modifled Newton-Raphson
technIque requIres the storage of one additIonal vector. The
extra computer time is neghgible.

Acknowledgement
The author would like to thank Professor D. W. Murray, Umversity
of Edmonton, currently at the University of Stuttgart, for valuable
discussIOns.
 85

References

[IJ Rlks, E.: The ApphcatlOn of Newton's Method to the Prob-

lemofElastlcStabllity. J. Appl. Mech. 39 (1972) 1060-1066.
[2J Rlks, E. : An IncrementalApproach to the SolutlOn of Snapp-
mgandBucklmg Problems. Int. J. Sohds Struct. 15 (1979)
529-551.
[3J Wempner, G. A. : Dlscrete Approxlmations Related to Non-
linear Theones of Sohds. Int. J. Solids Struct. 7 (1971)
1581-1599.
[4J Bathe, K. -J., Ramm, E., Wllson, E. L.: Flmte Element
FormulatlOns for Large DeformatlOn Dynamlc Analysis. Int.
J. Num. Meth. Engng. 9 (1975) 353-386.
[5J Bergan, P. G. : SolutlOn Algonthms for Nonlmear Structural
Problems. Int. Conf. on "Engng. Appl. of the F. E. Method",
H¢Vlk, Norway 1979, pubhshed by A. S. Computas.
[6J Wnght, E. W. , Gaylord, E. H. : Analysls of Unbraced Multi-
StorySteelRlgldFrames. Proc. ASCE, J. Struct. D1V. 94
(1968) 1143-1163.
[7J Shanfl, P., Popov, E. P.: Nonhnear Buckhng Analysls of
Sandwich Arches. Proc. ASCE, J. Engng. D1V. 97 (1971)
1397-1412.
[8J Ramm, E. : Geometnschmchtlmeare Elastostabk undflmte
Elemente. Habllitationsschnft, Umversltat Stuttgart, 1975.
[9J Argyns, J. H. : Contmua and Dlsconbnua. Proc. 1st Conf.
"Matnx Meth. Struct. Mech. ", Wrlght-Patterson A. F. B.,
OhlO 1965, 11-189.
[ 10J Plan, T. H. H. , Tong, P. : Variabonal FormulatlOn of Finite
DlsplacementAnalysis. IUTAMSymp. on "Hlgh Speed Com-
putmg of Elastic Structures", LH~ge 1970, 43-63.
[l1J Zlenkiewlcz, O.C.: Incremental Dlsplacement mNon-Linear
Analysls. Int. J. Num. Meth. Engng. 3 (1971) 587-588.
[12J Lock, A. C., Sablr, A. B. : Algonthm for Large DeflectlOn
Geometncally Nonhnear Plane and Curved Structures.
In "Mathematlcs of Flmte Elements and Apphcahons" (ed.
J. R. Whlteman), AcademlC Press, N. Y. 1973,483-494.
[13J Halsler, W., Stnckhn, J., Key, J.: Dlsplacement Incre-
mentatlOnmNonhnear StructuralAnalysls by the Self-Cor-
rectmgMethods. Int. J. Num. Meth. Engng. 11 (1977) 3-10.
[14J Nemat-Nasser, S., Shatoff, H. D. : Numencal Analysls of
Pre- and PostcnhcalResponse of Elastic Continua at Fmite
Strams. Compo Struct. 3 (1973) 983-999.
86

[15J Batoz, J. -L., Dhatt, G.: Incremental Displacement Al-

gorithms for Nonlmear Problems. Int. J. Num. Meth. Engng.
14 (1979) 1262-1267.
[16J Wessels, M.: Das statlsche und dynamische Durchschlags-
problem der imperfekten flachen Kugels chale bei elastlscher
rotahonssymmetrischer Verformung. DissertatlOn, TU Han-
nover, 1977, Mitteil. Nr. 23 des Instltuts fUr Stahk.
[ 17J Crisfield, M. A. : A F<l;st Incremental/Iterative SolutlOn Pro-
cedure that Handles "Snap_Through". Proc. Symp. on "Com_
putational Methods m Nonlinear Structural and Solid Mech. ",
Washmgton, Oct. 1980.
[18J Brendel, B., Hafner, L., Ramm, E., Sattele, J. M. :
Programmdokumentation - Programmsystem NISA. BerlCht,
Inshtut flir Baustatlk, Umversitat Stuttgart, 1977.
[19J Ramm, E.: A Plate/Shell Element for Large Deflections
and RotatlOns. Symp. "Formulations and Computational Al-
gorithms in F. E. Analysis", Cambridge 1976, MIT Press
1977.
[20J Sabir, A. B., Lock, A. C.: The Apphcatlon of Fimte Ele-
ments to the Large DeflectlOn Geometrically Nonlmear Be-
havlOur of Cyhndrical Shells. In "Vanational Methods m
Engng:' (ed. C. A. Brebbia and H. Tottenham), Southampton,
Umversity Press (1972) 7/66 - 7/75.
[21] Brendel, B., Fischer, D., Ramm, E. , Rammerstorfer, F. :
Lmear and Nonhnear 8tabihty Analysis of Thm CylindrlCal
Shells under Wmdloads. To be pubhshed, J. Struct. Mech.
1981.
 87

AppendIX I: The Arhflclal Sprlllg Method

Accordlllg to fIgure 3 the vector of the total external loads 1 G of the

modIfIed system is decomposed lllto the real load vector Ip and the
part resIsted by the s prlllgs 1 f

+ (A 1)

To retalll the desIred raho of specIfled loads It IS required that all

components of the real load can be obtallled by one common "load-re-
duchon-factor" ly

(A 2)

That IS, all components of conflguratlOn 1 have the same ratlO

1=1,2,3 n (A 3)

It follows that sprlllgs have to be attached to all loaded degrees-of-

freedom and all sprlllg shffnesses are coupled. The sprlllg stiffness
matrix C IS defllled by

'f = C . 'U (A 4)

Energy pnnclples reqUlre C to be a symmetrical matrIx (c kt = ctk ).

EquatIon (A 3) allows the elements c kt of the matnx to be determined
If one reference shffness c 11 is prescnbed

(A 5)

or If the reference load vector P IS llltroduced

ell T
C =-2-'P'P (A 6)
P,
The Iterahon equatIon, eq. (3 a), IS modifIed to

('K + C)· L1U(J) = JP + Jf - 'F - c 'u (A 7)

JG
88

The nght hand sIde expresses the out-of-balance forces. After itera-
tlOn (J ..., m + 1) the real loads are determined by eq. (A 2):

(A 8)

The "load-reductIon factor" IS obtallled by eq. (A 3):

(A 9)

It was found that an effectIve value of c ll IS one WhIch leads to

o< m+1y < 0.6 at the beglllmng of the analysis [7J, [8J.

AppendIX II: IteratlOn on a "Sphere"

The "sphere" wIth the center at m and the radius ds of the Imtlal
tangent vector t(1) (fIgure 13) IS defmed by

r(ll - ds 2 = 0 (A 10)

~
/ I
/ I

T~)/
'sphere'
-- /'
/
1.1/-.= 1
I
I
I / I
J /
O--;........oI:J/----- ____ J

IU
FIgure 13: IteratIon on a "sphere"
 89

If the radIus vector IS replaced by

(A 11)

and eq. (16) IS taken mto consIderatIon eq. (10)results in

(A 12)

or m matrix notatlon

(A 13)

AU (J) IS expressed by eq. (13a). Theneq. (A13) leads to a quadratIc

constramt equatlOn for the load parameter!::' A. (J) WhICh is the equivalent
to eq. (19)

+ c = 0 (A 14)

with the coeffIcIents

a = 1 +

(A 15)
Parameter Sensitivity of Nonlinear Structures Concerning
Stability Limit

F. G. RAMMERSTORFER, D. F FISCHER
Voest-Alpme AG, Linz, Austria

SurrmaJ:¥
'Ihe stability of nonlinear structures 1S considered, espec1ally hav1ng
in V1ew the pararreter sensitivity of the stability liroit load. I t is
shavn that under special cirCUIllStances, for which cri terions are de-
r1ved, nonlinear structures behave with a sudden, i.e. non-continuous,
<flange of the limit load if a certain system pararreter is varied. Such
phenanena belong to the f1eld of the catastrophe theory which allows
their classification and useful docurrentation.

1. Introduchon

Considermg the load displacem3Ilt behaviour in canbination with the

var1ahon of the determinant of the global tangent stiffness matrix and
perfo:r:ming supplerre:ntary e1genvalue-analyses at several load steps, the
stab1hty behav10ur of nonlinear structures can be detennined. Sare un-
expected effects which a linear analys1s would never reveal are found
by those nonlinear investigations.

The paper shONS hew structures can undergo effects like a non-continuous
parameter Sens1 tivi ty. This rreans that an infinitesimal variation of a
particular pararreter of the physical system can cause a finite change of
the stability limit. I t is shavn hew those problem:; are anbedded in the
catastrq;he theory.

As a particular example the described phenarena are discussed by the

cons1deration of the nonlinear behaviour of an elastically supported
1nitially curved beam wh1ch represents the behaviour of a W1de-Spanned,
daned roof of a hqw.d storage tank.
 91

2. Algonthms for nonlInear stab1hty analys1s

Assuming proportional oonservahve stahc loading the current load vec-

tor at the step n of the incrarental analysis can be expressed by

(1 )

with the reference load vector, !3-re f' and the current load nultiplier,
n A• The linearized 1ncrerrental eqw.libnum equation for the increrrent
n-n+1 follONwg the tangent stiffness concept (e.g. [1]) 1S

I]{ flU = n+1 AR -~. (2)

.. - -ref-
'lhe current tangent stiffness matrix of the system, ~, oons1Sts of

I]{=K +1]{ +1]{ (3)

~ .. e .. u .. g
with the constant linear elastic suffness matrix, ~e' the inihal dls-
placenent matr1x, I]{ , and the 1mhal stress or gearetnc matrix I]{ •
~u -g
~ denotes the vector of wternal forces in conflguration n. The solu-
tion of equ. (2) yields a first approXlmahon of the increrrental dls-
placerrent vector L\l! which can be ll'Cproved by eqw.librium interahon
[2,3] .

Acoordlng to the stahc stab1hty criterion, two adjacent equilibrium

m~':
configurations exist at the cr1 tical load level A. Hence,

m1:
§ 61,;! = Q (4)

has a nontr1v1al solution at the stab1lity lJ.nu.t. Equ. (4) leads to the
condition

det:
m'"
K~
= det K( . m'" A) =0 (5)

for the fundanental state rrr': in wluch the eqw.librium becares lnBtable.
Therefore, the behaviour of the detenrunant of the load dependent tan-
gent shffness matr1x, det ~(A)' can be used as an wmcator for the
stab1hty of the equihbnum. Denotlng the cntical load level by
m'': .;,
A = A , the follONwg relahons hold [4,5]:

a) hm, det K(A) = 0 (6)

A-A" ..
for buckling as well as for snap-through,

b) a sufflcient but not neCESSary condlhon for snap-through 1S

a
llIr! aI det ~(A) = - ro, (7)
A-X' -
92

c) buckling is indicated by a fini te negative value of the left-hand-

Slde of equ. (7).

All relations shawn for det K(A) are also valid for the normallzed de-
~

terrninant

det K (A) = det K(A) / det K (8)

n IV <'1;1 .<lfe

with ~e as explained in equ. (3). Due to practical reasons it lS useful

to consider the normalized rather than the original detenninant value.
Therefore, in the follcwlng part of this paper detn ~(A) is used in-
stead of det K (A).

As shawn in detail in [4,6,7] the relations (3) and (4) can lead to an
approximative (linearized) elgenvalue problem:

[~e + rn~ (~u + ~g)] 6~ = 9 (9)

Solving the elgenvalue problem (9) at a certain load level rnA < A;'
renders an approximation of the crltical load level:

( 10)

Wlth rn~. the smallest eigenvalue, and 6U. the corresponding eigenvector.
J -J
In equ. (9) a linear relatlonshlp between the load dependent parts of
the current stiffness matnx and the stiffness matrix at the stabihty
llffil t is assumed. ThlS approximatlve assumption becares rrore and rrore
accurate when the considered funda!rental equilibrium state rn respresen-
ted by rnA, approaches to the stability limlt conflguratlon, represented
by A"i In this case

. rn- rn
llID n·= n= (11 )
rnA_A'" J
with rnn the exact "nonlinear" elgenvalue, for winch
rn rn
A;' = n A ( 12)

is valid.

Equs. (9 - 12) describe the scherre of a supplementary elgenvalue ana-

lysis which renders - if i t is applied at a s~ce of incremental
steps up to the limlt load - a good lnslght lnto the nonllnear stablllty
behavlOur of the structure. Plotting the values of 11):,., (equ. (10)) ver-
sus an arbitrary cClTpOnent, rnup, of the displacement vector, rng , yields
a function ~ "'(uP) which starts at uP=Owith the value \"1'ln ,the clas-
 93

sical =1 tical load level of the linear structure, and lntersects the
load-dlsplacarent path, A (uP), at the stablhty limit point of the non-
linear structure (see flg. 2 of the follcwlng chapter) .

The curve wIuch lS obtalned by plotting rn~", versus rnA lntersects the line
f (A) = A at rn~",= A': The behavlOur of the ~'" (A) curve allCMS a distlc-
tion between buckling and snap-through, as shcwn in [6]:

lim A'" (A) = A'" (13)

A -A'"

holds ln every case, but specially for snap-through

lim aaA A'" (A) = - 00. (14)

A-A'"

In the buckling case the left-hand side of equ. (14) converges to a

finite value.

FUrther very useful algorithms for analysing the stability of nonlinear

structures are to be found in Brendel's thesis [6J. Special attention
will ncw be given to the "Current Stlffness Pararreter", S , recently in-
p
troduced by Bergan et al. [8 - 10] whlch represents the current stiff-
ness of the structure as a relation between a load increrrent and the
oorresponding displacerrent in=errent. I t indicates the snap-through li-
rni t load by

lim S (A) = 0 (15)

A-A'" P
SNAP
and
a
hrn, -a' S (A) (16)
A-A" 1\ p
SNAP

Hcwever S (A) cannot reveal the stablhty hrnit MUch is caused by bl-
P
furcatlon, l.e. buckling. Nevertheless the current stiffness pararreter
is ln any case a very useful and easily calculated pararreter which al-
lcws again in, canbination Wlth another stability indicator, e.g. the
determinant (equ. (6)), an investigatlon of the stability behaviour and
the rrechanism of stability loss as shcwn in [11 J •
94

3. A partJ.cular problem

Conslder the circular arch, elastically restrained at both ends, in fig.

1. Tlus slllple structure represents a highly nonlinear system and has a
very interesting Stabillty behaviour as will be sh.cMn ncM.

Fig. 1 The circular arch

The real background for the choice of this structure as an exarrple is

the consideration of the nonlinear behaviour of Wlde spanned roofs of
very large storage tanks. The follON'ing data are used in the consldered
9
exarrple: R = 105 m, r k = 1.922 m, rs = 34.6 m, EA = 1.741 x 10 N, EJ =
= 4.765 x 107 Nm2; the linear SprIDgs have the follON'ing suffness coef-
fi~ients c 1ref = 1.395 x 107 N/m, c 2 ~ 6.006 x 108 N/m, Y ref = 2.708 x
10 Nm/rad. The suffness coefficlents c 1 and y of the restraming
sprinJs are varied ID order to calculate the sensitivity of the struc-
ture with regard to this variation:

c1 =C c 1ref ' y =r y ref ' (17)

C and r are used as restraint paraneters.

'Ihe distributed load is location dependent:

r
q(r) = %
rs
with q
o
= 5.04 x 104 N/m at load level A = 1.
 95

Only dlsplacanents in the y-z-plane are allewed (plane proolem). The

arch was mcdelled by 30 nonllnear beam elements. The elastlc supports
were realized W1 th the ald of a crntnnatlon of linear beam- and truss-
elements. '!he calculatlons were performed using the NISA-program [12].

F1g. 2 shews the results of the nonl1near 1ncremental-iterative analysis

descrlbed above. '!he case W1 th C = 1, r = 1 1S cons1dered.

6
h
~*
/'
5 [= 1. I \
r=1. / \
/ \
4 I \
/
V ~'IUP
- \
\
/ \
3
I~. I

hlu~) "Iu~)

7 "'-. detn ~lh )

/..- NIII) ............
7 ' ,.
-.8 -.6 -.4 -.2 0 .2 .4 6 .a1. 1.2 1.4
u~f m) N f10SNI. de tn ~. u~ [m)

The nonllnear behav10ur of the arch (C = 1 r = 1)

- - load dlsplacarent path for ncdes K and F
---....... nonoal force N( >. ) 1n the cross sectlon at F
--- e1genvalue funct10ns ~:\u)
- - nonoal1zed determlnant of the current tangent st1ffness
matrix det K( >' )
n ~

A considerably nonlinear behavlOur can be ooserved. Especially the rro-

tlon of the support K which represents the motlon of the central ring
of the tank roof during load increrentation is remarkable. At the be-
ginning of the load 1ncrementatlon 1 t moves upNards and after reach1ng
a llffil t 1 t moves da.-mwards until the arch snaps through. This peculiar
behav10ur of the support motion is also reflected in the curious load
dependence of the nonnal caupression force N( >' ) which causes the alter-
96

natlng stiffness variation represented

_ by the det K (A) history. Consi-
n~

dering the eigenvalue functions A* (u) and the det K(A) CIllVe one gets
n~

the impression that the structure is almost at a stability limit atA~1.4

(the det K(A) CIllVe approaches the A axis and the A'" curves tend
n~
to
taIch the A(u) paths) but a further load increrrentation "stiffens" the
structure until A '" 2.5. Continuation of the load incrementation leads
finally to the snap-through at A "'_3: '!he detn~ curve reaches a limit
with horizontal tangent. When the A1, (u) curves cross the load chsplace-
ment paths, the detenninant of the stiffness matrix behaves correspon-
chnglyto equs. (6) anG (7). Spec~al attention will noN be given to the
detnK(A)
.. mstory. I t is cbvioos that after a certain pararreter variation
the d.et K (A) curve may cross the A ans just before the "re-stiff~g
n ..
pericd". ~s would render a sudden change of the critical load level A1,.
In other words, an infinitesiroal vanation of this certain system para-
meter would cause a finite change of the stability limit load, i.e. a
non-oontirruous pararreter sens~ tivity ooncerning stability limit.

3.1 Pararreter sensitivity

In order to find a sensibvity behav~our which was presurred in the pre-

vious chapter the sprmg stiffness c 1 , expressed by C according to equ.
(17) is var~ed. Sore results are shCMn m flg. 3.

Fig. 3 shONS the results of three typical cases (equ. (17»:

a) C = 1.0: Load incrementation leads to a stiffness-loss follar.red by a

re-stiffening, a further stiffness-loss and finally the snap-through.

b) C = 0.5: It is ranarkable that the determinant CIllVe crosses the A

axis before re-stiffening takes place. Considering the A(uK) behaviour
one can cbserve that the support K snaps upwards quite oontrary to the
case C = 1 in which it snaps darmwards. This value of C is already
beyond the critical value of C = C" at which the 1ll1U.t load j1.lIlpS (see
also fig. 5).

c) C = 0.35: '!he system moves tcwards the snap-through in a direct man-

ner just like the usual hinged-hinged arch.

'!he stab~hty behavl.our of these three typl.Cal cases can also be studied
by oons~deration of fig. 4 wluch represents the A1, (A) chagrams (see equs.
(13) and (14».
 97

r=1.
Ii
II
1\
\\
\ \ 4
\ \
\
-), (UK)
\ \ '
\

-.8 ·.6 -.4 -.2 .0 .2 .4 .6 8 10 1.2 1.4

u~lml

Flg. 3 Behavicur of the arch with dlfferent spnng stiffness pararreter C

- - load displacement paths

- - - eigenvalue functlons
- _.- determinant of the stiffness matnx

CorresponCill1g to equ. (14) one can pred1.ct that in all the =nsidered
cases the stabllity loss lS due to snap-through although the detnISP' )
curve for the case C = 0.5 (see flg . 3) does not allow thlS pred1.ctlon
(refer to EqU. (7». '!his uncertalnty In the use of the limit =ndltion
(7) (which lS only a sufficient but not a necessary =nditlon) is al-
ready pointed cut 111 [6].
98

OL----L____J -_ _ ~

o 2 3

-
Flg. 4 Elgenvalue functions )., ,', ()., l for three different C parameter
values

Further variation of the parameter C allaYS the presentation of the pa-

rameter Senslllvity of the limit load, ).,:" (el, in fig. 5.

3
~*

r= 1.
2

I
I
I
I
I
1(*
0
.0 .5 1. 1.5 2.
C
Fig. 5 Stability limit load as a function of the parameter C (dotted
line is the expected physically not realizable part of the curv~

Flg. 5 shaYS clearly the conslderable Jump of )., ,', (el at C = e'" '" 0.55.
ThlS jump behaviour has of course serious consequences for the design
 99

of such structures and the cntlClsm concerru.ng their stability safety.

Hence it 1S important to have a tool for classification of nonlinear
structures with regard to the parameter Sensit1Vity of the stabil1ty
lllnit load. '!he catastrophe theory may be this tool as it 1S shavn in
the follavlllg chapter.

4. Enbedding in the catastrophe theory

SystEmS like the considered one have catastrophic behaV10ur in a double

sense: F1rstly, the load displacement behaviour shavs catastrophic states indl-
cated by snap-through at II = x' and, secondly, the stabihty lirru.t load
1tself has a catastrophic parameter sensitivity.

4.1 General remarks

'!he rather young mathematical theory of catastrophes has its origin III

the endeavour to render mathematical models of natural processes having

a certain jump behaviour, especially of sum belonging to the field of
blOlogy [13]. In the meanbme this theory has found useful application
1n a w1de var1ety of other fields [ 14, 15]. I t is also used to clasS1fy
rhenanena concerning the stabih ty of the e:[Uihbrium of necl1anical
structures [15 - 17].

4.2 A short 1nsight into the catastrophe theory

Consider a necl1anical systen whim has a parameter dependent behaviour.

POlllts in the parameter danain at whlch the system manges its behaviour
1n a quali tative sense due to a variatlon of the parameters are called
b1furcation points [17J. Assume for SllllpliC1ty that the system has a
real scalar potential function, V (x,a,b), e.g. the carplete potential
energy of a loaded elastic structure W1th x the dependent var1able, des-
cnbing the state of the systen, and a and b the independent system pa-
rameters (control parameters ) . '!he surface 1n the a, b, x danain whim is
defined by

aV = 0 (18)
aX
represents the behaviour surface, e.g. the equ1librium states of the
elastic structure. '!he parameter danain is in this case the plane a - b
and is named the control surface. The system is structurally stable i f a
small var1atlon of the systen parameters does not change the qualltative
100

character of the potential function, otherwise the system is structural-

ly unstable [17J. Applying the more general classification theorem of
R. '!han [13] to this special case the follcwing statement holds: '!he
behaviour surface is - provided the system is structurally stable - a
smooth surface, whose projection into the parameter danain (control sur-
face) has only fold curves and cusps, see fig. 6. '!he fold curves are
the set of the bifurcation points (a,b)" in the control surface for
which the condition

a2 v I (19)
ax2 (a,b)"'= 0

is net, e.g. snap-through conditions for the elastic structure.

In the considered special case the potential, V(a,b,x), can be descri-

bed in the viclmty of the cusp qualitatively by one of the follewing
polynans [13, 17}:

v = .1.3 x 3 + ax (20)

or

v = .1.4 x4 + .1.2 ax 2 + bx (21)

In case of equ. (21) the behaviour surface is locally qualitatively de-

fined by equ. (18) which renders new:

av
ax
= x3 + ax + b =0 (22)

Application of equ. (19) and elimination of x yields under consideration

of equ. (22) the bifurcation set in the control surface:

(23)

which represents parabolas forming a cusp.

Catastrophes deflned by a potential corresponding to equ. (21) belongs

to the class of cusp catastrophes.

A simple exaIT\Ple taken fran [17], the v. Mises truss, may dem:>nstrate
the statanents given above. Flg. 6 shows the behaviour surface of this
structure, provided the Euler-bucklmg of the smgle truss elanent dces
not appear before the snap-through takes place, otherwise the system
would be structurally unstable.
 101

BEHAVIOUR SURFACE

CONTROL SURFACE tLJ:::;;:.;~&:;~::i?,L--= SET

Fig. 6 '!be behav~oor surface and the bifurcation set of the v. fuses
truss (taken fran [17 ] w~ th penni.ssion)

For ~s example the p:>tenhal ~s given by [17]:

(24)

In order to correspond with equs. (20 - 23) the follONing quantities

must be introduced:

x = ex
a = - ex 2 (25)
o
b = plEA

'!hus, the catastrothic behaviour of the v. Mises truss is descrilied by

the equs. (22) and (23) concermng ~libnurn states and stab~11ty 11-
mits, respechvely. 'lbe shaded area ~n fig. 6 ~s 11m~ted by the fold
curves; it has the follON~ng s~gniflcance: If a parameter value crosses one
of the fold curves caning fran the shaded area, then a continuous para-
meter variation renders a non-conhnuous system behaviour, Le. snap-
through.

4.3 App11cahon of the catastrophe theory to the problem urrler

considerahon

let us nON C<IIE back to the proolem described in chapter 3.

Oloosmg the load multiplier, A, and the restrainmg parameter, C, as

K
iooependent control parameters and the vertical displacanent, u z ' of the
102

support K as dependent behav1.OJr var1.able, the results of the above des-

cribed analysis (see chapter 3) renders the behav1.OJr surface as sha,.m
quali tatively in 'fig. 7.

_UK
z

r=1.
BEHAVIOUR SURFACE
(equilibrium states)

it
C

Fig. 7 The behaviOJr surface of the arch with regard to the equil1.brium
states

'!he potential V describing this behaviour is, naturally, the total po-
tent1.al energy . Consider1.ng the behav1.our surface in fig. 7 one ooserves
that it is cut into two different surfaces at a line along C = C". HenCE,
'!han's theor€lll concerning the exJ.stence of a smooth behaviOJr surface
(see chapter 4.2) is only valid f or two separated parameter sub-danains
which are connected along the l1.ne C = C'·' . Thus, the structure has a
structural instabil1.ty at C = C". This fact corresponds to the ooser-
vatl.On that for C < C', the support K snaps through upovards and for C > C',
dGmYards.

I t should be mentioned that the algori tiun used allONS the calculation of
the syst€lll behaviOJr only for load incr€lllentation up to the limit load.
'lherefore the continuation of the behaviOJr surface for >. > >.'" (C), Le.
the post-buckling behaviOJr, has only qualitatively predicted character.

Concerning the fold curve )/' (C) ill the control surface wh1.ch 1.S relevant
for the stability loss due to increased load, one should nonce the cor-
respondence with fig. 5.
 103

The described behaviour u~ (A,C) is the first catastrophic aspect, Le.

the snap-through of the structure. The second one is the non-continuous
variation of the stab1li ty linu t load for sane parameter variation. Fig.
8 sha.vs schematically the dependence of the cntical load multiplier,
A = l~ on the system parameters (=ntrol parameters) C and r. I t will
be shaND by the relevant potential funchon V (C, r , A) that we again are
dealing with a cusp catastrophe.

T'

F1g. 8 'Ihe catastrophic behaV10ur of the cn tical load level A'" (C, r)

In fig. 8 the behav100r surface is shaND for r:;;; 1. The cross sechon
through the behaviour surface A'" (C, r) at r=1 =rresponds W1th the A'" (C)
mrve m f1g. 5. Furthermore, sane pomts on the behav10ur surface for
C = 1.0, l.e. A'" (r ,C=1.) are calculated. Fran these infonnations the
behav10ur surface was drawn qualltatively. There is one ilrqJortant fact
to be noted: Although the behav10ur surface is a srrooth surface, the
folded part has only mathemahcal but no physical rnean1ng.

Always the la.vest critical value A'" 1S the stability limit. Hence, the
relevant part of the behav10ur surface is agaln slotted and only one
branch of the bifurcahon set in the control surface, namely the 0" (r)
curve, contalns the phys1cally relevant JUITp condl hons.

What potenhal funchon V (A ,C, r) belongs to this catastrophic beha-

viour? - V has to satisfy:
104

(26)

= o. (27)
i~
C=C (r)

A potential function which fulfills these requirerents is

A.

V= - Jdetno:K (X) dX + D, (28)

o
with detJ«5.) denoting the nonnalized detenninant of the tangent stiff-
ness matrix at the load level 5.. D is an arbitrary real constant.

'!he potential function (28) renders with equ. (26) the necessary condi-
tion (5) for the stability limit. Hence, V in equ. (28) yields the beha-
viour surface A = }/( fig. 8). The condibon for the jurrp of A''', equ. (27), leads
in canbination with equ. (28) to that situation in which the det K(;\.)
n~

curve tcuches the ;\. axis without crossing it. This is exactly the case
1f C approaches C;' (see figs. 9 - 11).

10

av
ar= 8

detn~ 6

58
Z

2 3
~
Fig. 9 '!he nonnallzed determinant of the tangent stiffness matrix as
a function of the load level (thin dotted lines are expected
curves)
 105

II
o. t 2. 3.
.0

-.2

- .4

-.6

V
-. 8

-1 .
/
/

-
hg. 10 The potentl.al function V (tlun dotted lines are expected
auves)

.6

.4
tv
a,,2=
.2
a
a" detn~ .o I-----.---r--h-Hllt--.--Hr-T-.--,

-.2

- .4

-.6

- .8

-1.

Fig. 11 Second derivative of the potential function

106

Fig. 10 shCMS that the potenhal V behaves like a polynaninal with or-
der four which is in correspondence with equ. (21). Referring to '!han's
theoran one can conclude that the assumption of a cusp ca.tastrq:he as
shCMl schanatica.lly in fig. 8 is justified. Furthenrore fig. 10 ShONS
clearly for decreasing e that for values e > C' indeed a po~nt of ~nflechon
in the VeAl anveexistsfora>. value srnallerthanX< butnothll e""C<a
second rninlIllUIll of V appears prior the origmal one. '!he appearance of a
second nu.nimJm of V renders the jump of "A* (e).

In order to generalize the results and considerations described above

follCMIDg theorans are stated:

'!heoran 1: I f a nonlinear structure which tends to a stability loss,Le.

3"A'~: lim det K("A) = 0, (29)

"A--"A*
n ~

has anyone systan parameter, Q, (control parameter) for

wmch is valid that

]~:s; "A'~: lim [~ detn ~("A) =0 (30)

Q "A-~ 3"A -
Q=Q

3QE~: with

[~ detn 1§("A)] :: 0 (31)

"A="A

wi th ~ denoting the feasible region of Q, then it is possib-

le that the stability limit has a non-continuous parameter
sensitivity w~th regard to a var~ahon of the parrureter Q.
For explanation, see the anve for e = 5.0 in fig. 9.

'!heoran 2: If the assumptions of theoran 1 are net and A fulfills the

condition

(32)

then the structure has a non-continuous parameter sensi tivi-

ty at

Qi' = Q, (33)
Le. a sudden jump of the crihcal load level "Ai, (Q) •
 107

'!he prcx:>f of these theorems is cbvious by considering equs. (6), (8) and
(26 - 28). Equ. (31) rooans that only minima at load levels smaller or
equal to the stability limit load have to be considered because of the
fact that usually a practical physical structure behaves in the unloaded
state, i.e. A = 0, with

detn K(A=O)
::: > ° (34)

and maxima of det

n K(A) at A values smaller than A* may only point to a
~

posslhle j~ of the stability limit load ~f a minimum exists prior the

~ (see fig. 9).

5. COnclusions

Usmg theorems 1 and 2 a nUll'erical procedure - which calculates in each

incremental step the behaviour of the detenm.nant of the current tangent
stiffness matrix and its first and second derivatives with respect to A -
allONS a classification of the considered structure concerning ~ts para-
meter sensitivity with regard to the stability limit. In a future paper
the use of theorew 2 in carianahon w~th a steepest descent method will be
presented willch renders a procedure wll1ch can be used to calculate the b~­

furcahon set in the control surface (see e.g. hg. 8). 'Ihis procedure
w~ll help to hnd out critical caromations of system parameters at which a
sudden j~ of the stab~li ty limit load takes place. The kno.vledge of
such a j~ is of great importance for the design of supporting struc-
tures especially in the field of licjhtweigt construction.

6. References

Bathe, K.-J.; Rantn, E; Wilson, E.L: Fimte Element Fonrulahons for

Large Deformation Analysis. Int.J .Nu!rer .Meth.Engng. 9 (1975),
353 - 386.
2 Bathe, K.-J.; Boloorchi, S.; Ramaswamy, S.; Snyder, M.D.: Sane Can-
p.ltational Capabilities for Nonlinear Finite Element Analysis. Nucl.
Eng.Design 46 (1978),429 - 455.
3 Matthies, H.i Strang, G.: '!he Solution of Nonlinear Finite Element
Equahons. Int.J.Numer.Meth.Engng. 14 (1979), 1613 - 1626.
4 Rantn, E.: Geanetrisch nichtlineare Elastostatik und finite Elemente.
Habilitationsscilrift. Univ. Stuttgart. 1975.
108

5 Gallagher, R.H.; Mau, S.: A Method of Limit Point Calculatioo in

Finite Element StruCtural Analysis, NASA CR-2115. 1972.
6 Brendel, B.: Geanatrisch nichtlineare Elastostabilitat. Bericht
Nr. 79-1, Inst. f. Baustatik, Univ. Stuttgart. 1979.
7 Brendel, B.; Ranm, E.; Fischer, D.F.; Ramnerstorfer, F.G.: Linear
and Nonlinear Stability Analysis of 'Ihln Cylirxlrical Shells Under
Wind ~. To be published in: J.Struct.Mech. 9 (1981).
8 Bergan, P.G.; Harrigrooe, G.; Kr9keland, B.; S~eide, T.H.: Solution
Techniques for Non-Linear Finite Element Problems. Int.J.Numer.
Meth.Engng. 12 (1978), 1677 - 1696.
9 Bergan, P.G.: Solution Algoritiuns for Noolinear Structural Problems.
Proc.Int.Conf.Engng.Appl. FEM, H¢vi.k, 1979.
10 Bergan, P.G.; Roland, Li S~eide, T.H.: Use of the current Stiff-
ness Paraneter in Solution of Nonlinear Problems. Energy Methods
in Finite Element Analysis. R. GlcM..nski, E. Y. Rodin, O.C. Zien-
kiewicz (editors). Jchn Wiley & Sons, Cllichester - New York -
Brisbane - Toronto: 1979
11 Ramlerstorfer, F.G. ; Fischer D.F.: Nonlinear Elastic-Plastic
Stability and Contact Problems Concerning the Straightening of a
Wave-Ll.ke Deforrred Strip Plate. To appear in Proc.1st Int.Conf.Num.
Meth. for Non-Linear Problems. Swansea: 1980.
12 Brendel, B.; Hafner, L.; Ranm, E.; Sattele, J.M.: Prograrmdok.umen-
tation - Programnsystem NISA. Univ. Stuttgart, Inst. f. Baustatik.
1977 •
13 '!han, R.: Topological Models in Biology. Topology 8, (1969),313-335.
14 '!han, R.: Structural Stability and Moqilogenesis. New York: BenJa-
min 1975.
15 Zeanan, E.C.: Catastrophe Theory. Scientific American, April 1976,
65 - 83.
16 Troger, H.: Zur Einteilung von Sprungeffekten in nechanischen
Systemen. ~ 54 (1974), T177 - T179.
17 Troger, H.: Ein Beitrag zum Durchschlagen einfacher Strukturen.
Acta Mechanica 23 (1975), 179 - 191.
Behaviour of Finite Element Solutions Near a Bifurcation Point

F. BREZZI
Universitil di PaVIa, Italy

Summary

We analyze the behaviour of finite dimens10nal approximation

of Galerkin type in a neighborhood of a simple critical pOint.
Error bounds of opt1mal type are derived. Some computat1onal
aspects are also treated.

1. Introduction

The aim of this paper is to present some recent results in the

theory of approximation of nonlinear problems, with a particu-
lar accent on the behaviour of the approximate solutions in a
neighborhood of singular pOints, such as normal limit pOints
and bifurcation pOints. Although some computational aspects
w111 be briefly sketched, the ma1n interest w111 be focused on
the problem of error bounds. The well known definit10ns of the
singularities will be given first, for the sake of simplicity,
in fin1te dimension. In the sequel we shall briefly present
the continuous problem and the abstract hypotheses on the fin!
te d1mensional approximation. Then the main results on the
error estimate will be stated. An example of application to
the model problem of the von Karman plate equations will also
be given

2. Regular and singular points in finite dimension

r
Let !(~,A)=(fl(xl, .•. ,Xn,A), ••. ,fn(xl, ... ,Xn,A» be a C -maE

ping (r~3) from Rn+1=RnXR into Rn and consider the problem

110

jfind pairs (~,A) such that

1 ! (~, A) =0.
(2.1)

Let (~O'AO) be a solution of (2.1), that ~s

!(~O,AO)=O. (2.2)

The total and partial Frechet derivatives of f at the pOint

(~O,AO) are defined as follows (Mms being the space of mxs
matrices and using the convention A=X n +1 ).

(Df O ) •• = ( af;~) {'~= 1 , ••• ,n

-~]
] - -
ax. X=XO,A=AO j=1, ..• ,nH (2.3)

(DfO) .. i,]=1, •.• ,n

- ~]
(2.4)

(2.5)

Remark - In the general case of a mapping F from a Banach

space El into another Banach space E2 the Frechet derivat~ve

of F at a point eOE El is defined as the unique linear map-

ping DFo from El into E2 that verifies

IIF(eo+e)-F(eo)-DFo.eli E ~ o( Ilell E) eEE 1 (2.6)

2 1

It is easy to check that (2.3)-(2.5) satisfy the general de

finition (2.6). In its turn the map xO+DFo (from El into
~ (E 1 ,E 2 » can be d~fferent~ated (~f F is smooth enough) i
this leads to h~gher order Frechet derivatives. In our fin~te
dimens~onal example we would have, for ~nstance,D fO:RnxRn+R n
xx-
given by

D fO: v+ (_z+ (2.7)

xx- .L

or equivalently
a2f~
D fO:
xx-
(:i,,_z) + L ax] aX~ z
r
y
j
(2.8)
r

Defin~t~on - A po~nt (~O,AO) solution of (2.1) is called a

regular point if the Jacobian matrix D fO is nons~ngular.
x-

The well known ~mpl~c~t function theorem ensures that if

 111

(~O,AO) lS a regular pOlnt then there eXlsts in a neigh-

borhood xxA of (~O,AO) a unlque branch A+~(A) such that:

'VA EA. (2.9)

Definition - A pOint (~o,Ao) Solutl0n of (2.1) lS called a

simple critlcal point of f if zero is a simple eigenvalue of
o fO.
x-

Assume now that (~O,AO) is a simple critical pOint of f, and

let (0 ){-fO)* be the transposed matrlx of 0 x-fO. It lS well
known that there exist a unique (up to a sign) pair of vectors
1°,1* such that

(2.10)

(2.11)

Among the simple critlcal pOlntsthere are pOints for WhlCh

the crlticality lS, in some sense, more apparent than real.
These are pOints which belong to some branch of solutions
that cannot be parametrized wlth respect to A.

Definltion - A simple critical pOlnt is called a normal li-

mit pOlnt lf DA!O lS not ln the range of Dxio, or, equivaleg
tly, lf
(2.12)

In order to simplify the notations, we shall assume from

now on that

..P. * = (0,0, .•• ,0,1) (2.13)

so that the last row and column of 0 f O are identically zero.

x-
ThlS wl11 not result in a loss of generality; in fact, since
zero is a slmple elgenvalue, we can always make a change of
coordinates (llke ln the Jordan normal form) such that (2.13)
is satisfied. It is also easy to see that ln such case the
submatrlx

s lJ. i,j=1, ... ,n-1 (2.14)

112

has to be nonsingular.

If (XO,A O) is a normal limit point condition (2.12) implies

(DA!o)n~O; interchanging xn and A=X n + 1 , we are now back to
the case of regular points.

Remark - In other words we could say that a normal limit

point is a point (~O'AO) such that the unique branch of so-
lutions through it has a tangent that is "orthogonal" to
the A axis. If in addltion one has

then (~O,AO) is called a nondegenerated turning pOlnt; it

is easy to see that in such cases the branch through (~O'AO)
verifies, in a neighborhood, A~AO (A~AO respectively).

Let us deal finally with simple blfurcation points. Assume

that (~O,AO) is a simple critical point and that (2.12) is
not satisfied. Wrlting

{
X' = (x -.... , ... ,xn-l ,0)
~=~'+a1° (2.16)
a = xn

equation (2.1) can be splitted as

i=1, ••. ,n-1,

~f
fi(~I+a1°'A)=0
(2.17)
(x'+a,f,O ,A)=O
n - .:t.

Since the matrix SiJ defined in (2.14) is nonsingular, the

implicit function theorem can be applied to deduce, from
the first n-1 equation of (2.17), the existence of a unique
mapping ~'=~' (a,A), defined in a neighborhood of (X~,AO)
such that

f. (x' (a,A)+a,f,O ,A)=O i=1, ... ,n-1 (2.18)

1 - .:t.

Hence (2.17) is reduced to the last scalar equation

f
n (x'
- (a,A)+a,f,O
.:t.
,A)=O (2.19)

in the unknowns a,A. We set for slmplicity

g(a,A)=f (x' (a,A)+a,f,O ,A). (2.20)

n - .:t.

It is an easy matter to check that

 113

(2.21)

and therefore for the study of the solutions of (2.19) we

have to look to the second derivatives. It is well known for
~nstance that ~f

det (2.22)

r-2
the solutions of (2.19) lie on two C branches crossing
transversally at (X~,AO). If on the opposite the determinant
in (2.22) ~s positive, then (X~'AO) is an isolated solut~on
of (2.19). The case det=O corresponds to a higher degeneracy
and w~ll not be studied here.

Defin~tion - Let (~O,AO) be a simple critical pOint of f such

that

We set

g(a,A)=f (x' (a,A)+a<p° ,A)

n - -
where ~'(a,A) is the implicit funct~on defined ~n (2.18). If
the determinant of the Hessian matrix of g at (X~,AO) is ne-
gat~ve then (~O,AO) is called a simple bifurcat~on pOint.

At a simple b~furcat~on point the directions tangent to the

branches can be computed as follows: let a O be the unique
solution of

~
DX!O£.O = -DA!o
(2.23)
(£. o ,1.* ) = 0,

If (~,A) RnxR ~s a tangent to a branch through (~O,AO), we

must have
II X112 + ~2 = 1 I (2.24)

(2.25)

Hence x must be of the form

(2.26)
114

and we have only to compute ~l and ~2' An easy computation

shows that

(2.27)

substituting (2.26) into (2.27) and using (2.24) and (2.25)

we may eas~ly compute ~l and ~2 and hence ~ and \. For more
details see rJl.
3. The infinite dimensional case

From now on we shall restrict ourselves to nonl~near problems

of the follow~ng form:

F(U,A)=O F:VxR+V (3.1)

where F can be written as

F(u,A)=u+TG(u,A); (3.2)

we make the following assumpt~ons:

V is a Banach space, (3.3)

G(U,A) is an r-times continuously differentiable

(3.4)
mapping from VxR ~nto another Banach space W(r~4),

T ~s a l~near compact operator from W ~nto V. (3.5)

Example - Let us consider the class~cal von Karman nonl~near

plate equations:

- ~[w,w] in n,
(3.6)
l¢+A¢,w] in n,

w = aw = ¢ = II = 0 on an . (3.7)
an an

with w=transversal displacement, ¢=stress function,

[u,v}=uxxVyy+UyyVxx-2UxyVxy and ¢=qiven funct~on. Setting

(3.8)

G: ((w'¢)'A)+(~LW'WJ,-[¢+A~,wJ) (3.9)
 115

T
(3.10)
~~~i = P1 ' ~1 H~(n) i=1,2
and applY1ng T to both sldes of (3.6) we are 1n the frame-
work (3.2) - (3.5) .

Let now go back to (3.2) and let (uo,A o ) be a solution of

(3.1). We denote by DuFo,DAFo,DuuFo,DuAFo,DAAFo the f1rst
and second Frechet der1vatives of F at the p01nt (uo,A o ).
We recall the following def1n1t1ons, Wh1Ch are the natural
extens10n of the fin1te d1mens1onal case.

Def1n1t1on - The point (uo,A o ) lS called a regular p01nt 1f

D FO lS an isomorph1sm from V onto 1tself.
u

Def1n1t1on - The p01nt (uo,A o ) lS called a slmple critical

p01nt if zero lS an e1genvalue of D FO w1th algebra1c mult1-
u
plic1ty 1.

In analogy w1th the f1n1te d1mens1onal case at a slmple crit~

cal p01nt the two eigenvectors ¢o and ¢* w1ll be uniquely de-

term1ned by the equat10ns

\DUFO¢O = 0 II ¢ °Il V = 1
(3.11 )
1(D FO)*¢* = 0 (¢0,¢*)=1
u

Defin1t1on - The p01nt (uo,A o ) is a normal lim1t p01nt if 1t

lS a slmple cr1t1cal p01nt and

(3.12)

If 1n addit10n one has

(DUU FO (¢o ,¢o) ,¢*);;o!O (3.13)

then (uo,A o ) is a nondegenerate turn1ng pOint.

Def1n1t1on - The point (uo,A o ) is a simple bifurcat10n p01nt

if it is a simple cr1t1cal p01nt where

(3.14)
and
(3.15)
116

with
jAo=(DUUFO (cpo ,cpo) ,cp*)
lBO=(OU:\FOcpO+OUU FO (cpo ,0°) ,cp*) (3.16)

° :\:\ FO+20 u:\ FOao+D uu FO (0° ' 0°) ,~,,*)

?C =(0

and 0° is the unique solution of:

(3.17)

4. Finlte dlmenslonal approximatlon

Assume again that hypotheses (3.2)-(3.5) of the previous sec

tion are satisfled. For the sake of simplicity we shall deal
only wlth approxlmations "of Galerkln type", although the
theory in [2J, [3J ,[4] is much more general. We assume then
that V is a Hilbert space and that the operator T is associa-
ted witn a bilinear continuous V-elliptlc form a(u,v) ln such
a way that for any f in V' we have

a(Tf,v)=(f,v) 'dVEV (W <;;V'). (4.1)

We also assume that we are given a sequence {Vh}h of finlte

dimensional subspaces of V and we define the operators Th:V'+V h
by
a (Thf, v h ) = (f ,vh ) 'dVhEVh . (4.2)

We set the approximate problem:

(4.3)

Example - In the case of the von Karman equations for a clam-

ped nonllnear plate we may set

a((w,cp), (0'~))=fQ(6W60+6CP6~)dQ. (4.4)

As a finite dimensional subspace Vh of V=H 2 (Q)xH 2 (Q) we could

°
take, for the slmplest case of a polygonal domain Q, pairs of °
some usual C1-conformlng flnite elements, like the composite
HSleh-Clough Tocher triangular element (12 d.o.f.) or the corn
plete P5 element (21 d.o.f.) or the reduced Bell element
 117

(18 d.o.f.) or any other that we may llke. For a nonpolygo-

nal ~ isoparametr1c elements should be considered. We noti-
ce that the operator Th defined 1n (4.2) assoc1ates to a
given pair (Pr,P2) the approximate f1nite element Solut1on
of the linear problem (3.10) so that the formulation (4.3)
1S the one that is used in pract1ce. We also p01nt out that
(T-T h ) (Pr,P2) measures the error 1n the fin1te element ap-
prox1mation of the linear problem (3.10).

We def1ne an "optimal order of convergence" by assum1ng that

there eX1st a subspace "~V (of smooth funct1ons) a posit1ve
constant c and an integer k>O such that

PEV' (4.5)

whenever the r1ght hand slde of (4.5) 1S f1nite.

Example - In the case of the von Karman equat10ns we have

k=2 for the HCT 12 d.o.f. element, k=4 for the complete P5
and k=3 for the Bell element (cfr. e.g. [5J). In any case
~=Hk+2 (~)XHk+2 (~).

It 1S well known that 1f (3.1) has a regular solution

(Uo,A o ) there exists a neighborhood U of U O such that for h
small enough (4.3) has a un1que Solut1on (Uh,AO) w1th UhEU;
moreover
[[UO-u h[ V'" O(h k )

prov1ded that uOE1f. For a slm1lar result 1n a more general

context see [2J.

Let us now state some results on the behav10ur of the solu-

tions of (4.3) 1n a neighborhood of a simple cr1tical pOint.

Theorem 1 - Let (UO,AO) be a normal lim1t point of (3.1).

There exists a neighborhood UxA of (UO,AO) such that for h
small enough both (3.1) and (4.3) have a unique branch of
Solut1ons 1n UxA. Moreover these branches can be parametr1zed
1n such a way that
k
[[u(a)-uh(a) [[ V + [A(a)-Ah(a) [ ~ ch [[uta) [['\1' (4.7)
118

Theorem 2 - If (uo,\o) lS a nondegenerate turning pOlnt of

(3.1) there exists a nelghborhood UxA of (uo,\o) such that
far h small enough (4.3) has a unique nondegenerate turning
pOlnt (uh'\h) in UxA and we have

(4.8)

wlth c independent of h, provlded that uo,T ¢ ** and the derl

vative of u in the direction of ¢o at (uo,\O) (tangent to
the branch) belong to ~.

We pOint out that we have a double order of convergence for

the critical values \0 and \~. The proof of theorem 1 and 2
can be found, in a more general framework, in [3J. PreVlOUS
results In this directlon where obtained for lnstance In
[8],[6J.

The dlfficulties connected wlth the actual computations of

the branch of approximated solutlons near a normal llmlt
point can be overcome wlth various techniques. See for instan
ce [10J, [12J and the references therein contained.

Theorem 3 - If (uo,\o) is a simple blfurcatlon pOlnt then the

re exists a nelghborhood UxA of (uo,\o) such that for h small
enough the set Sh of the solutions of (4.3) in uxA is composed
of two Cr - 2 branches which may not intersect (and, In gene-
ral, do not). In any case, denotlng by S the set of solutlons
of (3.1) In UxA we have
 119

(4.9)

provided that Uo,T*¢* and the two der1vatives of u 1n the

d1rection of the branches at (UO,AO) belong to~.

For the proof and for more general results we refer to [4J.
Theorem 3 states that the b1furcat1on pattern 1S not (ln
general) stable w1th respect to the f1n1te element discret1
zation. Except10nal cases are for 1nstance the bifurcation
from the tr1v1al branch and the syrnrnetry-break1ng bifurca-
t1on. See [4] for more deta1ls. PreV10US results 1n th1S d1
rect10n can be found for 1nstance 1n [9), [6J .
From the computat1onal p01nt of V1ew two classes of problems
ar1se when we follow a branch: one 1S due to the fact that,
1f we proceed too fast along the branch, we may jump over
the b1furcat1on p01nt without noticing 1t, and keep follow1ng
an unstable branch. Th1S usually results 1n a change in the
slgn of det(DuF h ); 1f the determ1nant 1S computed at each
step we are able to go back with a smaller size step and try
to locate the b1furcat1on point. Th1S 1S easy but not cheap
1n general. Another problem arises when we corne too close to
the b1furcat1on p01nt: then DuFh becomes (numer1cally) sing~

lar and the "d1rect1on of the branch" does not exist anymore;
we say that we are 1n a "black box". In order to go out of
the black box 1n the d1rect1on of the branches we may use
the follow1ng tr1ck: assume that we are at a p01nt (uh,X)
1n the balck box and let v h be the e1genvalue of DuFh Wh1Ch
1S smallest 1n modul~s. Let ¢- and -*
¢ be the corresponding
e1genvectors:
II ~ I =1 I (4.10)

(4.11)

We consider the following problem:

(I) _ The distance dist(A,B) of two closed sets 1S def1ned for

1nstance as

d1st(A,B)=max {sup 1nf Ilx-yll, sup 1nf Ilx-ylll.

x~A ye.B yt::B xEA
120

(4.12)
where

~ ~ ~* ~
(4.13)
-(A-A) (DAF h ,¢ )¢.

clearlyi h has a simple blfurcation pOlnt at (uh'~) (artif~

cial bifurcation). The tangent dlrectlons to the branches
of solutlons of (4.12) through (Uh'~) can be computed wlth
the procedure (2.24)-(2.27). If (u h ,\) is close enough to
the simple blfurcation point (UO,A O) the previous procedure
works and the tangent directlons to the artificial blfurca-
tion are good approximations to the tangent dlrectlons at
(UO,A O). In particular if II(Uh,~)-(uO'AO)II-:sO(hk) then i t
is easy to prove (using the technlques of [4J) that the tan
gent dlrections are computed with an error of O(h k ).
For other considerations on the computational aspects see
for instance the references of [1 OJ ' [12J .
For some extensions to the case of Hopf blfurcatlon or mul
tiple bifurcations see [1J and [11J respectively.

References

Bernardi C. (to appear)

Brezzi F. - Rappaz J. - Raviart P.A. - "Flnite dlmenslo

nal approxlmation of nonlinear problems. Part I: Bran--
ches of nonslngular solutions" - Rapport lnterne n. 52
(1979) Centre de Mathematiques Appliquees, Ecole Poly-
technlque, Palalseau, (submltted to Num. Math.)

Part II: Llmlt points." Rapport lnter

ne n. 64 (1980) Centre de Mathematlques Appliquees,
Ecole Poly technique, Palaiseau, (submitted to Num. Math.)

Part III: Simple bifurcation points".

Rapport interne n. 67 (1980) Centre de Mathematiques
Appliquees, Ecole Poly technique, Palalseau (submitted
to Num. Math.)
Clarlet P.G. "The finite element method for elllptic pr~
blems" North Holland (Amsterdam) 1978.

FUJll H. - Yamagutl M. "Structure of slngularltles and

ltS numerlcal reallzation In nonlinear elasticity"
Research Report KSU/ICS 78-06. Kyoto Sangyo Unlversity
(to appear in J. Math. Kyoto Univ.).
 121

Keller H.B. "Numerical solutions of blfurcatlon and

nonllnear elgenvalue problems" Appllcatlons of Bifur-
cation Theory (P.H. Rablnowltz ed.) Academlc Press
(New York) 1977

[8J Klkuchl P. "Plnlte element approxlmatlons to blfurca-

tlon problems of turnlng point type" Theoretlcal and
Applied Mechanlcs ~ (1979), 99-114.

Klkuchl P. "An lteratlve flnlte element scheme for bi-

furcation analysls of semlllnear elllptic equations"
Report Inst. Space Aero. Sc. n. 542 (1976), Tokyo Unl
versity.

[1 OJ Mlttelmann H.D. - Weber H. "Numerical methods for bi-

furcation problems - A survey and classiflcatlon"
Report n. 45 (1980) Universltat Dortmund, Lehrstuhl
Mathematlk III.

[11J Rappaz J. - Raugel G. "Pinite-dlmensional approximation

of blfurcation problems at a double elgenvalue" (to
appear) .

Reinhart L. "Sur la resolution numerlque del problemes

aux limltes nonllneaires par des methodes de continua
tion" These de 3 erne cycle - Universlte P. et M. Curle
1980.
Some Results in the Analysis of Thin Shell Structures

K.-J. BATHE, L.W HO

Massachusetts InstItute of Technology, Cambndge, USA

Abstract
This paper represents a progress report of some of the
research that we are conducting in the finite element ana-
lysis of thin shell structures. We consider our isopara-
metric displacement-rotation thin shell element and our
discrete-Kirchhoff-theory (DKT) plate/shell element, which
we are continuously refinlng for accurate and effective
geometric and materially nonlinear analysis. In the paper we
briefly discuss the locking phenomenon of the isopara-
metric element, the use of this element as a transition
element between shell surfaces and in shell-solid transi-
tions, and we give some results using the DKT element.

1. INTRODUCTION
Much research has been conducted during the last two decades
in the development of thin shell finite elements. However,
despite the large amount of research effort there do not
exist as yet what may be considered to be cost-effective,
reliable and general thin shell analysis capabilities.
We believe that an effective thin shell element should
satisfy the following criteria:
1) The element should yield accurate solutions when model-
ing any shell geometry and under all boundary and
loading conditions. In particular, the element should
not contain any spurious zero energy modes, so that
reliable results can always be expected: The theory
of the element formulation must be well-understood and
should not contain any "numerical fudge factors." These
conslderations are most important and many elements
that have been published do not satisfy thlS
criterion. Such element developments can represent
 123

interesting research but should not be used in actual

engineering analyses, because the generated analysis
results cannot be interpreted with confidence.

2) We should be able to use the element in the modeling

of general shell structures with beam stiffeners,
cut-outs, intersections, and so on.

3) The element should be cost-effective in linear as

well as in nonlinear, static and dynamic analysis.
In nonlinear analysis, the element should be applicable
to large displacement, large rotation, and materially
nonlinear conditions.

Various approaches have been advocated for the development

of thin shell elements [1-2]; however, based on the above
obJectives we have concentrated our research efforts onto
two procedures:

a) The use of displacement/rotation isoparametric elements

that can be employed with a variable number of nodes,
but are usually used with 9 or 16 nodes.

b) The use of a simple triangular element which is ob-

talned by superimposing the bending and membrane be-
haviors.

We have reported our formulatlons and flrst experlences

with these elements in previous publications [3-5]. Our
objective in this contribution is to present additional
results that we have obtained recently in our continued
research on the element formulations and implementatlons.
The aim in our research is to improve and refine the
performance of the elements to the maximum extent possible
but always subject to the constraints sUIT@arized in 1) to
3) above.

2. SOME RECENT RESULTS l"lITH OUR ISOPARAMETRIC DISPLACEMENT-

ROTATION ELEMENT

The use of an isoparametric displacement-rotation shell

element was earlier proposed by Ahmad et al. [6] for linear
124

analysis and Raw~ [7] and Kr~keland [8] for nonlinear

analysis. We have used the same concepts but refined
the formulation of the element in some details to enhance
1tS practical usage. The element tormulation 1S very
general because no specific shell theory is used; instead,
only the two following basic assumptions are employed on
the shell behavior:
(i) Particles originally on a straight line in the
direction of the "normal vector" to the midsurface
of the shell element remain on a straight line during
the deformation of the shell.
(ii) The stress in the direction of the "normal vector" to
the shell midsurface is zero.

The "normal vector" at any point of the midsurface of the

element is used in the description of the shell geometry
and displacements and is defined by the direction cosines
of the "normal vectors" at the nodal points. These
direction cosines are interpolated over the midsurface of
the shell elements. As previously discussed [3], the
element formulation reduces in the analysis of plates to
the finite element discretization of the Mindlin plate
theory. The formulation includes shear deformations which,
however, are assumed to be constant over the plate thick-
ness.
We have implemented the element with a variable number of
nodes, but in practical analysis the element is usually most effec-
tive using the 9-node (parabolic) or l6-node (cubic)
Lagrangian interpolations, as discussed further below. The
element can also be employed as a transition element between
solid and shell element idealizations or different shell
surfaces. Figures 1 and 2 illustrate some of tte features of
the element.
The major advantage of the element is its general applica-
bility, but it is also recognized that the use of the higher-
order elements can be expensive. For this reason
 125

/------.y

Figure I Nine-Node Isoparametric Shell Element

Flgure 2 Shell Elements with Variable Number of

Nodes and Transition Elements
126

much attention has been given to the use of the low-order ele-
ments. However, if the stiffness matrices of these elements
are fully integrated, the elements display a "locking phenomenon"
and therefore some reduced or selective integration techniques
have been proposed [1,2,9,10]. These techniques, which can also
be related to mixed formulations [11, 12, 5) , relieve the
"locking behavior" of the elements but because of the various
dlfficulties encountered have not resulted as yet into a shell
element that satisfies our three criteria of an effective prac-
tical shell solution capability. Indeed, based on the experience
available so far it appears that for the development of low-order
shell elements the approach used in the discrete Kirchhoff theory
is more effective [5].
In the following sections we consider two important ingredients
of the isoparametric shell element: the convergence of the
higher-order elements (when fully integrated) to a finite ele-
ment discretization of the Kirchhoff plate theory and the use
of the transition element.
Although the results of a number of investigations have been
published in which the locking phenomenon in conjunction with
reduced integration procedures was considered, there is still
a need for further insight into the problem of whether a mesh will
lock or not. The objective in our work was to address this
question first for the simpler case when using fully integrated
elements. We are currently extending the results obtained to
elements that are not fully integrated.

2.1 On the Use of the Higher-Order Elements

As shown earlier the low-order elements and the higher-order

serendipity elements can greatly overestimate the stiffness of
a thin plate when the element stiffness matrices are fully in-
tegrated [9, 13]. This phenomenon referred to as "element
locking" can be explained by considering the variational in-
dicator of the Mindlin plate theory,
 127

]]
h
3
""2 [1 AI5..T f.b 15.. dA + a £ yT ~Y J dA
(1)

L--TERM 1--' L----TERM 2-----'

-(potent~al of external loads)

where
dS X
ax
dW
dy
- Sy
dS
_ ---X
15.. y (2)
dY
dW
dX + Sx
dS X dS y
ay dX

['
1 v 0

~
E
12 (1-v 2 )
v 1

I-v
0 C
-s
Ek
2(1+v) 0 :}31
0 0 -2-

and
W transverse displacement of plate
section rotat~ons about x and y-axes,
respectively
h th~ckness of plate (assumed constant)
k shear correction factor
E,v modulus and Poisson rat~o

L characteristic length

The expression labelled "TERM 1" corresponds to the

bending stra~n energy and the expression labelled "TERM 2"
corresponds to the shear strain energy_ We notice that
128

for thin plates the constant a is large and can be

regarded as a penalty parameter, which enforces that the
shear strains are small and converge to zero as ~
de8reases. Hence, a Mindlin plate theory analytical
solution converges to the Kirchhoff plate theory solution
when the thickness of the plate becomes very small.
Considering now a finite element discretization of the
variational indicator, we can directly conclude that for
the finite element scheme to be applicable to the analysis
of thin plates, we must have that, with the interpolations
used on w, Bx and By' the two expressions for the shear
strains

(4)

(5)

admit very small values. Also, since we want to use full

numerical integration--to avoid the problem of generating
spurious zero energy modes in the element formulation--we
must have that the shear stra~n expressions in Eqs. (4) and
(5) can be very small throughout the element. Hence, our
criterion on whether an element or a patch of elements will
lock is simply that the finite element interpolations w~th
the available degrees of freedom must enable the cond~tions

o (6)

o ( 7)

throughout the element or patch of elements. However, this

means that a simple condition for a mesh of rectangular plate
elements not to lock is:
2q - k > np (8)

where
q total number of nodal points,
 129

k number of constrained degrees-of-freedom,maximum

corresponding to rotations about x and y-axes,
n number of finite elements,
p number of nodes (basis funct1ons) per element.
The relat10n 1n Eq. (8) 1S der1ved by consider1ng Eqs. (6)
and (7) when using the finite element displacement and ro-
tation interpolations. The maximum number of individual
polynomial terms that can occur in the shear expressions is
equal to np. If the relation in Eq. (8) is satisfied, the
coefficients of these terms can all be individually equal
to zero, Wh1Ch represents the worst case. Hence, the rela-
tion in Eg. (8) is a sufficient but not necessary condition
for the element mesh not to lock. Figures 3 and 4 illus-
trate the application of the condition in Eq. (8) for vari-
ous analysis cases and give computed results using ADINA
[14] (in each case one quarter of the plate was modeled) .

The relation in Eq. (8) is a sufficient but not necessary

condition because the np equations derived as described
above may be linearly dependent. In such case the patch of
elements may not lock although Eq. (8) is violated. Figure
5 summarizes some results obtained using tile- parabolic
Lagrangian element, for which the condition in Eq. (8)
predicted locking but the element meshes proved adequate
in the finite element solutions using ADINA. Hence, in
summary, the relation in Eg. (8) is a conservative rule for
telling whether a mesh locks.

We should emphasize that the slmple rule 1n Eq. (8) isstrictly

only applicable to rectangular (und1storted) plate elements.
The development of a similar criterion for qeneral plate and
shell eleMents appears to be much more difficult, since the element
distortions (measured on the natural element coord1nates)
enter into the calculation of the shear strains.

2.2 On the Use of the Transition Element

The transition element can be employed to model shell-

solid transltlons and the lntersections of dlfferent shell
130

NO OF
NO OF DO F POLYNOMIAL ADINA
ELEMENT TYPE
(2q- k) TERMS RESULTS
(np)

0 5 8 LOCKS

D 7 9 LOCKS

0 9 12 LOCKS

(a)
D o 0
17 16

Analysis of a Clamped Square Plate Modeled

DOES NOT LOCK

with a Single Shell Element

NO OF
ELEMENT MESH NO OF 0 OF POLYNOMIAL ADINA
CASE TERMS
CONSIDERED (2q-kl RESULTS
(np)

SIMPLY

·.
o • 78 64 DOES NOT LOCK
o 0 SUPPORTED

··..
o 0

o 0
o 0

CLAMPED 71 64 DOES NOT LOCK

SIMPLY
00 00 00 171 144 DOES NOT LOCK
00 o 0 00 SUPPORTED
00 o 0 00
00 o 0 00

00 o 0 00
00 o 0 00

CLAMPED 161 144 DOES NOT LOCK

(b) Analysis of a Square Plate Modeled with

Cubic Elements

Figure 3 Application of the Locking Criterion

 131

NO OF 00 F NO OF
ELEMENT MESH CONSIDERED POLYNOMIAL ADINA RESUIJ"S
(2q-k) TERMS
( np )

Ern 89 108 LOCKS

Em
•• LOCKS
97 112 ('2 - NODE ELEMENTS)
ARE LOCKING

••
•• •• • •
•• ••
••
••
•••• •••• 161 144
DOES NOT
LOCK
•• •• • •
•• •• ••

Figure 4 Application of the Locking Criterion

to the Analysis of a Clamped Square
Plate Modeled with Different Elements

NO OF
ELEMENT MESH NO OF DOF POLYNOMIAL ADINA
CASE (2q- k) TERMS
CONSIDERED RESULTS
(np)

EB
SIMPLY
36 36 DOES NOT LOCK
SUPPORTED

• •
CLAMPED 31 36 DOES NOT LOCK

SIMPLY
78 81 DOES NOT LOCK
SUPPORTED
• • •
• • •
• • • CLAMPED 71 81 DOES NOT LOCK

Flgure 5 Appllcatlon of the Locklng Criterion

to the Analysis of a Square Plate
Modeled wlth Parabollc Elements
1~

surfaces [2]. The use of the element is relatively easy

because no special constraint equations need be written at
the transition regions and yet full compatibility between
the elements is preserved. However, an important question
must be whether the element predicts displacements and
stresses accurately enough in the transition regions.
Figure 6 shows a simple folded roof structure that we
analyzed using the transition element. Figure 7 lists
the various finite element idealizations used. Since no
analytical or accurate numerical solution of the stresses
in the transition region could be located we used the
idealization in Fig. 7(a) to obtain an accurate prediction
of the stress components. The model in Fig. 7(b) corres-
ponds to the usual procedure of modeling shell intersec-
tions, whereas the model in Fig. 7(c) may be attractive
because of ease of program input of only one normal. The
definition of the model in Fig. 7(d) is equally effective,
but we need to question the accuracy of the results of the
stress predictions.
Figures 8 and 9 give the stresses predicted in the analyses.
It is interesting to note that the results with an average
normal at the shell intersection are not far from the re-
sults using the usual modeling procedure for shell inter-
sections (the model in Fig. 7(b». The results obtained
with the transition element are different near the transi-
tion region, but they show the trend of the 2-D fine mesh
results. Also, they indicate that the transition element
should be used to model only the actual transition region.

Figure 10 shows the model used in the next analysis of the

folded roof, which was established based on the above con-
siderations. Figures 11 and 12 give the results obtained
in the analysis of this model. It is seen that the pre-
dicted stresses are significantly more accurate than those
calculated with the usual shell idealization (in Fig. 7(b».
We should note that the number of degrees-of-freedom 1n the
analysis was about ~ th of those used in the two-dimension-
al model. Hence, although more detailed studies of the
 133

thickness = .124
E = 30,000, 000
8 = 30°

Figure 6 Folded Cantilever Plate Subjected

to a Line Load at Its Tip

(a) Two-Dimensional Finite Element Mesh

Figure 7 Models Used for the Analysis of the

Folded Cantilever Plate
134

(b) Shell Elements Using Constraint Equations

(c) Shell Elements Using the Average Normal Technique

(d) Transition Elements

Flgure 7 (Continued)
 135

600
Top Integration POints
(I e Olodlstonce of 0358

-
above the mid - surface)
500
4,4,2 I 3,3,2
l>-------<>
400 ------ 1
..--..
D- -(]
0---<)
b~ +---+
U)
U)
w 300
'"
I-
U)

200

100

025 20 225 25
OISTANCE ALONG FROM THE X-AXI S

Figure 8 Predicted Stresses with Folded Plate Models

of Fig. 7 (Gauss Integration Used)

- 200
b~

U)
U)
w - 300
'"
I-
U)
Bottom Integration POints
(I e at a distance of 0358
- 400
below the mid - surface)
4x4x2 3x3x2
Constraint Equation ~ l!r----ll.
Average Normal _________ o--~
- 500 TranSition Element _____ 0---0
2-D Fine Mesh +---+

Figure 9 Predicted Stresses with Folded Plate Models

of Fig. 7 (Gauss Integration Used)
136

- y

Figure 10 A Refined Model of the Folded Plate with

Transition and Shell Elements

600
Top Int'li!ra'ion POints
(I e 03~8 above the mid-surface)
0--0 Shell Elem.nt Mesh
500 +-+ 2 - 0 Fine Mesh

400

~
w 300
'"
~

'"
200

100

025 05 075 10 126 15 175 20 225 25

DISTANCE ALONG FOLDED PLATE MID-SURFACE FROM THE X-AXIS

Figure 11 Predicted Stresses with Folded Plate Models

of Figs. 7(a) and 10

- 100

- 200

b
'"'"<oJ - 300
'"
~

'"
- 400
Bolfom IntegratIOn POlnls
(I e 0358 below the mld- surface)
o---oShel1 Element Mesh
- 500 +-+ 2-D Fine Mesh

05 075 126 15 175 20 225 25

025
DISTANCE ALONG FOLDED PLATE MID-SURFACE

Figure 12 Predicted Stresses with Folded Plate Models

of Figs. 7(a) and 10
 137

transi tion element are still necessary t this example analysis

shows the potential in the practical use of the element.

3. SOME RECENT RESULTS WITH OUR DISCRETE-KIRCHHOFF-THEORY TRIAN-

GULAR ELEMENT
The triangular element developed has three corner nodes and
six degrees-of-freedom per node as shown in Fig. 13. The
element is flat and therefore a shell surface is modeled
with the element as an assemblage of flat facets in the way
some of the earliest shell analyses were carried out [1,2,15].
The primary obJective in the development of the element was
to have a very simple and cost-effective element of good
accuracy as an alternative to the higher-order isoparametric
elements. The three-node DKT shell element stiffness matrix
is constructed as follows [16]:

1) The bending behavior is described by a discrete Kirch-

hoff formulation. This description has been found to
yield the most effective three-node plate bending ele-
ment of a large nemmer of elements considered in a sur-
vey study [4, 5]. The element has high accuracy, has
no spurious zero energy modes, is computationally very
inexpensive and the programming is very simple (less
than 100 Fortran statements}. The element is very
likely the most effective simple plate bending element
currently available.
2) The membrane behavior is currently described by the
simple constant strain triangle. However, it would be
desirable to increase the order of accuracy of the
membrane description.
3) The above stiffnesses are combined in a local coordinate
system and the normal rotational stiffness is arbitrari-
ly set to llO~OO ) x (smallest bendlng stlffness) in order
to obtain 6 stlffness degrees-of-freedom per node.

The element can be used very effectively in geometric non-

linear analysis because it remains flat during the response
history. Also, for elastic-plastic analysis it is effective
138

x x

(a) Configuration and Degrees-of-Freedom

Z Z Z

8z 3
y y y
w3
-
v3 3
8Y3 v3
x x 8Y3
u3

!SM Ke K

~KM' I

K
=~~~~=a=
I I ]
-

I I Z

(b) Element Stiffness Matrix Components

Figure 13 The DKT Three-Node Shell Element

 139

to use in the material descr1ption stress resultants,which,

as in elastic analysis, circumvents the numerical integration
through the element thickness.

To indicate the effectiveness of the element we are present-

ing in the following the results obtained in the analyses
of three problems.

3.1 Analysis of a Square Plate

A square plate of side lengths 2a with clamped and simply-

supported edges was analyzed [5]. Because of the symmetry
conditions only one quarter of the plate was considered.
F1gure 14 gives the fin1te element meshes used in the
analyses for N = 2,4 and 8. F1gure 15 summarizes the
results obtained for the centre deflection of the plate.
We may note that, considering the computational effort in
the Solut1on, these results should be compared with 2x2,
4x4 and 8x8 mesh idealizations using four-node square
elements or lxI, 2x2 and 4x4 meshes with nine-node square
elements. Such evaluation and a comparison with other 3-
node elements show that the 3-node DKT element is indeed
very effective.

3.2 Analysis of a Pinched Cylindrical Shell

The structure analyzed and a typical finite element ideali-

zat10n used are shown in Fig. 16. In the figure the 10xlO
mesh used 1S shown, but the analysis was also carr1ed out
w1th 4x4, 6x6, 8x8 and 16x16 mesh topolog1es. Figures 17
to 19 give calculated displacement and stress resultant
distr1butions along the Ilnes DC, BC and AD of the shell,
respectively. It is seen that the finite element predic-
tions converge rapidly to the analytical solution as a
reasonable number of shell elements is employed in the
structural idealization. Table 1 summarizes the solution
times used in the analysis of the shell structure.

3.3 Large Displacement Analysis of a Simply - Supported Plate

The plate was subjected to a uniform pressure loading q. Figure

20 shows the finite element idealization used for the plate and
 ~
y

25
- - - - - - - S-S, CONCENTRATED LOAD
-----s-s, DISTRIBUTED LOAD
------CLAMPED, CONCENTRATED LOAD
20 .t' " "L o - - - - CLAMPED, DISTRIBUTED LOAD
• x
u 20
!i
z
o
~ 15
I&J
....J
"-
I&J
o
Z 10
It:
o
It:
It:
I&J
~ 5
ACZIZr A~B Alilf 0....,
............
'o- _ _ _ _ _ ~~_-
c~~121 C~D cello o o------1p===r-~_ 8
6 'r" 4 6
N=2 N=4 N=8 2 N

Figure 14 Element Meshes for the Analysis Figure 15 Percentage Error in the Predicted Centre
of a Square Plate using DKT Deflection of the Square Plate Using
Elements Element Meshes of Fig. 14
 U -displacement s and
rotations about Y, Z
axes zero along this
RIGID V-displacements and
DIAPHRAGM r.otatlons about X,Z
P axes zero along this
SUPPORT edge
(V=W=y=O)

B
Y,V
X aXIs zero along
Z ,W T this arc
RII = 100
LlR = 2
RIGID ... = 03
DIAPHRAGM x, Y axes zero along
SUPPORT this edge
(V=W=y =0) Y,V A

(a) Structure Considered (b) Typical Finite Element Mesh (lOXlO)

Used in the Analysis

Figure 16 Analysis of a Pinched Cylindrical Shell

Structure Using DKT Elements :::
142

D
0 6-0_,
-"'AO_i
- 0........

-50
0 .....
........
0

Etw ........... O,.a 0

p ~
AP~a 0

-100 --ANALYTICAL SOLUTION ~~...

o 4x4 MESH .......... L
.. 6x 6 MESH " ..
-150 ¢ a x a MESH \. ~
o loxlO MESH ~
6 16xl6 MESH

-200

D C
0 "·o-~-¢-6 -1I-6-¢_ I
itl_~~

"'-'1
o 'lI.~
-..¢
-5 o ~
RNa
p
. .....~
-10 --ANALYTICAL SOLUTION

..
0 4x4 MESH
6x6 MESH
~...
-15
¢
0
axa MESH
10xi0 MESH ~ 6
6 16Xl6 MESH

-20

Figure 17 Predicted Displacement and Stress Distributions

along DC of Shell in Fig. 16
 143

D C
o

-5

-10 --ANALYTICAL SOLUTION

o 4x4 MESH
• 6x6 MESH
-15 o 8x8 MESH
o 10xi0 MESH
• 16xl6 MESH

-20

-0.2 --ANALYTICAL SOLUTION •

o 4x4 MESH
• 6)(6 MESH
o 8x 8 MESH
o 10xi0 MESH
• 16 )(16 MESH

-0.1

I
C

Figure 17 (Continued)
144

50
,,~-.~'\.

Et Wn
-p-
a
B
.-~.~./, 1 1\1 i
C
I

-50 \
\
CI

-100 --ANALYTICAL SOLUTION

\
0 4x4 MESH "
• 6X6 MESH
exe MESH
-150 0
o loxia MESH
" 16 Xl6 MESH

-200

B
a
RNa
p
-5

-10 --ANALYTICAL SOLUTION

0 4x4 MESH
• 6x6 MESH
0 axe MESH
-15
CI lOxia MESH
" 16Xl6 MESH

Figure 18 Predicted Displacement and Stress Distributions

along Be of Shell 1n Fig. 16
 145

B C
0 ~..."..~-....~~~~c- I
)'.
~
RNji
P
-5

-10 - - ANALYTICAL SOLUTION

\. II

~
0 4K4 MESH
• 6K6 MESH
0 8 K8 MESH
-15
o 10KI0 MESH ...
... 16KI6 MESH
\
-20

...
0.2
--ANALYTICAL SOLUTION
0 4x 4 MESH
• 6x6 MESH
0 8x8 MESH
0

~ o 10XI0 MESH
P ... 16XI6 MESH 0
0.1 ...

Figure 18 (Continued)
146

-ANALYTICAL SOLUTION
o 4114 MESH
-4 " 6116 MESH
o 8118 MESH
o 101110 MESH
• 16 11 16 MESH
-2
Etu
p
O' A
o

0.003
--ANALYTICAL SOLUTION
o 4x4 MESH
0.002 " 6 X6 MESH
o 8x8 MESH
Ma13 o 10xi0 MESH
p
• 16 x l6 MESH
0.001

-0.001

-0.002

Figure 19 Predicted Displacement and Stress Distributions

along AD of Shell in Fig. 16
 147

the response predicted. In a first analysis, the edges were

constrained not to move in the plane of the plate, and in a
second analysis, the u and v edge displacemenmwere left
free. The predicted responses in these analyses are compared
in Fig. 20 with other solutions reported earlier.

4. CONCLUDING REMARKS

In this paper we have reported some recent results of our

research in finite element thin shell analysis capabilities.
Our present analysis capabilities can be employed in linear
analysis, in large displacement, large rotation and materially
nonlinear analysis, and for static and dynamic solutions. A
few demonstrative analysis results have been presented in the
paper, and the reader may refer to ref. 13-5, 14, 16] for addi-
tional analysis results.

We are conducting our research to obtain increasingly more

cost-effective and general solution procedures butwe noted that
primary emphasis is directed towards the reliability of the analysis
methods. This is so, because we are convinced that an ultimate
extensive usage of nonlinear finite element analysis capabili-
ties will depend primarily on the reliability of the solution
procedures available and much research work should be concen-
trated in this area.

Acknowledgment

We are grateful to the ADINA users group for supporting finan-

cially our research and development efforts, and we thank
C. Keilers and U. Tsach for having carried out some of the
analyses with ADINA.
148

500

400

I< 300

200
150
100
E =YOUNG'S MODULUS =107 PS I
~ = POISSON RATIO =l{O:i 300~~__~______L-.
h = THICKNESS =0.12 IN 4 12 22 TIME
a = WIDTH = 24 IN
q = UN I FORM APPL lED PRESSURE APPLIED PRESSURE LOAD
ALL EDGES ARE SIMPLY-SUPPORTED

J::.
.....
u PR ESE NT STUDY
~
LEVY
2
~
• ISOPARAMETRIC [3]
SHELL ELEMENT
II::
Z
2
.....
I.)

'"
..J
IL
IIJ
o
II::

'".....2
IU
I.)

LOAD PARAMETER l<=qa 4 /Eh 4

Figure 20 Large Displacement Analysis of a Simply-Supported

Square Plate Subjected to Pressure Loading Using
DKT Elements
 149

References

[lJ Zienkiewicz, O.C., The Finite Element Method, McGraw-

Hill, 1977.
[2] Bathe, K.J., Finite Element Procedures in Engineering
Analysis, Prenctice-Hall, Inc. in press.

[3] Bathe, K.J., and Bolourchi, S., "A Geometric and

Material Nonlinear Plate and Shell Element," J.
Computers and Structures, Vol. 11, pp. 23-48.
[4] Batoz, J.L., Bathe, K.J., and Ho, L.W., "A Search
for the Optimum Three-node Triangular Plate Bending
Element," Acoustics and Vibration Lab., Report 82448-8,
Dept. of Mech. Eng., M.I.T., Dec. 1978, 1980.
[5] Batoz, J .L., Bathe, K.J., and Ho, L.W., "A Study of
Three-Node Triangular Plate Bending Elements," Int.
J. Num. Meth. in Eng., in press.
[6] Ahmad, S., Irons, B.M., and Zienkiewicz, O.C., "Analy-
sis of Thick and Thin Shell Structures by Curved
Elements," Int. J. Num. Meth. in Eng., Vol. 2, pp.
419-451, 1970.
[7] Ramm, E., "A Plate/Shell Element for Large Deflections
and Rotations," in Formulatlons and Computatlonal Al-
gorithms in Finite Element Analysls, Bathe, K.J.,
Oden, J.T., and Wunderlich, W., (eds.), M.I.T. Press,
1977.
[8] Krakeland, B., "Large Displacement Analysis of Shells
Considering Elastic-Plastic and Elasto-Viscoplastic
Materials," Report No. 776, The Norwegian Institute
of Technology, The University of Trondheim,
Norway, Dec. 1977.
[9] Hughes, T.J.R., Taylor, R.L., and Kanoknukulchai,
W., "A Simple and Efficient Finlte Element for Plate
Bending," Int. J. Num. Meth. in Eng., Vol. 11, 1977,
pp. 1529-1543.
[10] Zienkiewicz, O.C., Bauer, J., Morgan, K., and Onate,
E., "A Simple and Efficient Element for Axisymmetric
Shells," Int. J. Num. Meth. in Eng., Vol. 11, 1977,
pp. 1545-1558.
[11] Lee, S.W., and Pian, T.H.H., "Improvement of Plate
and Shell Finite Elements by Mixed Formulations,"
AIAA J., Vol. 16, Jan. 1978, pp. 29-34.
[12] Malkus, D.S., and Hughes, T.J.R.,"Mixed Finite Element
Methods--Reduced and Selective Integration Techniques:
A Unification of Concepts," Computer Methods in Appl.
Mech. and Eng., Vol. 15, 1978, pp. 63-81.
[13] Pugh, E.D., Hinton, E.,andZienkiewicz, O.C., "A
Study of Quadrilateral Plate Bending Elements with
Reduced Integration," Int. J. Num. Meth. in Eng.,
Vol. 12, 1978, pp. 1059-1078.
150

[14] Bathe, K.J., "ADINA - A Finite Element Program for

Automatic Dynamic Incremental Nonlinear Analysis,"
Report AVL 82448-1, Mech. Eng. Dept., M.I.T., Sept.
1975 (rev. Dec. 1978).
[15J Clough, R.W., and Wilson, E.L., "Dynamic Finite
Element Analysis of Arbitrary Thin Shells," J.
Computers and Structures, Vol. 1, pp. 33-56, 1971.

[161 Bathe, K.J., and Ho, L.W. "A Simple and Effective
Element for Analysis of General Shell Structures, "
paper to be presented at the ADINA Conference, June
1981, M.LT.

Table 1 Total Solution Times in Analysis of Pinched

Cylindrical Shell (on a CDC Cyber 175)

Grid Solution time (sec. )

4x4 1.00

6x6 1.86

8x8 3.19

lOxIa 5.73

16x16 21.94
Nonlinear Dynamic Finite Element Analysis of Shells

TlR.HUGHES*, W.K.LIU", I.LEVIT*'

CalIfornia InstItute of Technology, Pasadena, USA

Summary
Dynamlcal aspects of a general nonllnear flnlte element shell
analysls procedure are descrlbed. The work extends preVlOUS
endeavors of the authors on quaslstatlc plate and shell anal-
YS1S. Several sample problems of a two- and three-dlmenslon-
al nature are presented WhlCh demonstrate the appllcablllty
of the methodology.

1. Introductlon
In thlS paper we are prlmarlly concerned wlth the nonllnear
dynamic flnlte element analysls of shells. The work de-
scrlbed hereln lS a sequel to a general nonllnear statlc
formulatlon presented In [20] and [21] and referred to In
thlS paper as "Part I" and "Part II", respectlvely. Due to
the extenslveness of Parts I and II, and In an effort to
keep the present ~ork as brlef as posslble, the detalls of
the statlc aspects of the theory wlll not be repeated here.
However, to enable the reader to better appreclate the "blg
plcture" we wlll brlefly mentlon the sal lent features of
Parts I and II.

Part I concerns ltself wlth three-dlmensional and fundamen-

tal a~pects of nonllnear flnlte element shell analysls.
Large straln and flnlte rotatlon effects are accounted for.
In addltlon, a falrly general class of nonllnear flnlte-
deformatlon constltutlve equatlons may be accommodated.
The elements are generallzatlons of reduced/selectlve lnte-
gration Mindlln plate elements WhlCh have been descrlbed In
prevlous works of the sen lor author and colleagues [15-17,

'Assoclate Professor of Structural Mechanlcs

**Graduate Research Asslstant
152

,24J. The extension of the selective integration procedure

to the nonl1near case is novel, the basic idea having been
described earlier for a related sltuation [12,13J. This in
particular facilitates the development of a "heterosis-type"
nonl1near shell element Wh1Ch exhibits high accuracy w1thout
ostensible defects. It has been noted in the past that the
reduced/selective integration process is equivalent to a
m1xed formulat10n [29J and thereby has some 1mplementational
advantages by v1rtue of the eliminat10n ab initio of the
auxiliary f1eld. However, some new thoughts on the matter
have recently appeared [30J and advantages of directly 1m-
plement1ng the mixed formulation are argued. Th1S and other
aspects of shell element behavior w111 no doubt remain active
areas of research for the coming years. Another noteworthy
aspect of Part I 1S the development of a const1tutive algo-
rithm which 1S "objective" for large rotation increments
(i.e., appropriately treats f1n1te rotation effects) and
maintains the zero normal-stress condit10n in a rotat1ng
stress coord1nate system. Lastly, the treatment of finite-
rotational nodal degrees-of-freedom precludes the appearance
of zero-energy 1n-plane rotational modes and is felt to be
espec1ally slmple and clean.
In Part II, two-d1mensional shell situations are dealt w1th.
The formulation appl1es to the practically important cases
of r1ngs, tubes, beams, frames, and shells of revolut10n.
The point of view adopted is to deduce the two-d1mensional
spec1al cases from the three-d1mens10nal formulation of
Part I. This results in a conceptual economy and perm1ts
much of the three-d1mensional software to be used for the
two-dimens10nal cases, particularly const1tutive routines.
Thus the implementational phase 1S considerably simplified,
however, some computat10nal overhead results due to the
degree of general1ty attempted in one program package. De-
tails are presented of how to modify kinematic quant1t1es
so that the particular constraints of the various two-
d1mensional cases are manifested (e.g., in- and out-of-
plane zero normal stress conditions). The elements employed
 153

are the uniform reduced integration Lagrange elements, which

have very nice properties for this class of problems. A
number of numerical problems are presented in Parts I and II
and the ~nterested reader is urged to consult these refer-
ences for further information.

In this work we begin where Parts I and II leave off. In

Section 2 we describe the construction of element mass ma-
trices. (The construction of all other element arrays is
described in Parts I and II.) Particular attention is paid
to the development of lumped mass matrices which yield large
critical time steps for explicit transient analysis. The
basic idea employed is that originally presented by Key and
Beisinger [25]. In Section 3 we present aspects of the
transient algorithm. For the most part the presentation is
very similar to our other recent works on this topic [14,18,
19,22,23] and 1S only br1efly sketched here. However, we do
p01nt out the spec1f1c mod1f1cat1ons Wh1Ch are necessary for
a successful trans1ent shell formulat1on, namely, those af-
fecting rotation-like nodal degrees-of-freedom. In Section
4, several numerical examples are presented which demon-
strate the capabilities. In addition to some dynamic elas-
tic-plastic calculations, we present some static elastic
buckling calculations which are of some interest.

2. Element Mass Matrices

Consistent mass

The element consistent mass matrix is defined by

+1

r::ab ~~ oJ -1J NT
-a Nbp
- 0j 0
d~ dO Sb
_
(2.1)

where 1 ~ a, b ~ n en , the number of element nodes; ~a is a

transformation matrix which maps rotational degrees-of-
freedom into so-called "fiber" directions; p is the mass
o
density in the initial configuration: J o is the Jacobian
154

determinant for the initial configuration; and

+1 +1

J J
... d~ dll (three d1mensions)

J ... do
-1 -1
(2.2)
+1
o
J ... dll (two d1mensions)
-1

N 0 0 N Z 0 0
a a a
(three
0 N 0 0 N Z 0
a a a dimensions)
0 0 N 0 0 N Z
a a a
N (2.3)
~a

N 0 N Z 0
a a a (two
dimensions)
0 N 0 N Z
a a a

in which N is the lamina shape function; and Z a = za (;;:) 1S

a
a thickness funct10n (see Part I for further deta11s).

Lumped mass

Element lumped mass matrices may be computed by performing

the following steps:

1. Calculate

+1
ds
Jo
J Jo
-1
(2.4)

rot
m
a J Na2
0
PoJ o do (2.5)

(po 1S assumed to be independent of s).

2. For elements other than heteros1s:

rot
m a = 1,2, ... ,n en (2.6)
a
For heterosis.

(2.7)
o
a = 1,2, ... ,8
 155

disp = 0
mg (2.8)

(N~er 1S the 8-node serend1p1ty shape funct1on, see

e. g., [34].)
-
3. V =
J Jo
0
dO M=
J PoJ o
0
de (2.9)

4. Normalizat1on
n
en
rot
Mrot = I m
a (2.10)
a=l
n
en
Mdisp
l mdisp
a
(2.11)
a=l

mrot
a
.;- (M/M rot ) mrot (2.12 )
a

disp
m
a
.;- (M/Mdisp ) mdisp (2.13)
a

5. Adjustment to rotational inertia:

<z > (z+ + z -) I 2 (2.14)

a a a

+
[za 1 z - z (2.15)
a a

+1
z2 dl;; I <z >2 + 1 [z 12
a
a J
-1
a
2
a 12 a
(2.16)

n
en
h I [za 1 I nen A V/h (2.17 )
a=l

a max{ a ,A/8} (2.18 )

a a

rot rot
m m a (no sum) (2.19 )
a a a

Remarks

a) In the axisymmetric case, j needs to be replaced

o
throughout by rJ , where r 1S the radial coordinate.
o
b) The rationale behind the lumped mass matrices has been
discussed more fully 1n Hughes et al. [171. The origlnal
156

idea is due to Key and Beisinger [25]. Although ad hoc in

nature, this technique can dramatically increase the allowable
time step in explicit transient calculations without adverse
effect on accuracy (see e.g., [11]).

3. Transient Algorithm
The transient algorithm is basically the implicit-explicit
predictor/multi-corrector scheme which has been described
previously (Hughes et al. [12,14,18,19,22,23]). However, a
slight reorganization is required to appropriately treat
rotational degrees-of-freedom. Let i be the iteration
counter; n be the time step number; ~t be the time step;
d ,v , and a be the (generalized) displacement, velocity,
~n ~n ~n

and acceleration vectors at step n, respectively; and define

the Newmark "predictors" by

d~n + ~ tv~n +.! (1 - 2 S) to. t 2 (3.1)

::n+l 2

::-n + (1 - y) to.t a~n (3.2)

::-n+l

where Sand yare the Newmark parameters. The global effec-

tive mass matrix is defined by

implicit
M* (3.3)
explicit

where ~ and ~ are the global tangent damping and stiffness

matrices, respectively. Various "mesh partitions" may also
be employed (see Hughes et al. [14,22] for further details) .
In each time step the following calculations need be per-
formed:
1. Set i = O.
2. Predictor phase:
d (0) (3.4)
~n+l ~n+l
v(O) (3.5)
~n+l ::-n+l
 157

a {oj = 0 (3 .6)
~n+l ~

3. Equation solution:

(residual or out-of-balance force)

(3.7)

4. In geometrically linear analysis go to step 5. In geo-

metrically nonlinear analys~s go to step 6.

5. Corrector phase for geometrically linear analysis:

(i+l) a{i) + !J.a (3.B)
~n+l ~n+l

(i+l) v + y!J.t a{i+l) (3.9)

'fn+l ~n+l ~n+l

d{i+l) = + S!J.t 2 (i+l) (3.10)

~n+l ~n+l ~n+l
If further iterations are required, increment i and go
to step 3. Otherwise, increment n and go to step 1.

6. Corrector phase for geometrically nonlinear analysis:

Employ (3.B) and (3.l0) .to obtain ~~!tl). The rota-
tional components of ~~~tl) are, ~n this case, viewed
as tentative, since they will not generally maintain
the fiber inextensibility condition. A radial projec-
tion procedure is used to appropriately redefine these
components (see Part I, eqs. 2.21-2.23 for details).
(i+l) . (HI)
The corrected ~n+l ~s then used to calculate ~n+l
via (3.10). Wlth this ln hand, (3.9) may be employed
. (Hl)
to deflne ~n+l . At this point, if further iterations
are requlred, lncrement 1 and go to step 3. OtherWlse,
lncrement n and go to step 1.

Remark

For static analysis purposes, the algorithm may be reduced

to a standard Newton-Raphson iterative procedure by setting
y = 0, S = !J.t 2 , d +1 = d , and ignoring all equations in
~n ~n

which the velocity appears, i.e., (3.2), (3.5), and (3.9).

158

4. Numerical Examples
The calculations described herein were performed on the
California Institute of Technology IBM 3032 computer in
double precision (64 bits/floating point word). In three-
dimensional situations 4-node quadrilaterals were employed
and in two-dimensional cases 2-node elements were used.
These elements are described in detail in Parts I and II,
respectively. In cases in which the material was completely
elastic, we used 2-point Gaussian quadrature through the
thickness ("fiber integration"). This has always been suf-
ficient in our experience. However, in elastic-plastic
situations we found it necessary to use 4-point Gaussian
quadrature through the thickness. In the calculations em-
ploying 4-node quadrilaterals, we used both selective (Sl)
and uniform (U1) reduced Gaussian rules for the lamina in-
tegration (see Part I for further details). For the two-
node element we used only the 1-point Gaussian rule in the
lamina direction (U1). The plasticity theory employed was
that due to Krieg and Key [27] which allows any arbitrary
linear combination of isotropic and kinematic hardening. A
plane stress radial return implementation was employed [26].
In all cases the corotationa1 hypothesis [1-3,5,7-9,33] was
adopted in the context of the constitutive algorithm pre-
sented in Part I. Newmark parameters were taken to be
S = 1/4 and y = 1/2 throughout.
4.1 Transient elastic-plastic response of a simply-
supported plate

This problem was solved previously by Liu and Lin [28] who
used small deflection theory and ignored membrane effects.
In our calculations we assumed full geometric as well as
material nonlinearity. One would anticipate some stiffen-
ing in this case due to the tensile stresses developed as
the plate deforms. Since the center displacement is of
the order of half the plate thickness, one would expect
small, but not insignificant differences. As may be seen
from the results of Figure 1, the expected behavior is
produced by our calculations which tend to be slightly
stiffer than L1U and Lln's wlth respect to peak response.
 159

2 .8
e lo s. ic - perfeclly
plcsl ic
2.4

L = 10 in
h :I 0.5 in (fh ic k]
E • 107 ps i
11 . 0.3
.8 .. ....... Liu end Lin 1979
p =
30X 10' ps i (y ie ld .Ire .. )
fTy =
2.588xI0-4 Ib-sec 2 /in4 .¥c - - - imp licit , u
l mped mass
• { 0 • <0 Q)
<) _ . - implic i t, consisten t mass
q 300 ps i 1>0 .4
- explicil, lumped moss
6. = { 2.230 x 10-5 sec implici t
5.575 x 10-6 .ec er.pli cil 0
sir. teen 51 elements
0 2 4 6 8 10
t i me ( 10 - 4 sec)

Fig \ Trans ient response of a sim ply-supported plate

This is more pronounced in the elastic perfectly plastic

case than in the purely elastic case as may be seen. This
is also reasonable, since larger membrane effects would tend
to develop due to the smaller bending resistance in plastic-
ity.
Comparisons between lumped and consistent mass matrices, and
implicit and explicit algorithms tend to be quite good.
Despite the fact that the explicit time step was taken to
be one-fourth the implicit, an economy of approximately a
factor of seven was noted for the explicit calculations in
the elastic-plastic case. In the purely elastic case this
was reduced to a factor of four, favoring the explicit tech-
nique.
4.2 Impulsively loaded elastic-plastic strip
Experimental results for this problem were given by Balmer
and Witmer [6], and results of a finite element calculation
were presented by Belytschko and Marchertos [10] who used
an elastic perfectly plastic model with artificial viscos-
ity. In our calculations we experimented with different
types of hardening, but did not include any viscous effects.
 160

Another difference between the Belytschko-Marchertos calcu-

lation and ours was transverse shear effects which were not
included in theirs, but were included in both the kinematics
and constitution of ours. A comparison of results is pre-
sented in Figure 2. The displacement is quite large, belng

0.8~------------------------------------------~

0.7

,
..--.
c 0,6 /''tf''
, 00'00
,../
+-
/ 0, 0 Ba Imer and Witmer 1964 (exp't.)
C
0) 0.5 I - -- Belytschko and Marchertos 1974
E I
,0
- - elastic-perfectly plastic
Q)
u , ···········isotropic hardening
a present
0- 0.4 P kinematic hardening study
(/)
I
I
------ combined isotropic-
\J kinematic hardening

!>-
Q)
0.3
+-
c
Q)
u
0.2
p = 2.61 x 1(j4Ib-sec2/in4 Vo = 5200 in/sec
O"y =4.14xI0 4 psi ~t = 10-6 sec
0,1 E= 1.04x107 psi ten UI elements
Ep= .667X 106 psi explicit, lumped mass
v = 0.3
O.OL-______ ~ ________ ~ ______ ~ ________ ~

0.0 0.4 0.8 1.2 1.6

time (msec)

Fig.2. Impulsively loaded elastic-plastic strip.

 161

approximately six times the strip thickness. Our peak re-

sponses tend to be in good agreement with the experimental
value, although they tend to occur at somewhat earlier times.
The closest agreement we are able to get is attained with
purely kinematic hardening, as may be seen. Considering the
ambiguities 1n material modeling, we feel the correlation is
quite reasonable.

4.3 Static axisymmetric buckling of elastic cylindrical

shells

Three types of boundary conditions were employed in this

study:
a) bottom fixed--top free;
b) bottom fixed--top fixed with respect to horizontal dis-
placement and rotation; and
c) bottom simply-supported--top fixed with respect to hor-
1zontal displacement.
Load versus d1splacement data and deformed profiles are
shown in Figure 3 up to the buckling loads at which t1me
the calculations were terminated. Comparison is made with
analytical results based upon eigenvalue analysis (see e.g.,
[ 32J ) .

d1 [t dl
50.---,---,---,---,---,---.---,
___ _c;,t~l~lr;_(J!,.a.ond c
r • 2 In
B ,._. C
:e 40
L . l O in
1"1-- 0 .05 In

'2 E • ~)( 107 Ib/ in2

,,:= . 33
<t:>
30 I , I
o classical, A
o

• losl conve rged

S.olulion
10
twenty 2 - node elemenf S

°0~--3~--~6---±9--~12~~15~~18~~21
verticat displacement 110- 2 in)

profile lood (10' Ibl profile lood 110' Ibl

5.0 I 5.0
10.0 2 7.5
15.0 3 10 .0
17.5
20.0
20.5
21.0 15 40.0

A B c
F lq 3 Static aXlsymmetrtc bucklmg of elostlc cylmdnco\ sh ells
162

"
".. AXIAL MOMENT 'is 1t3 AT • S lr

., q
, or
2
x
_0

,
~N
co
...J
_01~~~~ __ ~~ ____ ~ __ ~~ ____ 1
!z 0 f O 0.4 0.8 1,2 1,6
~~ AXIAL DISTANCE 11.3 (IN))( to- I

...J'

"XOr
Cl

LOAOING CONOITIONS AND NOME NCLATURE

AXIAL DISTANCE " IINI x 10-'

HOOP STRESS YS It" AT 4o " 'Ir

0.0 0.4 o',e 1.2 1.6
g:~----~----~----~=-~~----I
Q INSIOE
X N
xq OUTSIDE:
=6
-'"
iii
MEAN tn,
Q.
INSIDE
- - OUTSIIlE
"'- - -- MEAN
-0
til ..
til

!V)9
g
zgho~~--~O~
. 4----~O~.B----~I~.2~--~~---4 AX I AL STRESS YS .s, AT ~ • w
AXIAL OISTANCE .,IIN) x 10- 1

Fig 4. Tilted cylindrical shell sublected to Internal hydrostatic loading.

Stress results In the linear regime .

4.4 Cylindrical shell containing a fluid

An analysis was performed of a tilted cylindrical shell sub-
Jected to internal hydrostatic loading. The problem state-
ment is shown in Figure 4 along with stress results prior to
any significant nonlinear phenomena . Data employed in the
analysis are given as follows:
E 7.35X10 5 lb/in 2 (Young's modulus)
'J 0.3 (Poisson's ratio)
D 16 in L 20 in
R 4 in h 0.01 in
ex. = 30 0 Y 0.0361 lb/in 3 (specific gravity)
The effective specific gravity, Yeff , was defined to be
 163

~ ~ ~

I'--- r---.. r--"

10,/ / V
,/
t; - - / r-
V-
I-:
r-...
~
r-...
-- r-.
f(
L--

--
r---
k
V
f'-..
t<
,/

h?
~
!>

~~
I ~

--~'
~

~~ §~
g= 0,00 9 = 0 .6 0 g =1.20

Q "'~"
~

,..-,-1-- /
t7
t;po<: ~ - "..:
V-
~ I--'
~ t-
l\---I V F;'"
---
I"- ..... ~-- ......
r--- < <
....... r-
>< i><

>< K ~
~~'"
~~
.--
V~ ~
i=::

9 = 1.80
~~
g=1.82
~ g = 1.90
(START FROM 9 = 1.80) (START FROM 9 = 1.80)

Fig 5. Deformed configurations for the tilted cylindrical shell.

164

g · O.oo g ' 0.6 0

9 - 1.20 g : 1.80

9 - 1.82 g - 1.90
(START FROM g - 1.80) (START FROM 9 - 1.80)

Fig.6. Detail of lower regions for the tilted cylindrical shell.

 165

Yeff = yg. The value of g used in obtaining the stresses

shown in Figure 4 was 1.8. Comparison with a membrane theory
solution [31] confirmed the correctness of the numerical re-
sults outside of the boundary layer at the fixed end. Buckl-
ing calculations were performed by incrementing g. Results
are shown in Figures 5 and 6. As may be seen, the initial
bifurcation occurred in attempting to increment g beyond 1.8.
This corresponds to a maximum axial membrane stress of
547.53 psi at x = L, <P = 'IT, which is approximately half the
classical value. Although this result was not anticipated,
it bears a great deal of similarity to results obtained by
Argyris and Dunne [4] in their penetrating study of a com-
pressed cylindrical panel. They too found an initial
bifurcation of approximately half classical. When they
pursued the analysis into the post-buckled regime, a drop
in load of only 3% was noted at which time the load was
able to be increased again to a value in the vicinity of
classical. At this point a second, and more significant,
bifurcation ensued, and this was followed by several more
bifurcations. An important point in considering the re-
sults of Argyris and Dunne was that a preliminary eigenvalue
analysis, emp1oY1ng the same mesh used in the finite deforma-
tion case, yielded a buckling load approximating the classi-
cal value on the high side. This lends credence to the
analysis, but a full understanding of the initial bifurca-
tion in this case, and in ours also, does not yet seem to
exist. The results shown were obtained using Sl elements.
Similar results were also obtained with heterosis elements
(see Part I).

5. Conclusions
In this paper we have presented dynamical aspects of a
general nonlinear finite element shell capability. This
work is a sequel to stat1cal aspects Wh1Ch have been
reported upon in [20,21J. In particular, we have described
the construction of mass matrices and presented details of
the transient algorithms employed. Several sample problems
166

of a dynam1c and static nature have been included, wh1ch

demonstrate the app11cab111ty of the methodology.
Acknowledgments
I would 11ke to thank the follow1ng 1ndividuals and organ1-
zat10ns for the1r 1nterest and support of the research
presented here1n J. Crawford, C1v11 Eng1neer1ng Labora-
tory, Port Hueneme, Ca11forn1a, The Electr1c Power Research
Inst1tute, Palo Alto, Californ1a; and the Nat10nal SC1ence
Foundat1on.

References
1 Argyr1s, J.H.: Recent advances 1n matr1x methods of
structural analysis. Progress in Aeronaut1cal SC1ences
4. Oxford' Pergamon Press, 1964.
2 Argyris, J.H.: Continua and d1scont1nua. Proceed1ngs of
the Conference on Matr1x Methods in Structural Mechanics.
Ohio Wr1ght-Patterson Air Force Base, Oct. 1965.
3 Argyris, J.H., Balmer, H., St. Dolts1n1s, J., Dunne,
P.C., Haase, M.; Kleiber, M.; MaleJannak1s, G.A.;
MleJnek, H.-P., Muller, M., Scharpf, D.W.. F1n1te
element method--the natural approach. Computer Methods
1n App11ed Mechan1cs and Eng1neer1ng 17/18 (1979) 1-106.
4 Argyr1s, J.H.; Dunne, P.C .. Post-buck11ng, f1n1te element
analysis of circular cy11nders under end load. Report
224, Inst1tut fur Stat1k und Dynam1k der Luft und Raum-
fahrtkonstrukt1onen, Un1vers1ty of Stuttgart, Germany,
1977.
5 Argyr1s, J.H., Dunne, P.C., MaleJannak1s, G.A., Scharpf,
D.W.: On large d1splacement--small stra1n analysis of
structures w1th rotat1onal degrees of freedom. Computer
Methods in Applied Mechan1cs and Eng1neer1ng 14 (1978)
401-451, 15 (1978) 99-135.
6 Balmer, H.A., W1tmer, E.A.: Theoret1cal-exper1mental
correlation of large dynam1c and permanent deformat1on
of 1mpuls1vely loaded s1mple structures. Report
FDL-TDR-64-108, Air Force Flight Dynamics Laboratory,
Wr1ght-Patterson A1r Force Base, Oh1o, July 1964.
7 Belytschko, T., Glaum, L.: App11cat10ns of h1gher-order
corotat10nal formulat10ns for non11near fin1te element
analysis. Computers and Structures 10 (1979) 175-182.
8 Belytschko, T.; HS1eh, B.J .. Non11near tranS1ent f1n1te
element analysis w1th convected coord1nates. Interna-
t10nal Journal for Numer1cal Methods 1n Eng1neer1ng 7
(1973) 255-271.
 167

9 Belytschko, T., Schwer, L., Kleln, M.J.: Large dlsplace-

ment translent analysis of space frames. International
Journal for Numerlcal Methods In Englneering 11 (1977)
64-84.

10 Belytschko, T., Marchertas, A.H.: Nonllnear flnlte

element method for plates and ltS appllcatlon to dynamlc
response of reactor fuel subassemblles. Journal of
Pressure Vessel Technology (1974) 251-257.

11 Hinton, E., Pica, A.: Efficlent translent dynamlc plate

bending analysls wlth Mindlin elements. Earthquake
Englneerlng and Structural Dynamlcs, In press.

12 Hughes, T.J.R.: Recent developments In computer methods

for structural analysis. Nuclear Englneering and
Deslgn, in press.

13 Hughes, T.J.R .. Generallzation of selective integratlon

procedures to anlsotropic and nonllnear medla. Inter-
natlonal Journal for Numerlcal Methods in Englneerlng,
in press.

14 Hughes, T.J.R. Impllclt-expllclt techniques for symme-

trlc and nonsymmetrlc systems. Proceedlngs of the
Conference on Nonllnear Problems In Mechanics, Swansea,
U.K., Sept. 1980.

15 Hughes, T.J.R., Cohen, M.. The "heterosls" finite ele-

ment for plate bendlng. Computers and Structures 9
(1978) 445-450.

16 Hughes, T.J.R., Cohen, M.: The "heterosis" family of

plate finlte elements. Proceedlngs of the ASCE Elec-
tronic Computatlons Conference, St. LOU1S, Mlssouri,
Aug. 1979.

17 Hughes, T.J.R., Cohen, M. Haroun, M.: Reduced and

selectlve integratlon techniques In the flnite element
analysls of plates. Nuclear Englneering and Deslgn 46
(1978) 203-222.

18 Hughes, T.J.R., L1U, W.K.: Impllclt-expllclt finite

elements In translent analysis. stablllty theory.
Journal of Applied Mechanlcs 45 (1978) 371-374.

19 Hughes, T.J.R., L1U, W.K.: Impliclt-expliclt finlte

elements In translent analysis: lmplementation and
numerlcal examples. Journal of Applied Mechanics 45
(1978) 375-378.

20 Hughes, T.J.R., L1U, W.K.: Nonlinear flnlte element

analysls of shells: Part I: Three-dimenslonal shells.
Computer Methods In Applled Mechanlcs and Englneering,
In press.
168

21 Hughes, T.J.R.; Liu, W.K.; Nonlinear finite element

analysis of shells; Part II; Two-dimensional shells.
Computer Methods in Applied Mechanics and Engineering,
in review.
22 Hughes, T.J.R.; Pister, K.S., Taylor, R.L.: Imp1icit-
explicit finite elements in nonlinear transient analysis.
Computer Methods in Applied Mechanics and Engineering
17/18 (1979) 159-182.
23 Hughes, T.J.R.; Stephenson, R.A.; Convergence of imp1icit-
explicit algorithms in nonlinear transient analysis.
International Journal of Engineering Science, in press.
24 Hughes, T.J.R.; Taylor, R.L.; Kanoknuku1chai, W.; A
simple and efficient element for plate bending. Inter-
national Journal for Numerical Methods in Engineering
11 (1977) 1529-1543.
25 Key, S.W.; Beisinger, Z.E.: The transient dynamic analy-
sis of thin shells in the finite element method. Pro-
ceedings of the 3rd Conference on Matrix Methods in
Structural Mechanics, Wright-Patterson Air Force Base,
Ohio, 1971.
26 Krieg, R.D. Private communication, 1979.
27 Krieg, R.D.; Key, S.W.: Implementation of a time inde-
pendent plasticity theory into structural computer pro-
grams, pp. 125-137 in Constitutive Equations in Visco-
p1ast~city: Computation and Engineering Aspects (eds.
Stricklin, J.A., Sacza1aski, K.J.) AMD-Vol. 20. New
York: ASME, 1976.
28 Liu, S.C.; Lin, T.H.: Elastic-plastic dynamic analysis
of structures using known elastic solutions. Earthquake
Engineering and Structural Dynamics 7 (1979) 147-159.
29 Ma1kus, D.S.; Hughes, T.J.R.: Mixed finite element meth-
ods--reduced and selective integration techniques: a
unification of concepts. Computer Methods in Applied
Mechanics and Engineering 15 (1978) 63-81.
30 Noor, A.K.; Peters, J.M.; Mixed models and reduced/
selective integration displacement models for nonlinear
analysis of curved beams. International Journal for
Numerical Methods in Engineering, in press.
31 Shih, C.; Babcock, Jr., C.D.: Private communication,
1979.
32 Timoshenko, S.P.; Gere, J.M.: Theory of Elastic Stab1-
l~ty. New York: McGraw-Hill, 1961.
33 Wempner, G.: Finite elements, finite rotations and small
strains of f1ex~b1e shells. International Journal of
Solids and Structures 5 (1969) 117-153.
34 Zienkiewicz, O.C.: The Finite Element Method. London:
McGraw-Hill, 1977.
Elastic Collapse Analysis of Shells Via Global-Local Approach

AK.NOOR, J.M.PETERS
George Washington University Center at
NASA Langley Research Center, Hampton,USA

Summary
A two-stage glopal-Iocal approach is presented for predict-
ing the collapse behavior of shells. The first stage is
that of spatial discretization wherein the shell is dis-
cretized by using finite elements (or finite differences)
which cover the entire region of the shell. In the second
stage the vector of unknown nodal parameters is expressed
as a linear combination of small number of global functions
(or basis vectors). A Rayleigh-Ritz (or Bubnov-Galerkin)
technique is then used to approximate the nonlinear equa-
tions of the discretized shell by a reduced system of
nonlinear algebraic equations. For the case of loading
applied by means of axial end shortening, a scalar function
is introduced which measures the degree of nonlinearity of
the structure. Also, a quantitative measure for the error
of the reduced system of equations is proposed. The effect-
iveness of the proposed technique for predicting the col-
lapse behavior of shells is demonstrated by means of a
numerical example of the elastic collapse of a cylindrical
shell with a rectangular cutout.

Introduction
In spite of the significant advances made in numerical dis-
cretization procedures for nonlinear problems, the collapse
analysis of complex shell structures, having thousands of
degrees of freedom, is still not economically feasible on
present-day computers. Hence, increasing interest has been
shown in the application of a two-stage, global-local ap-
proach to these problems (see, for example, Refs. I to 7).

The first stage in this hybrid approach is that of cp~

~~~zation wherein the shell is discretized by using

finite elements (or finite differences) with local support

for the approximation functions of the individual elements.
In the second stage a cigniM-c.ant Jte.duc.:ti..on in :the. numbe.Jt 06 de.gJte.1U
170

06 6~eedom is achieved by expressing the vector of unknown

nodal parameters as a linear combination of a small number
of global functions (or basis vectors). A Rayleigh-Ritz
technique is then used to approximate the finite element
equations of the discretized shell by a reduced system of
nonlinear algebraic equations.

As would be expected, the effectiveness of the approach de-

pends, to a great extent, on the proper choice of the global
functions (or reduced basis vectors). Various choices for
global functions were proposed in the literature. These in-
clude linear bifurcation buckling modes (Refs. 3 and 4);
linear solution and corrections to it (Ref. 5); and nonlinear
solution and its various order path derivatives (Refs. 6
to 8). Numerical examples presented in Refs. 6 and 7 demon-
strated that the latter choice of a nonlinear solution, and
its various order path derivatives (which are commonly used
in the static perturbation technique) as global functions
results in an effective global-local approach. However, in
all the cited references, the structures were subjected to
externally applied loads. The case of loading applied by
means of axial end shortenlng, as would occur in a labora-
tory compression test was not considered.

The present study focuses on the application of the global-

local approach to the elastic collapse analysis of shells
subjected to prescribed edge displacements (or end shorten-
ing). Specifically, the objectives of this paper are: a)
to outline the modifications in the global-local approach of
Refs. 6 and 7 which are required in the case of prescribed
(nonzero) edge displacements, b) to demonstrate the effect-
iveness of the proposed technique by means of a numerical
example, and c) to give some insight into why and when the
reduced basis technique works.

The analytical formulation is based on a form of the geome-

trically nonlinear shallow shell theory wlth the effects of
transverse shear deformation included. A displacement form-
ulation is used with the fundamental unknowns consisting of
the three displacement components and the two rotations at
 171

each point of the middle surface of the shell. The numeri-

cal solutions were obtained by using 16-node quadrilateral
elements with blcubic Lagranglan interpolation functions
for each of the displacement and rotation components. The
computational algorithm presented herein is a modification
and an extenslon of that presented in Refs. 6 and 7.

Mathematical Formulation

Stage. 1 - Spatial V.u.,CJte.:Uzation:

A total Lagrangian formulation is used and the shell is dis-
cretized by using a displacement finite element model. The
governing finite element equations, for the case of a load-
ing applied by means of axial end shortening, can be cast in
the following compact form:

[K] {X} + {G(X)} - {Q} = 0 (1)

where [K] is the nxn linear global stiffness matrix; n is

the total number of displacement degrees of freedom; {X} is
the vector of nodal displacements; {G(X)} is the vector of
nonlinear terms which is cubic in {X} (see Ref. 7); and {Q}
is the vector of constraint and applied forces.

It is conv@nient to partition the vector of nodal displace-

ments as follows:

(2)

where subscripts f and c refer to free and constrained (pre-

scribed nonzero) dlsplacements, respectively. The constrain-
ed zero displacements and their associated equations are not
included in Eqs. 1. For simplicity, the constrained nonzero
displacements are assumed to be proportional to a single
parameter A, i.e.

{X } = A{Z} (3)
c

As the edge displacements are incremented, the value of the

parameter A changes, but the components {Z} remain constant.
172

Equations 1 can be conveniently partitioned into two sets of

matrix equations as follows:

fG l (Xf,X c ) }
+ o (4)
~G2(Xf'XC)
In the absence of externally applied loading (case of pre-
scribed displacements only), {Qf}=O and {Qc} equals the
vector of constraint forces associated with the prescribed
nonzero displacements {Xc}. The first set of Eqs. 4 can be
used to determine {X f } and the second set is then used to
evaluate the constraint forces {Qc}.

Stage. II - Re.duc.:Uon 06 the. NumbeJl 06 Ve.g.lte.e6 06 F.lte.e.dom:

In the second stage of the proposed technique, the vector of
nodal displacements {X} of the discretized structure is
approximated over a range of values of A, by a linear com-
bination of {X} corresponding to a particular value of A
and a number of its path derivatives (derivatives of {X}
with respect to the path parameter A) at the same value
of A. This can be expressed as follows:

(5)

where

1':1] (6)

and {~} is a vector of reduced unknowns (with r components

only) which are functions of the path parameter A. Note
that r is considerably smaller than the total number of
degrees of freedom n. The columns of the matrix [r] are
the global approximation functions.

The Rayleigh-Ritz technique is now applied to replace the n

equations obtained in the first stage, Eqs. 4, by the fol-
lowing reduced system of r nonlinear equations:
(7)

subject to the condition ~l = A.

 173

The coefficients of the reduced equations, Eqs. 7, are

given by:
[i] = [r]T[K] [r] (8 )

{G(lji)} = [r]T{G(lji)} (9 )

and
(10)

where the superscript T denotes transposition and {G(lji)} is

obtained from {G(X)} by replacing {X} by its expression in
terms of {lji}, Eqs. 5. Note the {Q} has only one nonzero
component, namely, Ql.

For a given value of A, the last r-l equations of Eqs. 7 can

be solved for lji2' lji3' ... ljir· The first equation is then
used to evaluate Ql.

The global functions in Eqs. 6 are obtained by successive

differentiation of the finite element equations of the dis-
cretized shell, Eqs. 4, and solving the resulting system of
linear algebraic equations. Also, as shown in Refs. 6
and 7, the evaluation of all global approximation functions
requires only one matrix factorization.

The criterion for selecting the number of basis vectors pro-

posed in Ref. 6 was adopted in the present study. This
criterion is based on monitoring the condition number of the
Gram matrix of the global functions, and terminating the
generation of these vectors when the condition number ex-
ceeds a prescribed value. Also, upper and lower limits for
the number of global functions were chosen to be seven and
two, respectively.

The range of A for which Eqs. 7 provides an acceptable ap-

proximation for the original discrete system, Eqs. 4, de-
pends on the degree of nonlinearity of the problem. This
range can be identified by monitoring the accuracy of the
solutions of the reduced system, Eqs. 7. When the error
exceeds a prescribed tolerance, the vector {x} generated by
174

the reduced system of equations is used as a predictor and

the Newton-Raphson iterative technique is used in conjunc-
tion with the original equations, Eqs. 4, to obtain a cor-
rected (improved) solution. Then a new (updated) set of
global functions is generated.

Computational Procedure
The computational procedure outlined in Ref. 7 for the ef-
ficient evaluation of the global functions, generation of
the reduced system of equations, automatic selection of the
displacement increments, sensing and controlling the errors
of the reduced equations is adopted in the present study.
However, the scalar S (current stiffness parameter) used to
characterize the nonlinear response (see Ref. 9) and the
error norm used in Ref. 7 are not applicable to the case of
prescribed edge displacements, and therefore, are modified
as described subsequently.

CUlV!.ent S.ti66nu-6 PM.a.mUe/L:

In the case of loading applied by means of prescribed edge
displacement, the following scalar S is introduced for
characterizing the nonlinear response:

S (i) = for the full system (ll)

for the reduced system (12)

where

I I
aaQ\c = [[Kef] +
(G2iJHaxfj
ax f . aA + [IKeel + [:::il]IZI (131
J J

r dljl j
aQl ael ]
---ax- E [Klj + aljl . d\
(14)

j=l J

and
 175

1
(15)

The subscript (i) in Eqs. 11 and 12 refers to point i of the

solution path and the subscript 0 refers to the point at
1.=0.

The current stiffness parameter defined in Eqs. 11 is similar

to that introduced in Ref. 9 for the case of externally ap-
plied loads. It has the major advantage of being easily
computed from the reduced system of equations. The selec-
tion of displacement increments and the frequency of error
sensing are then related to changes in this parameter (see
Ref. 6).

The parameter S(i) provides a global measure for the stiff-

ness of the structure at point i of the solution path. It
has an initial value of 1.0, increases when the structure
stiffens and decreases when the structure softens. For
stable equilibrium paths, S is positlve, for unstable paths,
S is negative, and at limit or collapse points, S is zero.

E!UWft No!tm:
To check the accuracy of the solution obtained by the re-
duced system of equations, Eqs. 7, the following error norm
is used:

(16)

where

(17)

If the error norm e is less than a prescribed tolerance, the

solution is continued; otherwise a corrected (or improved)
estimate of {X f } is obtained using Newton-Raphson technique
in conjunction with Eqs. 4. Then a new S8t of global
functions is generated using the technique outlined in
Ref. 7.
176

Numerical Studies
To test and evaluate the effectiveness of the proposed
global-local approach, a number of nonlinear shell problems
were solved by this approach. Comparisons were made with
solutions based on the full system of equations of the finite
element model. Typical results are presented herein for the
elastic collapse analysis of the cylindrical shell with rec-
tangular cutout shown in Fig. 1. The problem is one of three
problems used in Ref. 10 to assess the capability of various
programs to analyze shell structures. The load is applied to
the cylinder by means of a uniform axial end shortening which
is increased incrementally until the cylinder collapses .

E :: 6 .895 " 10 10 Newtons/m 2

.ip-"[9,-.,
ul --C__

0'0762m~
'<I:: 0 .3

R :: 0.1541 m
I 0 .1143 m
h = 3.556 " 10-4 m
a

Boundary conditions at x 1 :: 0 b

u 1 =). 0.2421 m

u2 = w ='1 ='2 =0 I II I

a
0r"
m

b
1--- -- - 0.1815 m- -- -----i

Finite element model

Figure l.-Cylindrical shell with cutout and finite element

model used in the present study.

Due to symmetry, only one octant of the shell was modeled

using the grid of shear-flexible elements shown in Fig. 1_
Bicubic Lagrangian interpolation functions were used to ap-
proximate each of the displacement and rotation components
(a total of 2996 nonzero displacement degrees of freedom).
Finite difference solutions to this problem using the STAGS
 177

(STructural ~nalysis of ~eneral ~hells) computer code are

presented in Refs. 10 and 11. Finite element solutions
for the same problem are presented in Ref. 12

2.5

2.0

1.5
Current stiffness
parameter S
1.01-----

0.5

o
Total axial load (Newtons)

Figure 2.-Variation of current stiffness parameter with

loading for the cylindrical shell with cutout
shown in Fig. l.
Figure 2 shows the variation of the current stiffness para-
meter S with loading. Figure 3 gives an indication of the
accuracy of the normal displacement w, at the centers of two
sides of the rectangular cutout, obtained by the global-
local approach. Figure 4 gives an indication of the accur-
acy of the total strain energy and Figure 5 shows the error
norms of the global-local approach with seven basis vectors
at various load levels. Figure 6 shmvs contour plots of
the normal displacements w at four different load levels,
each normalized by dividing by wmax for that load level.

In the present study the path parameter A was chosen to be

equal to the axial edge displacements, and therefore, the
components of the vector {Z} are equal to unity. The global
functions were first computed for the unloaded shell (A=O,
{Xf}=O). An error tolerance e ~ 0.05 was prescribed. The
seven global functions were used to advance the solution to
178

A=3. 048 x 10- 5 m. at which value the error norm was checked
and was found to exceed the prescribed tolerance (see
Fig. 5). New (updated) set of seven global functions were
generated and used to advance the solution until A=9.398x
10 -5 m. The predicted collapse load of the cylinder was
13 . 656 x 10 3 Newtons corresponding to A=9. 70 x 10 -5 m. The
collapse loads predicted in Refs. 11 and 12 are 10.008 x 10 3
and 12.792 x 10 3 Newtons, respectively.

linear solution
15

12

9
Tota l axial
load (Newtons)
6

3 - - - Fu ll system
+ Reduced bas is (7vectors )
" ' -_ _ _ ...l...-_ _ _--'-_ _ _----L_ _ _ -----I X 10-4
o 7 14 21 28
Normal displacement w (meters)

Figure 3.-Accuracy of normal displacements obtained by re-

duced basis technique at various load levels.
Cy lindrical shell with cutout shown in Fig. 1.
Roman numerals indicate points of generating
basis vectors.
The high accuracy of the normal displacements and strain
energies obtained by the reduced system of equations is
demonstrated in Figs. 3 and 4 . At A=9. 398 x 10- 5 m. the
errors in the maximum normal displacement wb and the total
strain energy U obtained by uSlng seven basls vectors were
0.53 % and 0.023 %, respectively.
 179

15

Bending
12 energy

9 Total strain energy

Total axial
load (Newtons)
6

- - - Full system
3
+ Reduced basIs (7 vectors)

o
strain energy (Newton' meters)

Figure 4.-Accuracy of strain energies obtained by reduced

basis technique at various load levels. Cylindri-
cal shell with cutout shown in Fig. 1. Roman
numerals indicate points of generating basis
vectors.

Hlgher accuracy of the reduced solutions can be achieved by

backtracklng the equilibrium path every time a new (updated)
set of global functions is generated. This amounts to
effectively reducing the error norm well below the pre-
scribed tolerance. When this technlque was used in the
present problem, the maximum value of the error norm using
seven functions reduced to less than 0.003 (see Fig. 5).
The computational expense involved ln the backtracking pro-
cess was insignificant.

In summary, the generation of the whole solution path up to

collapse of the cylinder involved a) generation of an ini-
tial set of global functions at A=O and b) updating the
global functions three times.

The. uoe. 06 ;(;he. g£.obal-£.ocal appfWach J..n ;(;fUJ., PJtob£.e.m Jte!.JuUe.d J..n Jte.-
du~ng ~he numbek on degJtee!.J on 6Jte.e.dom by a 6ac;(;oJt 06 ove.!t 400 (from
180

2996 degrees of freedom for the original finite element

model to seven degrees of freedom for the reduced system).

- - - Forward path
......... Backtracking

.05

.04

. 03
Error
norm
. 02

, 01
\

o
Total axial load (N ewtons )

Figure 5.-Error norms of reduced basis technique (using 7

vectors) at various load levels. Cylindrical
shell with cutout shown in Fig. 1.

Comments on the Potential of the Propose~proach

The following comments concerning why and when the global-

local approach works seem to be in order:
1. The success of the proposed global-local approach can
be mainly attributed to the separation of spatial dis-
tribution of the response quantities at any load level
from the variation of these quantities with loading.
The form of the spatial distribution of the fundamental
unknowns is given by the discrete finite element model
in the first stage. The use of the path derivatives
as global functions in the second stage permits the
known information about the nonlinear response in the
neighborhood of a point to be brought into the analysis.
 181

" 10 3
15.00

11.25

Total b
exlal N1 = 3.996 " 10 3 Newtons
7 .50
load
(Newtons) 2

3 .75

'--_--L_ _....l..._ _l..-_-.J " 1004

° 7 14 21 28
Normal d isplacement wb (meters)

N = 6 .224 " 10 3 Newtons

2

0.0762m~ au o.1143m

N3 = 9.138 " 10 3 Newtons

b
N4 =11.949 " 10 3 Newtons

Figure 6.-Contour plots for the normal d~splacements w at

various load levels (each normalized by dividing
by wmax for that load level). Cylindrical shell
with cutout (see Fig. 1).

In this manner substantially fewer degrees of freedom

will be required to achieve a desired overall accuracy
in comparison with that based on the finite element
technique.
2. Numerical experiments have shown that the spatial dis-
tribution of stresses and displacements varies slowly
182

with the loading. This can be seen from the normalized

contour plots of Fig. 6, wherein large changes in load-
ing are associated with small changes in the contour
plots. On the other hand, large changes in the contour
plots require updating the global functions.
3. If the proposed global-local approach is contrasted with
the static perturbation technique the following can be
noted. In both methods the vector of unknown nodal
parameters is approximated by a linear combination of
a small number of path derivatives, Eqs. 5 and 6. How-
ever, the coefficients of the linear combination {~} in
the static perturbation technique are fixed and are
(fiA)2 (fiA)r-2
equal to: A+fiA, 1, fiA, ~, •.• (r-2)!. By con-
trast, the coefficients {~} in the proposed global-
local approach are left as free parameters, and are
determined by applying the Rayleigh-Ritz technique.
Numerical experiments indicate that the use of free
parameters leads to accurate solutions not only within
the radius of convergence of the Taylor series but also
well beyond it.
4. If both external and internal forces are conservatlve,
the efficiency of the global-local approach can be in-
creased by increasing the error tolerance and mainta1n-
ing the accuracy of the reduced system of equations
through backtracking the solution path every time a
new (updated) set of basis vectors is generated.
5. In cases where rapid changes in the spatlal dlstribution
occur in a small region of the shell, the efficiency of
the global-local approach can be enhanced by using this
approach in conjunction with a substructuring technique
(or a partitioning scheme). The global functions in
the region of rapid change are updated more frequently
than the global functions in other regions.

Concluding Remarks
A two-stage global-local approach is used to predict the
collapse behavior of shells. The computational algorithm
presented previously by the authors 1S modified to handle
 183

the case of loading applied by means of axial end shorten-

ing, as would occur in a laboratory compression test. A
scalar function lS introduced for measuring the degree of
nonlinearity of the structure. Also, a quantitative measure
for the error of the reduced system of equations is proposed.

The potential of the proposed technique for predicting the

collapse behavior of shells is demonstrated by means of a
numerlcal example of elastic collapse of an axially com-
pressed cylindrical shell with a cutout, wherein reduction
in the number of degrees of freedom by a factor of over 400
was achieved. The example also provides an insight into
why and when the global-local approach works.

References
1 Besseling, J. F.: Nonlinear analysis of structures by
the finite element method as a supplement to a linear
analysis. Computer Methods in Applied Mechanics and
Engineering 3 (1976) 173-194.

2 Besseling, J. F.: Post-buckling and nonlinear analysis

by the finite element method as a supplement to a linear
analysis. ZAMM 55 (1975) T3-16.

3 Nagy, D. A.: Modal representation of geometrically non-

linear behavior by the finite element method. Computers
and Structures 10 (1979) 683-688.

4 Nagy, D. A.; Konig, M.: Geometrically nonlinear finite

element behavior using buckling mode superposition.
Computer Methods in Applied Mechanics and Engineering 19
(1979) 447-484.

5 Almroth, B. 0.; Stern, P.; Brogan, F. A.: Automatic

choice of global shape functions in structural analysis.
AIAA Journal 16 (1978) 525-528.

6 Noor, A. K.; Peters, J. M.: Reduced basis technique for

nonlinear analysis of structures. AIAA Journal 18 (1980)
455-462.

7 Noor, A. K.; Andersen, C. M.; Peters, J. M.: Global-

local approach for nonlinear shell analysis. Proc.
Seventh ASCE Conference on Electronic Computation,
St. Louis, Mo., Aug. 6-8 (1979) 634-657.

8 Noor, A. K.; Peters, J. M.: Nonlinear analysis via

global-local mixed finite element approach. To appear
in the International Journal for Numerical Methods in
Englneerlng.
184

9 Bergan, P. G.; Horrigmoe, G.; Krgkeland, B.; S¢reide,

T. H.: Solution techniques for nonlinear finite element
problems. International Journal for Numerical Methods
in Engineering 12 (1978) 1677-1696.

10 Hartung, R. F.; Ball, R. E.: A comparison of several

computer solutions to three structural shell analysis
problems. AFFDL-TR-73-15 (1973).

11 Almroth, B.D.; Holmes, A. M. C.: Buckling of shells

with cutouts, experiments and analysis. International
Journal of Solids and Structures 8 (1972) 1057-1071.

12 Argyris, J. H.; Hilpert, 0.; Hindenlang, U.;

Malejannakis, G. A.; Schelkle, E.: Flachentragwerke
im Konstruktiven Ingenieurbau. ISD-Bericht Nr. 263,
Universitat Stuttgart (1979).
Large Spatial Deformations of Rods Using Generalized
Variational Principles

W. WUNDERLICH, H. OBRECHT
Ruhr-Umversitat Bochum, Germany

Sununary

General equations and corresponding incremental formulations

for the analysis of rods undergoing large spatial deformations
are presented. The derivation is based on a generalized three-
d1mensional variational principle with the increments of the
Lagrang1an stresses, the deformed posit1on vector and the ro-
tation as independent variables. "Engineering" strains and con-
jugate "Jaumann"-stresses are related by a semi linear materi-
al law. The pr1nC1ple allows assumptions for the stresses and
for the kinematical var1ables simultaneously. This 1S used in
the reduction to the one-dimensional description for general
rods. Warping effects and arbitrary cross sect10ns are conS1-
dered in the fully spatial and geometrically nonlinear descr1p-
tion. The resulting generalized princ1ple in terms of seven
stress resultants and conjugate displacements and rotations 1S
given in its general and incremental forms. It could be used
for a mixed or hybr1d finite element approach. In the paper
"exact" stiffness matrices are obtained via the integration of
the local equat10ns of the principle in the form of a system
of first-order ord1nary differential equations. Numerical re-
sults of sample problems are given.

1. Introduct1on

To evaluate the actual load carrying capacity of slender rods

and of frame structures, nonlinear calculat10ns which account
for large deflect10ns and moderate or large rotat1ons are of-
ten necessary. This is part1cularly true 1f such structures
may fail by buckling into complex modes (e.g. by interactive
sidesway buckling) or if, as a result of initial imperfec-
tions the structure, or invidual members, show a pronounced
nonlinear behavior. Then a relevant decrease in overall stiff-
ness may occur before a "critical" load level has been reach-
ed. In this respect the effect of coupling between bending
and torsion as well as the influence of arbitrary cross sec-
tions (thin walled, open, and closed) have to be taken into
186

occount to real1stically model the actual behavior. Warping con-

straints and the cross-sectional shape sometimes play an impor-
tant role. It is obvious that the proper formulation of this com-
plex problem and its solution requires special attention.

The fin1te element approach to this subject has been treated in

a number of papers (see e.g. the lists of references in (1,2,3] ).
It 1S mostly based on the incremental form of the principle of
v1rtual work and a displacement formulation. Curved beams are
somet1mes approximated by straight elements, and the nonlinear
coupling of bending and torsion as well as the influence of ar-
bitrary cross-sections is rarely mentioned. In the approach of
[2,4] to the treatment of large deflections and moderate rota-
tions of spat1al rods a generalized variational principle with
Lagrangian stresses, displacements and rotations was used as
the starting point, and the effect of warping was included. The
derivat10n in this paper is similar in spirit, but embodies se-
veral 1mportant generalizations. It employs a more consistent
formulat1on, and is also applicable to large rotation problems.
The use of a generalized variational principle, with rotations
as unknowns, as the basis for the reduction of the three-dimen-
sional equations to the one-dimensional theory for rods facili-
tates the formulation significantly. Moreover, since in slender
structures - like rods, plates, and shells - moderate and large
rotations play an important role, it appears quite natural to
1nclude them in the description explicitly, rather than impli-
citly through the displacement gradient. This is accomplished
here through the use of the polar decomposition theorem and a
generalized variational principle of Hellinger-Reissner type
which, for the continuum, was first glven by Fraeijs de Veubeke
(5] and has been recast in incremental form by Atluri and
Murakawa (6]. In this paper the principle is modified further.
The three-dimensional principle, written in terms of so-called
"Jaumann" stresses and conjugate "engineering" strains, is redu-
ced to the corresponding principle for rods. It is expressed in
terms of components of the pos1t1on vector of the deformed con-
figurat1on, the angles of rotation about principal axes, and
the corresponding (Lagrangian) components of the "Jaumann"-stress-
resultants as fundamental unknowns. The general princ1ple is
 187

given as well as its total (T.L.) and updated Lagranglan (U.L.)

forms. The independent varlables chosen seem to be the most ge-
neral and consistent ones, and are also most approprlate for a
formulation which applles to arbltrarlly loaded general rods.

Due to its generalized character, the prlnclple allows assump-

tions on the dlstribution of stresses over the cross-sectlons
and, simultaneously, kinematic assumptlons. In the latter the
deformed posltlon of pOlnts outslde the aX1S are related to
the correspondlng position and rotation of the reference line
in such a way that arbitrary cross-sectlons can be taken into
account. For the warplng effect a kinematic unknown and the
correspondlng blmoment are additional parameters ln the formu-
lation. A somewhat lengthy, but stralghtforward, procedure of
reduction to the one-dimensional beam problem finally leads to
a generallzed variational princlple with seven kinematic vari-
ables and the same number of stress resultants as unknowns. This
principle could be taken as the basls of a usual mixed-type fi-
nlte element dlscretlzatlon. For one-dimensional problems, howe-
ver, a proven algorlthm which is based on the integration of the
correspondlng Euler-equatlons, written in the form of a system
of first-order ordinary differential equations [11] , is more ad-
vantageous, because it results in highly accurate element stlff-
ness matrlces. In this manner, the rapidly decaying functions
WhlCh account for warplng effects do not need speclal treatment.
The inltial dlsplacement and lnitlal stress matrices of the non-
linear problem are treated in the same way as the llnear parts of
the problem. Numerlcal calculations of sample problems have been
performed to demonstrate the effect of combined bending and torsion

2. BaS1C general relationshlps (contlnuum)

2.1 Reference configuratlons

The configurations of a body undergoing deformations may be re-

ferred to the undeformed initlal state with coordinates x A and
base vectors ~, to a current, deformed state with coordinates
xi and base vectors ~l' or to a global cartesian coordinate
188

system with unit vectors ~ (see Fig. 1). For the description
of curved members it is useful to use convected, curvilinear
coordinates, such that in all configurations a material par-
ticle is identi f ie d by the same set 0 f values e 1 , e 2 , e 3 . The
notation employed here follows that of [7,8].

current,
deformed state

F1gure 1

Denote the pos1tion vector of a particle and its neighborhood

1n the undeformed body by R = xA~A' and dR = dxA~A while its
location and ne1ghborhood in the current, deformed state may
be given by £ = xi~i and dr = dxi~i. The respective base vec-
tors and metrics are ~A' ~i and GAB = ~A ~B' gij = ~i ~j. The
use of capital and lower case indices 1S an aid which helps to
differentiate whether a tensor is associated w1th in1tial or
deformed base vectors, or both, in which case 1t 1S a two point
tensor. The orthogonal triads of the principal strain direc-
tions are denoted bv N = NaM -M
- -a G in the in1tial and by -a
n = n1ga-i
in the deformed state,respectively.

The deformation of the body causes the material lines of prin-

cipal strain to stretch and to rotate. The amount and orienta-
tion of the rigid rotation is given by the orthonormal two-
point rotation tensor g whose components sat1sfy
. N .
I a...~J a .. = 6~ (1)
N 1 1

Note that for vanishing rotation,g becomes unity. The initial

and current principal strain triads are then related by
 189

k
na a.k M NM
a and the following identit1es hold

k k oM
a· M n Na Na NK oa Na NM (2)
a M K J3 J3 K a K

K
A one point rotation tensor w1th components a. M may be defined
K . K K i a
by introduc1ng the shifter g.1 = GK .9:i to g1ve a· M = (g1 na)N W

2.2 Deformation quantities

As a result of the deformation the ne1ghborhood of a particle

is transformed according to

(3)

i
whera F.A denotes the components of the deformation gradient F.
It is a two point tensor and may be decomposed as follows

F 9: U (4)

In (4) 9: is the rotation tensor defined above and U = I + h is

the symmetric right stretch tensor. In the terminology introdu-
ced in [5] ~ is called the eng1neering strain tensor. As above,
a related one-point tensor may be defined as

( 4a)

The base vectors of the deformed and undeformed configurations

are related in a similar fashion

(5)

This shows that the rotation tensor can also be written as a

i i
shifter.9: QA = a. A and thus describes the rigid rotation of the
A

....
deformed base .9:i to the stretched base QA. Recall, that in terms
of the usual displacement grad1ent the base vectors .9:i are given by

(6)

where convected coord1nates have been used and ( )IA denotes

covariant differentation with respect to the undeformed metric.
190

2.3 Incremental relat10nships

To describe the change of quant1ties in going from a current

state to a closely ne1ghboring state, the notat1on given
below will be useu (see[7] and F1gure 2).

f.~__- -

F1gure 2

All quant1t1es which character1ze the current state are presu-

o
med known and will be denoted by the supercr1pt ( ), all
quanti ties of the neighbor1ng state by (-), whereas incremen-
tal quantities will not be labeled (e.g. ~ in Flgure 2 denotes
the current, ~ the incremental d1splacement vector). The perti-
nent relationships between the principal stra1n triad n
-a
, the
rotat1on tensor ~ the position vector £, and the deformation
gradient ~E of the adjacent state, and the quantities charac-
teriz1ng the current state as well as their 1ncrements are sum-
marized below in the1r T.L. and U.L. form:

T.L. : V.L. :

0
n
-a (9, + -a)N
-a
n
-a (I aiR
- + --a
0 o
r R + u + ~ r + ~

0 A i (7)
dr (~A + ~'A + ~'A)de (51 1 + -u,.)
1
de

0 0
vi (~ + ~) . (!. + g + g) vi = (;~. + ~) . (.! + £l.l

h~ + a~ .
1 1
 191

The components of the tensor ~ which describes the general rota-

t10n of an orthogonal triad, may be expressed in terms of trigo-
nometr1c funct10ns of sU1tably defined rotat1on angles. Euler
angles may be used, as was done 1n [1] but a formulat1on 1n terms
of angles about the axes of a reference triad are more appropri-
ate for the description of beam deformations. The full expression
of ~ is given 1n the Appendix (equ. Al) Wh1le ~ sat1sf1es the
ortho-normal1ty relat10n (1), the 1ncrements ~ do not. Rather,
from
_T
a a (~ + ~)T (~ + ~) I

it follows that for small increments ~ a skew symmetr1c rota-

t10n 1ncrement tensor ~ may be defined as follows:
T 0 oT T
w a a -a a -w (8)

For the case of a rotat1on through an increment <p about one

aX1S only, these relations take the special form

- 0
~+~
[
cos (<p + <p)
0
Sln (<p + <p)
o ] [
1 <p
'] [
cos<p
0
S1n~ ]

(~+ <p) (~+ <p)

~
= -Sln cos = -<p 1 -slng cos<p

0 0
or a ~ + ~ (~ + I)T ~ (9)

o
The corresponding general express10ns for ~ and ~ 1n terms of ~

and 1ncrements ~ are also given in the Append1x (equ. A2a).

Since ~ 1S skew symmetric, it may be expressed in terms of a
M
rotat1on increment vector with components w , such that

1 wKL
wM = -"2 EMKL

where EMKL denotes the permutat10n tensor. In turn, the rotation

rate of a corotat1ng coord1nate system may be expressed 1n terms
of the 1ncrements <P1 about the axes of the reference triad by:
192

-w3 2 1
0 w CI.' CP1
0 0
~ W3 0 -w1 ~ w
2 b CP2 b ~ (10)

-w 2 w
1 0 w3 CP3

o 0
Note that b becomes unity in the reference state. The matrix b
is given in the Appendix (equ. A3). A summary of the above rela-
tionships follows:
T.L.: U.L. :

wij u .. =-u
~J ji
u~J e: mji cPm

wi cP~

2.4 Strains and stresses

o
The Lagrangian strain tensor e: AB of the current configuration
is defined by
2 2 0 __ A B
ds - dS = 2e: AB d~- dX

where ds and dS are the deformed and undeformed lengths of a ma-

terial line element. It can be expressed in terms of the deforma-
tion gradient F~A or the displacement gradient u A I B by

( 11)

Substitution of the polar decomposition relationship (4) into the

first expression of (11) and use of the orthogonality property
. 0
(1) of a~M leads to the knownoexpression of e: AB in terms of the
"engineering strain" tensor hAB
o 0 1 0c 0
e: AB = hAB + 2 hA hCB (12)
o
Alternatively hAB may be expressed in terms of the displacement
o.A
grad ~ent and Uc :

O.A(",C OC,) _ ",A ( 12a)

Uc uB + u B uB
 193

The related incremental quantities are glven below in their T.L.

and U.L. forms. Note that the underlined terms are usually omit-
ted in a linear1zed 1ncremental formulation. They may, however,
be of significance in numerical procedures designed to satisfy
the full set of nonlinear equations iteratively.

T.L. : U.L. :

C 0C 1 C
(oA + hA + 2" hA)h CB Eik =
h ik +..!.
2 h~1 h jk

hA
B
= .A(oC +oCI + C 1 )+O.A C 1
aC BUB u B a C u B h~ a k.i ( 0.k + u k.i ll
) +u
J J __J_ J

As appropriate stress measures the so-called "Jaumann" stresses

rAB [5,8] are introduced. They are related to the "Lagrangian"
stresses tAB, to the 1. Piola-Kirchhoff stresses rrAK , -and
to the 2. P1ola-K1rchhoff stresses sAB by

0AB 0AC o.B

r = t aC (13)

o
AB
The symmetric part of r

~ AB
r (14 )

has been termed "engineering stress". Only this part contributes

to the work express10ns and the complementary energy density of
the funct10nals which are used for the derivation. It arises as
o
the conJugate quant1ty of the "eng1neering strain" tensor hAB [5].

From (1) it follows that the first equation 1n (13)may be inverted

un1quely to give
(15 )

The respective 1ncremental express10ns are given below where, for

completeness, the relationship between rAB and sAB has been includ-
ed.
T.L. : U.L. :
AB .B O.B ik mkr
r t°AC a C + t AC a C r t ik + °i
L E (jJm
r
oAC B oB ik 01 mkr
tAB r a· C + rAC a· C t ik r
r
-E L(jJm

rAB AB AC °B °AC hB ik 1k air hk

s -s hc + s r s + L
C r
194

oir o~
Here L , Lr denote components of Cauchy stress.

2.5 Constitutive relations

In nonlinear problems the cho~ce of constitutive relations is

of special significance. Generally speaking, such relations are
only meaningful if they relate objective measures of stress and
their conjugate strain measures. The following relat~ons, usu-
ally called hyper- and hypo-elastic, are appropriate if strains
rema~n small:

oB 1+v {OB v OK
EA ""E sA - 1+v sK oB}
A
(16 )

v K
B
EA = 1+v
E
B
{sA - 1+v sK oB}
A
( 17)

A linear relation between engineer~ng stresses and strains is

often called "semilinear":

( 18)

£AB
Since the engineering measures r °
a n d hAB are objective a
Legendre transformation

Q*(~)
Q.
= rB
£A hOB _
A
°
Q(g) (19 )

o 0
exists [5J , where Q (g) and Q*(~) are potent~al funct~ons corre-
spond~ng to the strain energy and complementary energy densit~es,

respectively. For ~sotropic materials ~t may be shown that

Q
Q*(~)
°
= Q*(~).
The corresponding incremental relationship employed in this paper
relates the increments in an analogous manner:

(20)

This strain-stress-relation implies that a transformation ana-

logous to (19) exists, and that the incremental complementary
energy density has the simple form

w* (~) (21)
 195

It is noted that the use of the slmple (nonobJectlve) lncrements

i in (20,21) is limited to problems with small strains ~. This
is assumed In the present formulation for rods. For large strains
approprlate corotatlonal increments of stress should be used In
the respective constltutlve laws.

Recall that the correspondlng Legendre transformatlon relatlng

the lncrements of the 2 nd Piola Kirchhoff stress and Green's
strain tensor reads [6]

(22)

C
where nAB = u
IA u CI Band W (f) lS agaln the lncremental strain
energy denslty. Substltutlon of (17) glves

(23)

It lS seen that in thls case W* may not be wrltten In terms of

sAB alone. Similar problems arlse In the attempt to formulate
Wand W* In terms of Lagranglan stresses and dlsplacement gra-
AB
dlents. Although t and u BIC are not obJective measures of
stress and straln, the sum of W and W* can slmply be wrltten as

W + W* -_ t·AB u BI A = sA B + 2s A n B
B EA (24)
B A

2.6 Generallzed varlatlonal prlnciple

A varlatlonal princlple of Hellinger-Relssner type uSlng the

rotatlon-tensor of the polar decompOSltlon expllcltly has been
derived by FraelJs de Veubeke [5] . In [8] Wempner used It
and a related principle In the context of buckling problems.
An lncremental verSlon was presented by Atluri and Murakawa
on several occaSlons, e.g. In [6] . The generalized prlnclple
used in this paper as the basls for the reduction to the beam
9 0
problem employs the stresses £ and the conjugate stralns h
as unknowns, whereby the latter is expressed In terms of the
rotatlon and the deformatlon gradient (see 12a):
196

oA °B 0A 0
6J CB 6{ J [-Q* C~) + r B hA - Po f uAJ dV
Vo

- J t~A u0A dS - J ~Adl.A

Q
u A) dS }
S Su
P (25)
OAB OC 0 OA °A
6{I[-Q*(E)+ r a.· B YC!A - r A - Po f SA] dV
Vo

+ boundary terms } 0

After the definition of stress-resultants and reduction,the

final principle for rods w~ll be written ~n terms of the defor-
mation gradient, the rotations, and nominal resultants, wh~ch

are components of the cross-sect~onal forces and moments ~n

the direction of the reference configuration. They are analo-

gous to tAB in three-d~mensional problems.

The incremental versions of (25) follow after sUbstitution of

.9: by ~ using (8). For the "total Lagrang~an" form one obtains

T.L. :

6{ I[-w* oL
(.!:) + a.. B r
AB OAB O.L K
YL IA + EMBL r ~ Y !A
M
W

Vo
o . L OK! rAB wM 1 OK! oA o.B M
+ EMBL ~ Y A 2"Y A rB ~ WM w
(26 )
+1 0 K! o.LOAB _P fA]
2 Y A ~ r wL "13 0 YA dV

The "updated Lagrang~an" form reads

U.L. :

~ik yj! . <p1

6J CB 6{ J [-w* (E) + riJ YJli + E lkj ~

V
1 oil
<p j ) - pf j yJJ dV
o~
-2 (r i <p.
J
- r <p~ <p l (27a)

-f:t j Yj dS - J tJ(YJ - yJ)dS} 0

Sp Su
 197

which can alternat1vely be expressed in terms of t ij instead

of rij to give
U.L. :

o{IC-w*(_r) +tiJ y I -E t 1J k+l(Ok r oks )

J 1 kiJ <P 2 'k <P <Pr -, <Pk <Ps
V
_pf J Yj]dV - It J Yj dS - It j (Y J - YJ ) dS } o
(27b)
Sp Su

The Euler equations result1ng from (27a,b) are the force equili-
brium (28a or 29a) and momentum balance (28b or 29b) equations,
the polar decomposit1on relation (28c or 29c) as well as all
boundary conditions:
U.L. :

1J 1 + E mkJ oi 1 + E mkJ 01 J o (28a)

r i r k i <Pm r k <Pm 11 + pf

riJ + ~kJ hi symmetr1c (28b)

k

m
h .. + (28c)
m1J <P
E
YJli J1
or

t1JI - pf J 0 (29a)
1

olk j
t 1J + <f jk hi + 1:' a· k symmetr1c (29b)
k

h + a (29c)
YJ 11 J1 J1

The 1nd1ces CB of the funct10nal should not be confused w1th

the cap1tal tensorial indices used above. They refer to a
classification of generalized variat10nal princ1ples 1ntro-
duced by the f1rst author (see e.g. [7] ). Some differences
in comparison with the versions given by Murakawa and Atluri[6]
should be noted. Firstl~ the deformation gradient, rather than
the displacement gradient appears. Although this d1fference is
minor in 1ncremental problems it seems more consistent to em-
ploy the deformation gradient Slnce it appears in the polar
198

decomposit10n theorem. Secondly, the increments of the rota-

tion angles have been introduced explicitly which helps to
further clar1fy the role played by the rotat10n of pr1nc1pal
axes 1n large displacement problems. W1th fifteen independent
var1ables the proposed forms also conta1n fewer unknowns. More-
over, th1s ch01ce of variables is particularly well suited to
the formulat10n of be~n theor1es, S1nce for beams the orienta-
tion of pr1ncipal axes 1S known and, with appropr1ate defin1-
t10ns of stress resultants, the rotat10n angles about these
axes ar1se as natural unknowns.

3. Genera11zed pr1nciple for slender rods

3.1 General remarks

In the deformation of slender structures, such as rods, plates,

and shells, large displacements and moderate to large rotat10ns
may occur. It is thus quite natural to include rotational de-
grees of freedom 1n the descr1ption explic1tly, rather than im-
p11citly through the d1splacement gradient. In the context of
shell theory th1s was recognized by Wempner [10] who derived
appropr1ate k1nemat1c relat10ns uS1ng the polar decompos1t10n
theorem. Here the theorem is also used for the description of
the kinemat1cal behavior of a rod. But un11ke in [10] it is 1n-
serted into the var1ational pr1nc1ple (25) from which, together
w1th add1t10nal assumptions on stresses, a consistent and gene-
ral princ1ple for rods underg01ng large displacements and rota-
tions 1S derived. Its U.L. incremental form gives local equations
which d1rectly reduce to those of well known small displacement
theories.

3.2 Kinematic assumptions

In the initial, undeformed configuration the position vector ~

of an arb1trary p01nt in a plane cross section wh1ch 1S perpen-

dicular to a sU1tably chosen reference axis may be expressed by

R (30)
 199

c
where 6 and n are the d1stances from the axis, R 1S the pos1t1on
vector of the p01nt on the aX1S, and ~k are the un1t vectors of
the orthogonal reference tr1ad at th1s p01nt (see F1g. 3). Vec-
tor ~1 1S tangent to the reference aX1S. The base vectors ~K

are then g1ven 1n terms of ~k by

(31 )

where the Chr1stoffel symbol r K11 accounts for the axial curvature
and tW1St of an initially curved rod.

cross - sect Ion

~
M= center of shear
S = center of gravity

F1gure 3

The number of the st1ffness parameters of the cross section can

be min1m1zed by referr1ng the tors1onal quantit1es to the center
M
of shear, w1th coordinates 6, ~, the bending quant1ties to the
center of grav1ty, and by orient1ng ~2 and ~3 such that they point
1n the d1rect10n of the pr1ncipal axes of inertia. This choice
1S most advantageous 1f complex, nonsymmetr1c cross-sections are
cons1dered .

After deformat1on the position vector, the vector of the orig1n

of the center triad and the base vectors will be denoted ba ~,
200

c
~, ~i and ~i respectively. From (5) it follows that 1n general

a. .c(",B+hB)A (32)
-1 ai U c C-B

where £c denotes the stretched reference triad. Employing now

the usual assumption that in bending normals to the axis remain
normal and unstretched we obtain
,.. ,..
~2 = ~2 ' ~3 = ~3 (33)

from which it follows that during deformation points normal to

the axis are only rotated but not stretched. In accordance with
(30) one obtains

(34a)

cB B B B ]
[Y +n a · 2 +6a.3 -wwa., ~ (34b)

The last term in (34a) accounts for warping of the cross section
due to torsion and w(n,6 ) depends on the cross-sectional shape.
Thus, the deformation of any point perpendicular to the referen-
ce axis is completely described by that of the respect1ve point
on the aX1S. Moreover, as a consequence of (33) and (32) the vectors a.
1
are seen to coincide with the tr1ad of principal strains and the
angles ~ and ~ appearing in (26,27) may be identified with the
increments of the rotation angles about the local coordinate
vectors a ..
-1

In the U.L.formulation expression (34b) reduces to the linear

relationship

c c
y, y, 0 6 -n <P, wW

c M c
Y2 Y2 + -(6-6) 0 0 <P2 0 (35)

c M c
Y3 Y3 n -n 0 0 <P3 0
 201

c c
It is noted that in (35) Z, y, and ~ denote incremental quanti-
ties, and that the sam~ relationshlp is appllcable in the T.L.
descrlption if rotat~ns remain moderate. Furthermore, (35)
differs slightly from (34) in that (35) provides for the fact
that the cent~s of shear and of gravity may not coincide. The
appropriate modifications of (34) have, however, been made in
all general expresslons whlch follow.

3.3 Stress assumptions and stress resultants

From the assumptions implicit In (33) and from the lnverses of

the constitutive relationship (18) it lS seen that the stress
state isessentlally uniaxial, and that the components of the
Jaumann stress corresponding to the vanishing strain components
must also vanish

022 033 023 ~32 = 0

r r r (36)

The remaining components ~1K are the normal and shear stresses
acting in the directions of the vectors ~ of the dlsplaced and
rotated cross section (see Fig. 3).

The reduction of the volume integrals In (25) to one-dimensional

integrals along the rod axis requires assumptions on the dlstri-
bution of the components ~1K across the cross-sectlon. Consistent
with the assumptlons of beam theory the dlstrlbutlon of the aXlal
stress ~11 is taken to consist of a constant term, two linear
terms in the coordlnates n and ~ along ~2 and ~3' and a billnear
term resulting from the "unit warping function" w(n,~):

0 0 0 0
011 N M3 M2 Mw
r
A F n+ ~ - F
w(n,~) (37)
nn F~~ ww

The dlstributlon of the shear stresses is taken as

0
J~
Q2 M M 0
012
r
A
[ (~-~) + w'n J T - r011 (~-~) 1fT (38a)

0 0
M 0
013
r
Q3 M
- [-(n-n) + w'1: ] ~
J
011
+ r (n-n) 1fT (38b)
A T

In (37) A lS the area of the cross-sectlon, Fnn and F~~ are the
202

pr1nc1pal moments of 1nert1a. Fww 1S the principal warp1ng

st1ffness, and J T is the tors1onal stiffness. It may be noted
that the representat10n (37) can also be deduced by substitu-
ting the deformation gradient from (34b) into (12a) and then
using the inverse of the const1tut1ve law (18).

Relat10ns (38a,b) d1ffer from those usually assumed 1n beam theory.

The third terms are due to warping of the cross section. They
account for the fact that the actual normal and shear stresses
act1ng on a warped cross-sect1on are no longer parallel to the
un1t vectors ~1. This d1fference 1S most pronounced so far as
shear stresses are concerned. The effect on the aX1al stresses
can be shown to be small.

0000 000
The quantit1es N, M2 , M3 , Mwand Q2' Q3' M1p 1n (37,38) arc 1dent1-
cal with the usual stress resultants. Th1S may be seen from the
var1at1on of the surface 1ntegral of (26). Not1ng that the in1-
t1al normal vector N is 1dentical w1th ~1(F1g. 3) one obta1ns

OAB
Jt NA 6YB dA =
oB
a· K
J rolK 6YB dA (39)
A A
By Subst1tut1ng (34) or (37) 1nto (39), carry1ng out the inte-
grations over the th1ckness, and comparing coeff1c1ents one
f1nds

011 0
r N
r 0
012 o
r dA Q2 2 (40a)
J 013 0
A r Q3

M M 011
0 -(k;-rJ (ll-ll) r

012
)
r k; 0 0 r dA M ( 40b)

A 013
-ll 0 0 r

011 0
Jw r
A
dA Mw (40c)
 203

0 0 0
N is thus the resultant aXlal force, M2 and M3 are the bendlng
o
moments about ~3 and ~2' and Mw lS ~he "warplng bi-moment". ; t
lS noted that the torsional moment M1 In (40b) dlffers from M1p
In (38a,b) by terms whlch are due to warplng. Slmllar terms
influencing 8 8
2 , 3 and ~2' ~3 have been omltted In relatlons
(40a,b) but are taken lnto account in the development of the
next chapter. In matrix notation (39) then reads

f °B
t 6YB dA ~'.Q
° 0 c 0 c
° 611'
6y + M 6w + Mw
(41 )
A
0c 0c 0
It is also seen that y , w , and ~ are the klnematic variables
000 0
which are conJugate to ~ 9, ~ and Mw' By further substitutlng
c 0 c c
6~ = 12 .6~ from (10), where ~ are the rotation angle lncrements
about the lnltial reference trlad, one may rewrite (41) as
follows:

0 0 c 0 0 c
° 611'
a .Q 6y + M b 6.!fJ + Mw
A (42)
c
° 6y
0 c 0
V + B 6~ + Mw 611'

o o 0
where V a.Q
°
B = b°T M
° ( 42a)
oc oc
Their conJugate klnematic variables are y and ~ . Equation (42)
o 0
may thus be interpreted as the vlrtual work performed by V, B
o - 0-
and Mw In the lnitlal conflguration. Flguratively speaking,~
o
conslsts of the sum of the components of .Q pOlntlng In the di-
rection of theinitlal reference system. The simple form of (42)
lmplles that the general flcld equatlons take on a partlcularly
slmple form if expressed In terms of these conJugate quantitles.

3.4 Generallzed varlatlonal prlnclple for beams and

correspondlng local equatlons

Introduclng the assumptlons made In the prevlous chapters into

(25) results, after lenghty but stralghtforward operatlons, In
a generalized variational principle for beams in terms of the
deformed posltlon vector of the beam aXls, the rotatlon angles
and the warplng parameter ~ as well as the stress resultants
glven above.
204

One obtains from (25):

6{ J[-Q* (2'~'~w) + "" ~All - Q1 + BA ~All

x
(43)
-M 1 W + M4 W'l]dx + boundary terms} o

where

(43a)

which may be expressed in terms of the transformed resul-

o 0
tants ~ and ~ by using (42a). Similar expresslons for the T.L.
and U.L. formulation may be derived from (26,27) or directly from
(43). This completes the reduction to the one-dlmenslonal problem.

The incremental versions associated with (43) are represented

schematlcally below. Writing them in this form most clearly dis-
plays their common structure.

I
T. r
c
X l E
: "Irtual war c
c Q !19.Y) jlIU!) 'f ill
If I

J6
Y'"
---~---~- - ----- • --------,--
,, 'Y" ds.
(44)
5
kmemallc I malenal
~Trj.~)
,
~I~) M
t:1 equallons law
I

/til" /til",

or J{6~T (!2.L + !2.G) z - 6z T E} ds + ... (44a)

S

In (44) the components of ~ are the incremental kinematical

quantities and stress resultants, and E denotes the vector of
external forces and moments. Matrix !2.L contains constant mate-
rial and geometric parameters as well as differential operators
which operate on the components of z. All terms containing quan-
titles of the current state are grouped together in !2.G.
 205

It is noted that the components of ~ representing stress re-

sultants are the lncrements of the quantltles ~ and ~ rather
than B. Wlth this choice of lncremental varlables the field
equat.lons are partlcularly amenable to the numerical solutlon
scheme outlined in the next chapter. For the case of the T.L.
formulatlon and an inltially stralght beam,expression (44) is
glven in detail In the Appendlx (A4, A4a). In the respectlve
expresslons for curved beams additional terms accounting for
the lnltial curvatures and twist appear [4] .

The Euler equatlons assoclated with (44) are obtalned by In-

tegration by parts along the aXlS. They are again expressed
In terms of z as defined above, which leads to a system of
flrst order ordinary differential equatlons which is again
given schematically below:

c c
r kmeff'allcal molena I 1 E
c relallons law -2 T(j,'l!) -!1/~) c
ill
'!. I if
vr
I

vr
Ix
I

•
I I
------~-r------- --------1-- (45)
¥ I
I .l(
force -
eq~rllbnum f (9,(1) 21j, ~)
M momenl- I M
M., Mw

or z' (~L + ~G) ~ - E ( 45a)

As before all nonlinear terms are written separately in ~G.

Due to the particular choice of lndependent variables the

matrices i, ~ and Q are the same as those appearing in ~

of (44).

Finally it is noted that the corresponding expressions in

the U.L. description follow from (44) and (45) if kinematic
variables of the current state vanish, and the covariant de-
rivatives are taken wlth respect to the updated, deformed
206

geometry. Frequently th1S 1S not done 1n practice. Instead, the

deformed conf1gurat10n is often approx1mated by straight ele-
ments. In the full express10n (45) Wh1Ch is glven 1n the Appen-
dix(A5) these influences are incorporated via the terms k2, k3
and h which represent the initial (current) values of the cur-
vatures and the tW1St of the beam ax1S [4] .

4. Numer1cal Solut10n technique

The variational principle (44) and the corresponding local equa-

t10ns (45) can now be used as the bas1s of the d1scret1zat10n
process, either in the T.L. or 1n the U.L. descript10n.

One way 1S to approximate the unknowns by 1nterpolation func-

t10ns, resulting 1n a fin1te element approach of m1xed type.
Another poss1b1lity 1S to start from the local equations (45).
In this approach the system of f1rst-order ordinary d1fferen-
t1al equations (45) 1S 1ntegrated numerically over each element,
and the resulting 1ntegral matr1x 1S transformed to a st1ff-
ness matrix. It can then be treated 1n the same way as 1n the
usual displacement finite element approach. Th1S method is par-
t1cularly well sU1ted to the treatment of the one-dimenS10nal
rod problem and has been shown to give very accurate results.
It has the add1tional advantage that a certain relat1ve error-
bound can be given a-priori for the element matr1x. It also
appears to better account for the rap1dly decaY1ng warping
funct10ns, and the1r sometimes slgn1ficant effect on the load-
carrying behavior of slender beams, than st1ffness matrices
developed from a usual displacement formulat10n. For further
details of the approach, which has been used successfully by
the f1rst author and his coworkers, see [2, 4, 11] . The
examples glven below have been solved 1n th1S way.

5. Numerical examples

The examples were ma1nly chosen to demonstrate the interaction

of bending and torsion during large spatial deformat10ns. It
1S assumed that the material 1S suff1c1ently elast1c so that
 207

material nonlinearities need not be cons1dered. The calcula-

tions are based on the D.L. expression of (45), i.e. (A5).

5.1 cantilever under end moment

In this standard example an 1nit1ally stra1ght cant1lever 1S

deformed 1nto a full c1rcle by an 1ncreas1ng bending moment
at the free end. The numerlcal results glven 1n Figure 4 show
good agreement w1th the analyt1cal Solut1on, and demonstrate
the usefulness of the above approach.

~-c~~__~L~~~_ _ _ _.~
M= ;mEl analytical
2L
~ numerical
(10 elements
160 load-stepsl

F1gure 4
M --Jill.
l

5.2 Two span beam under bending and torsion

In th1S example a two-span beam loaded by a slngle load 1n

midspan, a constant d1str1buted load as well as a d1str1buted
torsional moment 1S analyzed. It was chosen to demonstrate that
warp1ng constra1nts over the central support may be important,
and that near the load caus1ng sidesway buckl1ng the torsional
rotat1ons are no longer small. The configurat1on is shown in
Figure 5. The cross sectional parameters were taken as follows:
M 6 2
~ = - 0.088 m, EF = 1.008 ·10 kN, EFnn 1575 kNm , EF~~
2 9.882 kNm 2 . The
1 6902. 9 kN m , EF ww = 15.057 kNm 4 , GJ T
boundary condit1ons are ObV10US from the f1gure and all loads
- 0
were increased proportionally, i.e. E = A· E, where the values
o o 0 a /
o£ E were taken as: P~1 = 5kN/m, P~2 = 1kN/m, m1 = 0.2 kNm m
10 kN.
208

cross-section

R -1 II 1111111101 jJ II I H

"g;" 6:
II I 0

1" om
II H t

1c
~2

"~rf~10cm~
Figure S

The numerical results are given in F1gS. Sa-d. In Fig. Sa the

load parameter A is plotted against the vertical and lateral
deflections at the midpoint M. It shows a clear nonlinear be-
havior as the critical load is approached. The dashed lines in
Fig. Sa and all subsequent figures indicate the results obtai-
ned from a version of the above equations which is restricted
to moderate rotations. It is seen that the discrepancies are
small at low load levels but that they may become considerable
as higher loads are approached. It is interest1ng to note that,
depending on the loading range, these simplified equations may
over- as well as underestimate displacements and moments.

Figure Sb shows that since failure 1S due to tors1on the rota-

t10n angle ~1 may become quite large at higher load levels.
Finally, the torsional and bend1ng moments M1' M2' M3 as well
as M at points A, B, and M are plotted in Figs. Sc, d as func-
w
tions of A. From Fig. Sc it is seen that the bi-moment Moo at
the center support 1S of the same order of magnitude as the
bending moment, and that at M the bending moment M2 about the
n ax1S even decreases after a certain load level has been reach-
ed, while the vertical deflection continues to increaserap1dly.
A

--- ---
8
7

6
5
-large rotation theory
4
-- - moderate rotation theory
3
2

2 4 6 8 10 12 14
Figure Sa: Displacements u 2 , u 3 at M
it 209

8
7 ~--

-- --
......., --- <P.
6 1
./
/'
5 /'
/
4 /
/
/ - large }
3 / rotation theory
- - - moderate
2 h
1/

1,01
10° 20° 30° 40°
Figure 5b : Torsion angle <p at M

it
8
7

6
-- ~-
"-

/
/
, M3

/
/'
5 /'
/'
/'
4 /"
~
3
2

Mz,M 3
50 100 150 200 [kNm]
250

Figure 5c: Bending Moments M2 , M3 at M

it

-- ----
M1 _
8
~
...-
7

5
4

3
2

M1 [kNm]

5 10 15 20 25 Mw [kNm4
Figure 5d:Torsional moment M, and warping bl-moment M
w
at B
210

5.3 Cantilever under bendlng and torslon

The cantllever shown in Figure 6 has the same cross sectlon

as the beam ln the previous example. It was agaln loaded pro-
portlonally by a vertical and a lateral end load as well as
an end moment with relative magnl tudes ~ k: = 2.5 kN,
o
t? k: = 0.8 kN,
Ml = 0.2 kN m.

cross-section

Figure 6

In FlgS. 6 a -d the same quanti tles as before are plotted ver-

sus the load parameter A. It is seen that as the failure load
is approached, both vertlcal and lateral displacements lncrease
very rapldly. The torsion angle however, increases much faster
and becomes quite large at loads of less than 50 % of the fail-
ure load. Flg. 6d shows that at the clamped end the warplng
constraint is so severe that its influence may not be neglected.

A-
u3 u2

-- -- --- ---
8

-- ---
7
6
./
/'
5 V

4
3
large }
rotation theory
mo d era te

5 10 15 20 25 30 35

Figure 6a: Dlsplacements u 2 , u 3 at B

 211

A
8 - - Iorg(' }
rotation theory
- - -moderate
7
6

-
'1\
- -
- -- - -
5
4

3 ""
2

5° '00 15° 200 ,,0 30° 35"

~
'

Figure 6b : Torsion angle ., at B

A
8
/

"
7

5
4
3
2

~-~--'---r--~--'---r---~ M"M 3
20 40 60 80 100 12 0 [kNm]
Figure 6c : Bending moments H , M3 at A
2

A
8
-- -------- - ---
7

6
-M.
5
4

J
2

M,(kNm]
2 4 6 8 10 12
• 2
M... {kNm J
Figure 6d : Torsional momen t M, and warping
bi- moment Mw at A
212

6. References

Bathe, K.J. and Bolourchi, S: Large displacement analys~s

of three-dimensional beam structures. Int. J. Num. Meth.
Eng. 14 (1979) 961-986.

2 Wunderl~ch, W. and Beverungen, G.: Geometrically nonl~near

theory and analysis of curved rods (in German). Bau~nge­
nieur 52 (1977) 225-237.

3 Argyris, J.H. et al.: On large displacement-small strain

analysis of structures with rotational degrees of free-
dom. Compo Meth.Appl. Mech. Eng. 14 (1978) 401-451,
15 (1978) 99-135.

4 Beverungen, G.: Geometrically nonl~near stress- and sta-

bility analysis of spat~ally curved rods (~n German) .
Mitteilung Nr. 76-13 (1978),Techn.-wiss. M~tteilungen,
Inst. f. Konstrukt~ven Ingenieurbau, Ruhr-Un~versitat
Bochum.

5 Frae~js de Veubeke, B.: A new var~at~onal principle for

finite elastic d~splacements. Int. J. Eng. Sci. 10 (1972)
745-763.

6 Murakawa, H. and Atluri, S.N.: Fin~te elast~city solut~ons

using hybrid finite elements based on a complementary
energy principle. J. Appl. Mech. 45 (1978) 539-547.

7 Wunderlich, W.: Incremental formulation for geometrically

nonl~near problems. In "Formulat~ons and Computational
Algorithms in F~nite Element Analys~s". K.J. Bathe,
J.T. Oden, W. Wunderl~ch, Eds. MIT Press (1977) 193-240.

8 Malvern, L.E.: Introduction to the mechanics of a cont~nu­

ous medium. Prent~ce-Hall, Englewood Cliffs (1969).

9 Wempner, G.: Complementary theorems of solid mechan~cs.

In "Var~ational Methods in the Mechan~cs of Solids".
S. Nemat-Nasser, Ed. Pergamon Press (1980).

10 Wempner, G.: Fin~te elements, f~n~te rotat~ons and small

strains of flexible shells. Int. J. Sol~ds Structures 5
(1969) 117-153.

11 Wunderl~ch, W.: Calculation of transfer matrices, applied

to the bend~ng theory of shells of revolution. In "The Use
of Electronic Dig~tal Computers ~n Structural Eng~neer~ng".
Proc. Int. Symp. Newcastle upon Tyne (1966).

Acknowledgement - The support of the "Deutsche Forschungs-

geme~nschaft" and the valuable help of Dr.-Ing. H.G.L~ekweg and
D~pl.-Ing. V. Schrodter ~n carry~ng out the numer~cal cal-
culat~ons is greatfully acknowledged.
 213

AppendlX

0
[ e,e, 8182C3-C183
(A 1)
.9,.= C2 8 3 C1C3+818283
-8 2 8 1 C2

0 o
where C cos <Pl 8i sin <Pl (i 1, 2,3)
1

(C 1 8 2 C3 +8 1 8 3 ) <P1 (C183-8182C3)<P1
+8 1 C2 C3 <P2 +C1C2C3<P2
-(C1C3+818283)<P3 +(81C3-8283C1)<P3

(C18283-81C3)<P1 (-C1C3-818283)<P1
+8183C2<P2 +83C1C2<P2 (A2a)
+(8182C3-83C1)<P3 +(8183+82C1C3)<P3

o C1<P2+81C2<P3
(A2b)
U) = -81<P2+C1C2<P3 -<P 1 +8 2 <P 3

-C1<P2-81C2<P3 o

1 o
o
E. o C1
(A3)
o -8 1
 c c c c c
YI YZ >; WI W z W3 vr I MI I M2 I M3
~ ~ ~ I MOl '"
r-- ""
,~

~
1(:) .
I~

o 0
lel "3 -"2
k oA \0)1.. \00
~ 0 ~ IX31Q3 0 0
£A +IX13 EA +ct.?3 EA + IX33i -"3 I~ I ;;
00
00 I of}. o
IXI/OZ 0 0
~" 1 IX3102 0 I
- EA HX 12 - £A -(XZ2 - EA -a32 "2 -vr I~

-1 ,OJ
6~T'~--~--~----T---~--~--~--~--~----+-------~-------r--o--o---+----t----+----t---~ (A~)

-~
EA
0 o 0
o 0 oR0 0 0 oZ
~ _~2_ 0 _ (XI/(XZI _!!i1 _ (X21 IX31
~I EA +IX23.
I EA Ci22
I I EA EA EA
o 0 0 0 02
(X31030 0
I Ci31 0 Z _ &/1&31 ~I !!J.I
~, £A +Ci33 - £A -a32 -EA
EA £A
o o
~I - "3 "2 -1
o 1
f)1 -v, -u.,t
"3
o o _L
-"2 l/I ~I E~~
_--L
0)1 E~",
L..-
 215

c-

• •
o 0 j
~"2+ ""21/1
0 0
~~3+}~3Y, M. M•
.b. : M,1/1 ~ 0 0 0 0 d 1!Q2+~Q3 (A4a)
- '1i1/l "3 - Q 2 + fi1.,1/1 "2- 3
} • 0.

° , fi121/1
~ "2+ ° ° -~y,+~Xt'~X;
° °"2°
J ° 0"3-
- fi1.,1/1 .
o 0 02
<52 +M.,"., -QlEA+Q,
0 M3+M.,

lit, "3°+l~ °
2 31/1 -~"+~X2+~~
-;:'2 +t:t., *'3
+1't:tw1/lX2- t! 3
o 0 02 0
0
• +M.,".,-Q)'EA+G,
M •
-1! Q2+ ~ Q;
M.
•
M; + M.,"2
•• a
-M2+ Mo.";
•• GJT+ 1:'0.,
 c c c c c
-'"
'"
Y, Y2 Y3 </, 'P2 ~J I 1/1 I VI I l<j I ~ I /vi, I /vI2 I /vI3 I /vi", }

c
- Q3 Q2 1
-
Y, k] -k2 -a ~ EA
J2 -k3 h 1
c
Y3 k2 -h -1

4>, +k3 -k2 1

c -L
~ -k] h EFtt
c ---L
'P3 IS h EF7I7I
---L
EF",.,
dH- (AS)
axv: -
I k3 -k2
1
~ -k3 h
~
~ k2 -h
-'1/ Q2
~
N, - 13 2 -Q 3 ~.~ k3 -k2
0 0 ""
/vI2 -Q2 Q, ~ 1 -h
M3 ii -k3 fil2
D
"'13 - 153 Q, - M2
• -~ -1 k2 -h fil3
EA
/vi", GJ
-~a2 ~ - t:12 -1 m",
", b. 3 "'1 104 ,
'2&
-- -
Large Deflection Finite Element Analysis of Pre- and Postcritical
Response of Thin Elastic Frames

D. KARAMANLIDIS, A. HONECKER, K. KNOTHE

Techmsche Universltat Berlin, Germany

Abstract

A finite element variational approach and a computat1onal

algorithm are presented for pred1cting the nonlinear static,
pre- and postcritical response of elast1c framed structures.
An updated Lagr~ngian formulation 1S used and the structure
is discret1zed by uS1ng curved, mixed-hybrid bearo elements.
Throughout the development correct1on terms are maintained
in the 1ncremental energy functional to reduce the dr1fting
of the approximdte Solut1on from the true solution. Th2se
correct1on terms correspond to checks on the stress resultants
equilibr1um and compatibility in the reference state. To tra-
c1ng the unstable postcr1t1cal load deflection path in snap-
through and bifurcation problems a general1zed incrementation
procedure has been developed and implemented. The high accu-
racy and effect1veness of the proposed approach 1S demonstra-
ted by means of well-selected numer1cal examples.

1. Introduction

Large deflect10n nonl1near analys1s has become 1n the last

decade the focus of intensIve efforts. Considerable pro~ress

has been made 1n the development of versatile and powerful

f1ni t:e element discret1zat1on methods as well as of 1mproved
numer1cal methods and programm1ng techniques for nonlinear
analys1s of structures. Since the first applicat10ns of com-
puters to nonl1near structural analysis a number of nonlinear
beam elements have been presented. 'rhe large number of publ1-
cat10ns on th1S topic is due to the fact that various kine-
matic formulat1ons as well as different energy functionals
can be employed.

In the development of a geometrically nonl1near f1nite element

procedure an updated Lagrangian or a total Lagranglan formula-
tion can be employed. Although both formulations lead to iden-
218

tical results (as pOinted out by Dupuis et al [1J and Bathe

and Bolourchi [2J ) their computational efficiency is diffe-
rent. In the case of geometrically nonlinear framework ana-
lysis the updated Lagrangian appears to be more effective than
the total Lagrangian formulation.
The most popular finite element procedure for geometrically
nonlinear analysis is still the so-called assumed displace-
ment method based either on the principle of minimum poten-
tial work or the principle of virtual work (see, for example,
Refs. [1 - 5J). Pirotin [6] was propably the first 1nvest1-
gator to study the nonlinear response of a sinusoidal arch
utiliz1ng curved beam elements. In that work three d1fferent
models have been implemented: an assumed displacement model,
a m1xed Reissner model and what he called a modified stress
hybrid model. The idea underlying the latter model has been
introduced by Pian's pioneering work [7J. Pirotin's arch
problem was being analysed later by Atluri [8J and Boland [9J.
Geometrically nonl1near analyses of plane frames using mixed
var1at1onal procedures have been carried out also by Backlund
[10] and more recently by Noor et al [11J, Wunderlich and
Beverungen D2J and Karamanl1d1s et al [13J.
The objective of this paper 1S to present a consistent large
rotation updated Lagrang1an formulation for two-dimensional
curved beams. The finite element procedure is based on an
incremental hybrid mixed variational principle ITmR contain1ng
both displacements and stress resultants as independent
variables. The beam elements are assumed to be curved and the
conventional beam displacement functions are employed as
shape functions formulated 1n local, convect1ve co-ordinates.
The deformed element geometry has been for convenience 1nter-
polated in the same fashion as the beam lateral deflect1on.
Finally, stress resultants have been approximated by tr1al
functions which a priori satisfy the incremental homogeneous
part of the equilibrium equations for the curved beam. The
element has been 1mplemented for use 1n elastic and elastic-
plastic analys1s 1n the pre- and postbuckling range (see, Refs.
[14 - 16J). In the paper some demonstrative sarople solutions
are presented.
 219

2. A geometrically nonlinear theory for plane curved beams

The equat~ons of the two-dimensional, arbitrarily curved

plane beam element are formulated based on the general con-
tinuum mechanics e quations, using an ~ncremental updated
Lagrangian approach. These equations have been outlined by
various authors previously (see, for example, Refs. [2,4,9J).

Fig. 1 shows a single curved beam element ~n three different

states, ~n the ~n i tial configuration ~ (o) ~n the r efe rence
configuration n( N) and in the incrementally adjacent c u rrent
configuration n ( N+1 ) .

" ct.'Of'fIWCf
CUI'fe"t stol e "
()Uhll

I I
I I
I
I I
I I
I, I
I
I undeformed I,
I in itialslo le
glCiI

I I
I I
I
j'L ____ . _ __ _
~, .'
!,L _______ _
!, xl

Figure 1. Local, c onv ec t i ve Figure 2. De scrip t i o n of

co- ord~nate - systems mot i o n by the u p da ted
Lagrangian formu la t i o n

The updated Lagrangian formulation of the governing equations

of a deformable body requires the use of Euler stresses
( E)a ~~) as initial stresses, and Truesdell stresses ~ *a~~)
~ J ~ J
as incremental stresses. Consistent to these are the Almansi
stra~ns a~N) as initial strains and updated Green strains
~*a(N) as ~ incremental strains. In the case of a two-dimen-
lJ
sional curved beam, these tensor quantit~es are reduced to
the scalar state variables (E)a(N) , ~ *a (N ) , a(N) and ~*a(N) .
220

Following Washizu [17J the updated Green strain increment

~* a{N)can be for convenience decomposed into a llnear and
into a quadratic part. The linear part as well as the Almansi
initial strain a{N) can be further decomposed in their exten-
sional and curvature parts:
* a (N)
~ + "2 (~w 'x(N)) 2 (1a) ,
(N)
e (1b) .

where (see Fig. 2)

* (N) ~u~N) z~N)~w~N)
~ e + (2a) ,
x x x
(N) (N) (N) (N) (w ~N) )
2
e u'x + z'x w'x (2b) ,
2 x
* (N) ~w(N)
~ K (2c) ,
'xx
(N) (N)
K - W
'xx (2d) .

To obtain these strain displacement relations the assumptions

of J.I1arguerre's shallow shell theory have been adopted (see, for
example, Refs. [9,13,14J).
Integration of the normal stress (E)a(N)resp. ~* a(N) over the
depth h of the beam leads to the stress resultants (see Fig.3):

(E)N(N) + ~*N(N) h/2 (E) (N)

= J [ a + ~*a(N)Jd1; (3a) ,
-h/2
h/2
(E)M{N) + ~*M(N) J [(E)a(N) + ~*a(N)hdl; (3b) .
-h!2

In this paper stress resultants and beam strain quantities

are assumed to satisfy the linear constitutive law:

(4) ,
 221

where E A and E I are the extensional and flexural beam

stiffness, respectively.

Finally, the following stress resultants equilibrium e0.uations

can be formulated

o (Sa) ,

o (Sb)

S(N) denotes the shear force

are the external distributed

"t
I
I
i
I
K.L global nodes
I
L_ -----------
"
Figure 4. Element external
loads

Figure 3.Beam stress resultants Figure S.Incremental solution

within the updated Lagrangian schemes, a) Euler method
formulation b) self-correcting
method
222

3. F~nite element variational approach

The finite element procedure utilized in this paper is based

on a generalized incremental variational pr~nciple in which
the displacements ~u and ~w and the stress resultants ~*N and
~*M are the independent variables+. Assuming the two-d~mensio­
nal frame to be discretized into (n) individual elements the
generalized energy functional ITmR is defined as follows ~4J:

(~*N) 2 *
1~
ITmR - L{ f(N) [- 2EA
+ (E) N ~(~w, x)
2EI
2J dx
n 1

+ [~ * N ~u + ~ *S ~w + ~ * M ~8Jl(N)}
0

I f(N)(~P1~u+~P3~w) dx - L (~N~u+~S~w+~M~8)
n 1 k

+ L
nIx
f(N) [(E)N(~U, +z,
x
~w, )_(E)M~w,
x xx
-P1~u-P3~wJ dx

U,
x

* (E) M
+ ~ M ( EI + w'xx)} dx

stati onary (6 )

Eqs. (6) is subject to the following constraint conditions

/

* N,x = 0
~ in l(N) (ia) ,

~ * M,xx + (~ * N z, ), = 0 in l(N) (ib) ,

x x

which represent the linear incremental homogeneous parts of

the stress resultants equilibrium equations (5). On the other
hand the kinematic boundary condit~ons and the ~nterelement

displacement compatibility must be satisfied exatly.

+ Since no confUSlon is possible the state label 'N' will be

dropped In the following.
 223

The functlonal ITmR ' eqs. (6), contalns two types of correction
terms, an equlllbrium correction term (2) and a compatibility
correctlon term ~ . These correction terms are due to the
fact that equllibrlum equations for the initial stresses as
well as strain dlsplacement relations for the initial strains
are nonlinear equations which cannot be fulfiled completely
by a sequence of llnearized incremental equations. Durlng the
incremental procedure presented here the correction terms in
the load step (N) are taken into account ln the following step
(N+l). Accordlng to the nomenclature introduced by Pian [18J
and Stricklin et al [19J the geometrically nonlinear procedure
presented ln thlS paper can be thereafter identified as a
self-correctlng lncremental procedure (see Fig. 5).

4. Interpolation functions

It is convenient to introduce matrix notation in this chapter

to replace the scalar notation above. Also, the independent
variables of the functional ITmR shall be interpolated in terms
of trlal functlons appropriately chosen to satisfy the afore-
mentionend constraint conditions of the functional.

The displacement lnterpolatlon functions are constructed

assuming CUblC variations for ~w and llnear varlations for ~u

(8 )

where ~uis the element nodal displacement vector. Integration

'On
of the ordinary differential equations (7) with respect to the
local, convectlve co-ordlnate x leads to the stress resultants
interpolation functions

B1

(9) •

Introducing eqs. (8) and (9) in eqs. (6) yields

224

AT
- .12 feT ;E!;nfe + 1 LliC n
IImR - ~
n
{
n n 2 Jt~ LliC n

AT T AT
LliC n LI~'n +
fen ~n Llltn + LliC n ~n - feTn!(,n }

T
- ~ LliCk Lltk stationary ( 10)
k

where 'V
H is the flexib~lity matrix
/fO is the initial stress stiffness matrix
LI~ ~s the incremental element lUIllped load vector

~ is the equilibr~um imbalance term and

R, is the compatibility mismatch term.

T -T - - -
LI!C k = {Llu k , Llw k , Ll0 k } and LI~ k= {LINk' LlS k , LlM k } denote the
displacement and load vector of the global nodal point k,
respectively.

Taking the variation of II mR , eqs. (10), with respect to the

independent coefficients ~ yields a solution for the ~ in
n A n
terms of the nodal displacement vectors LliCn :
-1 A -1
~n = ;E!;n ~n LliCn-;E!;n !(,n ( 12) •

Placing this into IImR results

IImR = ~ {~LI~~(~~ ;E!;~1 ~n + /f,~)LI~n - LI~~(LI~n

n

stationary ( 13) •

F~nally,taking the variatio~ of ITmR with respect to the global

nodal displacement vector LliC yields the matrix equation

where

~ is the assembled (global) tangent stiffness matrix

LI~ ~s the global load vector and
~ is the conSistently generated equivalent correction
load vector.
 225

5. Algorithms for tracing the complete deformation path

Eqs. (13)is usually applied in an inver sed form such that the
incremental displacement vector ~~ is obtained by premulti-
plying the right hand side vector ~~ -! by the inverse of the
matrix ~. However, difficulty arises if the load displacement
relation presents a snap-through or bifurcation behaviour. In
these cases, in the neighborhood of the critical pOints the
matrix ~ becomes singular, so that its inversion is impossible.
In the past, several schemes have been suggested for tracing
the postcritical deformation path of instability problems. The
most popular of these procedures are:
- displacement incrementation(Argyris, Pian/Tong, et al),
- ficticious springs concept (Whitman/Gaylord, Sharifi),
- " arc length"incrementation (Riks, Wempner, Wessels) ,and
- current stlffness parameter concept (Bergan).
The common feature of the first three methods is to introduce
modifications or constraints that render the matrix ~ positive
definite in the postbuckling range, while the procedure by
Bergan simply "jumps" over the instability points.
In solution of practical problems it is highly desirable that
an algorithm tracing the complete deformation path fulfils
the following appropriate requirements:
- The method should not fail in the case of load displace-
ment curves having vertical tangents.
- The method must be capable of detecting that a bifur-
cat~on point has been reached in order to trace the
secondary equilibrium path.
- Symmetry and sparse population (bandwidthness) of the
coefficient matrix ~ must be preserved.
- In order to keep the truncation error of the incremental
procedure nearly the same for each loading step, the
load steps should be varied automatically according to
the local curvature of the load displacement curve.
In a recent work [20J by the junior author (A.H.) a new in-
crementat~on scheme for trac~ng the complete deformation path
226

has been developed and implemented. The numerical results

for several test examples lead to the conclusion, that the
new incrementation scheme, to be outlined and discussed in
the following, satisfies nearly all of the requirements
above.

The basic idea behind this new step-by-step algorithm ln to

impose, in each load step, external work increments ~S(N)
(see Fig. 6)

(16)

rather than load or displacement increments. For proportional

loadlng an arbitrary load increment may be expressed by scal-
ing an appropriately choised reference load q
.. re f(e.g. the
estlmated buckling load):
~-(N) ~A (N)- (17)
~ ~ref
where A denotes the load parameter. It is convenient to scale
the incremental displacement vector ~~(N)in the same manner
as the incremental load vector, i.e.:
A(N) ~A(N) A(N)
~!C !(,ref (18)
... (N)
where the reference displacement vector ~ref is defined by
(see Fig. 6)
(N) A (N) -
~ !Cref = ~ref ( 19)

Placing eqs. (17) and (18) into eqs. (16) yields:

(20)

so that

~A(N) ± (21 )
-T A(N)
~ref ,!(;ref
In our numerical studies, carried out by the new incrementation
algorithm the following choices have been made:

~S (N) = I- T A ( 1) I (22)
p ~ref !Cref

p 0.025 • 0.100 % (23)

 227

(24) .

In the ne1ghborhood of 1nstab1lity points, application of

egs. (21) is still pOSSible, producing very small load in-
crements and 1n some cases local dr1ft1ng from the true load
deflect10n path. On the other hand, it is seen from egs. (21)
that the algorithm fails in regions of the load deflect10n
path where the denom1nator qT
_re
f ~(Nf)
""Te
becomes close to zero.
For problems where the loading consists of a slngle concen-
trated load only, th1S sltuation ar1ses 1f the load deflec-
t10n curve has a nearly vertical tangent. To overcome this
difficulty, 1n the numerical stud1es to be presented in the
next chapter, 1t has been found to be conven1ent to replace
in these regions (A ~ Band C ~ D, Fig. 6) eqs. (21) by
LlA(N) = Slgn{LlA(N-1)}. LlA(1) (25).

The applicability of the proposed incrementat10n algorithm

wll1 be demonstrated in the next chapter.

F1nally, it 1S worthnotlng that Bergan's current stiffness

parameter concept appears to fail not only in the case of
the load deflection curve hav1ng a vertical tangent, but
also for problems where the current stiffness parameter S
p
1S nearly constant (e.g. Euler buckling problems) .
The reasons for these shortcornrn1ngs are:

- In the case of a vertical tangent the current stiffn8ss

parameter S becomes inflnite (see, for example, egs. (6)
1n Ref. [24J).
If S is nearly constant (Lls(N)close to zero) egs. (10) in
p p
[24J becomes mean1ngless.

Indeed, inspection of the avallable literature on the current

st1ffness parameter concept [21-26, 28]reveal that nonl1near
structural problems having one of the properties above (i.e.
vert1cal tangent or S
p
= constant) have never been considered.

6. Numerical test studies

In order to test and evaluate the beam element resp. the step-
228

by-step algorithm proposed, a large number of arch and frame

problems were solved. In this paper results of some selected
problems are presented and discussed. These problems are:
1) Symmetrical buckling of Williams'toggle frame, 2)Shallow
sinusoidal arch under gravitiy loading, 3) Bifurcation of a
right-angle frame, and 4) Nonlinear response of a nonshallow
clamped-hinged circular arch.

a.
.,"0
a
0
-'
70
I aM' I a .., I
60

E
50

40

3D

o,L-~~~-----~-------r------- 20

10

Q1 02

Figure 6. Idealized load de- Figure 7.Load deflection

flection curve with horlzontal curves of Williams'toggle
and vertlcal tangents frame

6.1 Williams'toggle frame

This problem has been treated by Frey [4] and Wood and Zien-
kiewicz [27J previously. In both references the assumed dis-
placement flnite element approach together with a modified
Newton/Raphson method have been used. The finite element re-
sults using the mixed-hybrld beam element and the self-
correcting step-by-step procedure, are shown In Flg. 7,
where very good agreement with the numerical results of
Ref. [27] is obtalned. The variation of Bergan's current
stiffness parameter S as functlon of load level is shown
p
In Fig. 8.
 229

6.2 Shallow slnusoidal arch under gravlty loadlng

The shallow sinusoidal arch shown in Fig. 9 has been fre-

quently used to demonstrate the appllcabillty of Bergan's
current stiffness parameter concept (see Refs. [21-24J).
As lt is seen from Figs. 7, 8, 9, 10 and 11 the nonlinear
response of thlS structure is very similar to that exhiblted
by Willlam's toggle. The analysis presented here has been
carried out using the new incrementation algorithm. Fig. 9
indlcates very good agreement ofl the obtalned Solutlon with
the numerical results by Bergan et al [22J. The varlatlons
ot the axial thrust N and the current stiffness parameter S
p
as functions of load level are shown in Figs. 10 and 11,
respectively.

\0

Q. Q35

74=
Q9 I-

I--

;--- - - u
o
.3 025
t----

f- 020

Q2

01
015

010
-]T
- present analysIs .o.S .Q025
- f-- • Sergan tt aI 1221 .o.~ =0 05

60 80 100
0~0---+--~tO~-~'~5--~2~0-~
20
load P Deflechon Parameter w/z.

Flgure 8. Variation of S for Figure 9. Load deflection

Wllliams'toggle frame P curve of shallow arch

6.3 Rlght-angle frame

The rlght-angle frame under point load shown in Fig. 12 has

become nearly a classical case for demonstrating experiment-
ally or analytically structural imperfection sensitivity.
The problem has been analysed previously by Argyris and Dunne
230

[5J and by Frey [4] using the finite element displacement

method together with a modified Newton/Raphson iteratlve
procedure.

10

+---4---+-----'..-...-.... --+---4
vf
~
.
K
a.Q35
~

Ii
/"
~ 030 /,
e.

~
-g
.9 025 t - - - - - j \
-~
020t---+
L \
015
02 -

01 t----
r-

r
~ r--
'---

-t±
010 I-
I
005
-01

-02
t----

t----
-; I---

h
~

I
1000 2000 3000 4000 o 005 010 015 020 025 030 035
Axial Thrust N Load Parameter p

Flgure 10. Varlation ofaxlal Figure 11. Sinusoidal shallow

thrust N for a shallow arch arch: Variation of S
P
The flnlte element results of the present analysls compared
wlth the reference Solutlons [5J and [4J (see FlgS. 12 and
13) clearly indlcate the following:
- The proposed nonlinear mixed-hybrld beam element produces
accurate results even if the deflections arising in the
problem under conslderatlon are very large.

- Using the pure incremental solution algorithm large devia-

tions of the approximate solution from the true solution
can arise, even if the loading steps used have been kept
very small.
- The step-by-step algorithm proposed in thlS paper is capable
to trace the complete nonlinear load deflection path of this
complex structure. It should be remarked ln thls connectlon
that accordlng the variation of Sp as function of load level
(not shown in this paper), it seems to be not possible to
analyse this problem completely using Berganls current
stiffness parameter concept.
 231

0.002

a§ 11
PfM • Eul. , buc.lth n9 lood
ii: l ' _OJ
1.0
'0
8
-' 0.9
0,8

0,7

0,6

O,S

0,4 1.1(XXj

E. rl' , .... 600 , !.6()(XX)

0,3
18 elements

0,2
- Argyris/Ounne (5)
Q1
• present work

0
o
Rotation e
Figure 12. Right-angle frame: geometry and
imperfection sens itivity

~~ \!
~ 1~ -f~'f--t- J -1--+--+-- +--+---1
~12OO~' \-'\ \ ~
1M» ~ / \ \- - t ----l- -t----t--l

or----+----+--!/
400

J I
1\/

\
-" /)
'
--~----1_--~----~__l

<OO+-----i,---+/ / \ jvt---t--+---- I-----f- --i

eoot---+---+,i I~
( \;/
• / ~I ,! It---t---'---"'------'----j
I.. UO 24
1__ \ /- -
f ..

-1200+---+----' Eo7.2 . '0 ' , A. 6 , 1· 2

II \ j' Helem.nl,
' :~~} NtwtonlAophson (Ft~141 1
-16OO'+ -----if---li' \ / -t----j _ . f--cOtT"K ting inc.remenla l
I V - -PUN ir.crMlental
-~,t---_+----~--_+---~--_+--~

20 60 eo )0 II)

Deflections u,W

Figure 13. Complete load def le ction curve s of

right-angle frame
232

6.4 Nonshallow clamped-hinged circular arch

Because of its extreme complexity, the complete nonlinear

pre- and postcritical response of the clamped-hinged circular
arch shown in Fig. 14 has not been studied previously.
In this paper the example was selected to test the useful-
ness of the proposed new algorithm to trace a very complex
load deflexion path. An illustration of the nonlinear be-
haviour of this arch is provided in Fig. 14 which shows
a plot of v ertlcal central deflection versus applied load.
It is seen that the clamped-hlnged arch with single load
at apex displays a bifurcation as well as a snap-through
singularity. In the example under consideration it was poss-
ible to trace the two braches of the

-0' -02 a 02 OL 06 08 10 12 " '6

Deflection ~romete r w/R

Figure 14. Cl amped - h i nged ci r cula r arch :

l oad deflection curves and deformed
configurat i ons

load deflection curve by off-Ioadlng after passing the bi-

furcation pOint . It should be noted that as no reference
solution e xists, accuracy of the numerical results presented
 233

here can be ~ndicated by real~s~ng that for all load steps

the bending moment at left end as well as the norm of the
correction vector k are very small (ca. 10- 4 :1) in compari-
son w~th the stress resultants. Regardless of that, there is
no doubt that further studies using an iterat~ve scheme are
necessary. Only then, one has the possibility to make the
numer~cal drifting error of the approximate solution from
the true solution as small as possible.

7. Conclus~ons

In th~s paper it has been demonstrated that

- the mixed-hybrid beam element prov~des an adequate tool

for geometrically nonlinear analyses even if d~splacements

as well as rotat~ons are very large,

the proposed self-correct~ng, step-by-step procedure is

capable to trace automat~cally the complete pre- and post-
critical equ~librium path even in very complex nonl~near

s~tuations. Nevertheless, further studies are needed to im-

prove the capab~l~ty of the scheme to trace the passage
from the pr~mary to the secondary path in general cases of
b~furcation instability.

References
Dupu~s, G.A.; H~b~tt, H.D.; McNamara, S.F.; Marcal, P.V.:
Nonl~near material and geometr~c behav~or of shell
structures. Compo & Struct. 1 (1971) 223-239.
2 Bathe, K.-J'i Bolourchi, S.: Large d~splacement analys~s
of three -dimens~onal beam structures. Int. J.Num. Meth.
Engng. 14 (1979) 961-986.
3 Ramm, E.: Geometrlsch n~chtl~neare Elastostatlk und
f~n~te Elemente. Hab~litat~onsschrift, Un~vers~tat
Stuttgart, 1976.
4 Frey, F.: L'analyse stat~que non lineaire des structures
par la methode des elements f~n~s et son application a
la construct~on metall~que. Ph. D. Thesis, Universite
de Llege, 1978.
5 Argyr~s, J.H.j Dunne, P.C.: On the appl~cation of the
natural mode technique to small stra~n large displacement
problems. Proc. World Congress on Finite Element Methods
ln Structural Mechanlcs, Bournemouth/England, 1975.
6 Pirot~n, S.D.: Incremental large deflection analyses of
elast~c structures. Ph. D. Thes~s, M.I.T., 1971.
234

7 Pian, T.H.H.: Derivation of element st1ffness matrices

by assumed stress d1str1butions, AIAA J.2 (1964)1333-1336.
8 Atluri, S.: On the hybrid stress finite element model for
incremental analysis of large deflection problems. Int.
J. Solids Struct. 9 (1973) 1177-1191.
9 Boland, P.L.: Large deflection analysis of thin elastic
structures by the assumed stress hybrid finite element
method. Ph. D. Thesis, M.l.T., 1975.
10 Backlund, J.: Finite element analysis of nonl1near struc-
tures. Ph. D. Thesis, Goteborg, 1973.
11 Noor, A.K.; Greene, W.H.; Hartley, S.J.: Nonlinear finite
element analysis of curved beams. Compo Meth.Appl.Mech.
Engng. 12 (1977) 289-307.
12 Wunderlich, W.; Beverungen, G.: Geometrisch nichtlineare
Theorie eben gekrlimmter Stabe. Bauingenieur 52 (1977) 225-237.
13 Karamanlidis, D.; Tsuzuki, 0.; Knothe, K.: A study of the
geometrically nonlinear behaviour of plane curved beam
structures using a mixed hybrid finite element procedure.
ILR Mitt. 54, Berlin, 1978.
14 Karamanlidis, D.: Finite Elementmodelle zur numer1schen
Berechnung des geometrisch nichtlinearen Verhaltens ebener
Rahmentragwerke im unter- und Uberkrit1schen Bereich.
Fortschr. - Ber. VDI - Z 1/63, DUsseldorf, 1980.
15 Mast, St.; Karamanlidis, D.: Elastoplast1sche Berechnungen
von ebenen Rahmentragwerken nach einem gemischt-hybr1den
Finite-Element-Verfahren. DER STAHLBAU (in press).
16 Karamanlidis, D.; Knothe, K.: Geometr1sch nichtl1neare
Berechnung von ebenen Stabwerken auf der Grundlage eines
gem1scht-hybr1den F1nite-Element-Verfahrens. To appear
in: Ingenieur Archiv.
17 Washizu, K.: Variat10nal methods in elasticity and plasti-
C1ty, 2nd Edition. Pergamon Press, 1975.
18 Pian, T.H.H.: Variat10nal principles for incremental finite
element methods. J. Franklin Inst. 302 (1976) 473-488.
19 Stricklin, J.A.; Haisler, W.E.; von R1esemann, W.A.;
Evaluation of solution procedures for material and/or
geometric nonlinear structural analysis. AIAA J.11 (1973)
292-299.
20 Honecker, A.: Entwicklung, lmplementierung und Austestung
von Losungsalgorithmen fUr Gleichungssysteme mit singular
werdender Funktionalmatrix. Diploma Thesis, Berlin, 1980.
21 Bergan, P.G.; Horrigmoe, G.; Krakeland, B.; Soreide, T.:
Solution techniques for nonlinear finite element problems.
Int. J. num. Meth. Engng. 12 (1978) 1677-1696.
22 Bergan, P.G.; Holand, I.; Soreide, T.H.: Use of the current
st1ffness parameter 1n solution of nonlinear problems.
Energy Methods in Finite Element Analysis,ed1ted by
R. Glowinski, E.Y. Rodin and O.C. Zienkiewicz. John Wiley
& Sons Ltd., 265 - 282,1979.
 235

23 Bergan, P.G.: Solution algorithm s for nonlinear struc-

tural problems. Proc. Int. Conf. Engineering Application
of the f1nite Element Method, 13.1 - 13.38, Hovik!Norway,
1979.
24 Bergan, P.G.; Soreide, T.H.: Solution of Large displace-
ment and instabi11ty problems using the current st1ff-
ness parameter. Proc. Int. Conf. F1n1te Elements 1n Non-
linear So11d and Structural Mechanics, Geilo!Norway, 1977.
25 Remseth, S.N.; Holthe, K.; Bergan, P.G.; Holand, S.:
Tube buckling analysis by the finite element method.
Ib1dem.
26 Krakeland, B.: Non11near analysis of shells using degene-
rate isoparametric element. Ib1dem.
27 Wood, R.D.: Zienkiewicz, O.C.: Geometrically nonlinear
finite element analysis of beams, frames, arches and
axis symmetric shells. Compo & Struct. 7 (1977) 725-735.
28 Noor, A.K.; Peters, J.M.: Reduced basis technique for
nonlinear analysis of structures. AIAA J. 18 (1980)
455 - 462.
Geometrically Correct Formulations for Curved Finite Bar
Elements Under Large Deformations

K. BRINK" W. B. KRATZIG**
Ruhr-Universitat Bochum, Germany
Dr.-Ing., Hochtief AG Essen, formerly Research Group
Leader, Ruhr-University Bochum, Germany
*> Dr.-Ing., Professor of Structural Engineering,
Ruhr-University Bochum, Germany

Summary
Based on an energy-consistent, geometrically nonlinear theory
of rods several curved finlte bar elements are derived. Large
deformations and large rotations are considered in a kine-
matically rlgorous sense. The solution-procedure - a modified
Newton-Ra~hson-method within the modular software-system
FEMAS - demonstrates the excellent convergence of the derived
elements compared wlth standard examples of the literature.

1. Introduction
The state of the art of computer-hardware and computational
techniques enables the analysis of arbitrarily complicated
structural responses. The finite element method as the most
frequently applied approximative solution technique is often
combined with further (sometimes very tough) mechanlcal or
mathematical approximations. In this contribution the authors
intend to demonstrate the advantage and simplicity of the
use of approved mechanical modells and geometrically complete
formulations.

For the subsequent treatment we use the following assumptions

within the framework of the classical theory of curved rods
[6, 16]:

a) The thickness h of the rod is small compared wlth its

length.
 237

b) Stresses and stralns normal to the axis are neglected,

thus the cross-section preserves its shape during de-
formation.

c) Any straight normal to the undeformed axis wlll remain

stralght and retain its rectangularity during defor-
mation.

d) Homogeneous, isotropic and hyper-elastic material will

be assumed.

e) Only plane deformations are considered in this paper.

f) Self-equll1brating stresses over the cross-section and

pre-imperfections of the aX1S of the structure will not
be considered.

2. Basic nonlinear equations

Within these assumptions we now derive an energy-consistent

theory of plane curved rods subjected to arbitrary large
displacements and rotations. Due to Fig. , any bar element
will be described by the system of convected co-ordinates
z' in the axial and z3 in the normal direction. The unde-
o 0*
formed configuration P, P of the structure, fixed by the
spatlal co-ordinate system x', x 3 , shall be used as the ini-
tial state of reference for the future Total-Lagrange-For-
, 0
mulation. Introduclng for z the natural curve length s of
the undeformed state of reference we achieve a formulation,
o 0
which contains only physical tensor components: la,1 = la31
= ,.

In order to relate the base vectors of the deformed state

to those of the undeformed initial state of the structure
we start with the position vectors of any point PIP * ) lying
on (outside) the deformed axis (see Fig. '):
o
r = r + u r * (1)

Considering
238

u*
3
u + Z't!' ' Z = Z (2)

we find for the components of the displacement vectors

0 01 0 03 0
u = u°i a. = u e1 + u ~3 ,
-1.
01 01 0 03 03 0
u * = (u + zw ) ~1 + (u + zw ) .e3 (3 )
-
0
where the superscript marks the decompos1.tion with re-
spect to the initial state. With the help of

undeformed
initial state z
t
I

deformed state

spaftal system of cartesIan co-ordinates

curvtlmear convected aXIal co-ordinate
straIght convected normal co - ordinate
natural curve length of the undeformed aXIs

Fig. 1: Undeformed and deformed bar element

 239

o 0
o
~ a ~
~1 ;::'1 = ~1 (4 )
a z1 a s

we further find from (1) the relation

o
e1 = .el + ~, 1 ' ( 5)

which can be transcr~bed by virtue of Frenet's equations

0 0 0 0 0
b 11 .e3 )(
.e3
e1' 1
0 01 0 0 0
~3'1 -b 1 ~1 -)(
.el (6)

into the transformation for the base vector ~1:

a~1 =

(7 )

This expression holds for arbitrary large displacements. A

s~milar treatment for the normal base vector a 3 leads to
01 0 01 03 01
+ u b 11 o 1 + u '1 - u b 1 o
+
~3 ~1
y-;;;; .e3

03 01
a1 0 a1 0
+ (8 )
y-;;;; ~1
y-;;;; ~3

using the abbreviation:

01 2 03 2
all (a 1 ) + (a 1 ) (9 )

In order to evaluate the vector components of w for arbitrary

large rotations we decompose (see Fig. 1):
01 0 03 0 0 0
w = w a 1 + w a 3 = sin cp a 1 + (cos cp -1) a 3 • (10)

Using the orthogonal~ty condition

~1
. 0
e3 = (~1 + ~, 1 )
. 0
(~3 + ~)
0
~1
0
~3 0 ( 11 )
240

we flnd after some additlonal transformatlons for the exact

angle of rotation ~

l03 1 03
a1

l- j'
_ U'l + u°1 b°11
~ arc tan arc tan
01 03 01 01
1 + u '1 - u b 1 a1

~'1= a 11 [- 03 01
(a 1 ) , 1 a 1 +
03
a 1 (a011 ) , 1] (12 )

The adequate relations for small rotations (index L: linea-

03
rized values) take the form (w ~ 0) :

L L
~ ~ '1 (13 )

Green's well-known three-dimensional straln tensor

1 * * 0* 0*
El'J' = -2 (r ,l, ' r
~ ~
"] - r "l . r, ],), l,j = 1,2,3
~ ~
(14)

furnlshes wlth the assumptions made only one (axlal) component

Ell' whlch can be broken down into a constant part a 11 and
another one, varylng linearly over the thlckness h:

(15 )

By virtue of the previously derived formulae and under the

01 0 03 0
abbreviations u u, u w we find for the complete ex-
pressions of the 1. and 2. strain tensor:

0 0 2 2 0 0 2
0 1 0 1 0 1
a 11 u'1 - w 2" (u
x +
2" (u, 1)
x)
+
2" (w, 1) +

0 0 2 0 0 0 0 0 0
+
1
"2 (w x) - u'1 w x + U
w'1 x ,

o o o o
X + cos ~ (- X + ~'1 - u'1 + ~'1 u'l
------------------------------------------
o 0 o o
- ~, 1 w x) + sin ~ (w, 1 - ~'1 w'1
o 0

- ~, 1 u x). (16 )
 241

The approximatlon for small rotations changes the 2. strain

tensor into

- -
0 0 0 02 0 0 0

I3 SR
11 CP'1 u'1 It + w It + CP'1 u'1 CP, 1 w It

0 0 0
,
02
+ cpw'1 It + CPu It (17)

while the assumption of inflnitesimal small deformations

leads to the most far-reaching simplification (underllned
terms) .

Defining the hyper-elastlc potentlal energy function

(18 )

the global rate of energy equation, in which inertia terms

are neglected, can be broken down as [6, 9]:

D
Dt

(19 )

Carrying out the time derivatives in (19) and combining body

forces f with surface forces t into a general load vector p
this statement transforms into

/0
s
(EF a 11 a 11 + EI 13 11 ~ 11) d~
s
ds
o
(20)

where dotted quantlties represent material tlme derivatives

or increments.

With this functlonal we conclude our summary of non-Ilnear

equations as a basis of the lntended flnlte element approxi-
mation. For more complete informations the presentation [3]
may be recommended.
242

3. Incremental nonlinear relations

These just derived nonlinear equations can be used only in

iterative solution techniques. To include the ability of
incremental procedures we now define in Fig. 2 a neighbouring
- -*
position P, P of our previously considered deformed state P,
P * by adding an infinitesimal small state of deformation. This
neighbourlng state shall be reached from P, P * by a linear
step of computation. Under this assumption the lncremental
angle of rotation ~, superposed upon ~, has to be considered
sufficiently small, e.g.: Sln ~""~' cos (p""l.

deformed state
undeformed
imflal state

P,P
components of the
difference vectors
l:Y and !¥ for z = 7

Fig. 2: Initial state, deformed state and neighbouring

position
 243

With this linearization we find for the incremental differ-

ence vector ~ (Fig. 2), decomposed again with respect to
the initial state:
o o
w = ~ cos qJ ~1 + ~ sin qJ ~3 (21 )

From the orthogonality condition, appl~ed to the neighbour-

ing position
o 0 0 0

(e1 + ~'1 + ~'1) • (e3 + ~ + !) = 0 , (22)

a series of transformations similar to those following (11)

for the incremental rotation leads to:

~
a 11 [- 0 ~ 0 01
(w'1 + u x) a 1 +
0 0 0 03
(u'1 - w x) a 1 ]'
(j) , 1
a 11 [- 0 ~ 0 01
(w'1 + u x) , 1 a 1 -
0 ~ 0 01
(W'1 + u x) a 1 ' 1

0 ~ 0 03 0 ~ 0 03
+ (u'1 - w x) , 1 a 1 + (u'1 -w x) a 1 ' 1 J
[ ]
0 0 0 01 0 0 o 03
a 11 ' 1
---2 (w'1 + u x) a 1 + (u'1 -w x) a 1
(a 11 )
(23)
o
Again we have introduced the abbreviations
~1
u
0
u, u
~3
w.
The incremental d~splacement vectors Q, ~ are defined by
the difference of the position vectors -*
!, rand (1):
0 0
r = r + u + u
'V
,
0* * 0*
r* r + u
*
+ u r + u + U + z (~+!) (24)

Continuing with the incremental strain tensor (14)

(25)

and the corresponding strain energy increment

244

.
+
(a ++
11 + a 11 )
s
(a 11
o
+ i~,)J ds , (26)

1n which all products up to quadratic nonlinear terms are

considered consistently:

(27)

we can find after considerable evaluations all basic equations

of the neighbour1ng position, related to and decomposed with
respect to the undeformed initial state, in Table 1. This
consistently linearized set of equations forms the fundament
of the future finite translation.

The global rate of energy equation in Table 1 co~tains on

the left hand s1de - +
in the terms a'1 a+11 , P"
;I;- *P11 - the
contribut1ons of the incremental and orig1nal displacements,
in the terms a 11 g~1' ~11 ~~1 the contributions of the 1nitial
strains and the quadratic d1splacement-increments. The r1ght
hand s1de contains the energy rates of the load-increments
and - emphazised by underlin1ng - the rate of energy equation
of the original deformed state. In any state of equilibrium
these underlined terms have to vanish; the remain1ng state-
ment would correspond with a purely incremental solut1on
technique. During a certain solution process, if equilibrium
has been reached not yet, they can be interpreted as unbalanced
external of internal forces. This leads to the advantage of a
very efficient coupling of incremental and iterative solut1on
techniques.
 245

Incremental global rate of energy equatIon (homogeneou

hyper-elastIc material, conservatIve loads)

.J
+'+ +.:. +'+ + .:.
[EF(Ci. 7? Ci. 77 + Ci. 11 Ci. 77 )+EI((J11 (317 + (311 (311)J ds =
5 o. • • +
J[(p' +/,') 6, + (m + In) ip - EFCi. i1 all - EI (371 (311 Jds
5 - - (, :;: 1 3, U, = U I U3 = w)

KinematIc equatIons for Incremental strain vartables

for large rotatIOns

o 0 _ 0 03 0 1)
+ Vi,1 Sin tp(x- tp-t) + tp (x -tp'l }(COS tp 01 +S", tp 01

+ CP'1 (COS tpot -Sin tp aT)

xcos tp (x - tp'I) - 6 ip'l XSin tp + 6, ip Sin tp ( x. -tp"
++
(317 = 6 ip 1 )

+ 6, 1 rp'l COS tp - wrp X. Sin tp (x. - tp'l ) - W ip,r xCOS ip

+ wip COS ip (x - tp'I)- W'I CP'1 Sin ip

for small rotatIOns

+ 2. 02 2. 0 ~ 0 0 1.2 0

(317 = U ip X - U ,7 (x - tp,: ) + W X (x - ip" J + W,1 tp X

CoeffICIents of transformatIOn
a/ = 1 + (;,1 - WX
0"3
= W'1 x
0 0'

1 + U

Table 1: Incremental basic equations of the neighbouring

posltion, related to the initial state
246

4. Matrix formulation of basic equations

For the numerical treatment we select as primary kinematic
o
variable the vector y of the complete incremental displace-
ments, given in the first line of Table 2. This vector
,00
contains the displacements a, w as independent quant1ties,
o
the rotat1on ~ and all first derivatives as dependent ones.
Addit10nally Table 2 shows the linear parts of the incremen-
tal strain-displacement relations; the non-linear parts may
be found in Table 3. In both cases small and arbitrary large
rotations are distinguished. For details of the evaluation
procedure of these matrices and in connect1on w1th the1r
numer1cal computation see [3]. The specific d1scretizing
prozess of the incremental matr1x equations is similar to
that one of any linear theory and may therefore be passed
over here.

5. C1rcular bar elements

In order to check the derived incremental theory of curved
rods, arbitrary arches are - to a large extend - the most
suitable slmple structures. Their ability to b1furcation as
well as to snaps-through combines two completely d1fferent
non-linear response phenomena. The encouraging results of
our start, a 2-node circular bar element with a total of
6 DOF, finally led to the systemat1c development of several
element famil1es [8], WhlCh will be reviewed briefly.

In the case of (straight) beam elements the interpolation

functions for the axial and normal displacements in general
can be chosen eas11y such that r1gid body modes are descri-
bed properly and will cause no stra1ns. For curved bar ele-
ments, e.g. circular bar elements, both requirements are
difficult to fullfill. Three different ways are known to
solve this problem:

a) A set of n functions, which descr1be the rigid body

modes correctly, 1S superposed upon the m original in-
 247

Complete mcremental dIsplacement vanables

y = { 6 I 6,1 I ~ I ~., I rp I CP. I }

Matnx of mcremental strams

_ [all]
-e: =
(3- = e:
-
.,.
II

Lmeanzed mcremental stram-dIsplacement relation.

+ 0 0 0

g = (Q .,. Qv ) Q

Operator matnces.

Q = [+H 1-:, 1~ I ~ I j 0

for large rotatIons

~
, xO
"3 l 0'I1 -1 X(1-0
" 11 ) 0'31 0 0 ~
Qv = 0 0 0 0 0 0 0 0 01
X rx-cp.,J x-fX-tp,.) x rX-C{J.I) (x-'P.,) (x -ip'l) 0lcos!p-1
sm!p cos!p cos!p-F sm!p (alcos!p.,.a!sm!p) -atsm!p

Qv =
[..,
for small rotatIOns
x

ip
01

ic 2
oJ -1
C{J.,
x(1-0;)

-1fJ•• X
'3
0 1

epic x'
0

01
'3 'I
o ]
0 1 -1

Table 2: L~near incremental strain-displacement relations,

related to the initial state
248

Non-Imear mcremental strom-dIsplacement relatIon

r [Os ~ ] [Os ~ ] I r [Os Q]
= T = 2 T [Q5 ~ ]
Non-linear operator matrtx.
for large rotahons
r-
1 • 1 1 • I
1 -
0
2 x 0
"2 "2(X - IP.I} coscp -2 Stn cp
1 1•
0
'2 -'2 x 0 ; (ic -CP'I) Stn cP ; coscp
Qs =
0 1
'2
1 x'
-'2 0
1 •
'2(x -CP,I}Stn cp j coscp
-x
1 ' 1 1· }
,-2
0 0 '2 '2 (x -CP'1 cos cp - ; Stn cp _

for small rotatIons

1
-
o o 2 o
o 1
2 o o 1
2
1 1
o '2 o o '2
1
o o 2 o
-
Matrtx product Q: Qs

, T'
~
I-f-
I
if
_0
I
ill
I 6(3-
1-0
I '
Qs Qs =
---+-
1-~6(1
I
kt--I'

o-J
I
I =Qo

L -Q
I
I-
I i
0Cf.
Qo . Matrtx for mltlal stresses due to Cl11
Q(1.. Matrtx for mltlal stresses due to (311
o neglected sub- matrix (elemenrs depend on Z2)
Table 3: Non-l~near ~ncremental strain-displacement relations,
related to the initial state
 249

terpolatlon functions. The (m+n) free parameters then

are determlned such that boundary as well as inter-
mediate conditions are fullfilled and rigid body mo-
tions do not lead to strains.

b) The requirement of strainless rigld body motions

transfers the (llnearized) kinematic equations for
~11' ~11 in Table 1 into a homogeneous differentlal
equation of order 2. The solutions describe rigid
body modes; they have to be completed by additional
interpolation functlons, which satisfy the inhomo-
geneous klnematic relations.

c) The third way is the separation of the displacement

fleld into rlgld body modes and strain-producing
displacements [14].

Finally, for the interpolatlon functions we may select arbl-

trary polynomials or Hermite-polynominals; the latter safe
a time-consuming matrix inversion.

The interested reader may find 13 different clrcular bar

elements in [8], further more in [3]. In the present paper
we sketch the evaluation of one of the best elements contai-
ning 5 nodes and 12 OOFs (see Fig. 3). The requirement of
vanishing strains transforms the (linearized) klnematlc
equatlons (16)
o 0 0
o - w )(
o o 0 02
13 11 = 0 = CP'1 - u'1 )( + w )( (28)

under consideration of (23)

o 0 0

cP, 1 = -w, 1 - u )( (29)

o
and for constant (circular) curvature )( = -1/R into the
harmonic differential equation:

o 1 o
w'11 + 2 W o (30)
R
250

CIrcular bar- element

wtfh 5 nodes

Vector of nodal dIsplacement

~=I~I~I~I~I~I~I~I~I~I~I%I~1
Vector of free parameters
Q=I~I~I~I~I~I~I~I~I~I%I~I~1
Boundary conditIons and intermedIate conditIons
I/I,W,U for z'=O and z'=I, w,u for z'= 1' z'=f and z'= il

Fig. 3: Circular bar element with 2 external and 3 internal

nodes

As a first solution of (30)

Os
w (z 1 ) z1
a 1 sin R+ a 2 cos z
1
R
Os z 1 1
u (z 1 ) a 1 cos R - a 2 sin ~+ a 3R
R
s (z 1) a3 (31 )
ql

we confirm the rigid body modes found by [ 1 3] . A sllghtly

extended solution of (30) is formed by:
Os
w (z 1 ) a 1 sin ~+
1
z1
a 2 cos R
z1
R a 3 R sin R
1 1
Os
u (z 1) z
a 1 cos R - a 2 sin ~+ z1
a 3 R( 1 - cos R-)'
R

ql
s (z 1 ) a3 . (32)

Furthermore, as interpolation functions for the strain

variables a 11 , ~11 we select polynomials of order 3 and 4
with a total of 9 free parameters. Assuming trial solutions
° w° their undetermined constants can be evaluated by
for u,
comparlson from the corresponding relations of Table 1.
After suitable comblnatlons of all parameters we end up
with the lnterpolation functions of Table 4. Based on the
vector of the complete incremental displacement varlables in
Table 2, the edge and intermediate conditions of Flg. 3 we
 251

conclude this brief lnformation with the baslc matrices ~,

~ in Table 5. For further detalls concernlng the derived

elements see [3, 8].

Interpolation functions descnbmg ngld body modes

1 ZI ZI ZI )
US ( z) = a, cos R -a2 S/n R +a3 R (1-cOS1f

S( 1)
W Z = a, Sin RZI + a2 COS RZI - a3 R Sin RZI
tpS (ZI)

Interpolation functions descnblng strains

u£(z') = a L +a (ZI)2 +a (Z'~ (Zl)' (ZI)S
8 R 9"2fi7 I03R +°,1 4R' +0 125fi5
£ , ZI (ZI)2 (ZI)3 (z ')'
w(z)=o, +oSR+ 06 R2 +°7 ~ -0'2~
£
tp (z
')
=-05 R1 -°6 2WZl -°7 fi3
3(ZIJ2
+°8 W
Zl
+ °9
(ZI)2
2R3
k!f.... (ZI) , (4(ZI)3 (ZI)5)
+a'0 3R' • 01/ ~ + °,2 ~ + --sfi6
Corresponding strains for the
5 node circular bar - element
• 1 1 1_,,2 I Z I)3
au =(a,.09)-+(OS+09)~ +(06+01O)~3 +(07+ 0 11)T
R ~ R R
• 1 ZI (ZI)2
{311 = (0, - 206. 208) R2 .(05- 60 7 .209 )fj3 .(06 +2010+12012)7

(ZI)3 (Zl)'
• (°7 • 20,1 )RS" + 012 fi6
Combined mterbolatlon functions
for element displacements
u (Zl) = US(ZI) + U£(Z')

W (Zl) = WS(ZI) + u€(z')

tp (Zl) = tpS(ZI) + tpe(zl)

Table 4: Interpolation functlons for circular bar element

252

cos tf Sin Ii R(1-cos tf) z

R
(Z')l
2R'
(z'j3
3R'
(z')~

4R'
(z't
SR'

~~nfi -~cos~ smf R

1
R"
z'
rz£
R
~
R' 14-
R

Sin f COSf -Rsm ~' 1

R
Z' (ZI)2
--;;r Ifr
(Z"J
-7
IZ')1.

ficosfi -irsln~ -COSfr" R

1 iz'
R' ¥R _t.Jd
R'

1 _L
R
2z'
-7 -1£4-
R 7
z' fz!):!
2i'? 1!t
3R ~
4R
!.Jii'..k't
Rio SRi>

_...L 6z 1
-fj'f
1
qr
z'
fj'f
1ft. ~ ~.1?J:-
R' R' R R' R

COS-fR
5 5 5' 5' 5'
5m- -Rsm :R
4R --:R 16R 64fiT - 256R'

cosfR -SlfJiR Rf1-cosm 5

7;R
5'
J2R! ~
5' 5'
~
5'
~12aRS

-fR
cp" =
2- 5' 5' 5'
Sin2~ COS -RslO ;R 2R 4R'" 8R' -76fT
cos~
2R -srniR RO-cosiR 5
2R
5'
8R'"
5'
Tt:fiT
5'
6i:iF"
5'
160R 5

35 95' 275 1 81!,-

Sin ]5
4R
C05&
4R -RSln :~ t:R 16fiT 6ifiT - 256R-

COS}} -sln~ 1m-cos !J? 35

ZR
95'
J2R1
275 1
mAT 1ff2ZRL
87 s~ 21.3s 5
512DRf

1 _.2.l 35' 5 5' 5' 5' 45' 5'

-R R -7 ~2 7fil 3R' 4.' -,:;z-.5fi6
-tr
sin cos It- -Rsm-} 5
71- V
5' $ ,
(If" Ii'
5'

C05+ -bin -f> RfI-wsi, 2-

R
5'
:?R'
5'
JRJ
5'
~
5'
5ii'

Table 5: Additlonal matrices for circular bar element

6. Software implementation and examples

The previously descrlbed basic incremental relatlons have

been implemented into the FEMAS-system [2] together with
several circular bar elements forming a combined incre-
mental-iterative solution procedure. FEr~S, an in-core
 253

modular software system, has been designed for the handling

of banded, symmetric matrlces; It has been proved to be a
highly efficlent tool for the present purpose.

It should be emphasized, that the internal control of any

future load-step in FE~~S has been carr led out automatically
- for every state of equilibrium - by the computed increment
of the mechanlcal work. We could show, that this physlcal
quantity permlts a far better estlmate of the future stiff-
ness of the structure than a certaln displacement component:
the computed work-increment controls the magnitude of the
next load-lncrement. By thls option different load-lncrements
easlly and automatically can be chosen as the state of de-
formation demands.

Finally, two often cited examples of the literature will be

recomputed and their results compared. The flrst one, given
in Fig. 4, shows a very shallow circular arch, clamped at
both ends and loaded by a point-load P In the apex. The struc-
ture deformes continuously into a symmetric, stable post-
buckling configuration. During this path very small rotations
appear, thus small and large rotatlon theories lead to
identlcal results. Our computation has been carried out with
only 2 bar elements for one half of the structure. Comparable
results were obtained with 4 3-dimensional elements [10],
12 8-node plane stress elements [1, 11]and 4 to 32 beam
elements [4] as shown in Fig. 4. Introducing a small imper-
fectlon pattern of 0.05 In of amplltude also the experimental
result of [5] could be verified.

The second example - astee~ pin-jointed circular arch

bearing again a pOint-load P in the apex - shows an un-
stable symmetric bifurcation: The stable post-buckling con-
figuration appears as the first unsymmetric buckling mode,
reached at a considerable lower limlt load. This structure
(Fig. 5) has been lnvestigated analytically in [7] and
numerically by [12, 15], both using 20 plane stress ele-
ments. In order to force the iteration lnto the unsymmetric
254

P[/b}

40+--~-+

30-t---

20+-~~r---~-----+----~

10 +---H--+- ~=----.-

O.-------~--------~I--------~I--------r-~I
0,0 0,1 0,2 0,3

R = 133.114 in (3 = 73397°
h =0.1875 In E = 107pSI
b= 1 In A = (32 ~ = 1165

Fig. 4: Shallow clamped arch

buckling path, a horizontal perturbation load P/1000 has

been applied successfully.

Using the same remedy in our computation the analytical so-

lut10n could be verified with 6 bar elements and 60 load-
increments. As Fig. 5 demonstrates for this highly non-
llnear response any small rotation theory, lead1ng to an
error-margin of about 40%, 1S completely 1nsufficient. These
and many further results prove the applicabillty of the
 255

w [15i
16+--
own solutlonJ7l.
large rotatIOn theor.

0 I
[In}
0 6 12 18 24wo
0 03 06 09 12 U o
R =100 in =80 In
L
h= 11n=b (J =5313°
H = 40 In E =107 pSI

Fig. 5: Steep pin-jolnted arch

derlved theoretlcal concept and its software lmplementation.

Acknowledgement: Conceptual ideas and substantial support

in the FEMAS implementation of our colleagues Beate Lecht-
leitner and Hermann Beem are gratefully acknowledged.
256

7. References

[1] Bathe, K.J. / Ozdemlr, H. / Wilson, E.L.: Static and

Dynamic Geometric and Material Nonlinear Analysis.
UC-SESM Report Nr. 74-4, UC-Berkeley 1974.
2] Beem, H. / Brink, K. / Kratzig, W.B.: FEMAS - Benutzer-
anweisung und Programmdokumentatlon (in print) .
3] Brink, K.: Theorie und Berechnung in einer Ebene ge-
krummter Stabe unter groBen Verformungen. Institut
fur Konstr. Ingenieurbau, Ruhr-University Bochum,
Techn. Rep. No. 79-1.
[4] Dupuis, G.A. et al.: Nonlinear Material and Geometrlc
Behavior of Shell Structures. Compo a. Struct. 1
(1971) 223-239.
[5] Gjelsvik, A. / Bodner, S.R.: The Energy Crlterion and
Snap Buckllng of Arches. Journ. Eng. Mech. Div. (1962)
87-134.
[6] Green, A.E. / Naghdi, P.M.: Non-isothermal Theory of
Rods, Plates and Shells. Int. J. Sol. Struct. 6 (1970)
209-244.
[7] Huddleston, J.V.: Flnlte Deflection and Snap-Through
of Hlgh Circular Arches. Journ. Appl. Mech. (1968)
763-769.
[8] Kratzlg, W.B. et al.: Elementbibliothek, Lehrstuhl III
des Instituts fur Konstruktlven Ingenleurbau, Ruhr-
Universltat Bochum.
[9] Kratzig, W.B.: Elnfuhrung In die Thermodynamik der
Deformationen I. Seminarbericht Nr. 73-1, Lehrstuhl
fur Baumechanlk, T.U. Hannover, 95-115.
[10] Mallet, R.H. / Berke, L.: Automated Method for the
Large Deflectlon and Instability Analysls of 3-Dlmen-
sional Truss and Frame Assemblles. AFFDL-TR-66-102
(1966) .
[11] Papenhausen, v.: Elne energlegerechte, inkrementelle
Formulierung der geometrisch nlchtlinearen Theorie
elastischer Kontinua. Inst. f. Konstr. Ingenleurbau,
Ruhr-University Bochum, Techn. Rep. No. 75-13.
[12] Ramm, E.: Geometrlsch nlchtlineare Elastostatlk und
finite Elemente. Habilltatlonsschrlft Stuttgart 1976.
[13] Sabir, A.B. / Lock, A.C.: Large Deflectlon, Geometrl-
cally Nonlinear Finlte Element Analysis of Clrcular
Arches. Int. J. Mech. ScL, 15 (1973),43-47.
[14] Schrader, K.-H.: Genaulngkeitsprobleme bel Ansatzen
mit Starrkorperverschlebungen. Finite Elemente In
der Baupraxis, Verlag W. Ernst & Sohn, Berlin 1978,
231-245.
[15] Sharifi, P. / Popov, E.P.: Nonlinear Buckllng Ana-
lysls of Sandwich Arches. Journ. Eng. Mech. Div. (1971)
1397-1412.
[16] Wunderlich, W. / Beverungen, G.: Geometrlsch nlcht-
llneare Theorie und Berechnung eben gekrummter Stabe.
Bauingenieur 52 (1977) 225-237.
Part ill:
Physically Nonlinear
Problems
Some Nonlinear Problems of Soil Statics and Dynamics

O. C. ZIENKIEWICZ
UmverSIty College of Swansea, U K.

ThlS brlef presentatlon dlscusses the approprlate formulatlon for dynamlc.

quasl-stat1c and statlc problems In sOlI mechanlcs. Some constltutlve relatlons
of plastlc1ty are presented and numerlcal computatlon lllustrated by an example
of an earth dam subject to earthquake forces.

Introduct1on
SOlI. and lndeed rock and concrete are porous mater1als wlth the vOlds generally
fllled wlth water WhlCh exerts a pressure on the SOlld partlcles and mlgrates
through the materlal. Any numerlcal computatlon must therefore allow for the
essent1ally two phase behavlour - and on occaSlon when the saturatlon lS not
complete for a posslble three phase sltuatlon.
In thlS paper we shall conf1ne our attentlon to the purely two phase
sltuatlon and dlSCUSS the formulatlon necessary to obtaln the statlc or dynamlc
response of 'solI structures' up to the pOlnt of collapse. Practlcal problems
to WhlCh the numerlcal Solut1on must be appllcable lnvolve such tYPlcal sltuatlons
as the settlement of bUlldlngs. lncremental collapse of offshore platform structures
due to the actlon of sea waves or the posslble earthquake collapse of earth dams or
nuclear power stat10n foundat1ons. For many problems In such categorles qUlte
large d1splacements and local stralns can be tolerated so approprlate cons1derat1on
has to be glven to both materlal and geometrlc nonllnearlty.

The effectlve stress concept

If we deflne the total. Cauchy. stress In terms of all the forces actlng on an
lsolated portwn of the sohd-fluld ensable as a,) • then the "effectlve"
stress lS deflned as

(1 )

where the symbol p stands for the fluld pressure 1n the pores of the SOlld skeleton

In above the usual conventlon of posltlve tenslle stresses lS malnta1ned

although In sOlI pract1ce It 1S often useful to reverse th1S.
260

The ~dea of using the separat~on g~ven ~n (1) stems from Terzhag~ [1]
who observed that un~form pressure changes alone could not deform the sol~d

mater~al apprec~ably or lead to ~ts failure. The only not~ceable effect of such
pressure change~ can be to cause a volumetr~c stra~n change assoc~ated with the
compress~on of solid gra~ns - th~s we can wr~te as

d~=-dp/KS (2)

where Ks ~s the average bulK modulus of the sol~d phase.

{he volumetric stra~n ~ is usually neglected in analysis of so~ls but
may be of ~mportance ~n rocK or concrete mechanics where the total stra~ns of the
sol~d sKeleton are small. We shall thus reta~n it ~n subsequent derivations.
W~th the very lim~ted effect of un~form pressure changes ~t is qu~te clear
that the major deformat~ons are associated w~th the effect~ve stress o'
Whatever the spec~f~c form of the const~tut~ve law we can write the incremental
(or rate) relat~onsh~p as

(3)

where the f~rst term corresponds to the constitut~ve law assoc~ated w~th the Jaumann,
corrotat~onal, stress increment and the second term accounts for effect~ve stress
changes due to 'rig~d body' rotat~on of an element of volume. In above dE~
stands for any '~n~t~al' or creep stra~n ~ncrements and

p _ 1 P
dE'1 -3 dE,i (4)

The stra~n and 'sp~n' ~ncrements, dE;1 , are def~ned ~n the usual
manner as

dE'1 • .1. (du +du ) ( 5)

2 I,J hi

and

(6)

~n wh~ch dUj are the d~splacement ~ncrements.

At this stage we should note that qu~te generally the tangent modulus
D'Jkl w~ll depend on the current effective stress levels g' ,accumulated
plast~c stra~ns or other h~story or damage parameters and ~n the case of an~sotrop~c

mater~als on the or~entat~on of the mater~al elements. In the last case it is

necessary ~n computat~on to Keep account of th~s or~entat~on by a su~table

~ntegrat~on of elementary rotations d2

 261

Equillbrium relatlons
Following Blot we shall use the vector w, to denote the mean displacement of the
fluld relatlve to the solld skeleton. Thls lS defined by taKlng the total quantlty
of fluld and dlvldlng by the total cross sectlon. Thus the average dlsplacement
in pores lS W,I n where n lS the poroslty.
Wlth thls we can wrlte the total equlllbrlum of the solld-fluld volume as

a"d + P9, = Pu, + Pfw , (7)

Wl th 9, belng body force acceleratlon. ~ fluld denslty and P denslty of the

total 'mlxture'.
To above equatlon the fluld equlllbrlum condltions have to be added. Here
the essential effect lS that of the VlSCOUS forces exerted by the solld boundarles
on the flowing fluld. These can be deflned In terms of a permeablllty tensor. ~j

and we can wrlte

(8)
- P"

where ~~ are the 'reslstlvlty' coefflclents whlch are the lnverse of the
permeabillty matrlx.

Mass balance
The flnal equatlon coupllng the behavlour of two phase materlals lS that of mass
balance. Clearly durlng strainlng of the solld matrlx the volume of pores changes
and the resultlng change has to be compensated by the outflow and compresslbllity
of the fluid.
. .p
= - P n IK f - Ell (1 - n I (9)

In the above equatlon we have neglected a very small term arlslng due to the change
of pore shapes when the materlal lS subject to pure effectlve stress - assumlng thus
that durlng such actlon the solid gralns are essentlally lncompresslble. The
lncluslon of thls term In a slmllar framework of analysls lS posslble as shown by
Zlenklewicz et al [4] but represents a negllglble effect.
Alternatlve derlvatlons of relatlonshlps governlng the two phase behavlour
are posslble using a mlxture theory. Some examples of such formulatlons are
avallable In llterature [5] - generally for llnear constltutlve relatlons - but
except for ultra-hlgh frequency phenomena In practlce these reduce to the above
format.
262

An approxlmatlon
Before attemptlng a flnlte element solutlon of the equatlons 1-8 It lS convenlent
but not essentlal to lntroduce an approxlmatlon. In thls we take the effect of
the addltwnal fluld acceleratwn wil as negllglble. Such an approxlmatwn allows
us to use ul and p as prlmary varlables rather than ul and WI thus reduclng In
general the number of unknowns. Wlth that approxlmatlon It lS easy to ellminate
W from (8) and (9) glvlng now In thelr place the well known equatlon of flow
In porous medla

(10)

The above equatlon together wlth (1) through (7) In whlch effects of WI have been
omltted governs the solutlon for all statlc and dynamlc problems of sOlI/rock or
concrete behavlour.
Though the above formulatlon dlsregards one of the dynamlc effects It lS
appllcable for most englneerlng dynamlcs sltuatlons assoclated wlth sOlI behavlour
even If permeablllty values are relatlvely hlgh and the frequencles are of the
order of those occurrlng In earthquake analysls.
In Flg. 2 we show results of a study on a Ilnear problem of a sOlI layer In
whlch the zone of analysls In whlch the approxlmatlon lS not valld lS lndlcated [6]
In terms of two non dlmenslonal parameters.
The same flgure lndlcates zones In whlch the effects of permeablllty are
of no lmportance l.e. the zone of undralned behavlour; and a zone In whlch dynamlc
effects can be totally dlsregarded l.e. the zone of consolldatlon behavlour.

Large dlsplacements and deformatlon

The baSlC equatlons presume that the dlfference between the natural coordlnates of
a materlal pOlnt l.e. ~ =~ + ul and those spatlal coordlnates XI lS such that

.Q.L = 1
a~ -
If dlsplacements are large thls assumptlon lS stlll valld provldlng an updatlng
of the spatlal coordlnates is carr led out In reasonably small lncrements of dlS-
placement ul Computatlon procedures whlch we shall now descrlbe can follow such a
pattern and the changes of domaln are taken care of automatlcally.

Solutlon of dynamlc and quasl-statlc problems

As In purely statlc sltuatlons the coupllng of the two phases dlsappears and the
fluld flow and solld equatlons can be solved separately there lS llttle to be sald
about such solutlons for whlch any of the 'standard' flnlte element procedures
 263

avallable can be used. Interest therefore focusses on dynamlc problems of the

klnd lnvolved In earthquake analysls of foundations or In a speclal case of
problems In WhlCh dynamlc effects are negllglble and the translent nature of the
phenomena 1S due entlrely to the seepage of the flu1d. It 1S on such classes of
problems that the methodology expounded here 1S focussed.
Proceedlng 1n a standard manner of f1n1te element d1scretlsatlon for d1S-
placements U, and pressures p we can wr1te [now uS1ng a vector1al notat1on)

.!:! = .!:!.iI [ 111

where Nand f:! stand for appropr1ate shape funct10ns and'y and E. for nodal
quant1t1es. In above
1/ =( u • v • w )T

where u,v,w are the Carteslan dlsplac8ment components. Eq.7 can now be wr1tten
uS1ng standard Galerk1n methods and Subst1tut1ng [1) as

J§'"g'dQ - 9i? + ~~ + t =0 [12 )

\I
where M = [/'lPNdQ
1S the usuaVmass matr1x, Q = ~TT ~ dQ
and f the force vector lnclud1ng body forces p~ and boundary tract10ns [7]. In
above ~ 1S the standard small stra1n matr1x ar1slng from [5).
Slm1larly [10) 1S d1scret1sed to glve

MIi =9 [13 )

where
lj = k~ ~ )T ~ Y~ dQ

?= b &T(n/Kf +ll-n )/Ks) ~ dQ

b
~ = Ef:t~Pft:idQ
In all of the above equat10ns Q represents the Current domain and lf
d1splacements are large an updat1ng of coord1nates 1S necessary at each step of
the computat1on.
Equat10ns [12) and [13) are supplemented by the const1tut1ve relat1onsh1p
[3J and an appropr1ate def1n1t1on of the sp1n vector of [6J. In the preVlOUS
express10ns the "standard" stra1n matr1ces were used. Now we have varlOUS
alternat1ves and the wrlter prefers to retaln the notat1on consistent w1th wr1t1ng
the stresses 1n a vector form. We thus deflne follow1ng Nayak [In 3 d1mens1ons)
264

o o o 0 -Wy .Wz

o o o ,Wx 0 - Wz

o o o -Wx .W y 0
0 -~Wz + rWy
-twz 0 -1Wx
-tWy -twx 0
(14 )

where the angular velocltles are

(15)

reducing to

( 16)

in 2 dlmenslonal plane problems. Now the stress lncrement in vectorial form is

wrltten as

dQ' (17)

'!IT = (1,1 .1 .0 .0 ,0 l
The above equatlons wlth a speclfled constltut1ve relat10n complete the
d1scret1sed Solut1on Wh1Ch has to be solved now 1n the time doma1n.

T1me stepp1ng algor1thms

The nonllnear system of ord1nary d1fferent1al equat10ns (12) to (17) can be solved
by a var1ety of algor1thms. However 1n all 1t 1S convenlent to use some form of
stagger1ng or part1t1on1ng to avo1d very large equat10n systems. In such staggered
Solut1ons one set of equat10ns 1S stepped forward at a tlme uSlng extrapolated
values of the second var1able. Thus for 1nstance we can apply a central d1fference
 265

algorithm to the dynamlc equatlon (12). Now we have for the n-th step

(18 )

from whlch ~ n.l can be found expllclty if a lumped matrlx form lS used and
Q~ • en .~n , ~-1 etc. are known.
The stress lncrement can be evaluated uSlng eqs. 16 and 17. Thus first
we compute ~B uSlng a difference form and then flnd

(19 )

(Note that In above calculatlons we have used a slmple Euler rule of forward
lntegratlon although wlth llttle addltlonal effort average values of Q~V2,~~1/2

etc. could have been lnserted.

lS known the pore pressure changes can be found uSlng an
lndependent algorlthm on the seepage flow equatlon lnsertlng approprlate approxl-
mahons for g and Q
Qn.l =(!:In.l - iin )/~t

(20)

We choose here to use an unconditwnal stable algorlthm lI2<e..; 1 (generally

with the upper value)
o

From this the new set En.l lS easily obtalned and computation can proceed
further.
Stabllity of such staggered algorlthms has recently been lnvestlgated
thoroughly by Park-Felllpa.
In th~present case we note that
(a) as the permeablllty lncreases l.e. ~_CD the equatwn (13) becomes
approxlmate

and if further the frequencles lnvolved are such that ~y =0 then

uncoupling of the two systems occurs. In such a case the stablllty of
computatlon of the explicit scheme depends on the wave speed In the deslred
medium
and
266

(b) as the permeab~l~ty decreases i.e. ~- 0 then the seepage equat~on reduces
to • T.
?1!+9~=O

Now we can evaluate ~ as

and elim~nating th~s from eq. 12 results ~n a st~ffness term

from which we conclude that stabil~ty ~s governed by undra~ned cond~t~ons.

As generally the rat~o of 'dra~ned' to 'undra~ned' wave speeds ~s of the

order of 1/10. the advantages of us~ng an undra~ned assumpt~on d~sappear.

In pract~cal computat~on we generally f~nd that for cond~t~ons not

correspond~ng to the above l~mit~ng cases an ~ntermed~ate assumpt~on for cr~t~cal

~t needs to be used. A study is now be~ng completed on the cr~t~cal stab~l~ty

of the comb~ned algor~thm.

The expl~c~t-~mplic~t. staggered scheme Just outl~ned ~s advantageous and

practical for relatively short t~me span phenomena such as explos~ons or earth-
quakes. For longer term computat~on such as those involving consol~dat~on or
wave loads on offshore structures. impl~c~t-~mplicit schemes must be used. We
shall not d~scuss the detail of such schemes except to ment~on that now. ~f large
d~splacement form of equat~ons ~s to be used. the computat~on w~ll ~nvolve the
calculatlon of the der~vat~ves

",0_
uY Q
dIlg' dn I
These yield tangent st~ffness matr~ces due to geometr~c (~n~tial stress) effects
and deta~ls of such calculat~ons can be found elsewhere.
It should be noted that here generally one determinat~on of such st~ffness

matr~ces w~ll only be necessary ~f quas~-Newton (B.F.G.S.) methods are used.

Const~tut~ve law
The br~ef scope of th~s paper permits only a very brief d~scuss~on of the
const~tut~ve relat~ons appl~cable to so~ls under dynam~c or quas~-stat~c cond~t~ons.

The bas~c models used generally fall into one or another of the two general
classes:
(1) elasto-plast~c or elasto-v~scoplast~c

(2) endochron~c.

For statlc or quasi-stat~c monoton~c load~ng problems ~t ~s generally conceded that

models of the f~rst category are more eff~c~ent requ~ring the least number of
parameters for the~r descr~ption. Here the best models are those of the "cr~t~cal
 267

state" family developed by Roscoe in WhlCh a hardenlng process occurs wlth

lncreaslng plastlc compactlon. However, quite adequate models are those wlth a
non assoclatlve ldeal elasto plastlc behavlour. Here a Yleld surface correspondlng
wlth the Mohr-Coulomb failure lS usually used.
A survey of some models of thlS type has recently been presented by the
author and thelr performance compared on some typlcal problems.
The need to examlne the behavlour under load reversals and to lncorporate
rate effects has caused much effort to be devoted to the subject recently. In a
conference devoted to the subject of behavlour of sOlIs under translent and CyC11C
load [12] varlOUS approaches have been analysed.
It appears that once agaln elasto-plastlc models alone Dr In comblnatlon
wlth a "deterloratlon parameter" of an endochronlc type are most advantageous. In
the flrst the concepts of anlsotroplc klnematlc hardenlng of the type flrst lntro-
duced by Mroz [13] comblned wlth plastlc modulus lnterpolatlon suggested by Popov
and Defallas [14] are effectlve [15]. Here the crltlcal state eillpse lS used as
a "boundlng" surface wlth an lnterlor eillpse of smaller Slze representlng a
klnematlc Yleld surface.
Wlth such models the phenomenon of "denslflcatlon" WhlCh occurs In cycllcally
loaded sOlIs and whlch lS responslble for pore pressure rlses and occaslonally
llquefactlon of the sOlI can be modelled. Flg. 3 shows such a model In the stress
lnvarlant space and the resultlng volumetrlc straln lncrease.
In the second approach an establlshed elasto-plastlc, lsotroplc model lS
used and the denSlflcatlon lS added as an lnitlal straln ~. dependent on the
'endochronlc' deterloratlon parameter

d~ =J del) del]

where el ) lS the devlatorlc straln.

Such a model was used wlth success by the author ln varlOUS practlcal
computatlons. Flg. 4 shows for lnstance the computatlon carrled out by the
scheme descrlbed to slmulate the partlal collapse of San Fernando dam [16] [17].
Other uses of the deterloratlon parameter have been reported and some very
successful models are now belng produced, though much work ln thlS area lS yet
necessary.

Concludlng remarks
The fleld of problems descrlbed In thlS paper lS one In WhlCh both the formulatlon
as well as physlcal modeillng present some of the greatest challenges. It lS
hoped that this lnterlm statement of the state of the art wlll stlmulate further
contrlbutlons WhlCh wlll lncrease the efflclency and wldth of appllcation of the now
existing methodology.
 268

Acknowledgements
The author is ~ndebted to research co-workers and students ~n the var~ous phases
of th~s work. Dr. D. Naylor. Dr. E. Hinton. Dr. G. Pande. Dr. C. T. Chang.
Mr. K. H. Leung and Mr. V. Norris have made part~cularly important contr~but~ons.

Thanks are also due to the Science Research Counc~l (U.K.) for support~ng some
of the members of the team.

References

1. Terzhagi, K.l Erdboumechan~k. Franz Deuticke, Vienna. 1925.

2. B~ot. M. A.l Theory of three-dimens~onal consol~dat~on. Journal of Appl~ed

Physics, Vol. 12. pp. 155-164. 1941.

3. B~ot, M. A.l Mechanics of deformation and acoust~c propagat~on in porous

media. J. Appl. Physics. Vol. 33. pp. 1483-98. 1960.

4. Z~enk~ew~cz. O. C., Humpheson, C. and Lewis. R. W.l A un~f~ed approach to

so~l mecha~ics problems. Chapter 4 of Finite Elements ~n Geomechan~cs.
pp. 151-178. ed. G. Gudehus, J. W~ley. 1977.

5. Bowen. R. M. l "Theory of mixtures", Contwuum Phys~cs. Vol. III.

editor Eringenl Academ~c Press. New York. 1976.

6. Zienk~ew~cz. O. C. and Bettess. P.l Soil and other saturated porous mater~al
under trans~ent. dynam~c cond~tion. general formulation and validity of
various simpl~fied assumptions. Soils under Cycl~c and Transient Load~ng.
editors D. C. Zienkiewicz and G. N. Pande. John Wiley (to be publ~shed).

7. Z~enk~ew~cz. O. C.l The Fin~te Element Method, third ed~t~on. McGraw-Hlll.

1977 .

9. Park, K. C.l Partitioned transient analysis procedures for coupled f~eld

problems - stab~lity analysis - to be publ~shed.

10. Park. K. C. and Felippa. C. A. Partltioned transient analysis procedures for

coupled fleld problems - accuracy analysls - to be publlshed.

11. Z~enkiewicz. O. C.l Constitut~ve laws and numerical analys~s for so~l
foundat~on under stat~c, translent or cyclic loading. Proceed~ngs of Second
Internat~onal Conference on Behavlour of Offshore Structures (BOSS), London 1979.

12. Pande. G. N. and Z~enkiewicz, O. C. Proceedings Conference on SOlIs under

CyCllc and Trans~ent loading, Swansea 1980. Balkema Press 1980. (See also
special volume to be published by J. Wiley & Sons, 1981).

13. Mroz. Z. On the descrlpt~on of an~sotroplc work hardenlng. J. Mech. Phys.

Solids, 15. pp. 163-75. 1967.

14. Dafal~as. Y. F. and Popov. E. P. Plastic internal variables formallsm of

cycl~c plasticity. J. Appl. Mech. 98 (4). 645-650. 1976.
 269

15a Mroz, Z., Norrls, V. A. and Zlenkiewlcz, O. C.l An anlsotroplc hardening

model for salls and ltS application to CyCllC loadlng. Int. J. Num. and
Analytlcal Meth. Geomech. 1978, 2, 203-221.

15b Mroz, Z., Norris, V. A. and Zienkiewicz, O. C.l Appllcatlon of an

anlsotroplc hardening model ln the analysis of elastlc-plastlc deformation
of soils. Geotechnlque, 1979, 29, 1-34.

16. Seed, H. 8.l Conslderatlon ln the earthquake-reslstant design of earth and

rockflll dams. 19th Ranklne Lecture of the 8rltlsh Geotechnlcal Society,
Geotechnlque, 1979, 29, 215-263.

17. Zlenklewicz, O. C., Leung, K. H., Hlnto~ E. and Chang, C. T.l Earth dam
analysls for earthquakes. Proceedings Conference on 'Oeslgn of Dams to
reslst Earthquakes', Inst. C1V. Eng. London, October 1-2, 1980.
270

12

pore pressure:-P pore pressure : - P pore pressure: 0

Tot al s tress Hydrostaic stress . Effect ive st ress
o -p 0'

F:Lgure 1 Total and effect:Lve stress :Ln a porous mater:Lal

Zone I B=bC Slow phenomena (wand u can be neglected)

Zo ne II B=Z>!C Moderate speed (w can be neglected)
Zone III B;tZ;C Fast phenomena (w can not be neglected)
on I y 1ull Bio t eQuat ion valid

I-- - - - - Ora i ned (influence

01 n:, negligible)

- - -...-n:,
_ kP V; _ 1.2..L
n, - w l2 - n: f 2
2 2 T ?
n =J!l.l.. =n:? ( - )
V~ T
k= kl p, g k - kinemal ic permeability
T =2L/V, ,(=( 0' K,ln II p
= P K, /P, n
z K,/ p, (speed 01 sound In waler)

n " 0.33

F1gure 2 A Ilnear elast:Lc sem:L-:Lnf:Ln:Lte space subject to a per:LOdlC

surface load:Lng -

Zones of appllcablllty of varlOUS assumptlons

B· Blot's equation , Z: Zlenklewlcz's approXlmatlon,
C: Conventlonal (Terzhagl) equatlon
 271

c
101

A "-- -- - f -- - -....-

.0
It I
ib l

£v~O-l

(d )
2 6 8 10 12
N°ol cycles " 16 18 20 100

5'
0 10 20 30 ~o KYO

Flgure 3 Crltlcal state plas t lcl t y

a Slngle surface b two surface

CyCllc loadlng under dralned trlaXlal condltlons

[ c) shear stress vs . volumetrlc straln
[d) volumetrlc straln s vs . number of cycles
[ e) shear stress vs. shear strai n
(f) shear straln (at end of cycle) vs . number of cyc l es
 272

( b)

Materia l zone EI~t i c modulus V

4>' C ~.0 i l
~ight p:* (I. Bere )
MN/M2 bPareE KNM TMl A B '(
1 Alluvium 200 0,4 38 10 2.09 / / V
2a Hydraulic fi II - sand 90 Q41 37 10 2,02 2914 1904 4,6
2b . . .. 110 041 37 10 tW~
~ c1r_
29t~ t901 4.6

3 Clay (ore 90 0.41 37 10 2.02 /' /' V

4 Ground shale -hyd rau li e fi II 90 0·41 37 10 r:9?~~ / / V
5 Rolled fill 60 03 25 126 2.0 / / V
(a )

o·75 r--------------------------------------------------------,
0·50 ~------------------------------__tt---------------------_I

0.50 ~-------------...:.....----.l--_t-----------_j

0 . 750~-------------------75------------------~1~
0------------------~15
Ti me: s
( b)

Figure 4 The Lower San Fernando Dam analysls

[a) Data assumed for the analysis and the mesh of quadratlc
flnite element
[b) Base motlon input
[c) Deformed mesh plot [displacement x 2.0)
[d) Contours of excess pore water pressure bUlld up
 273

3s .

6s .

10s .
( c)

3 s.

6 s.

10 s.
Contour unit =kN/m2
Contour SPQClng=25 kN/m2
(d)
Numerical Methods in Elasto-Plasticity -
A Comparative Study

A. SAMUELSSON, M. FROIER
Chalmers Umversity of Technology, G6teborg, Sweden

Summary
Problems involving nonviscid, quasistatic, small strain
elastoplasticity are discussed in different aspects: the
role of an objective rate of stress, the choice of varia-
tional formulation, the choice of time step procedure.

Introduction
The development of the mathematical theory of nonviscid,
quasistatic plasticity started with Tresca, Levy and St.
Venant in the middle of the 19th century. In the first part
of our century v Mises, Prandtl and Reuss established a
theory for small strain plasticity for an isotropic mate-
rial with isotropic hardening. Their work is still the basis
of the mathematical theory of plasticity. After the second
world war the theory was generalized and extended in many
directions by among others Hill, Prager, Drucker, and
Koiter.
Lately, the theory was redefined in the framework of func-
tional analysis, especially the theory of variational in-
equalities by among others Duvaut and Lions [1], and
Johnson [2]. It was, wlthin this theory, possible to prove
the existence of a solution to the cases with elastic -
perfectly plastic and elastic - linear hardening properties.
 275

Before the development of the finite element method very

few practical solutions were obtained. In late 60ies the
first finite element applications to twodJmenS10nal boundary
value problems governed by Prandtl-Reuss equations appeared
ln literature. An lmportant contribution here was due to
Yamada, Yoshimura, and Sakurai [3] who deduced an explicit
tangent stiffness matrix according to the Prandtl-Reuss theo-
ry. The solution was obtalned ln a step by step method 5uch
that in each step a linear problem wlth an elastoplastlc
stiffness matrix was solved. It was found that due to the
&tep approximation the Yleld condition became violated. The
method was then improved by proJection to the yield surface
and iteration. Important contributions from this period were
due tu Marcal et al [4] and Zienkiewicz et al [5].

Following the development of the application of functlonal

analysis to the Prandtl-Reuss problem equillbrium and
mixed finite element procedures suggested by the analysis
were proposed by Johnson [6] and others. A mixed method
has been applied to problems with plane stress by the pre-
sent authors [7], [8]. Llnear programming techniques used
for Ilmit load determinations have by Maler and others
been shown to be useful also for the Solutlon of elasto-
plastic boundary value problems [9].

The Prandtl-Reuss theory has been extended to problems

with finite strain by Hill [10], Lee [11], Hutchinson [12],
and others. In this theory stress and rate of stress have
to be defined in a strict and useful way. When specializing
this general theory to small strain elastoplasticity the
result is found to differ from Prandtl-Reuss theory.

In this paper some questions concerning small strain plas-

ticity for isotropic material and monotonous loading will
be discussed.

Questions

1. Is it of practical importance to use objective rate of

stress also in small strain plasticity?
276

2. Which are the essential d1fferences between the con-

ventional f1nite element method with an elastoplastic
stiffness matrix and the method based on a variational
inequality?

3. The transient problem is solved in the two methods with

a forward respectively a backward Euler method, so the
order of error is O(~t). Small time steps are certainly
needed. Would it be worth trying a higher order method?

The tangent modulus and tangent matrix in small strain plas-

ticity

Let an 1nfin1tesimal unit element be loaded in principal

direction 1 with a force P so that the nominal stress vec-
tor is {00' 0, OJ, 00 = P. After deformation the side
lengths of the elements are 1+e 1 , 1+e 2 , 1+e 3 • Then the
Cauchy stress vector is {01' 0, O} = {00/(1+e 2 ) (1+e 3 ), 0, O},
or with use of volume constancy, {00(1+e 1 ), 0, OJ. Deriva-
tion with respect to time of 00 = 01/(1+e 1 ) gives

dO O d0 1 de 1
°1 (1)
dt 1+e 1 dt - 2 dt
(1+e 1 )

For small strain, e 1 « 1 , then

°1
e1
.
-°0 + °1
e1
(2)

The nom1nal and real tangent modulus thus differ by the

amount of 01. For low or disappearing tangent modulus the
effect of 01 is significant.

Hutchinson [12] gives the following elastoplastic constitu-

tive relation in stiffness form for the large strain case

*1j *ijkl.
T = S nkl (3)

with

s1jkl 2G{~(gikgjl+gilgjk) + (4)

1 kl i' kl
+ [vi (1-2v)]g J g - (a/g) s J s }
 277

* ..
where ,1J is the Jaumann rate of a contravar1ant component
of the Kirchhoff stress, n
kl 1S the rate of a covariant com-
ponent of the Lagrang1an strain, ~ 1S the metric tensor in
the deformed body and

1J 1J kl/ 3 (5)
, - g gkl'

Further

kl.
a = 1 if s n kl > 0 and (6)

_ 1j kl
J 2 - gikgjlS s /2 = (J 2 )max

a = 0 else,

and g is a generalized harden1ng modulus

1/g = (E/E t -1)/[E/E t - (1-2v)/3]2J 2 (7)

with E t and J 2 taken from a un1axial case.

The Jaumann rate of the Kirchhoff stress ~ is related to

the convected rate by

,* 1 j =,. i j ik, j 1 + g
+ 12( ' g j k, i l + g
i l, j k +
(8)

+ g Jl , 1k ) n. kl

1k sik lk lk
For small strain g ~ 0 ,nkl ~ E kl , , 00 ' the nomi-
nal stress, here denoted by a 1k • Combinatlon of (3) and (8)
and restriction to small strain and plane stress glves the
followlng tangent elastoplastic relation

(2G/g) iA11 (9)

l SYM

where G is the shear modulus and

278

2 2 2 (10 )
A11 = A - Sx (B-SxS z ) /(A-S z ) - O'x g / G
2 (B-S y S Z}2/(A-S;) - O'yg/G
A22 A - Sy

A12 B - S S (B-s S ) (B-s S ) / (A-s 2 )

-
x Y x z Y z z
2
A13 (B-SXSZ}SZTXy/(A-Sz) - SxTxy - , xy g/2G
2
A23 (B-SySZ}SZ'xy/(A-S Z) - s YTxy - 'Xyg/2G
2 2 2 2
A33 g/2 - 'xy z - (O'x +O'y) g/4G
s z Txy/(A-s)

with

A = g(1-v}/(1-2v}, B = gv/(1-2v}
The numerical experiment reported below indicates that the
effect of using the more correct stress rate definition is
significant also for small strain if the tangent modulus
also is small. There seems to be no practical reason not
to include the extra terms in the tangent matr~x.

Constitutive equations in two formulations

In the Prandtl-Reuss theory with isotropic linear harden~ng

the admissible set is a convex

(11 )

where

F(O'
~J'
~p} = 13/2\s .LJ I - (a Y+HS P ) (12)

and EP = 12/31s P
~J
I ~s proportional to the Euclidean length
of the 9-dimensional plast~c strain rate vector.
It is assumed that during plast~c flow = 0, that is F
~ •• - Hs'P
~v
= 0 ( 13)
O'ij ~J

and that the rate of plastic strain takes place in the di-
rection normal to F in the pOint O'ij:
 279

(14 )

The rate EP of effective plastlc strain can be eliminated

between (13) and (14) giving an elastoplastic relation va-
lid for E t » a

( 15)

In the alternative formulatlon the constitutive equation

is given in two equations,

(16 )

V(T .. ,n)€P (17)

1)

The latter equation expresses for F(a .. , E P ) = 0 that the

1)
vector {El j ' -HE P } is a normal vector to F at (a ij , E P ) so

E ••
.p 1 ( 18)

-H~~
[

Now 1aF/aa.·1
1)
= 13/2 and ClF/dE P = -H so

(19 )

It follows that A = EP so (14) is deduced from (16), (17).

Variational equations in two formulations

Let a regular region n c R3 be loaded at pOlnt Xl and time t
by volume load Ui • At the part anT of the boundary an the
region is loaded by T and at the remainlng part an of the
1 u
boundary the displacement u i is equal to zero. An initial-
boundary value problem is then defined as:
Find the displacements u.l (t) and stresses cr 1)
.. (t) for a gi-
ven volume load U. (t) and surface load
1
T.1 (t) satisfying
280

1
(a) -OJi,j = Ui (20)

1
(b) -2 (u
E l' J'
1,J. +u.J,l.)

(c)
- . .p on Q
EiJ - AijklOkl + Eij

'1
(d) -EI?(T . . -0 . . ) + HEP(n-E P ) >
lJ lJ lJ

(e) u. (0) 0 . . (0) = 0

1 lJ

(f) u 1 (t) = 0 on Q
u

(g) T. n.o ..
1 J Jl

The constitutive relations (c), (d) may be substituted by

the stiffness relation (15)

(c') (d') °lJ. (20)

In both procedures (20a) is replaced by a virtual work

equation

Iu . U . dV + I u. or . dS, Vu. E V (21 )

Q 1 1 dQ 1 1 1
T

where V is a space of functlons that satisfy (20f) and

are regular enough. A regular finite element mesh is then
chosen and a subspace Vh C V of finite element functions is
deflned. Let a general such function be

T~
u = ¢ u (22)

where ¢ is a rectangular matrix of basis functions and u is

a column matrix of node values. With the functions in ¢ as
u. in (21) the following system of equations is obtained
1

!¢UdV + (23)
Q

where 0, U, and T are matrices with components 0ij' Ui and

T. and V lS the linear strain operator.
1
 281

In the first procedure (23) is time differentiated and

combined with (20c',d') giving

{f(V¢T)Ts(a,H)V¢TdV}~ f¢UdV + f ¢TdS (24)

rI rI d rlT

This transient system is to be integrated under constraint

of initial conditions, variational equilibrium (23) and
yield condition (11).

In the second procedure (20c) is scalarly multiplied by

(T .. - a .), integrated over the domain and combined with
lJ lJ
(20d) glving a variational inequality

(25)

Let in a finite element mesh the stress components and the

hardenlng functions be dlscretized to values at the Gauss
points. Collect all these values in column matrices 0 and
EP . The varlational inequality (25) then becomes a set of
uncoupled equations

(26)

Together with (23) a trans lent system of equations is de-

fined which should be integrated under constralnt of initial
conditlons.

We observe that the equilibrium equation is in rate form in

the first formulation and in total form in the second and
that the hardening function E P is eliminated in the first
formulation by use of the consistency condition (13) but
is kept as unknown in the second.
282

Time discret1zat1on 1n two formulations

In the first formulation, (24), 1t is natural to use a

slmple forward Euler procedure. Starting at a known state
at time t, {u t ' 0 t }, set

(27)

where 6t is a short time-step. In (24) both the coefficlent

matrix and the right hand side are supposed to be known so
~
u t can be solved. Then

(28)

and

(29 )

For a given space discretization this procedure tends to

the correct solution for 6t ~ O. For a fin1te 6t the con-
slstency cond1tion (13) becomes v10lated so stress points
may fall outside P. The values of (27) and (29) therefore
need to be adJusted. P01nts outs1de P are projected onto
F = 0 in an approx1mate way. Since w1th the proJected
stresses the equll1brium eq~at1on is not sat1sfied the re-
sidual and its effect upon Ut and at have to be calcula-
ted.

W1th the second formulat1on, (23), (26), a backward Euler

method may be used. With

(30)

at a time-step 6t, (26) becomes

The solution to (31) 1S the minimum of a functional, namely

 283

~f ~ T ~f ~ ~p ~ 2
(0 -T) A(o -T) + H(Et_~t-n) , (32)

"(T,n) EP

~f
o (33)

1S a fictitious stress obtained by adding to 0t-~t an elas-

t1C stress increment.

For plane stress the var1ables of (32) are T 11 , T 22 , T 12 , n.

In order to satisfy also the Y1elding cond1tion this is
augmented according to Lagrange's method with a multiplier
A. The following non-linear system of equations is obtained

o (34)

f f
2T22 - 2VT 11 + 2(v0 11 -0 22 ) + A(2T 22 -T 11 ) 0

f
4(1+v)T 12 - 4(1+v)012 + 6AT 12 = 0

A.2H(Oy+Hn) = 0

2 2 2
+ T22 + 3T 12 - (Oy+Hn) = 0

This system 1S easily solved by Newton-Raphson's method for

known 0~1' 0~2' 0~2' These are obtained from (33) where,
however, ~t is unknown. The problem has 1n each time-step
to be solved by 1terat10n between the equil1brium equation
and (34). Stresses obtained from (34) glve a residual Rj
when put 1nto (23). This corresponds to a change 6t(U.-U. 1)
] ]-
1n d1splacement which can be obtained approximately from
an elast1c calculation

[f(V¢T)TA-1V¢TdV]6t(~.-~. 1) = pRo (35)

n ]]-]
where p is a constant that should be chosen in a way that
speeds up the calculation.

Small time steps 6t are needed in order to describe the

transient problem in a sufficiently accurate way, particu-
larly 1n the highly non-l1near reglon near the lim1t load.
284

Small time steps are also needed ~n order to get sat~s­

factorily good convergence in the iterative procedures.

It should be observed that higher order time step proce-
dures than the simple Euler method leads to complicated
solution agorithms.

MPa 101
56.9
III
~ 5107 II

I~ lum
5~.5

!~~
52.5

3.0m
K )II 'IIDTH·0.2m 1€ 1
750 2250 m~crostra~n

(a) (b)

y
u ; 0 FOR ALL NODES ALONG THIS Ll';E
x

~
J
...
.....
--I COMPUTA~ION~L SLPPORTS
x
(POI~T A
l'IITH Nt:LL REA-:~ION UP TO
I{OUNDOFF EI{I{O'lS

(c)

Fig 1. a) Selfequilibrating load on cantilever

b) Assumed uniaxial stress-strain relations
c) Finite element model

Numerical experiment

A cantilever beam with self-equilibrating load, Fig. 1a,

made of an isotropic material with a uniaxial stress-
strain relation according to Fig. 1b, is analyzed accor-
ding to the conventional method with an elastoplastic
stiffness matrix on an IBM-computer with double preci-
sion, [13]. The results obtained with an objective stress
 285

is compared wlth the results obtained with a nominal stress.

Due to the antlsymmetric condltions only half of the struc-
ture is studied. The elements are eight-node isoparametrlc
with 3 x 3 Gausspoints according to Fig. 1c.

~"e strain EX ln the indicated Gauss-polnt A is calcula-

ted with the two stress definitions, the three stress-
strain relatlons I, II, III, according to Fig. 1b, wlth uni-
form load-steps ~P/PE 0.025 after the elastic limit load
P e is obtained, Table 1.

PIPE 1 025 1 050 1 075 1 100 1.125 1 150 1 175 1 200 1. 225 1 250 1.275

OBJECTIVE 784 882 982 1152 1803 4335

STRESS
I
NOMINAL 784 882 982 1152 1803 4328
STRESS

OBJECTIVE 782 874 966 1071 1352 2019 2830 4473 12803
STRESS
II
NOMINAL 78~ 874 966 1071 1352 2019 2814 4376 11742
STRESg

OBJECTIVE 781 866 952 1038 1249 1583 2056 2549 3434 5220 11846
STRESS
III
NOMINAL 731 866 952 1038 1249 1583 2056 2540 3403 5045 10687
STRESS

Table 1. Strain in point A in microstrain for three stress-

-E
x
straln relatlons and two dlfferent theories, ~P/PE =
= 0.025

In Table 2 results from calculations with three different but

uniform load-steps, 0.100, 0.050, and 0.025, are displayed.

The solutions were ln each load case obtalned by solving the

equilibrium equatlon and satisfactlon of the yield condition
iteratively. The procedure was stopped when no monotonous
convergence was obtained.
286

PIPS 1. 025 1. 050 1.075 1.100 1.125 1.1';0 1.175 1.200 1.225 1 250

0.1000 1029 3865

liP 1055 2006 3823 14113
0.0500 857
P;; 0.0250 782 874 966 1071 1352 2019 2814 4376 11742
0.0125 791 883 975 1080 1362 2019 3001 4q47 13321

Table 2. Stra1n -&x 1n p01nt A in m1crostra1n for three

d1fferent load 1ncre~ents. Nom1nal stress 1S
used.

In Table 3 and Table 4 results from calculations with the

two methods descr1bed above are compared. The calculat10ns
are made for the problem In F1g 1 w1th the straln-stress
relat10n II, F1g 1b, but changed to a bilinear form by ex-
tenslon of the second line of the graph. The obta1ned
values of the effect1ve stra1n &p in p01nt A are displayed
1n Table 3. As shown 1n this Table the convergence of the
time-step procedure 1S fast for this blllnear case. In
Table 4 the development of the plast1c zones can be followed.
The labels of the integration points are given in F1g 2.

As can be seen from Tables 3 and 4 the numerical values of

the straln in point A as well as the development of the
plastic zones olffer considerably between the methods after
10% 1ncrease of the elastic llmlt load. The different time
step procedures in the methods do not seem to effect th1S
conclusion.

PIPE 1 021i 1 050 1 075 1 100 1 125 1 150 1 175 1 200 1 22t; 1 250 1 275 1 300

o 0;00 lOS 3.' 1HO "aln 4780 1108

§~
I", ~
o 0250
" "I '" 317 S., 1.!5J i'O]'j ?QOR 104:l' ~OOIJ 6132 7320

(l 0125 4.
I" 221 12S 60' 1253 2038 21:)10 1'146 5 '104 6115 7311

..
o 0500 127 795 7./ 1702 n''J 1811

'8'8
8"f;
,ji~
o 0250 45 127 21. ,"S
". / 1254 17)] 2208 2713 3286 3FJ19

o 012; '5 I" 21. ,., '03 700 1264 1132 2706 2735 3287 3811

Table 3. Effective strain &p in point A in micros train

for three load-steps and two methods
 287

E.. .
SYM·y·r.~.--.r---~----.----'--~

D ••• ~
i
C • • I
B • !
2 3 8

FIg 2. Order of yielding In integration points

PIPf,.
, 025 , 0>0 , 015 , 'DO
, 12'> , 150 , 17' , ,"0 , ,", , 2')0 , '" , JOO
rirst 1A ,0
"
mcthod lA
Jr.
'" " C,'
It
" 7A <B

'"
li'

Second
" lA
'0 '" " <T'
1< 3B
mt·thod
2. 1, S. 7A AA

Table 4. Order of yieldlng in integration

pOlnts for the two methods

Concluslons

It is lmportant to use the more correct stress-strain

definitlon obtalned wlth a Jaumann rate of stress for low
tangent modulus. It can easlly be lmplemented in a flnlte
element program.

In the conventlonal method for elastoplastlc numerlcal

analysls the constltutlve rclatlon lS glven in form of an
equatlon between rates of stress and straln. The yield
condltion is utlllzed and the rate of the hardenlng para-
meter E P lS ellmlnated at the deduction of thls equatlon
by use of the conslstency conditlon.

In the method wlth use of a variatlonal inequality the

constltutive equatlon lS satlsfied by a restricted mini-
mlzatlon of a functional with EP kept as a variable.

Ou~ nurecrical experiments show that large differences in

the development of the plastic zones can be obtained in
flnlte element calculations wlth the two methods.
288

It does not seem to be practlcal to use a hlgher order

dlfference method than the slmple Euler method. ThlS
means that small tlme steps are needed partlcularly near
the llmlt load. The number of lteratlons to obtaln
equlllbrium ln each tlme step will ln general be low when
short tlme steps are used.

References

Duvaut, G.; Lions, S.L.: Les inequations en Mecanique

et en Physique, Dunod, Paris, 1973

2 Johnson, C.: Existence theorems for plasticity prob-

lems, J. de Math. Pures et Appl., Vol. 55, 1976,
pp. 431-444

3 Yamada, Y.; Yoshimura, N.; Sakurai, T.: Plastic stress-

straln matrix and its application for the solution of
elastic-plastic problems by the finite element method,
Int. J. Mech. SCi., Vol. 10, 1968, pp. 343-354

4 Marcal, P.V.; King, L.P.: Elasto-plastic analysis of

two-dimensional stress systems by the finite element
method, Int. J. Mech. Sci., Vol. 9, 1967, pp. 143-155

5 Zienkiewicz, C.C.; Valliappan, S.; King, I.P.: Elasto-

plastic solution of engineering problems; 'inital
stress', finite element approach, Int. J. num. Meth.
Engng, Vol. 1, 1969, pp. 75-100

6 Johnsson, C.: A mixed finite element method for plas-

ticity with hardening, SIAM J. Numer. Anal., Vol. 14,
1977, pp. 575-584

7 Froier, M.; Samuelsson, A.: Variational inequalities

in plasticity, recent developments, Finite Elements in
Non-linear Mechanics, Tapir, Trondheim, 1978, pp. 63-85
 289

8 Samuelsson, A; Froier, M.: Finite elements.~n plast~­

city - A variational inequality approach. The Mate-

matics of Finite Elements and Applicat~ons III, Academic
Press, London, 1979, pp. 105-115

9 Maier, G.: A minimum principle for incremental elasto-

plasticity with non-associated flow-laws, J. Mech.
Phys. Solids, Vol. 18, 1970, pp 319-330

10 Hill, R.: On the classical constitutive relations for

elastic-plastic solids, Recent Progress in Applied
Mechanics, the Folke Odqvist Volume, Almqvist &
Wiksell, Stockholm, 1967, pp. 241-249

11 Lee, E.H.: Elastic-plastic deformation at finite

strains, J. Appl. Mechs, Vol. 36, 1969, pp. 1-6

12 Hutchinson, J.W.: Finite strain analysis of elastic-

.plastic solids and structures, Numerical Solution
of Nonlinear Structural Problems, ASME, AMD, Vol. 6,
1973, pp. 17-29

13 Backlund, J.; Wennerstrom, H.; Axelsson, K.: PIFEM,

Computer program for elastic-plastic structures
(in Swed~sh), Skrift LiTH-IKP-S-067, Dept. of
Mech. Eng., University of Technology, Linkoping,
Sweden, 1976
Finite Element Elastoplastic and Limit Analysis:
Some Consistency Criteria and Their Implications

L. CORRADI, G. MAIER
Pohtecmco dl Milano, Italy

Sunnnary

The fonnulation of element elastic-plastic law::. IS considere:l in the

framework of the dIsplacement approach. The finite element model is re-
garde:l as an actually discrete system, composed of a finIte number of
parts whose individual behaviour is described in terms of generalized
stresses and strains. A general procedure for formulatIng the element
laws is proposed, and it is shown that sane camlonly used fomulations
of fInite element elastic-plastIc analysIs can be recovered on the ba-
SIS of partIcular assumptIons. It IS pointed out that these formula-
tions may vIolate sane "consistency" requirements, WhICh are discussed
in the paper; these violatIons may explaIn some of the inaccuracIes ex-
perienced in computations. A fairly general method for restoring con-
SIstency is proposed. The implications of sane of the approximatIons
involved are also discussed, with particular reference to the limit
analysis problem. Simple examples illustrate the effects of lack of
consistency.
Introduction
When material non-linearities began to be considered in finite element
analyses, rather straightforward adjustments and generalizations of the
well established elastic methods were first adopted. Because of the
clear impossibilIty of enforcing the non-linear material law throughout,
this law was imposed only at a selected number of points within each
element. The '~auss integration points" used to compute numerically the
element matrices proved to be convenient choice 111 . However, inaccu-
rate solutions were occasionally experienced. In order to avoid such
occurrences, more or less empirical rules, based on computational ex-
perience, were proposed in the literature (see, e.g. 121 ).
In the author's opinion, the implications of this approach were not can-
pletely explored so far. Recent papers, e.g. 131 ,have re-considered the
properties of finite element models within the framework of the algebra
of finite dimensional spaces. In this way is was possible to assess pre-
cisely the number and kind of independent strain and stress distributions
 291

to assume in association with a given displacement model. The approach

gave evidence to certain "unconsistencies" which may arise with tra-
ditional procedures, e.g. when an attempt is made to enforce the mate-
rial law at an excessive number of points 141. Actually, these concepts
are not completely new; with1n the framework of the displacement method
they can be regarded as a more precise and r1gorous formulat1on of the
ideas on which the so-called "natural" approach is based 151; as such,
they were known since long and exploited in order to achieve operative
and computational advantages in elast1c analyses 161.
When this point of V1ew is adopted, the flllite element model is regar-
ded as an aggregate of constituents each of which has an 1ndiv1dual
behavior described in terms of generalized (element) variables. Thus,
the non-linear material law has to be replaced by an "element" law con-
cern1ng generalized quant1t1es assoc1ated with the assumed displacement
field. However, difficulties may ar1se in the explic1t der1vation of
the non-l1near constitutive relat10ns for high order elements. With
reference to piecewiselinear elastic-plast1c laws 171 a general proce-
dure was proposed in lSi, and generalized in 19,101 to regular yield
functions. In particular, in the latter paper 1101 it was shown that
the values assumed by strains and stresses at Gauss points within the
element can be regarded as element generalized var1ables, provided that
certain "consistency requirements" be complied with. In some instances,
cons1stency requires the use of "reduced" 1ntegratlOn procedures which,
in fact, had turned out to be computationally advantageos in some cases;
however, the use of reduced integration imposes some restrictions on
the finite element models which can be adopted 1n plastic analysis.
This paper prov1des a review and a general1zation of the aforementioned
results and proposes a fairly general method for fonnulating "consis-
tent" elastic-plastic laws for flllite elements. For the sake of brevity,
only perfectly plastic mater1al behavior 1S considered. Reference 1S
made to the displacement approach alone. The approximations implied by
the procedure proposed are discussed, with part1cular reference to the
limit analysis (collapse load) problem. It is shown, in particular, how
an upper bound to the collapse load is not always obtained, III spite of
the use of compatible models.

Elastic-plastic element laws

In the displacement formulation of finite element analysis, the displa-
cement field within each element is expressed as a function of the nodal
292

values by means of suitably defined interpolation functions. Then, the

strain~isplacement relations and the principle of virtual work provide

the following equilibrium and compatibility equations

(la,b)

where: ~ and ~ are S-vectors of the strain and stress distributions,

respectively, within the element; ~ and i are I-vectors of element nodal
displacements and forces, respectively; V is the element volume. Sub-
script e denotes quantities variable over the element and will later
permit to drop the specification (?S) in order to simplify the symbology.
Eqs.(l) must be supplemented by the material constitutive law. For bre-
vity, only elastic-perfectly plastic behaviors will be considered here
and the material properties will be assumed as constant within each
element. The constitutive laws first imply that

o (~
--e = --e
D e (x)
--e - (2)

(3)

where: ~ is the elastic matrix; ~ and Ee denote the elastic and plas-
tic portion, respectively, of the total straIn. The plasticity condition
(here assumed to be "regular", i.e. to define a smooth yield locus) and
the associated incremental flow rule can be written as follows 17 I:

(4a,b)

~ > 0
e= ~e ~e 0 (4c-e)

Since for high order (nonconstant strain) elements it is difficult to

enforce the constitutive relations everywhere, they are usually imposed
only at some selected points within the element.
An alternative approach to formulation of numerical (finite element)
elastoplastic analysis is based on the definition of element generalized
strains ~ and stresses~, in such a way that the "virtual work equiva-
lence" requirement is fulfilled 1111:

(5)
 293

In this case, the role of constitutive laws is played by "elEment" laws,

which can be written as follows

() = D e (6)

(7)

(8a,b)

(8c-e)

In the above relations, symbols have the same meaning as in eqs.(2-4) ,

except that they refer now to generalized (element) variables.
In order to give explicit expression to the quantities involved in
eqs.(6-8) , a relation between generalized strains and stresses and the
corresponding field values is to be established. This relation can be
written in the general form:

E
-'e-
(x) = b(x)
---
E (9a,b)

Equations similar to (9a) , with the same interpolation function matrix

Q(~' hold for elastic and plastic strain fields. Eq.(9b) is a conse-

quence of eqs.(9a) and (5). Eqs.(9) can be inverted, to give

(101 ,b)

with
(lla,b)

An explicit expression for the element elastic matrix ~ is readily ob-

tained from eqs.(9) and (2) and reads:

(12)

In order to formulate the plastcc portion of the elEment law, let a row
matrix .!l.(~ of interpolation functions be defined such that

~ (x) = .!l.(~ A (13a,b)

e-
294

Then, by applying to eq.(13b) the chain rule of differentiation and ma-

king use of eq.(lOa), one obtains

(14)

Thus, as soon as matrix ~(~ is chosen, all the element matrices appe-
aring in eqs.(8) can be defined. Dlfferent element laws are obtained
with different ~(~, each involving different approximations 191. The
implications of some assumptions will be discussed subsequently. Howe-
ver, it is worth noting here that some fOYrrlllations earlier proposed
for the element plastic laws can be regarded as particular cases of the
present approach; e.g. the "yield criterion of the mean", successfully
used for limit analysis and shakedown problems I 121, is recovered from
the preceeding equations when matrix ~(!) reduces to the unit scalar.
The element properties must next be expressed in terms of nodal varia-
bles, as required for the assemblage. To this purpose, the equilibrium
and compatibility equations for the element have to be used. They read:

E: = C U (15)

where ~ is a constant matrix. Eqs.(6) ,(7) and (15) permit to express

the element elastic properties in the form:

-f=k-uu -u-k
-up n
L (16)

having set:
k Ct D C k = Ct D (17a, b)
-uu -up
These matrices represent the elastic stiffness matrix and the initial
strain matrix, respectively, consistent with the present approach.

The ''natural'' formulation of the displacement method

Vector ~ of generalized strain parameters cannot be chosen independen-

tly of the assumed displacement model. As formally shown in 141 , ln a
finite element whose displacement field is governed by I nodal dis-
placement components ~, there are precisely D = I - R independent
strain states, R being the number of the element rigld modes. A most
spontaneous choice is to assume these independent straining modes as
generalized strains for the element.
As first pointed out by Argyris 151, vector ~ can always be expressed
as a linear combination of the R-vector ~ of element rigid body motions
 295

and the D-vector 0 of the "natural" or straining modes, through a non-

singular transformatlon

(lSa,b)

Let ~ and i denote the force quantities corresponding to £ and i. The

expressions dual to eqs.(lS) read

f = [~ ~~J {~} (19a,b)

If vectors ~ and i are chosen as strain and stress parameters, respec-

tively, it follows from eqs.(lS) and (lS)(19) that ~ = ~o. Since ~o is
a matrix of full row rank, it follows from eq.(lSb) that if Q lIDplles
i f Q, i.e. the element generalized stress vector i does not contain
self-equilibrating (or "redundant") stress states. Moreover, as by the
very definltion of rlgid body motion it must be ~(~ ~ = Q, from eqs.
(la) ,elSa) and (9a) one obtalns

(20a)

A comparison of eq.(la) to eqs.(9a) (lSa) shows that the following equa-

lity holds

(20b)

Eqs. (20) provide the relations between matrices ~(20 and Q(20, which
must be fulfllled when ~ = i if eqs.(la) and (9a) (lSa) have to represent
the same strain field within the element.
However, the formulation of the element elastic-plastlc behaviour requi-
res also some connections between matnces Q(20 and M20, since they
both contribute to the representation of the strain field within the
element. In fact, on the basis of the relations of the preceeding sec-
tion, one can write both the following expressions for the plastic strain
rate field over the element

Ee (~ = Q(20 E = Q(~) [d.<tia £1 i (21a)

D (_x)
"'-€
= {3¢ e--e (x) = {3¢ e--e--
/3G } ~ e- /3G } A(x) ~ (21b)
296

Clearly, it would be desirable to choose !(~ so that the alternative

distri1:utions of EeC20 provided by eqs. (20) are as close as possible to
each other. Unfortunately, the rather formal definition of the natural
modes makes it difficult to discuss the implications of different choi-
ces of ~ (20 when.£ = i is assumed. In order to overcame this difficulty,
alternative definitions of the generalized variables will be proposed.

Alternative definitions and consistency requirements

Let vectors i and i collect the corresponding field values at J suita-
bly chosen "strain points" within the element, i.e.
. ·t t·
£ { . . . --e£ (x.) ... } ; A
-J -
= {... A•e-J
(x.) ... }
t
(22a, b)

where x· (j 1, •.• ,J) are the strain point coordinates. Then, the fol-
-J
lowing interpolation matrices can be assumed

b(x) = [... -
--
Ieob.Jex)
-
... J (23a,b)

b.~.) = 0 .. , i,j = 1, ... ,J (Z3c)

J 1 J1
Here, !s denotes the SxS identity matrix and 0ji the Kronecker symbol.
With the above assumptions, the dimension of vectors ~, ~ is SJ and that
of i, 1, i is J.
A camplete discussion of consistency between matrices Q.(~ and ~(20,
as defined by eqs.(23) , cannot be presented in this paper because of
. .
space limitations. However, a certain degree of consistency can be
reasonably expected, since both Ee(x) and Ae(20 are governed by
their values at the same points. As shown in 1 9 1 , complete consisten-
cy in the sense that both eqs.(ZO) lead to the same distrl1:ution of
u
~-
(x) , can be achieved only if $ is a linear function of stresses
e
(or a piecewise linear one, an assumption which could be easily be in-
cluded in the present formulation by generalizing it to singular yield
loci). Same approximations are implied in other cases; nevertheless,
the above assumptions lead to element laws which represent the most
natural extension to high order elements of the formulation first pro-
posed by Hodge 1131
Same restrictions on the number and location of strain points must be
complied with in order to ensure that the strain field defined as func-
tion of ~ through eq. (ZZa) is consistent with the one expressed by
 297

eq.(la) in terms of~. To this purpose, note first that from eqs.(la),
(lSa) and (2Za) one obta1ns

(Z3)

On the other hand, if eqs.(la) and (9a) (lSa) have to provide the same
strain field, the follow1ng equal1ty must hold

(Z4)

Slnce -C-A
p
= 0, from eqs.(lS) and (18) (19) it follows that:
-

E = M 0 (ZSa,b)
where
M=CA
- --0
(26)

is a SJxD constant matrix. When SJ = D and ~ is non-singular, there

1S a one-to-one correspondance between ~ and i. Then, eq.(ZZa) can be
considered as a particular choice of natural modes and a valid defi-
n1tion of general1zed stra1ns. It can be eas1ly shown that in th1S
case eqs.(ZO) are replaced by eq.(Z4) and by the following relat10n

(Z7)

For linear elements, S = 1 and it is always possible to select J = D

strain points where generalized strains are defined. However, in plane
or space elements it seldom happens that D is a multiple of S
Then two alternatives are possible.
(a) SJ > D (rank of ~ = D). In this situation, eq.(ZSa) imposes
SJ-D constra1nts among the components of ~. Hence, these components
cannot be conceived as generalized strains for the element. Moreover,
eq.(2Sb) shows that a "fict1tious redundancy" (inconsistency of the
first type) 1S generated, in the sense that a (D-SJ)-dimensional
subspace of the stress parameters ~ corresponds to ~ = Q (and ,hence ,
to i = Q).
(b) SJ< D (rank of ~ = SJ). In this case, 1f the element strain
state has to be governed by vector ~ ,D-SJ fictitious modes of
rigid body motion (in the sense of deformations without strain energy
changes) must be attr1buted to the element, in addition to the R true
rigid modes. The fictitious, additional rigid modes are represented
by these vectors i for which ~ = Q is obtained from eq. (2Sa) and
can be generated by replacing in eqs.(l) the original expression of
298

matrix ~(~ with the one obtained in tenns of Q(2S) from eq. (24) .
In thIS way, a consistent element is arrived at; however, its elastic
stiffness matrix ~, eq.(17a), has only rank SJ ,instead of D
If the additional rigid modes are not introduced, two different ele-
ment behavIors are obtained from eqs.(l) and (9), the latter beIng
"stiffer" than the fonner. This second type of inconsistency will be
called "artificial stiffening".
The preceeding discussion shows that, at least in principle, consis-
tent element fonnulation can be achieved starting from the assumptions
(22) (23) ,provided SJ ~ D . Note that, while the components of
vectors -E and -A coincide with the relevant field values at the strain
poir.ts, the same is not true, in general, for the components of 2".. and
~ . However, this coincidence can be enforced by selecting as strain
pomt locations -]x the coordinates of the points that would be used
to evaluate numerically by GaussIan quadrature an mtegral over the
element volume. This result can be readily arrived at if eqs.(9b) and
(13b) are numerically integrated, account taken of eqs.(23).
The above special chOIce of x turns out to be particularly conve-
-]
nient. In fact, the element constitutive law, eqs.(6-S), can be shown
to reduce, in this case, to the material law, eqs.(2-l) , enforced at
the strain (Gauss) points 110 I . Thus, the element properties defined
by eqs.(13b) and (14) need not to be evaluated, with a substantial
saving in computational effort. It can also be shown that the element
stiffness matrix ~ coincides with the customary expression, provi-
ded a numerIcal integration based on the J Gauss (strain) points is
used to evaluate It ,10 , . When SJ < D ,a "reduced integration" is
performed/which automatically generates the required additional rigid
modes in the element behaviour.
Thus, the familiar procedure for elastic-plastic finite element ana-
lysis which enforces the material law at the Gauss integration points
,l , ' can be regarded as a special case of the present approach, based
on the assumptions (22) (23) and on a particular choice of the strain
point coordinates.
The consistency requirements previously discussed might explain some
aspects of the numerical experience described in the literature. For
instance, the inaccurate solutions sometimes obtained when using an
excessive number of Gauss points (such that SJ > D) could be justified
by the fictitions redundancy implicitly introduced in the element
 299

behaviour. Also the advantages provided by the use of reduced integra-

tion 121 can be understood as due to the fact that consistency is res-
tored by making the rank of matrix -tiU
k equal to the number of lffiplicitly
assumed generalized strain parameters. Practical limItations to the
applicabIlity of thIS procedure arise from the fact that D is seldom
exactly a multiple of S and, hence, reduced Integration is usually re-
quired in order to achIeve consistency. Reduced integration, even if
always possible In principle, must be used with some care; in fact the
reduced rank of matrix k may cause singularIty of the assembled stif-
-tiU
fness matrix 121 or decrease the possibilIty of stress redistribution
In the plastic range due to the diminished degree of redundancy of the
assembled system 110 I.
On the other hand, inconsistency of the fIrst type cannot be avoided
with assumptIons (22)(23) when SJ > D; in fact, no matter how refined
is the numerical integration, the rank of the resulting stiffness ma-
trix cannot exceed D. In this case, however, it is possible to avoid
fIctitious redundancy by assuming £ = i and i according to eqs.(22b),
(23b,c). If x. are Gauss pOlnt coordinates, vector ¢ will still contaIn
-J -
as components the yield function values at these points; however, now
these values are functions of the element generalized stresses ~ = ~;
eq.(lOa) permits to write

o (x.) = s(x.) 0 (28a,b)

-€ -J - -J -

where matrix ~(~ IS evaluated, through eqs.(ll) , on the basls of the

expression (20a) of Q(~'
In this case, the elastic element matrices in eqs. (17) are the "exact"
ones. However, matrix [ai/a~l, eq. (14), must now be computed on the
baSIS of the current values of the material state variables. Therefore,
whereas In principle the plasticity condition can be enforced to a high
level of accuracy by selecting for J a large number, the computatio-
nal effort Involved in evaluating the element properties will increase
apprecIably wIth increasing J.
The present procedure cannot ensure consIstency when SJ < D. In this
case, the plastlc strain rate field would be actually governed by a sub-
set of the set of available parameters and the element will "artificial-
ly stiffen" as plastic flow occurs.
300

Assemblage relations
The equations governing the behaviour of the entire system, composed of
N finite elements, are derived below. Subscript n will denote varia-
bles pertaining to element n (n = 1, ... ,N); e.g. vectors ~ and in will
collect the nodal displacement and force components of element n.
Let ~ and f indicate the vectors of the nodal displacement and force
components, respectively, for the assembled system. For each element,
the followlllg customary connectivity relations can be written

u
-n
= -n-
L U (29a,b)

If -n
u and U are defined In the same reference frame, -n
L IS a Boolean
matrlX, containing at most a single unit entry in each row. The compa-
tibilIty and equIlIbrium equations for the discrete system, relating
the nodal variables of the dlscretized structure to the element genera-
lized strains and stresses, can be obtained from eqs.(29) and (15).
They read

~
= -n
C -n-
L U n = 1, ... ,N (30a)
N
F = L -n-n~
Lt Ct (30b)
n=l
Eqs.(30) and the element law, eqs.(6-8) , written for each element n, pro-
vide a complete formulation of the incremental elastic-plastic problem.
It is worth mentioning that, as shown in 191, the present formulation
fulfIlls all the requirements for the validity of some extremum proper-
ties which, in turn, permit the use of numerical methods based on mathe-
matical programming (specifically quadratic programming) for the solu-
tion of the analysis problem 17,141. This aspect will not be discussed
here; attention will be focused on the llffiit analysis problem alone.

Limit analysis
Let the finite element model in point be subjected to a given distribu-
tion of nodal loads fa. For perfectly plastic systems, the "collapse
multiplier" s is defined as the maxlffiUffi load multiplier t/I such that
loads t/lF-{) can be carried by the structure. The static theorem of limit
analysis 1111 states that s is the maximum value of t/I for which in
all elements generalized stresses ~ can be found which are in equili-
 301

brium with loads ~Eo and do not violate the plasticity condition:

s = max ~ subject to :

(31)

The kinematic theorem of limit analysis can be stated as follows: s is

the mininrum of the (''kinematically admissib1e'ry load factors such that
the rate of external work equals the internal disslpation rate for some
compatible field of velocities and strain rates. The explicit formula-
tion of limit analysis by the kinematic theorem as a minimization pro-
blem is generally more complicate than that by the static theorem, unless
piecewise1inear yield conditions are assumed 17,141.
Eq.(31) provides the collapse load for the discrete system. It is of in-
terest to compare the value of s obtained in this way to the collapse
load of the original continuous structure which was modelled by finite
elements. Since the present approach concerns displacement models, in
which compatibility is ensured throughout, s would be expected to re-
present an upper bound to the actual collapse load. This condition does
not necessarily hold when reduced integration techniques are used to
impose consistency, since in this case rigorous intere1ement compatibi-
1itymay not be ensured. However, even when compatibility is strictly
imposed, the upper bound nature of s cannot be "a-priori" guaranteed,
because of the approximations introduced when writing the plastic por-
tion of the element constitutive law.
In fact, except for very simple (constant strain) elements, eqs.(8) en-
force the material law only in an "average" sense. In order to clarify
this point, consider the expression (23b) of ~(~). The functions bj(~
are usually negative over part of the element volume; thus, eqs.(8a) and
(Sc) do not prevent possible local positive values of ~e(~ or negative
values of ~ (x). Moreover, eqs.(4d,e) are imposed only globally by their
e-
element counterparts, eqs.(8d,e), since from eqs.(13) it merely follows
that

!t i = Iv ~e(~ ~e(~ dV (32)

In general, when -]
x· represent Gauss
the collapse load of the continuous
is less than the exact value of the
tip1ier associated with the assumed
point coordinates, upper bounds on
structure are obtained, even if s
kinematically admissible load mu1-
displacement model. This remark ex-
302

plains why accurate estimates of the collapse load are often obtained
even by means of canparatively "stiff" finite elEment models.

Examples
Sane very simple examples are considered below in order to illustrate
sane of the concepts discussed in what preceeds. Consider first a can-
tilever beam of uniform cross section (with flexural rigidity EI and
limit moment M ). The beam is acted upon by a concentrated force F at
p
the free end. The beam is modelled with a single finite element, with
a cubic displac~lent field; thus, the strain (curvature) distributlon
in the beam is linear.
15 r-------r-----~------_.------_,

0.. J-3(con.)
~
...... 1.0 1-----I-+--~---e-k4-o-c-t- - - I

.5 L-._...L---.! _ _ _- ' -_ _ _ _-'--_ _- - '

o 1.0

Fig.l. Force~isplacement diagram for a statically determinate,

perfectly plastic cantilever beam and different elEment laws.

Since the sectional behaviour was assumed as perfectly plastic, the "ex-
act" load-dlsplacement relationship results in a linear plot up to the
value FL = Mp/£; at thlS point, the load cannot further increase and
collapse occurs. This relation is depicted in fig.l, together with the
diagrams obtained with different element laws.
Since I = 4 and R = 2, a consistent element is obtained with the use of
J = 2 strain points. They were located at the Gauss point coordinates
and matrices QC20 and ~(20 were defined according to eqs. (22) (23). The
plot indicates the same behaviour as the exact one, even if the collapse
load is overestimated by about 29%.
 303

When J = 3 is adopted in connectIon with eqs.(22) (23), the effects of

the inconsistency of the first type become apparent. For F = 1.127~p/~
the first strain (Gauss) point yields; however, the fictitious redundancy
generated in the model prevents the beam from collapsing. Collapse oc-
curs for F = 1.39S~ /£), when the secon strain point also yields
p
(fig.l). This result IS obtained with traditIonal procedures, byenfor-
cing the ''material'' constItutive law at three Gauss pomts within the
element.
For J = 3, consistency can be restored either by assuming a fourth order
displacement distributIon through the addition of a displacement type
nodal variable (I = 5), or by using the natural modes as generalized
strains for the element, while retaining the expressIon (23b,c) for
~(~) and expressing the element plasticity condition through eqs.(28).

The two procedures are fully equIvalent and the dashed line in fig.l
shows that the fictitious redundancy has thus disappeared.

--
20

1 5 L1
cubic element:
x l ,2 = £(1 + 13) /2;

Co
:!; 1 0
LJ./ ~ ~ Q1
quartic element:
....
'- V 'quartic x l ,3 £(1 :;: 13) /2,
...-'

/ x2 £/2.

VCUb'l
o 1 00
00 25 50 75

Flg.2. Collapse load for the cantIlever beam of flg.l, as a

function of the straIn point locations.

Since Gauss point coordinates were chosen as strain point locations,

an upper bound on the actual collapse load was obtained. However, it
is not always so with either choices of strain point coordinates, in
sp~te of full compatibility. This circumstance is illustrated in fig.2

(from ref. lSi), where the collapse loads obtained for consistent
304

cubic (I = 4, J = 2) and quartic (I = 5, J = 3) elements is depicted. In

the latter case, unsound constitutive laws are obtained for S < 1/1.3,
since negative dissipation becomes possible. It appears that with cubic
elements, the finite element collapse load is below the exact value of
FL = Mp/£ when S < 1/3. Points L2 and Q2 indicate the values obtained
when strain points and Gauss points coincide.
For linear elements, there are no difficulties in determining, on the
basis of the displacement model, the number of Gauss points where the
constitutive law is to be enforced without violating consistency. This
is not always possible with more complicated elements and the procedure
based on eqs.(28) should be used to avoid inconsistencies. As an example,
consider an axisymmetric, tronco-conical thin shell element (fig.3) 1151.
By neglecting transverse shear deformations, the strain distribution
within the element is represented by the 4-component vector

-e
£ (r) = {£ s (32)

where £s' £~ denote, respectively, the meridian and circumferential mem-

brane strains and xs ' x~ the corresponding curvatures; r is the radial
distance from the shell axis.
The simplest displacement model able to ensure interelement compatibility
consists of a linear expression for ue(r) and a cubic one for weer). Thus,
the displacement field depends on I = 6 parameters, which can be iden-
tified with the nodal components indicated in fig.3. Since the only rigid
mode for the element is a vertical translation, we have D = s. A possible
expression for matrix Q(r) , in terms of natural displacements, is (e de-
noting the angle between generatrix and shell axis):

2 cose sine 1 ctge ~ 2~

r r sine r . 2
Sln e
0 1 0 0 0

Q(r) (34)
0 0 - -sine
r- - -2 -3 r
sine

0 0 -2 -6 r
0 sine

Fram eqs.(34) and (11), the expression of ~(r) can be readily derived.
 305

Two strain points can be assumed in the element and matrix ~(r) can be
defined accordingly. If the constitutive law were directly imposed at
the strain points, a fictitious redundancy would show up, since SJ = 8
is greater than D. However, the ,inconsistency can be removed through
the use of eqs.(28); if strain points are located at Gauss points, the
element yield function will result in the following two-component vector

(35)

where cr is the 5-vector of generalized stresses associated to the assu-

med natural modes.

Fig.3. Tronco-conical thin shel]

element.
5

Conclusions
Elastic-plastic and limit analysis are areas of non-linear structural
mechanics where the finite element method turns out to be, once again,
a powerful tool for accurate numerical solutions in complex situations,
but exhibit some aspects and requirements quite distinct and partly
novel with respect to elastic analysis. Some of these peculiarities,
considered in the paper, basically concern non trivial, sometimes over-
looked "consistency" conditions, to be fulfilled in modelling the ele-
ment fields, or, equivalently, in formulating element plastic laws in
(element) generalized variables; it has been shown that these should be
taken into account in order to avoid sistematic and possibly signifi-
cant source of inaccuracy.
306

Aknowledgment
The financial support of C.N.R. (National (Italian) Research Council)
is gratefully aknowledged.

References
1 Nayak, G.C.; Zienkiewicz, O.C.: Elastic-plastic stress analysis. A
Generalization for various constitutive relations including strain
softening. Int. J. Num. Meth. Engrg., 5, (1972), 113-135.
2 Zienkiewicz, O.C.: The finite element method. McGraw-Hill, London,
1977.
3 Besseling, J.F.: The force method and its application in plasticity
problems. Computers and Structures, 8, (1978) 323-330.
4 Besseling, J.F.: Finite element method. 53-78, in: Trends in Solid
Mechanics, Besseling and van der Heijden edts., Sijthoff & Noordhoff,
1979.
5 Argyris, J.H.: Continua and discontinua. 11-189, In: Proc. 1st
Conf. on Matrix Methods in Structural Mechanics, Dayton, Ohio,
AFFDL TR. 66.80, 1966.
6 Scharpf, D.W.: A new method of stress calculation in the matrix dis-
placement analysis. Computers and Structures, 8, (1978) 465-477.
7 Maier, G.: A matrix structural theory of piecewiselinear plasticity
with interacting yield planes. Meccanica, 5 (1970) 54-66.
8 Corradi, L.: On compatible finite element models for elastic plas-
tic analysis. Meccanica, 13 (1978) 133-150.
9 Corradi, L.: On some "consistent" finite element approximations in
non-linear structural analysis. Proc. 5th AIMETA Conference,
Palermo, Italy, Oct. 1980.
10 Corradi, L.; Gioda, G.: On the finite element modelling of elastic-
plastic behaviour with reference to geotechnical problems. 1st. Int.
Conf. on Numerical Methods for Non-Linear Problems, Swansea, U.K.
Sept. 1980.
11 Martin, J.: Plasticity. MIT Press, 1975.
12 Dang Hung, N.; Konig, J.A.: A finite element formulation of the
shakedown problem using a yield criterion of the mean. Compo Meth.
Appl. Mech. Engrg., 8 (1976) 179-192.
13 Hodge, P.G. Jr.: A Consistent finite element model for the two-dimen-
sional continuum. Ingenieur Archiv, 39 (1970) 375-383.
14 Cohn, M.Z.; Maier, G. Eds.: Engineering Plasticity by Mathematical
Programming. Pergamon Press, New York, 1979.
15 Dang Hung, N.; Trapletti, M.; Ransart, D.: Bornes quasi inferieures
et bornes superieures de la pression de ruine de coques de revolution
par la methode des elements finis et par la programmation nonlineare
Int. J. Nonlinear Mech., 13 (1979) 79-102.
A Finite Element Method for Problems in Perfect Plasticity
Using Discontinuous Trial Functions

C. JOHNSON, R. SCOTT
Chalmers UniversIty of Techno}ogy, Goteborg, Sweden

Introductlon. It lS known (see e.g. [3],[5]) that the dlS-

placements in an elastlc perfectly-plastlc body may be dlS-
contlnuous. Conventlonal flnlte element methods (dlsplacement
methods) for plastlclty problems are based on uSlng continuous
trlal functlons and are thus not partlcularly well adapted to
the nature of the true solution. In thlS note we propose a
flnite element method of dlsplacement type for problems in
perfect plasticlty where we use a flnite element space Vh of
plecewlse polynomlal functlons with no requirement on inter-
element contlnulty. In order to be able to approxlmate a dlS-
continuous solution u of a plastlclty problem accurately wlth
functions ln Vh , the flnite element mesh wlll have to flt the
dlscontlnultles of u. Thus, Slnce the locatlon of these dlS-
contlnultles is ln general not known in advance, one would
llke to use some kind of adaptive technlque where according
to the results of computatlons the flnite element mesh lS
succesively modlfled. In thlS note we do not consider this
more general problem but concentrate on analyzing the proposed
method ln the case of a given mesh.
An outllne of the note lS as follows: In Sectlon 2 we
formulate the problem, ln Sectlon 3 we lntroduce the flnite
element method and prove a convergence result, and flnally
in Section 3 we glve the results of some numerlcal experiments
ln one space dlmenslon.
308

1. The plasticity problem

For simplicity we shall consider a stationary (static)

problem corresponding to the so-called Henky's law. The quasi-

static rate problem corresponding to Prandtl-Reuss' flow rule

does not present any new serious difficulties and leads to a

problem of Henky type at each incremental step.

Let Q be a bounded region in 1R3 occupied by a body £..

made up of an elasto-plastic material. The displaaements of

E is given by u = (u i ), i = 1,2,3, where u i is the displace-

ment in coordinate direction xi. The strains corresponding to

u is given by the strain tensor E(U) = {Eij (u)}, i,j = 1,2,3,

where
aU i
Hax:- + ~}.
E iJ' (u) = ax
J ~

Further, let the stresses in £ be given by the symmetria

stress tensor °= {O~j}' O~j 0ji' i,j = 1,2,3. Let the

elastic properties of the material, wh~ch we assume to be in-

dependent of x for simplic~ty, be given by the symmetric posi-

tive definite transformation A: R9 ~ ~9 so that

E = Ao (Hooke's law),

is the elastia strain corresponding to the stress o. The set

of plastiaally admissible stresses is given by the closed

convex set B C ~9 with 0 E interior of B, where R9 is the set

s s
of symmetric 3 x 3 tensors t= {t,~J,}. For instance we may

take B = {t E IR?:
s
It I < 1} , where 1·1 is the Euclidean norm
,
~n
/R9 , i. e.,

3
Itl= ( L t~j)'.
~, j =1
For simpl~city, let us assume that B is bounded. Together,
 309

A and B define the properties of the elasto-plastic material.

For simplicity, we assume below that A = Identity.
We can now formulate (~n a non-precise way) the plas-
ticity problem as follows: Find a and u such that
a = lIE (u) in n , (1 .1 a)

-diva =f in n , (1.1b)

u = 0 on r , (1.1c)

where II:m; + B is the proJection onto B with respect to the

usual scalar product in m9 , f = (f 1 ,f 2 ,f 3 ) is a given volume
load and diva ((diva Ii) with

da ij
(diva) i dX j

Here and below we use the follow~ng summation convention:

repeated indices indicate summation from 1 to 3 so that in
part~cular

3
E
j =1

For simplicity, we assume that the boundary conditions are

given by (1.1c). Note that (1.1a) can also be written

E(U) = (E(U)- IIE(U))+ a::: EP+E e ,

which corresponds to a splitting of the total deformation E

into an elastic part EP and plastic part Ee •

In order to give a precise formulation of (1.1) let
us introduce some function spaces. Let

H = {or = ('[ ~J
.. ):'[ ~J
.. = '[ J~
.. E IJ 2 (n)} ,

with scalar product and norm given by

(a , '[) = ~ 0 i j '[ i j dx , 11'[11 (0,0) !,

p = {'[ E H: '[ (x) E B a.e. x En} ,

310

v = [H~(~)]3 and finally W = [L2(~)]3 wlth the usual scalar

product (.,.) and norm II ·11 •
The stress a can, according to the Principle of the
complementary energy, be characterized as the unique solutlon
of the problem
2
Sup - 111 Til , (1 .2)
TEP
-dlVT=f
where fEW. Recalling Green's formula:

~'rljSij(V)dX = fVi'rijnjdS -~ViTij,jdX ,

where n = (n.) denotes the outward unit normal to r , we have

J
if v E V and T E H wlth div'r E W,

(T,S(V» = -(div T, v).

Thus
if -dlv T f,
Inf [(T,s(v»-(f,v)] = {O
vEV -co otherwise,
and hence (1.2) can be wrltten

Sup [Inf L(t,v)] , (1 .3)

TEP vEV
where
2
L(T,V) = -liITII + (T,dv»-(f,v).
The corresponding dual problem obtained by lnterchanging the
order of Inf and Sup reads

Inf [Sup L(t,v)]. (1 .4)

vEV tEP
Here it is easy to evaluate the Sup: We have

Sup L(T,v) = Sup[-lIIT-s (v)11 2 + ills (v)11 2 - (f,v)]

TEP TEP

-llls(v)-IIs(v)11 2 + llls(v)11 2 - (f,v) (1 .5)

(S (v) ,IIdv) )-111 IIdv)11 2 - (f ,v) ::: F (v).

 311

Thus the dual problem can be formulated

Inf F(v), (1 .6)

vEV
with F def1ned 1n (1.5). In the case B {T E /R~: I T I < 1}

one has

F(v) fg(ldv)l)dx - (f,v),

Q
where g:R + /R is given by

,r 2
Irl < 1,
1
if
g (r) = t
r-! if Irl > 1.
The funct10nal F:V + /R 1S not coerC1ve on V and thus the pro-

blem (1.6) does not have a Solut1on 1n V in general. In order

to f1nd a solution we have to extend the space V and the

natural space 1S then the space of "functions of bounded de-

o
formation" BOW) defined as follows ([1] , [5]):

where Q 1S a cube containing D, v has been extended by zero

outside Q, M(Q) 1S the set of bounded measures on Q and M(Q)s

1S the set of symmetr1c tensors with components 1n M(Q). The

o
norm 1n BO(Q) 1S glven by

Furthermore, one can prove (cf. [3], [4], [5]) that for

< p ~ 3/2,

IlvjlL (Q) ~dls(v)IIM(Q)'

p s
o
and that ([5]) BO(Q) 1S compactly imbedded in [L (Q)]3 1f
P
< P < 3/2.
One can prove that under a safe load hypothesis, to be
o
stated precisely below, the functional F is coerC1ve on BO(Q)

and since ~O(Q) 1S the dual of a Banach space (cf.[l]) the

312

problem
Inf F(v), (1.7)
vEB~ (n)
admits a solut~on u E rlb(n). Here the domain of F is extended
to B~(n) by defining for v E rlb(n)
2
F(v) = Sup [-,IITII +<T,E(V» - (f,v)], (1 .8)
TEPnC(Q)s
where c(n)s denotes the set of symmetr~c tensors with compo-
nents in C(Q) = {cont~nuous functions on Q} and <','> de-
notes the duality between C(Q)s and M(Q)s extending the scalar
product (T, E (v) ) •
The safe load hypothesis reads as follows:

3 XEP and 0 > 0 such that -div X = f in nand

dist(X(x),dB) > 0, x En, (1.9a)

X E C{~)s' (1.9b)

where dB denotes the boundary of B in ~:.

The finite element scheme to be considered below may
be thought of as a discrete analogue of (1.7) obtained by
replacing the space Bb(n) by a finite-d~mensional space con-
sisting of piecewise polynom~al discontinuous functions. In
the analysis of this scheme below we shall use the following
alternative formulation of the plasticity problem in both
stresses and displacements where the somewhat non-standard
space ~D(n) is not utilized. Define
3
[L 3 / 2 (n)] ,

P {T E P:div T E [L 3 (n)]3},
L (T ,v) = -,II Til 2_ (div T ,v) - (f ,v) ,
and assume that (1.9a) holds and f E [L 3 (n)]3. Then, one
can prove (see [3], [5]) that the functional L:P x ~ ~ ~

adm~ts a saddlepoint (cr,u) E P x ~, i.e. 3(cr,u) E P x ~ such

 313

that

L(T,u) < L(o,u) < L(o,v)

~

V(T,V) E P x V, (1 . 10)

or, equivalently,

(0, T-O) + (u,div T -div 0 ) ~ 0 VT E P, (1.11a)

- (divo,v) = (f,v) 'Iv E V. (1.11b)

o
Moreover, u E BD(~) and we note that 0 is uniquely determined.

From (1.10) it follows that 0 E P and u E V are solutions of

the following analogues of (1.3) and (1.4):

Sup [Inf L(T,v)], (1.12 )

TEP vEV

Inf [Sup L!T,v)]. (1.13)

vEV TEP
To see the connect~on with our or~ginal formulation (1. 1 ) , we

note that if u E V then by Green's formula, (1.11a) can be

written

(O-E (u) ,T-O) > 0 VT E P,

i.e. ,

o = lIE (u) •

Thus, lf u E V (1.11a) is equivalent to (1.1a). However, ~n

general u i V and then ~t is not immediately clear how to

~nterpret (1.1a) (or 1.11a».

2. The finite element method

Let us for simplic~ty consider a problem in two space

d~mensions. We use the same notation as in the prev~ous sec-

tion with the obvious modifications (the only notable diffe-

rence is that the spaces L3/2(~) and L3(~) ~n the definition

of P and V may be replaced by L2(~) in the two-d~mens~onal

314

case). Thus, ~ is a bounded region in the plane which we

assume to be polygonal to slmpl1fy the presentation. Let now

C
h
= {K} be a triangulation of ~ and introduce
2
{v E [L 2 (~)] :vl K E iPk(K), K E C h },

where ~k(K) 1S the set of polynomials of degree at most k de-

f1ned on K. Thus, Wk cons1sts of piecew1se polynomial func-

tions w1th no cont1nuity requirement on interelement bounda-

ries. We consider a fam1ly of triangulat10ns {C h } indexed by

the positive parameter h such that U Wh is strongly dense

h
in V. Clearly, Whc B~(~) and we now pose the following d1S-

crete analogue of (1.7):

Inf F (v) • (2.1 )

vEW h

Let us evaluate F(v) for v E Wh accord1ng to (1.8). We have

for T smooth, v E Wh using Green's formula

<T, t: (v) > = - f V d1V T dx

~

Z (T,t:(v))K - Z f[V.]T .n.ds

KECh S S 1 1J J

where we sum over all sldes S of the tr1angulation Ch ,

[vi] 1S the jump of v 1 across Sand

f T t: .. (v)dx.
K 1J 1J

Thus, by varying T in the inter10r of each element K and in a

th1n strip around each slde S, it follows that for v E Wh

F(v) = Z ((t: (v) ,ITt: (v) )K-~IIITt: (v)II~)

K
(2.2)
+ Z Js(v) - (f,v)
S
where

f sup [V.]T .n.ds .

S TEB 1 1J J
 315

The functional F:W h + ~ 1S clearly continuous and by the safe

load hypothes1s (1.9) it follows replac1ng f by -divX and

uS1ng Green's formula that

F(v) = L:((s(v),I1s(v))K - (s(v),X)K - !IIIIs(v)lli)

K

+ L: !(sup[v.)1 .n. - [vJX .. n.)ds.

SST EB 1 1J J f 1J J

US1ng the fact that dist(X(x),dB) > 0 it 1S then easy to see

that

F(v) ~ ~lvjlBO '

for IlvjlBO sufflciently large and thus F(v) + "" as IlvjlBO + "".

Th1S proves that (2.1) admits a Solut1on u h E Who The func-

t10nal F:Wh + ~ 1S convex but not strictly convex and thus

u h may not be un1quely determined.

4
Let us not1ce that in the case B {T E IRs: IT I < 1}
we have

F(v) L:! g(ls(v)l)dx + L:1[vll- (f,v).

K K S

2.2 Convergence

To prove convergence of the method (2.2) it is con-

ven1ent to use the follow1ng character1zation of the solution

u h E Who Let us 1ntroduce the Lagrangian ~:~h x Wh + ~

def1ned by
2
.!t(T,V) -!IITII + [T,s(v)lh - (f,v),

where

L:(T,s(V))K- L: f[v 1 T.. n.ds,

K S S 1 1J J

~h = [T (i, 'T') : T E P, TEll L2 (8) s

S
with T(X) E B a.e. on S, VS}.
316

Thus, a T E O'h has an "interior component" T and a "boundary

component" 1 and these components are independent. Consider

now the problems

Inf [Sup }f, IT ,v)], (2.3)

VEWh TEO'h

(2.4)

The problem (2.3) is clearly the same as our original discrete

problem (2.1) and thus has a solution u h E Who Since B is

bounded in fR4 ~h is bounded and thus the problem (2.4) has

s'
a solution (Jh E tDh (in fact we may consider tiPh to be a
fin~te dimensional since Wh is finite-dimensional). Thus (cf.

[2])~: ~h x Wh + fR has a saddle point ((Jh'u h ) E ~h x Wh and

the extremality relations can be written:

V T E !Ph' (2. 5a)

(2 • 5b)

We shall prove the following result:

Theorem 1. If the safe load hypothesiS (1.9) is sat~sfied,

then for a subsequence {h} tending to zero

(Jh + (J weakly in H,

uh + u weak* in Bb W) I

uh + u in [Lp W) )2 for ~ p < 2I

0
where (o/u) E !" x BDW) satisfies (1.11).

Proof. Since -div X = f where X is g~ven by (1.9) we have by

Green's formula

(2.6)

where we consider X to be an element of ~h in the natural

way i.e. by letting

 317

x xlu int K inter~or of K.

int K'
K
X xlu s .
s
Now, tak~ng ,= X in (2.5a) we obtain

(2.7)

so that

IloJI ~ C, (2.8)

with C independent of h. Next taking, =X + 0, l' where

1,,1 < 1, in (2.5a) we get using (2.5b) and (2.6)

[o'1,E(u h )]h ~ (oh'X + OT 1 - 0h) - [X - 0h,E(uh)]h

wh~ch shows that

II uJI BD (n) < C. (2.9)

By (2.8) it follows that 30 E P such that for a subsequence

{h} tending to zero

weakly in H.
o
Further, by (2.9) it follows, using the fact that BD(~) is

compactly contained in [L (~)]2 for < P < 2 (see [5]), that

p
3u E ~(n) such that

weak* in Ifu(~),

in [L (~)]2.
P
It rema~ns to pass to the l~mit in (2.5). Given w E V there

ex~sts a sequence wh E Wh n V such that wh + W strongly in V.

If wh E Wh n V then
318

and thus

Th~s proves that

(a,E(w)) = (f,w) Vw E V,

Le.

-(diva,w) = (f,w) Vw E V.
F~nally, by choosing T ~n (2.5a) smooth we f~nd since then

that
- 2
o~ lim{-,llahll + (Oh,T) + (d~vT,uh) + (f,u h )}

< -!llaI1 2 + (a,T) - (divT,u) + (f,u)

(a,T-a) - (U,d~VT - diva).

o
By dens~ty we then obtain (1 .11a). Thus (a,u) E P x BD(~)

satisfies (1.11) and the proof is complete . •

3. A numerical experiment

Let us cons~der the follow~ng analogue of (1 .7) ~n the

case of one space d~mens~on:

Inf F (v) , (3.1)

o
vEBD(I)
where I = (0,1) and

F (v) = f g ( I v' I ) dx - f fv dx,

I I
with g as in sect~on 2, correspond~ng to taking

B = {1 E~: 111 ~ 1}. Alternatively this problem can be formu-

lated as follows (c. f. (1. 11) ): F~nd (a, u) E P x \7 with

P = h E L2 (I): l' E L2 (I), 11 (x) I ~ 1 a.e. x E~} and

 319

v L2 (I) such that

t (O,T-O) + (U,T' - 0') -> 0 'liT E P, (3.2a)

(0' ,v) = (f ,v) 'Iv E V. (3.2b)

Here l'
=
dT
dx and ( ., .) denotes the scalar product in L2 (I) .

Let us choose the load f to be

fIx) = Af(x) , x E I,

where

,
for 0 < x < 1
:2 ,
fIx)
-
{ :-1 for < x < 1.

It is easy to check that the (un~que) solution of (3.2) is as

follows for 0 < x < ~ (o(x) is symmetric and u(x) ant~symme-

tric around x ~) :

A 3
u(x) '6(x - ~)
4
for 0 < A < 12,
1
~(x2 - TI)
o (x)
2

3 A
AX
u(x) -6- + ( 1 - "8)x,
for 12 < A < 16,
AX 2 A
o (x) -2- + 1 - "8 ,
and that no solution exists for A > 16. We note that the

solut~on is elast~c (~.e. 10 (x) I < 1 for x E I) for 0 < A <

< 12 and elasto-plast~c for 12 ~ A < 16 with the plast~c

region concentrated at the po~nt x =! (o(~) = 1 wh~le

lo(x) I < 1 for x f ~). Further, we note that the displacement

A
u has a jump at x = ~ of magnitude TI - 1 for 12 < A < 16
(see Fig. 1).

We shall now compare the conventional finite element

method for (3.1) using cont~nuous tr~al functions with the

finite element method proposed above using discontinuous

320

trial functions. Let

o = Xo < xl < ••• < xn = 1, I.] = (x.] -

l' x]. ) ,

be a partition of I, introduce the spaces

Vh = {v: v is continuous on I and viI. linear,
j = l , ••• n}, ]

Wh ={v:vI I . lJ.near, j = 1, ... n},

]
and consider the following discrete analogues of (3.1):

Inf F(v), (3.3)

vEV h

Inf F(v), (3.4)

vEW h
Here
n n
F(v) E f g ( lv' I) dx + EI[v(x )]I-(f,v)
j =1 I j j=O ]

for v E Wh , where [v(x j )] denotes the jump of v at node x j .

To solve (3.3) numerically we used Newton's method on the non-
linear equat~on F' (u h ) = 0 characterizing the exact solution

u h . In order to apply the same method for (3.4) the non-

differentiable functional F was first regularJ.zed by repla-
cJ.ng the term I [v (x]) ] I by
l[v(x·)]1
e: g ( e:]) ,

with e: small. The problem (3.4) was then solved approximately

by applying Newton's method to the regularized problem. In

the computations we used e: = 10- 7 and the iterations were
terminated when the relative error between two consecutive
iterations was less than 10- 7 . The number of iterations re-
qUJ.red was then in the range 5 - 10.
We considered some different partitions. By o(N) we
denote a uniform partition with N intervals. Then if N is odd
the jump discontinuity in the displacement at x = , will
 321

fall in the middle of an interval in o(N). G1ven a partition

o(N) with N odd we introduce four partitions ok = 0k(N),

k = 0, .•. ,3 obtained by adding to o(N) the mesh point
1
Yk =i + k 12N' k = 0, •.• ,3.
Thus, 00 has a mesh point at the jump in u and ok' k = 1, ... ,3
are modifications of 00 obtained by shifting the mesh point
at the jump sl1ghtly (c f. Fig. 2).
The error between the exact solution (u,a) given above
and the solut1on (uh,a h ) of the discrete problems (3.3) and
(3.4) (here a h = u'h in each subinterval) is as follows for

A = 15:

Ilu-uJI L (I) II a-a J IL (I )

2 2
cont. (V h ) discont. (W h ) cont. (V h ) d1scont. (Wh )

° (9) 0.64 -1 0.64 -1 0.23 0 0.23 0

°0 0.64 -1 0.46 -2 0.23 0 0.12 0
°1 0.58 -1 0.31 -1 0.22 0 0.13 0
°2 0.55 -1 0.48 -1 0.21 0 0.18 0
°3 0.55 -1 0.55 -1 0.21 0 0.21 0

° (19) 0.34 -1 0.34 -1 0.12 0 0.12 0

0.34 -1 0.11 -2 0.12 0 0.-2 -1
°0
0.32 -1 0.18 -1 0.11 0 0.70 -1
°1
0.30 -1 0.28 -1 0.10 0 0.90 -1
°2
0.30 -1 0.29 -1 0.10 0 0.10 0
°3
0(39 ) 0.19 -1 0.19 -1 0.57 -1 0.57 -1
°0 0.19 -1 0.28 -3 0.57 -1 0.31 -1
°1 0.18 -1 0.12 -1 0.53 -1 0.35 -1
°2 0.17 -1 0.18 -1 0.50 -1 0.45 -1
°3 0.17 -1 0.17 -1 0.50 -1 0.50 -1
322

We note that in the case of discontinuous trial func-

tions we have for the partition 00 with a meshpoint exactly
at the jump of u at x = ! an O(h 2 ) convergence in displace-
ments and O(h) in stresses. We also note that the reduct~on

~n the error in displacements is considerable when adding the

mesh point at x = ! (that is when passing from o(N) to 00(N))
while the reduction in the error in stresses ~s less pronoun-
ced. However the error in displacements is very sensitive to
the location of the mesh point close to the jump ~n u as ~s

clear from partitions 00 to 03.

Let us also give the jump in u h at the meshpoint clo-
sest to x =! for the partitions 00 to 03. The jumps in u h at
other meshpoints are very small. The jump ~n the exact solu-
t~on u at x =! is 0.25.

N =9 N = 19 N = 39

°0 °1 <5
2 °3 °0 °1 °2 °3
°0 °1 °2
°3
jUlll' 0.25 0.18 0.11 .0 0.25 0.22 0.18 0 0.25 0.23 0.21 .0
in~

Thus, we see that the discrete solution u h w~ll have a

jump at the meshpoint closest to ! if this meshpoint is suffi-

ciently close to !.
 323

u
->.--___ 1

x=1 x

Fig. 1

<5 (3) ,

0 3 (3)

Remark. The l~m~t load problem (cf. [1]) can be formulated

as follows:

~nf G(v) I (3.5)

o
vEBD W)
(f ,v) =1

where

G(v) = sup <,/€(V».

,EPnC(Q)
s
If B is the unit ball and v E V we have

G(v) = fl € (v) Idx.

rl
A natural approach for the numerical solution of (3.5) 1S as
follows
324

l.nf G(v), (3.6)

VEWh
(f,v)=1

with Wh consisting of piecewise polynoml.al functl.ons of

degree k (k = 0,1, •.. ). If B is the unit ball we have if

G (v) =E J I E: (v) Idx + E J I [v] Ids.

K K S S
For k = 0 (which is probably the most l.nterestl.ng case) thl.s
method corresponds to the Yl.eld ll.ne method for ll.ml.t load
problems, a classl.cal method in engineerl.ng computations (note
that in the elastic-plastl.c case considered above we must have
k ~ 1 to achieve convergence) . •

References

E. Chrl.stl.ansen, H. Matthies and G. Strang: The saddle

point of a differentl.al problem l.n "Energy methods l.n
fl.nl.te element analysl.s" ed. Rodl.n, Zl.enkl.ewl.cz and
Glowinskl., Wiley (1979).

2 I. Ekeland and R. Temam: Convex Analysl.s and Variatl.onal

Problems, North Holland, Amsterdam (1976).

3 C. Johnson: Existence theorems for plasticl.ty problems,

J. Math. pures et appl. 55 (1976), 431-444.

4 G. Strang and R. Temam: Functions of Bounded Deformatl.on,

Universite de Paris-Sud (1979).

5 P. Suquet: EXl.stence et regularl.te de solutions des

equatl.ons de la plastl.cl.te parfal.te, These 3me cycle,
Universite Parl.s VI (1978).
Fmite Element Analysis for Combined Material and Geometric
Nonlinearities

M. A CRISFIELD
Transport and Road Research Laboratory, Crowthorne, Berkshire, U. K

Introductl.on
For a number of years, the author has used the finite ele-
ment method to investigate the collapse strength of thin
plated steel structures [1-3]. The work has been directed
primarily towards steel bridges which are usually fabricated
from engineering steel for which the stress-strain curve
exhibits a significant plateau. The collapse behaviour
usually involves an interaction between material and geo-
metric non-ll.nearities and is influenced by initial geo-
metric imperfections and residual welding stresses. The
present communication descrl.bes a number of numerical
techniques that the author has developed in order to analyse
such structures. The topics covered l.nclude approximate
yield criteria, accelerated iteratl.ve methods and incremen-
tal solutions using a 'length constraint'.

Yield criteria involving stress resultants

When analysing steel plates and shells, the most accurate
way of treating plasticity is to apply von Mises yield
crl.terion in conJunction with a numerical integration
through the thickness [4-6]. Up to eleven layers of inte-
gration stations have been used [4]. Unfortunately such an
approach can lead to large requirements for computer stor-
age and time. Since only modest computer facilities were
aval.lable, the author adopted an alternative approach and
applied an approximate yield criterl.on that is a direct
functl.on of the Sl.X stress resultants [1]. The savl.ngs
326

in computer storage are obvious but there was also found to

be an appreciable saving in computer time [1]. This fol-
lowed partly from the reduced number of 'disc accesses' but
also from the reduced number of iterations required to
achieve equilibrium. The latter reduction was caused by the
smaller number of 'stresses' moving on the yield surface
during the iterations.

Using the von Mises yield criter10n and assuming that the
equivalent stress is at yield throughout the depth of the
plate or shell, Ilyushin [7] derived a yield surface involv-
ing the six stress resultants M ,M,M ,N,N and N To
x Y xy x Y xy
this end, he defined the follow1ng non-dimensional quadratic
stress intensities.

1 (N2 + N2 - N N + 3 N 2) N
Qt X Y X Y xy
N2 N2
0 0

l:..(M 2 + - M M + 3 M 2)
M
Qm H2
xy
M2 X Y X Y M2
0 0 ••• (1)

1 - 10M N
Qtm N"11(Mx Nx + My NY 2 x Y
o 0

102MYNx + 3 M N ) MN
xy xy MN
o 0

where N is the uniax1al yield force 0 t and M 1S the uni-

o o 0
axial yield moment 0 t 2 j4. The stress intensities were
o
related by means of complex parametric relationships of the
form:

Qt fl (<P ,V), Qm f2 (<P ,V) , Qtm f 3 (<P,v) (2)

where the two parameters <P and 11 are

e i2 e iO
<P , 11 ... (3)
eil eil
 327

and e il is the equivalent stra1n on the top surface, e i2

the equ1valent strain on the bottom surface and e io the
minimum value of e i . Unfortunately, equat10ns (2) are not
1n a form that are suitable for computat1on. Consequently
a number of attempts have been made to approximate equations
(2) by eliminat1ng the parameter ¢ and~. Ilyushin proposed
the approximation:

1 (4)

Robinson [8] has shown that this is the best of the various
approximations that are llnear in Qt' Qm' Qtm space. The
author [1] invoked normality and treated the yield function
in a flow sense to derive elasto-plastic modular matrices
relating the increments of the stress resultants to the
increments of the total 'strain'. The matrices were incor-
porated in an incremental/iterative finite element formu-
lation and appl1ed to a number of problems involving the
collapse analysis of imperfect steel plates. Reasonable
agreement was obtained with both alternative solutions
using a layered von Mises approach and with experimental
resul ts [ 1] .

For problems involving sign1ficant elements of compress1ve

loading, the computed collapse loads tended to be too large.
This follows from the basic assumption that the equivalent
stress is at yield throughout the full section. Consequent-
ly, no allowance 1S made for the softening 1nduced by 'fibre
plasticity' before the attainment of 'full section yield'.
For bending dominant situations, this 'full sect10n yield'
can strictly only occur as the equivalent plastic curvature,
xps , tends to 1nfinity. Consequently, an approximate method
was proposed [9] to allow for this 'fibre yield' by involv-
ing the equivalent plastic curvature in the yield function.
To this end, the Mo term in equation (4) was replaced by
a(XpS)M O where a was chosen so that a(Xps)M o follows the
uniaxial moment plastic curvature relationship w1th
328

increasing equivalent plastic curvature. i.e. equation (4)

is replaced by

1 ( 5)

The following expression for a was proposed in Ref. [2].

a = 1.0 - 0.4 EXP(- 2.6

VIxApS ) • •• (6)

where Xps is the non-dimensional equivalent plastic curva-

ture obtained by summing the incremental equivalent plastic
curvatures

~x~s =(3EO:Y

= (3EO:r
E is Young's modulus and ~ '~X and ~Xpxy are the incre-
px py
mental plastic curvatures.
As X +00, a + 1 and F2 (equation (5)) coincides with the
ps
original Ilysushin approximation of equation (4). When
Xps = 0, a = 0.6 and equation (5) gives

1 (8)

which is a reasonable approximation to the true expression

for first fibre yield:

(9)

For imperfect plates subject to uniaxial compression,

application of this modified yield criterion gave computed
collapse loads that were about 5% lower than those obtained
using the original Ilyushia yield function. The maximum
loads were very similar to those obtained using a layered
 329

von Mises approach. However, following collapse, the com-

puted load/shortening curves dropped too sharply. In an
attempt to improve this situation, the author replaced
Ilyushin's approximate yield surface (eqn. (4)) by a more
accurate approximat~on due to Ivanov [10]. Ivanov's yield
surface is given by

- ~ 1 ... (10)

The equivalent plastic curvature may be ~ntroduced into the

yield funct~on [11] using a s~ilar technique to that
described for Ilyushin's yield criterion. The method has
been used to study the collapse behaviour of simply suppor-
ted imperfect steel plates subJect to uniaxial compression.
The results are compared ~n Fig. 1 with alternative solu-
tions obtained [6] using a five layered von Mises approach.
The agreement is very satisfactory.

Accelerated ~terative solut~on procedures

In his early work, the author used the standard solution
procedure combin~ng ~ncremental load~ng with mod~fied
Newton-Raphson (mNR) ~terations. Recently, the mNR method
has been replaced by accelerated iterative techniques
[13-14] that are related to quasi-Newton methods [15,16].
Two such procedures are related to the BFGS up-date [16,17]
and will be discussed here.
Using the BFGS method, the displacements p are up-dated
according to

p. + n; /). ... (11)

-~ -,--~

_K.- l
where /).
-~
_~ 2'i ... (12)

g. is the current gradient (of the total potential energy)

-~ out-of-balance force vector and K.-
or _~
l is up-dated from
 w
w
o
DimenSions and properties
alb = 0875
E = 207000 N/mm 2
v
- - - - Present solutions (JY1od,fled Ivanov) v 003
Lu - - - Little's solution ( 6 ) Ivon MlSes) 00 = 247 N/mm 2

v :: free

10, ~ ~ 10rl----------------------------------------~

09 09

08 08
"C
] 07 E 07
.t:: .t::

~ 06 ~06
0- .g
~ "C
"C
.3 05 !!l05
...J

04

02

01 01
OLr__-L__-L__ ~ __ __'L-~L- __L-__ ~ __ ~ __ ~ __J
o·
o 02 04 06 08 10 12 14 16 18 20 o 02 04 06 08 10 12 14 16 18 20
Shortening/squash shortening Shortening/squash shortening
la) IMPERFECTION CJ. o= 0001b (b) IMPERFECTION CJ.o = 0005b
FIg 1 NON-DIMENSIONAL LOAD SHORTENING RELATIONSHIPS FOR UNIAXIALLY COMPRESSED SIMPLY SUPPORTED PLATES
 331

-1
~i-l according to
-1 T
K.
-1 -1 ~i-l ~i-l Yi ~i-l
-l ~i-l T
y. o T
~i-l _l -l-l Yi

+~ +
h
T

II i-l
-1
~i-l y.
~i-l
-l
T
Y
Y'
' 1T
l -l-

~l-l
~i-l
y.
_l
T
.. (13)

where y.
_l g.
-l
- 2i-l
... (14)

lli is a 'step length parameter' which may be found using a

line search [15) but has usually been taken as unity in the
author's work. If

*
o. -1
• .. (15)
_l - ~i-l 2i

is the standard mNR iteration neglecting any up-dating of K,

sUbstitution of equation (13) into equation (12) gives

o. * *
-l hl~i + filli-l~i-l + (1 - hi)~i-l • •• (16)

*
where ~i-l is given by

* -1
- K
-1
• •• (17)
~i-l ~i-l 'h-l -a

fixed matrix K The scalars h. and f.

and
~i-l becomes a -a l l
are given by

a.
l
h.
l ~
l
... (18)
d. t.
f.
l
- ~
l
+ (h.l - 1) (1 - -.2:)
b.
l l

ai' b i , d.l and t.l are scalars glven by the inner products
332

a.
~

b.
~

• •• (19)
d;
*
=(0. -
* T
0'1) g .
.,. -~ -~- -~

*
(~1'
* 1) Ty .
o.
.-
-
-~- ~

Equation (16) defines a memoryless 'one-step' version of a

vectorised BFGS method given by Mathies and Strang [17] .
The procedure is also a scaled vers~on of a new conjugate
gradient method due to Shanno [ 18] .
By making the approximation

••• (20)

substitution of equation (13) into equation (12) gives

* . .. (21)
~i = h~~i + elni-l~i-l

where hi has already been defined (equation (18)) and

c.
e.
~
hi(l - ~)
b. - 1 ... (22)
~

and c. * T
o. . .. (23)
~ -1 Yi

Equation (21), which is Crisfield's faster mNR iteration

[ 12], requires the storage of one less vector than does
equation (16).

The two methods (equations (16) and (21)) have been applied
to a number of elastic and elasto-plastic large-deflection
problems [13]. Both techniques have been found to require
substantially less iterations than the standard mNR method.
The extra computation per iteration is almost negligable,
the only penalty being the extra storage. In general,
 333

equation (16) has been found to be marginally more efficient

than equation (21) but this slight advantage is compensated
by the reduced storage required by the latter equation.
The two accelerated methods and the standard mNR procedure
have each been used to analyse a clamped cylindrical shell
subject to uniform pressure loading. A five-by-five mesh
was used to ideal1se a quarter of the shell and line searches
were ommitted (i.e. n i , equation (11) was set to unity).
The number of iterations required for the accelerated method
of equation (21) is compared with the number of iterations
required by the standard mNR techn1que in Fig. 2 (fixed
load increments). Clearly the accelerated procedure has
led to a very sign1ficant improvement in the convergence
characteristics. The performance of the accelerated method
of equation (16) was found to be very similar [13] to that
of equation (21).

Incremental procedures using 'length control'

Generally in structural analysis, the iterative solution
procedures are not used to find a single equilibrium posi-
tion for a given f1xed load. Instead, a complete equil1-
brium path is often traced. Clearly advantage should be
taken from the adjacent equilibrium positions. This can be
achieved by automatically adjusting the sizes of the load
increments to account for the degree of non-l1nearity [ 3,19].
Alternatively, the load level can be continuously adjusted
as the iterations proceed. Such approaches have been pro-
posed by Bergan [19] and Riks [ 20]. The author has had
considerable success in adopt1ng a modified version [ 14]
of Riks's procedure [20]. The method 1S not only faster
than the standard fixed load mNR procedure but also allows
limit points to be passed [14,20]. The technique can be
used in conjunction with the standard mNR method or with
the accelerated techniques that have previously been dis-
cussed [13,14] The former application will be br1efly
described.

g will be taken as the total fixed applied load vector

while the scalar A will represent the 'load level'. The
334

0020

Fixed load

0015

N
E
E
Z Fixed length Increments
0-
,j
..... I = no of accelerated Iterations
~ 0010 ii, JI J = no of standard m N·R Iterations
"a.
E
~c
:::)

0050

R = 2540mm, L = B = 508mm
E = 3105 N/mm2, v = 0 3, t = 3 175mm

OL-----------------~--------------~~--------------~
o 10 20 30

Central deflection ratio I::. I

ct

Fig. 2 FULLY CLAMPED CYLINDRICAL SHELL WITH

UNIFORM PRESSURE LOADING
 335

standard equilibrium equations can then be expressed as

o (24)

If A is now considered to be an add1tional var1able, a

further governing equation is required. R1ks prov1ded this
equation by means of the 1ncremental length constraint

(25)

where ~£ is a given fixed length 1n N+l dimensional space.

(N is the number of displacement var1ables). The author has
found it advantageous to replace equat10n (25) by the simp-
ler expression

(26)

Riks directly added the constraint of equation (25) to the

N equilibrium equations of equation (24). He then solved
the complete system using the Newton-Raphson procedure.
Unfortunately this destroyed both the symmetry and the
banded nature of the original equilibrium equations. An
alternative approach is possible and follows from a differ-
ent presentat10n of the equil1br1um equations. i.e.

g(A + OA) g(A) - OA g o (27)

Following from equation (27), the 1terative change given by

the mNR method (for the unknown load level Ai + oA i where i
1S the iteration number) is

O.
-1
-K
-1
g(A. + OA i)
- 1 _o.1 (A.)
1 + OA.1 ~T ... (28)

where o (A.)
-1 1
-K
-1
g. (A .)
-1 1
... (29)

is the standard (fixed load level) mNR change and 0T is the

f1xed (and stored) tangential displacement
336

-1
~T K g • •• (30)

OA i is as yet unknown. However,

lip. + O. (31)
-1 -1

and substitution from equation (28) into equations (31) and

(26) (for lIPi+l) leads to the following simple scalar quad-
ratic equation for oA i .

o (32)

(33)

The appropriate root is chosen with a V1ew to maintaining a

positive 'angle' between the incremental displacement vector
at the present and past iteration. Details are given in
[14] which also describes the application of the length
constraint to the accelerated iterative methods.

As an example, the clamped shell of Fig. 2 has been re-

analysed using 'fixed length' rather than 'fixed load'
increments. The lengths were chosen to coincide with those
obta1ned from the previous 'fixed load increment' analysis.
The results are shown below the curve in Fig. 2. Clearly
the imposition of the 'length constraint' to the mNR method
has dramatically 1mproved its convergence characterist1cs.
Further improvements are gained (Fig. 2) by adding the
acceleration (of equation (21)). However, the accelerat10n
is relatively less benefic1al once the length constraint
has been added. The addition of the length constraint only
 337

marginally increases the computer time required for each

iteration.

Further work is required before the most efficient solution

strategy can be defined. However, the simpllcity and effect-
iveness of the standard mNR method, when combined with the
'length constralnt', make thls technlque a strong contender.

Acknowledgements
The work descrlbed in this Paper forms part of the programme
of the Transport and Road Research Laboratory and the Paper
is published by permission of the Director.

References

1 Crisfield, M.A.: Large deflectlon elasto-plastic buck-

ling analYSis of plates using finite elements. TRRL
Report LR 598, Crowthorne, England, (1973).

2 Crisfield, M.A. and Puthli, R.S.: Approximations in the

non-linear analysis of thin plated structures. Finite
elements in non-llnear mechanics, Vol. 1, Tapir Press,
Trondheim, (1978) 373-392.

3 Crisfield, M.A.: The automatic nonlinear analysis of

stiffened plates and shallow shells using finite
elements. Proc. Instn. Civ. Engrs., Part 2, Paper 8335,
to be published (Sept. 1980).

4 Marcal, P.V.: Large deflection analysls of elastic-

plastic shells of revolution. AIAA Journal 8 (1970)
1627-1634.

5 Backlund, J. and Wennestrom, H.: Flnite element analy-

sis of elasto-plastic shells. Int. J. Num. Meth. in
Engng. 8 (1974) 415-425.

6 Little, G.H.: Rapid analYSis of plate collapse by live

energy minimisation. Int. J. Mech, Sci. 19 (1977)
725-744.

7 Ilyushin, A.: Plasticite. Edltions Eyrolles, Paris,

(1965) .

8 Robinson, M.A.: A comparison of yield surfaces for thin

shells. Int. J. Mech. Sci. 13 (1971) 345-354.

9 Crisfield, M.A.: On an approximate yield criterion for

thin shells. TRRL Report LR 658, Crowthorne, England
(1974) .
338

10 Ivanov, E.V.: Inzhenernyi Zhurnal Mekhanika Tverdogo

Tela. 6 (1967) 74-75.

11 Crisfield, M.A.: Ivanov's yield criterion for thin

plates and shells uS1ng finite elements. TRRL Report
LR 919, Crowthorne, England (1979).

12 Crisfield, M.A.: A faster modified Newton-Raphson

1teration. Compo Meth. in Appl. Mech. and Engng. 20
(1979) 267-278.

13 Crisf1eld, M.A.: Incremental/iterative solut10n proce-

dures for nonlinear structural analysis. Int. Conf. Num.
Meth. for Nonlinear Problems, Swansea (Sept. 1980).

14 Crisfield, M.A.: A fast 1ncremental/iterative solution

procedure that handles 'snap through'. Symp. on Compo
Meth. in Non11near Struct. and Solid Mech., Washington,
(Oct. 1980) .

15 Denn1s, J. Jnr. and More, J.: Quasi-Newton methods,

motivation and theory. SIAM Review 19 (1977) 46-84.

16 Broyden, C.G.: The convergence of a double-rank min1-

m1sation 2: The new algorithm. J. Inst. Math. Appl. 6
(1970) 222-231.

17 Matthies, H. and Strang, G.: The solut1on of nonlinear

finite element equations. Int. J. for Num. Meth. in
Engng. 14 (1979) 1613-1626

18 Shanno, D.F.: ConJugate grad1ent methods with inexact

searches. Math. of Operations Res. 3 (1978) 244-256.

19 Bergan, P.G.: Solution algorithms for non11near struc-

tural problems. Eng1neering Appls. of the F1nite
Element Method, Computas, Hovik, Norway, (1979) 13.1-
13.39.

20 Riks, E.: An incremental approach to the solution of

snapping and buckling problems. Int. J. Solids & Structs.
15 (1979) 529-551

Crown copyright 1980. Any views expressed in th1s Paper

are not necessarily those of the Department of the Envir-
onment or of the Department of Transport. Extracts from
the text may be reproduced, except for commerc1al purposes,
prov1ded the source is acknowledged. Reproduced by per-
mission of Her Britannic Majesty's Stationery Off1ce.
Incremental Elastic-Plastic Deformations of Stiffened Plates
in Compression and Bending

J PAULUN, E. STEIN'
Techmsche Umversltat Hannover, Germany

Summary
The deformation process of stiffened plates (e.g. box gir-
ders for bridges) with increasing stat1c compression loads
until critical states is investigated in the frame of geo-
metrical nonl1near theory. Instead of geometrical imper-
fections a constant normal load is applied on the plate.
The elastic-plastic mater1al is described accord1ng to v.
Mises with an isotropic hardening. Partly plastifY1ng over
the thickness of the plate 1S considered in the FE-descript-
ion by using elements with the full plate thickness. So,
neither layered elements nor smeared elements for plate and
stiffeners are used. Two methods for the numerical treatment
of incremental elast1c-plastic compression and bend1ng are
developed and compared, the first follow1ng a proposal of
Pfluger and the second improving the first by an equilibrium
iteration 1n order to reduce the errors with respect to the
Prandtl-Reuss equations. Up to now a Tresca-material was
mostly used for plate bending. Concering the FEM, rectangu-
lar elements (12-parametric for bending and 8-parametric for
plane stress) are used for the plate and beam elements for
the stiffeners. Calculat10ns w1th various st1ffness para-
meters give insight into the load carry1ng behaviour near
critical deformation states.

Introduction

It is wellknown from tests that so-called orthotropic plates

have large load-capacities in the region of geometrically
and physically nonlinear deformations. The postcritical
states and especially the state of Ultimate load should be
known in order to estimate a construction in the frame of
safety-analysis. Those constructions are very complex be-
cause of the stiffeners which are necessary for stabiliza-

• In honour of Prof. Dr.-Ing. Th. Lehmann on his 60th birthday

on August 10, 1980
340

tion, and the statical systems change during the deforma-

tion process.
Because of the nonlinearities tests need to be large scale,
and they are very expensive. Therefore numerical methods
- especially FE-methods - are necessary for the engineering
treatment. The discret~zed system must be investigated by
an incremental and iterat~ve technique with special con-
sideration of critical, postcritical and ultimate loadings.
The first problem arises from the discretization itself. Due
to elast~c-plastic defor.mation different directions of main
stress axis are present in the plate-layers for combined
membrame and bending forces. Therefore layered plate ele-
ments would be recommendable but too expensive for practical
calculations. On the other hand fully integrated elements
for the plate including the stiffeners are only possible
for dense patterns of stiffeners. So, integrated plate ele-
ments over the full thickness of the plate and seperate
beam elements for the stiffeners seem to be recommendable.
Those integrated plate elements for small elastic-plastic
strains but moderately large rotations are developed in
this paper. The second problem concerns the ~ncremental
calculation of critical and postcritical states because of
complicated interactions of local buckling and plastifying.
So, local unloading is possible for increasing total load-
ing.

2. Geometrically nonl~near treatment of deformation states

The geometrically nonlinear description of thinwalled plates

and shells allows in most cases - even for elasto-plastic
deformations - the assumption of small strains combined with
large rotations and displacements. For the elastic part a
linear relation of the spatial strain rate tensor D and the
o -
Jaumann stress flux! is oftenly chosen, see Lehmann [1].
For ~sotropic material it has the form
o v ]
~ = 2G[~ + 1-2v tr(~)~ (2.1)
 341

(G: shear modulus, \I: Poisson's ratio).

For small strains it becomes a linear relation between the
Green strain tensor E and 2. P~ola-Kirchhoff stress tensor
S

(2.2)

Correspondingly a flow rule for large plast~c deformations

with plastic incompressibility (compare Paulun [a])
(2.3)

simplifies for small strains into

• •[ 1"3
~p = A ~ - tr (~)!
] • (2.4)

For proofing the above statement the material deformation

gradient
dx (lx i
F := ~X N.tS Nk
o k -~ -

(with: ~i fixed orthonormal base vectors in space, Xk La-

grangian and xi Eulerian coordinates) is split by polar
decomposition

F R U (2.5)

into the rotation tensor R and the stretch tensor U. For

small strains it holds

U = 1 + E~ , (2.6)

with I~II ::~~I :: I~I and E -+- O. Then, because of the

transformation

F R det F (2.7)

the ob j ecti ve s tress flux ~ of Truesdell [2]

342

b -1 • T
T := (det F) F S F (2.8)
- -
has the relation with the Jaumann stress flux ~
b 0

- --
T = T - T D - D T + tr(D)T (2.9)

Introducing the material law (2.1) into (2.9) it follows

(G1 - T)D + D(G1 - T) +[2GV 1 +

1 - 2 v~Jtr(~)
--_ . (2.10)

Due to the above assumptions, ~ lS of hlgher order small

compared with G1 so that (2.10) simplifies to

0
b
T = 2G~ v
+ 1-2V tr (!?) 2J = T. (2.11)

On the other hand from (2.7) and (2.8) follows

b
T = R 5 RT (2.12 )

and from the known relation

(2.13)

we finally get

(2.14 )

Introducing (2.12) and (2.14) into the material equation

(2.11) the prediction (2.2) comes out immediately. Corre-
spondingly, the assumption of small plastic strains FTF
-p-p
= RTR
-p-p
= 1 is used in (2.3), and leads to the material equ-
-
ation (2.4).

The tensors E and S allow a Lagrangian representation of

large deformations. A linearization of Green strain tensor
E is only possible incrementally as described below. Other-
wise large rotations can not be considered. For bending of
a plane plate in the initial configuration Kirchhoff hypo-
thesis is used. Then the kinematic variables are reduced
to the displacements and the rotatlons of the middle sur-
 343

face of the plate. The numer~cal treatment of plates ~n

compression and bend~ng uses the FE-method. The k~nematic

var~ables of the discretization process are the nodal dis-

placements, row matrix y. The corresponding load vector
result~ng from the given load is R.

3. Incremental representation for discretization methods

The total load, given by the column matrix~, is applied

with ~ncrements ~RN (N = 1, 2, .•. , n). A total strong non-
linearity ~s mostly weak ~n the ~ncrements, so that the
deformat~on process only needs few iterations for a single
load step. Furthermore load history of an elastic-plastic
material can be better descr~bed. The material tensor is
changed after each load step and improved during the iter-
ation process. For the calculat~on of the ~ncremental dis-
placement vector ~Y from N to N + 1 we postulate that the
whole conf~guration ~s known after step N. The principle
of virtual work for calculating the incremental displacement
vector ~V can be written ~n the form

Jtr[o~~(~N + ~~)] dV _ o~yT(~N + ~~) o (3.1)

Vo

This"total Lagrangian representation (T.L)" is wellknown

from Bathe, Ramm, Wilson [3] and was used also by Bathe,
Borlourchi [4] , Washizu [5] , Atluri [6] , Larsen, Popov
[7] a.o. For shorter representation matrix notation is used
for the v~rtual work of the g~ven load in (3.1). With

~E sym {~FTFN + .1. ~FT~F}

2 - - (3.2)

o~E sym {o~FTFN} + sym {o~FT~F}

we get fran (3.1) by using the symmetry of SN and ~S

344

f tr(o~E ~S)dV + f tr(o~FT~F SN)dV

Vo Vo
(3.3)

= [o~yTBN - f tr(o~~T~N~N)dV] + o~VT~R

Vo

The two terms in brackets on the righthand side of (3.3)

vanish as follows. We replace In the prlnclple of virtual
work for state N, namely

o (3.4)

T
with oE N = sym {oF N FN }, the variation of the displacements
by the-variation ;f ~he increments, oyN + o~y, and corre-
spondingly o~N + sym {o~FTFN}. From (3.4) follows the sim-
plification in (3.3) as was shown by Klee, Paulun [9]

The collection of (2.2) and (2.4) leads to the Prandtl-ReuB

material law which is given here in incremental form

~s 2G~E + 2Gv tr(~E) 1 + 1 tr(SD~E)SD (3.5)

1-2v - a
2
with the hardening parameter a sk2 + tr(SDSD), the plastic
G - -D
tangent modulus s and the stress deviator ~ := ~ -
-j tr(~)2' This constitutive equation is a linear depen-
dence between the components of ~E and ~~ if the quantities
S, k and ~ can be kept constant by suitable average values.
For relatively small load steps we can use the initial
N N N
values s , k , ~ • In the case of pure elastic deformations
one gets 1a + O. Introducing the linear incremental material
law (3.5) into the principle (3.3) and discretizing the
system by finite displacement elements we get a cubic system
of algebraic equations for the displacement increments ~y

(o~~ contains ~y linearly, and ~~ contains ~y quadraticly).

A linearization of the strain increments (3.2) in the form

L L
~E o~E (3.6)
 345

leads to discrepances for large strains which can be cor-

rected by iteration of the equillbrium conditions with a
modified Newton Raphson process

J tr(o~pT pN ~S)dV + J tr(o~pT~F SN)dV

Vo Vo (3.7)
J ~~T(BN + ~B) - J tr(o~pT ;N+1 N+1)dV. S

Diverging from (3.3) we have to replace pN and ~N by the

N+1 -
unknown tensors pN+1 and ~ on the rlghthand side. One
can chose pN, SN- as initial values and gets iteratively
- - -N+1 -N+1 -
improved tensors ~ ,~ by the correction terms ~~ and
~s. At the end of the iteration process the righthand side
of (3.7) vanishes. Then principle of virtual work - repre-
sented by the lefthand side of (3.7) - descrlbes the equi-
llbrium conditions of state N+1 exactly. This is corre-
sponding with the vanishing bracket terms for state N in
(3.3) •

Within a theory of small strains - as presumed in this pa-

per - the correction values due to equilibrium iteration
for linearized strain-increments are small of higher order.
Then (3.3) can be reduced to

L L T N
J tr(o~E ~S)dV + J tr(o~P ~p S )dV (3.8)
Vo Vo
Por purely elastic deformations also the second term in
(3.8) vanishes. In the case of elasto-plastic behaviour
this second term in (3.8) can only be neglected if the
achieved yield stress Y is small compared with the plastic
tangent modulus ~

J tr(o~pT~p §N)dV .... 0 for Y«~. (3.9)

Vo

Por the following we restrict the formulation to plane

plates in the undeformed state. After N load increments
346

the resulting rotations of the middle surface are presumed

to be large. The FE-discretization is supposed to be adapted
in such a way that the rotation tensor RN corresponding to
(2.5) and (2.7) is approximately constant within each ele-
ment. Then we can introduce local orthonormal base vectors

(3.10)

which nearly coincide with the base vectors of the con-

vected material coordinates Xi. We define ~3 as the unit
normal vector of the undeformed middle surface and corre-
spondingly ~3 of the deformed middle surface at state N,
see fig. 1.

fig. 1. Global and local base vectors and displacements for

a load increment from state N to state N+1

Then, the displacement increments

N
- x (3. 11 )

can be represented in directions of the local base vectors.

In the flrst step,L~ is decomposed into the undeformed direc-
tions

+ + T
~_ = ~N-1 + v~2 + w~3 RN Lu. (3.12 )
+ dt at.
:= l N. (\/) N.
For the local dlsplacement gradlent H dX axJ -~ -J
i t follows

+
LF := FN+ 1 - FN RN H (3.13)
 347

and with the assumpt10n of a 11near mater1al behaviour 1n

the 1ncrements ~~ = E[~~], prlnciple of v1rtual work takes
the form
+T + +T + N ]
L [ f1.tr(o~ E[~l )dV + f tr(oH H ~ )dV
1 V vi
o 0

w1th summat10n over elements 1.

US1ng Kirchhoff plate theory for comb1ned bending and plane

stress we get from the first term in (3.14) the wellknown
matr1x representat10n of the geometrically 11near problem,
+.T+·+.
namely the expresslon o~l ~ly1 for element 1 as will be
explained in chapter 5. The row matrix ti contains the local
nodal displacements and rotations. The transformation into
global coordinates

(3.15)

leads to the element st1ffness matrix k 1 1n global coord1-

nates

The calculat10n of transformation matrix T needs special

+-
cons1derations. The nodal d1splacements uk are transformed
accordlng to (3.12) by using

RN
~k ® ~k ::
N
Rjk N.®
-]
N
~k w1th Rjk N.
-]
. ~k·
One gets
N + + N
~Uj Rjk uk or uk Rkj ~Uj . (3.16)

The local incremental derivatives in the nodes are assumed

to be small so that we approximately get, see fig. 2,

n • +
~a(a= 1,2).(3.17)
-3
348

--7-- -.,.
T( +
!J2
n. local base vectors
_1 in state N

rio local base vectors

2" -'132 -1 in state N+1

+ + +
w,ac =U3,0I. = !J3 • !Joe

fig. 2. Local base vectors in states Nand N+1

+
For the local rotational tensor R it yields

+ +
<P -w ,1
+ + + + +
+ H -<P -w :::: R (n. ·n.)N.®N
,2 -1 -J _1 -J
+ +
w,1 w,2
N.® N.
-1 -J

. + 1 + +
with the average inplane rotat1on <p 2(u 1 ,2 - u 2 ,1) :::: O. =
For the increments with respect to global directions we get

with lIa 1 = ~1
+.
(~2 -
~2) , lIa 2 = ~2
+
(~3 ~3) and lIa 3 =
. -
~3
.+
(~1 -
~1) . These rotational increments are calculated
with the transformation

RN +
H. (3.18)
 349

The components are

N + +
La 1 R11 ¢ + R 13 w, 2
N + N + (3.19 )
La 2 -R 21 w , 1 - R22 w,2
N + N +
La 3 -R 32 ¢ + R33 w , 1

So, the transformation matrix Ti for each element i can be

computed. The initial values for state N+1 are obtained using
the nodal displacements uN + Lu for the calculation of RN+ 1 ,
not (3.18).

The solution of systems of linear algebraic equations for

the increments of displacements and rotations needs special
techniques for postcritical states. Riks l10, 11] has given
a modification of Newton-Raphson iteration using a normal
to the tangent in each load-displacement diagram. So the
equation-matrix remains regular beyond the snapplng pOints.
Wessels [12] has developed a similar method wi th symmetrlc
matrices, and he has obtained good results for postcritical
buckling of spherical shells.

4. The elasto-plastic deformation of a plate in bending

A closed solution of elasto-plastic bending of plates for

ductile metals (J 2 -materlal with isotropic hardening) is
only possible for special cases. This results from the
fact that the yield condition etc are formulated in stresses
and strains whereas in the plate theory we need integrated
quantltles like bending moments and curvatures. Former in-
vestigations are restricted either to the simpler problem
of brittle metal materials (Mohr, Coulomb, Tresca) or to
approximated yield conditions for the bending moments
(B!cklund [13], Eggers, Kr6plin [14] e.a.). In the first
of the following calculation methods for ductile materials
we use a proposal of PflUger [15, 16] where a "smearing"
over the thickness of the plate is also used. In the second
method given here, the first method is improved in a sense
350

that the Prandtl-ReuB material law 1S fulfilled by a post-

iteration of equilibrium conditions.

4.1 The assumption of Pfluger is that the plastic strain-

rates are proportional to the total deviatoric strain-rates

·D (4. 1 )
nE

with
:= E -"31 tr(E)1.

The Prandtl-ReuB material equation is fulfilled by Pfluger's

assumption if the stress deviator is proportonial to its own
·D D
rate, ~ = K~ • For the example of a rectangular plate, see
fig. 3, the d1rections of main stresses and the stress rates
coincide in regions which are responsible for the buckling
behaviour.

--
*+ + *+ ---- + +
.-
.-
.-
.- / "- / , ..-
..-
directions of main
stresses before

--
.-
buckl1ng

-
.- directions of main
---- --
stresses due to

* +
*
buckling
r- I-
r-
e.... / "- / , l-
~

fig. 3. Buckling of a simply supported rectangular plate

The proportional factor n in (4.1) can be derived from the

relations

(4.2)

cos a (4.3)
1/2

3 (E - T) (4.4)
3E - (1-2v)T
 351

Here E is Young's modulus, T the tangent modulus in the

unlaxial tenslon test, and a the angle between the change
of stress and the outer normal at the yield surface, see
fig. 4.

Yleld surface f a

fig. 4. Y1eld surface, outer normal and angle a

For a neutral change of stresses we get a = 2' and therefore

If

·pl
= o. The relations (4.2 - 4.4) lead after some trans-
-
E
-
format10ns to mod1fied elasticity constants, compare Pfluger
[15] ,
lI/ 3 (1-n)
E ":::"2-;-(-::-1+:-v-;)~+""':"""(71:--n"""")'-:-:(1'-"2'--v') E , (4.5)

1+v - (1-n) (1-2v)

2(1+v) + (1-n) (1-2v)

These constants change for each load increment so that the

plastlfying plate 1S calculated as an elastic system In each
step from N to N+1. An evaluation of (4.4) is given in fig.s.
One can see the weak change of no with respect to v. For
v = 0,5, (4.4) is reduced to no = 1 - In case of an affine i.
change of stress (~ 1) we flnally get

T
n = no = 1 - E". (4.7)

For the case ~ we also refer to a representation of

Lehmann [1 7]
352

10
, 110

1,0 T
E
T
fig. 5. no in dependence of E and v

*
The values E and illv are variable over the thickness and the
coordinates of the middle surface as well.

Using FE-techniques a Gaussian-paint-integration is re-

commendable. The average values for a Gaussian-paint can
be calculated by the help of n layer-paints, see fig. 6.
* *
From the values E and v for each layer-point one can calcu-
late a weighted average. The weighting coefficients can
be obtained from a beam model.

2
I
I SP middle surface
-+-.- of the plate
I
I n
GP

fig. 6. n layer-points, LP's, at the normal of the middle

surface for one Gaussian-point GP

4.2 The second method is based on Prandtl-ReuB's constitu-

tive equation which is fulfilled iteratively using the first
method. For pure bending this procedure was described by
Paulun [18] ...Ii thin a geometrically linear theory. We want
 353

the fictitious load increment * which

~R effects the same
displacement increments ~V for the "elastic" plate with
*E *
and v as the real load increment --~R acting on the elasto-
plastic plate. *
~R
-
is calculated with the following iterat~on

scheme (j: ~teration index):

'*
a) ~R. from h), initial values ~B,
-] !II'
b) ~V.
-]
with ~R
-]
using the f~rst method according to 4.1.
c) strain- and curvature increments of the middle surface,
using ~V. and initial values depending from the inter-
-]
polation functions of the FE-model,
d) strain ~ncrements in the layer-points from displacement
increments of the middle surface,
e) stress increments from strain increments and initial
values due to Prandtl-ReuB equation,
f) bending moment- and normal force increments by numerical
integration of stress increments over the layer-points,
g) load increment ~R. by Gaussian integration of the incre-
-]
ments f), transformation and assembling of the element
terms,
h) correction of fictit~ous load increments
-lIE
~R.
-]
- (~R
- - ~R.),
-]

and back to a).

The iteration is finished if II~B - ~Bjll < E for given E«1
is fulfilled. After the iterat~on we calculate the final
increments of comparative plastic strains in each layer and
add them to the previous values.

5. Comprehensive representation of the ~ncremental- ~terative

process

We use matrix notation for the example of a 12-parameter

rectangular plane bending element and a 8-parameter rectan-
gular plane stress element. The incremental nodal displace-
ments and rotations are denoted in fig. 7. We consider the
* *
"elastic" plate with E and v and use the work principle in
the form (3.14). Matrix notation of the lefthandside yields
354

(5.1)

with the global stiffness matrix K.

., !!1

tD ~I: @ +
v
+ +
1:13 w,1 w,2

flg. 7. Rectangular element with 24 parameters (nodal dis-

placements and rotations)

+'T+' +'
The element oriented inner virtual works ov 1 kl v 1 can be
developed very similar to the geometrically linear elastic
case, see Stein, Wunderlich [1~, e.g. Puttlng all kine-
+
matic variables together in vector h we get the lntegral
- +T* +
of the inner virtual work for an element i f o~ ~ ~ dA,
;\- + + + + + +
with n {u,1 v,2 u,2 v,1 w,11 w,22 w,12 w,21}

ill
li
[<
tf1

t3
Q 12 f2
Q
*
j *
f1
ill
f2 *fp + *fG (5.2)

and the plate thickness t.

* E* *\I 0 0
* N
0
N
0
fp ~ ~ °11 °12
1 -\I
*
\I 0 0
N N
0
1-~ ° 11 0 °21
0 0 -2- 0
N
0 0 0
1-~ °21
N 0
°22 0
-2-
N N
0
°12 0 °22

influence of physical influence of geometrical

nonlinearity nonlinearity
 355

From the decoupling of bending and plane stress follows

*
that f does not change along the plate th~ckness. The f~ct~-
tious "elastic" plate only yields the init~al displacements
for the iteration so that no mistake comes out from the
above separation.

The discretization ~ g(X1,X2)~ with the unit displacement

states Q and 20 nodal displacements in ~ = {~1 ~2 ~3 ~4};
+ {+ + + + + .
~k = uk v k wk w,1 w,2} leads to the relat~on

(5.3)

with submatrices for bending and plane stress in ~according

to[1~ • The transformation into global coordinates, namely

+i (5.4)
v

follows from (3.16) and (3. 19). For a node k the transfor-
mation matrix has the form, see (3.16) and (3.19)

+
u flu 1
+
v flu 2
+
w flu 3
+ (5.5)
lIa 1
w,1
lIa2
+
w,2 l',a 3

with N

N can vanish in special cases of large rotations. Then,

instead of fla i , other relevant components of the incre-
mental rotation tensor (3.18) with the corresponding trans-
formation matrix are chosen.
356

Additional nodal variables, especially higher kinemat1c

quantities (like t,12 for the 16-parameter bending element)
need further consideration for transformat10n and assembling.
One can average or introduce torsional springs, e.g.

Considering plastic material behaviour according to chapter

4.1, we have to store the stresses u N= {u N11 uN22 u N12 } of each
layer of the plate, see fig. 6., furthermore the comparative
plastic strain £Pl and the actual yield stress Y of state N.
v 1
W1th the tangent modulus T = T(£P ) out of the characteristic
E with (4.5)v and *v with (4.6) for each
material line we get *
layer pOint. For one Gaussian pOint with n layer points we
can calculate the average values

(5.6)

with the weighting factors Kk (k = 1, ... ,n). The weighting

factors are determined seperately for plane stress and
* *
bending and lead to the matr1ces f1 and f2' see (5.2), for
each Gaussian point (GP).

After calculating the element st1ffness matr1ces, assembling

and solution of the system of algebraic equat10ns we have
the incremental nodal displacements and rotations, and
finally with (3.15), (5.2), (5.3) the local incremental
displacement grad1ent, the changes of curvatures and the
membrane forces and bending moments for each Gaussian point.

With the strain increments for a layer-point k

+
o 0 0 X3k 0
~P
000 (5.7)

o o

we get the stress increments uS1ng Prandtl-ReuB equation

 357

[~ - (5.8)

with the elasticity matrix~, the plastic tangent modulus

ET
~ = E-T and the row matrix s of the deviatoric stresses

A S!. A
(5.9)

The plastic strain increment 6S PI

calculated as

(5.10)

and yields the increment of comparative plastic strain

With these datas the determination of initlal values for

the next load lncrement is possible.

Following the more accurate method according to chapter

4.2, we have to compute the membrane and bending forces
~ = f 60dz and ~ = f 60zdz by numerical integration over the
II
- (t) - - (t) - .
plate thickness. Furthermore fictitious nodal forces of
element l (acting according to the nodal displacements !i)
are determined with principle of virtual work as follows

(5.12)

with a Gaussian integration on the righthand side. After

transformation into global dlrections

(5.13 )
358

and assembling the element terms in each node we get the

g lobal load increment bR corresponding to step g) in the
-J
iterat10n procedure, chapter 4.2. During the iteration one
can slmultaneously 1mprove the material matrix in (5.8),
US1ng adequate average values for ~, ~, Y from the last
1 teration loop.

6. Examples
In the f1rst example, f1g. 8, membrane stresses and hori-
zontal displacements of the m1ddle surface of a slmply
supported square plate under uniform load are calculated.
144 square elements in a quarter of the plate were applied.
These elements have 4 nodes with 5 DOF (degrees of freedom)
in each node. The compression stresses in a ringlike outer
reg10n effect the supporting of the inner tens10ned plate
1n states of large deflections. Comb1ned w1th this behav10ur
we see hor1zontal displacements of the support lines which
1S known as the pillow effect.

E = 21 000 kN/cm 2

v = 0,3

t = 0,.5 em
00076 load
[cm[QOt;07;/;;6--'----'---'---'---'--'---'---L-'-'-+---------'~
q = 1 N/cn,2

r
120cm
load step
6q = 0,2 q

fig. 8: simply supported square plate under uniform load

 359

In the second example, fig. 9, a simply supported rectangular

plate under in-plane load and dead weight is considered.
108 square elements (6 x 18) as used in the first example
were applied. The deflections of point A and B of the ob-
tained buckling shape, fig. 10, are compared with results
by Girkmann [21] ,see fig. 11.
2
E 21000 kN/cm a 300 cm
0.3 b 100 cm
4
19.0 cm girder t 1.4 cm
°0/25000 bending load
7.0 - 20.3 kN/cm in-plane load

girder

s
(

~7
I
w{",,,.'

fig. 9: Example and results by fig. 10: Plots of the undeformed

Girkmann for the unstiffened plate the deformed stiffened and the
buckled unstiffened plate
Go t Il<N/mmJ
<>--- '. - .. , - " - ' , - " --0

/;. Poin'
D
A} Glrkmann
Point 8

--0- FEM nonl inear

- .- F EM linear
0- .. - F'EM stiffened
w
-5 o 10 [mm)

fig. 11: Deflection of point A and B (fig. 10)

under given loads 00t
360

The improved plastic calculation concept is applied in the

third example, fig. 12 . Triangular 21-parameter elements
with 6 DOF in the corners and 4 layer-points over half of
the thickness were applied. The i ncreasing of yielding
domains over the plate area and thickness is calculated due
to uniform load normal to the plate applying 8 load steps.
The lower part of fig. 12 shows the increasing of yielding
zones over the thickness (in the upper half) for a diagonal
section. One can recognize the wellknown fact that plasti-
fying begins in the corners of the plate and from the 4th
load step on also in the middle . The results correspond to
those given by Laudiero, Tralli (20) .

";::===::::::::j E • 21 000 kN 2
ern
v = 0.3

y = 2 1. kN 2
o ern
(. C" 100. kN 2
ern
8 load i nc rements

--....;;::j APi = 0 . 08 ~P 1
(1 = 2, . . . , 8)

[em) 11'-- - - - - - - 120 - - - - -----t

1 2 3 , ,~_ ;--"-

~YER
POINTS _ . - , -~y-=- ._.-.-armlr SECTION I-I

fig. 12 simply supported square plate under uniform load.

Extending of the yield-domain during increasing load.
 361

7. Conclusions

The incremental numerical concept for the calculation of

elastic-plastic deformations of plates in bending and com-
pression given in this paper has been proved to be problem-
adequate and effective. The assumptlon of large displacements
and rotations of the middle surface of the plate but of small
strains is valid for elasto-plastic deformations up to ulti-
mate loads.

In two examples correspondence with known results can be shown.

In the third example which treats the extension of yielding
zones only the tendency of agreement with the cited paper is
possible because no comparable results were available where
Prandtl-Reuss equations are fulfilled exactly for v. Mises
material over the whole thickness of the plate.

An optimization of the number of layer- and Gauss-points for

some important elements has still to be done. This needs further
practical calculations with the method.
362

References

Lehmann, Th.: Einige Bemerkungen zu einer allgemeinen

Klasse von Stoffgesetzen fUr groBe elasto-plastische
Formanderungen, Ing. Arch. il, 297 - 310 (1972).
2 Truesdell, c.: Hypo-elasticity, J. Rat. Mech. Anal. i,
83 - 133 (1955).
3 Bathe, K.-J.; Ramm, E.; Wilson, E. L.: Finite element
formulation for large deformation dynamic analysis, Int.
J. Num. Meth. Eng., ~, 353 - 386 (1975).
4 Bathe, K.-J.; Borlourchi, S.: A geometric and material
nonlinear plate and shell element, Int. J. Compo Struct.,
.!..!.' 23 - 48 (1980).
5 Washizu, K.: Variational Methods in Elasticity and Plasti-
city, 2nd ed., Pergamon Press (1975).
6 Atluri, S.: On the hybrid stress finite element model for
incremental analysis of large deflection problems, Int.
J. Sol. Struct.~, 1177 - 1191 (1973).
7 Larsen, P. K.; Popov, E. P.: Large displacement analysis
of viscoelastic shells of revolution, Compo Meth. Appl.
Eng. ~.' 237 - 253 (1974).
8 Paulun, J.: Plastische Inkompressibilitat bei groBen
Formanderungen, Acta Mechanica 37, 43 - 51 (1980).
9 Klee, K.-D.; Paulun, J.: On numerical treatment of large
elastic-viscoplastic deformations, Archives of Mechanics
1, 333 - 345, (1980).
10 Riks, E.: The application of Newton's method to the
problem of elastic stability, J. Appl. Mech. 39, 1060 -
1065 (1972).
11 Riks, E.: The incremental Solution of some Problems in
Elastic Stability, Nat. Aerospace Lab. NRL TR 74005 U,
The Netherlands (1973).
12 Wessels, M.: Das statische und dynamische Durchschlag-
problem der imperfekten flachen Kugelschale bei elasti-
scher rotationssymmetrischer Verformung, Mitt. Inst. f.
Statik Nr. 23, Universitat Hannover (1977).
13 Backlund, J.: Mixed finite element analysis of elasto-
plastic plates in bending, Archives of Mechanics, ~,
319 - 335 (1972).
 363

14 Eggers, H.; Kroplin, B.: Yielding of plates with hardening

and large deformations, Int. J. Num. Meth. Eng., ~,
739 - 750 (1978).
15 PflUger, A.: Zur plastischen Beulung von Flachentragern,
ZAMM 47, T 209 - T 211 (1967).
16 PflUger, A.: Zur plastischen Beulung der Rechteckplatte,
Ing. Arch. il, 258 - 269 (1972).
17 Lehmann, Th.: Zur Beschreibung groBer plastischer Form-
anderungen unter BerUcksichtigung der Werkstoffverfesti-
gung, Rheologica Acta ~/3, 247 - 254 (1962).
18 Paulun, J.: Zur numerischen Berechnung elastoplastischer
Plattenbiegung, ZAMM 60, T 144 - 146 (1980).
19 Stein, E.; Wunderlich, W.: Finite-Element Methode als
direkte Variationsverfahren, in Finite Elemente in der
Statik, W. Ernst & Sohn, Berlin (1973).
20 Laudiero, F.; Tralli, A.: Finite element incremental
analysis of elastoplastic plate bending, Meccanica ~,
190 - 202 (1973).
21 Girkmann, K.: Traglasten gedrUckter und zugleich quer-
belasteter Stabe und Platten, Der Stahlbau ~, 57 - 61
(1942).
A Study of Some Generalized Constitutive Models for Elasto-
Plastic Shells

O. M. EIDSHEIM, P. K. LARSEN
The NorwegIan Institute of Techno!ogy, Trondheim, Norway

Abstract
Constitutive laws for elasto-plastic shells based on strains
and curvatures of the middle surface and the corresponding
stress resultants and stress couples are investigated. The
stress-strain relationships are obtained using the classical
theory of plasticity for work-hardening material, expressed
in generalized variables. Two expressions for the consequ-
tive Yleld - condi tion are studied, and modified to lmprove
the behaVlour. The models are verlfied against results ob-
tained by numerical integration through the thickness for
the special case of bending and stretching of beams.

Introductlon

Durlng the last decade our capabllity of predicting the load-

deformatlon behavlour of large engineering structures has
been significantly improved. Computer codes based on the fl-
nite element method and the finlte difference method, incor-
porating both the effects of large displacements/rotations
and material nonllnearitles, are available for use in the
analysis. However, in practical design work the computatio-
nal cost and computer storage requirement may render these
codes intractable. Extensive research efforts are hence di-
rected towards the improvement of the codes with respect to
computational efficiency.

Two alternative methods are available when describing the

nonlinear material behavlour in shells. In the flrst
approach the constitutive equations in terms of stresses and
strains are applied to shell layers and integrated through
the thickness. In order to reduce the computer storage re-
 365

qU1rement and reduce the computational effort the second

approach formulates the const1tutive law 1n terms of middle
surface strains and curvatures and the1r conjugated stress
resultants. Such const1tutive laws have been proposed by
Crisfield /1/, Bieniek and Funaro /2/, and Eggers and
Kropl1n /3/.

While reduc1ng the computational effort, the latter approach

requires add1tional assumptions regarding the stress distri-
bution and the propagation of the plastic zone over the shell
th1ckness. The subJect of the present discussion is the
accuracy and the eff1c1ency of such generalized const1tutive
models. For the sake of simplicity the discussion is restr1ct-
ed to the case of infinitesimal d1splacements and rotations.

Variat10nal formulation for elasto-plastic shells

In the context of a hybrid stress finite element model the

governing equat10ns of the d1scret1zed system are derived
from an incremental var1ational principle of the Hell1nger-
Reissner type /4/. Introducing the incremental complementary
energy function ~U and relaxing the interelement continuity
c
requ1rements the follow1ng funct10nal 1S defined /5/

- f ~u
T-
~Tds
S
an

(1)

Nand M S are here the general1zed stresses (stress result-

as a
ants), and U and T are the vectors of boundary displacements
and tractions respectively. Prescr1bed values are "bared",
and the summation is extended over all elements.

For a d1Scussion of the actual element model reference is

given to /4,5/. With respect to the material behaviour only
the complementary energy funct10n 1S of further 1nterest, and
366

for the present purpose ~uc is taken on the form

(2)

where

(~M ~M ~M) (~N ,~N ,~N )

11 22 12 11 22 12

The elasto-plastlc compliance matrix D is given by

l:: ::1
Note that Egs. (1) and (2) are based on Love's first order
shell theory, and that no explicite dependence on the trans-
verse shear forces is indicated in Eg. (2).

Generallzed constltutlve models

Beyond the assumption of the eXlstence of a compliance rela-

tionship the precedlng variational functional was derived
without further knowledge of the constitutlve law of the
materlal at hand. When applied to elasto-p1astlc shells the
compllance matrlx D relating increments of membrane forces
and moments to increments of strains and curvatures is need-
ed.

Elastic materials

For elastic materlals the compllance relationship is given

{::} +e ]{: }
by Hooke's law, which on matrix form is written as

{:: } [:~n :~1 (3)

wlth
 367

.! 0
h 0

and

-v
1
o

Flow theory of plast1city

The plastic part of the comp11ance relationship is derived

from the flow theory of plastic1ty 1n terms of genera11zed
stresses. The development of the theory parallels that of
the classical 1nfinitesimal theory, incorporating the con-
cept of add1t1ve decompos1t1on of elastic and plastic de-
format10ns. The Y1eld cr1terion takes the form

o (a. 1,2)

where K 1S a hardening parameter, the flow rule becomes

dA (4 a,b)

The cons1stency cond1t1on and the loading/unloading cr1terion

are der1ved from the Y1eld cr1ter1on as 1n the classical
theory.

Analogous to the classical theory it is necessary to develope

an in1t1al yield criter10n descr1b1ng the onset of plastic
deformation 1n an 1nf1n1tes1mal shell element. The hardening
rule now describes the subsequent yield surfaces as the plas-
tic zone propagates throughout the element. Note that this
harden1ng 1S related to the propagation of the plastic zone,
368

and occurs even for elastic-perfectly plastic material. F1-

nally, for materials of the latter type and for mater1als
where the stress-strain curve goes asymptotically toward a
limit1ng value a limit yield surface is defined, at which the
entire section is fully plastified.

Initial and limit yield surface

For elastic shells the stress distribut10n over the thickness

is linear, and the extreme fiber stress is

( 5)

Introduction of Eq. (5) into the von Mises Y1eld cond1t1on

g1ves the following condition for initial yielding 1n the
extreme fiber

+ 2 NM)~ 1 (6)
NoMo

The 1nvar1ants of the genera11zed stresses are here def1ned

by

2
N N2 2 - N
+ N22 N + 3N 2 (7a)
11 I I 22 12
_2
M M2 + M2 - M M + 3M 2 (7b)
I I 22 I I 22 I 2

NM N
II
M
I I
+ N M
22 22
- ~(N
I I
M
22
+ N M ) + 3N M
22 I 1 I 2 I 2

(7c)

and
1
N ha M h 2a (8)
0 0 0 "6 0

a 1S here the uniaxial yield stress. Note that the effect

o
of transverse shear stresses and the normal stress a 1S
3 3
neglected here. This 1S in accordance with the plane stress
assumption of Love's first order theory.
 369

The initial yield condition 1S g1ven by Eq. (6) when the posi-
tive value of ± NM is chosen

f F
o
- 1 o ( 9)

The most common of the 11m1t Y1eld condit1ons is that develop-

ed by Ilyushin /6,7/ for elastic-perfectly plast1c materials
2 2
FL (N + M + -1- \NM\ ) ~
= 1 (10)
N2 M2 NLML
L L 13
where

N
L
= ho
0
M
L
= "23 M
1 2
4" h 0 0
(ll)

This cr1ter1on 1S based on von Mises' yield cond1tion and

the Kirchhoff-Love assumption, and neglects the effect of
transverse shear stresses. It should be noted that Eq. (10)
is an approximation, and is as such str1ctly valid only for
bending dom1nant situat10ns /7/.

Crisfield's subsequent yield surface

Crisf1eld's subsequent Y1eld surface 1S derived for elastic-

perfectly plastic materials, and has the form /1/

1
f F - 1 (12)
c c
13

where M models the moment-curvature relat10nsh1p. For the

uniax1al case the exact relationship is g1ven by

(13)

where K = 20 /rnh). For a shell element in a multiaxial

o 0
state of stress Crisf1eld gives M as a function of the equ1-
-p
valent plastic curvature K

M (14)
370

with
-P
K
J_2_(dKP2
o 13 11

FC approaches FL in the 11mit when KP + 00, while it is in

error at init1al yielding due to the coefficient 1/13 for
the INMI term. In order to improve the behaviour at in1tial
yielding a modif1cation of Eq. (12) on the form
2 2

- + _1_ INMI) ~
(~ + -M
F*
c N2
L
a 2 M2
L 3
r.:
bNLM L
1 (16)

is now introduced where the parameter b is taken as

b(KP ) = 1 - (3 - _1_) (1 - a(KP )) (17)

13
In the further discussion the original model is denoted
Modell and the mod1fied one Model 1*.

Bieniek's subsequent yield surface

Based on the concept of translation of the Y1eld surface in

the moment hyperspace Bieniek and Funaro g1ve the follow1ng
Y1eld surface /2/
_ 2
+ M* + _1_ o (18)
M2 13
o
_ 2
Here, M* is the invariant of the components (MaS-M~S) as
given by Eq. (7b), with translations M~S determ1ned by

dM* for load1ng

as
for unloading/neutral loading

The proport1ona11ty factor B 1S g1ven by

F2
S
B (19)
F2
m
 371

where Fs is the absolute value of the gradient of the yield

surface in a dimenslonless formulatlon, while Fm is the part
of Fs contributed by bending.

The Yleld surface deflned by FB 1 reproduces neither the

correct initial yield surface nor the limit surface. In the
former case agreement may be obtained by the introductlon of
-P
the parameter b(K ), as in the modified Crisfleld model, Eq.
(16). In the numerlcal studies the orlginal Bieniek model
is denoted Model 2.

Fig. 1 depicts the various yield surfaces for the degenerate

case of a beam-column subjected to moment M and axial force
N. At initial yielding both Crisfield's surface F (KP=o)=l
c
and Bieniek's surface FB(M~S = 0) = 1 deviate from the exact
surface Fo = 1, causing delayed onset of plastic deforma-
tlons. For this slmple case It lS also observed that
Ilyushin's surface is in error compared with the exact one.

08+-----~--~~--~----~--~

02+---~-----+-----r~~+-~~

o2 o4 o6 o8 1 0
N
"N..
Fig.l - Initial and limit yield
surfaces for beam-column

Subsequent surfaces for hardening materials

The formulations of Crisfield and Bienlek are both derived

for elastic-perfectly plastic materials. It is obvious that
an initial yield surface eXlsts also when the stress-strain
curve has a hardenlng regime. The limit Yleld surface, how-
372

ever, is only defined when the stress-straln curve reaches

an asymptotic value 0L' Fig. 2. In the following the Yleld
surfaces F 1 and FB = 1 are modified to take account of
c
material hardening, using the concept of isotroplc hardenlng.

Horizontal plateau

Hardening regime

--~-
Elastic - perfectly plastic

Flg.2 - Unlaxlal stress-straln curve

Model
-------1
Crlsfleld's Yleld surface, Eq. (12), may be rewrltten as
2 2
(N + M 4s NM )!:>
16 + - 0 (20)
h 2 a 2h 4 13 ah 3 0

where
s = INMI
NM

Introduclng the equivalent stress 0c deflned by the left

hand slde of Eq. (20) and a hardenlng function H(E~S'K~S)
the yield function takes the form
- p P
f
c
= 0
c
(N o,M o,a)-H(E o,K 0)
a~ a~ a~ a~
o (21)

The plastic compliance matrlx DP is easily determlned from

the flow rule, Eq. (4) , uSlng the relatlonship

ao dOC
c
~
dN + ~ dM as
as
a.S as
dA (22)
dO c dH
dO c __ dO C
__ aa dK
-P
dH
P + -P- -P
dE dN dK dM da dK aA
as as as as
 373

For futher development a plastic work functlon lS deflned by

whlch Ylelds

H' • 1
( 23a)
ho
c

1
H' ( 23b)
ho
c

Furthermore, one has

da

-p
The equlvalent plastlc straln lncrement dE may be expressed
in terms of the invarlants of the plastlc strain and curva-
ture lncrements. Alternatlvely, the following expressions
are appllcable /5/

dsP = 1:.h die (24)

or
- -P 1
dE (25)
he;:
c

Model 2

Followlng the prevlous development Bienlek's Yleld surface

for a hardenlng material lS chosen as

o (26)

where
2 _ 2
N M*
( - + 36 (27)
h2 h4
374

Again, the comp11ance relationsh1p 1S determined from the

flowrate, using the following expression for the proportion-
ality constant

daB daB
NaS dN aS + dM aS
d>"
~S
(28)
dH
p
daB
+
dH
p
daB
- --
daB
dM*
daB
. B
d£aS dN aS dK aS dM aS
as
dM aS

The equivalent plastic strain 1ncrement is again given by

p P
Eq. (25). dH/d£aS and dH/dK aS are g1ven by Eqs.
(23), replacing 0c by 0B.

Numerical studies show that the original Bieniek model in

certain cases violates Drucker's postulate, leading to a
nonposit1ve definite comp11ance matrix. This is noted by
the fact that the parameter B, Eq. (19), becomes negative.
The reason for th1s defic1ency is that Bieniek's subsequent
Y1eld surface exceeds the 11m1t yield surface FL = 1 in cer-
ta1n regions of the stress space. In the present invest1-
gation this problem is avoided by using FL = 1 as plastic
potential 1nstead of 0B whenever this occur.

A further mod1f1cation of the model is 1nvest1gated, where

Eq. (19) is replaced by

B (29)

Here, the effect of the sca11ng factor 0o/OL depends on the

shape of the stress-strain curve. In the following numeri-
cal studies the model based on Eq. (29) and includ1ng the
parameter b as given by Eq. (17) 1S denoted Model 2*.
 375

Numerlcal studles

In the present investigatlon the materlal models are veri-

fied by numerlcal slmulatlons for the degenerate case of uni-
axial bending and stretching. Comparison is made with the
results obtalned from the classical flow theory of plasticity
uSlng numerical lntegratlon through the thlckness.

Elastlc-perfectly plastic materlals

In the following the results are' presented on a nondimension-

al form, with the scaling factors NL and ML , Eg. (11), and

Fig. 3 shows the moment-curvature relationship for a beam

subjected to pure bending, where both Crisfleld's (Model 1)
and Bleniek's (Model 2) models show good agreement wlth the
"exact" solution. In Figs. 4 and 5 the beam is subJected to
a "pre-straln" E = 0,75 E L , which is kept constant whlle the
curvature lS varled. The delay in lnltial yieldlng is here
noticeable for both models 1 and 2, while models 1* and 2*
show lmproved behavior for K < 1,2 KL . The interaction
(M-N) curves for the models are given in Fig. 6, where again
the deflclency of the origlnal models are observed. Simula-
tion studies with other values of the pre-straln show Slml-
lar results with those given here.

The results obtalned when an lnltial curvature K = 1,4 KL is

imposed before the beam lS stretched are deplcted in Flg. 7.
Here, the Ilmit surface FL = 1 was exceeded for E = 0,33 E L ,
resultlng in a non-posltive deflnlte stress-straln relation-
ship. For the remaining part of the deformatlon fB = 0 was
replaced by FL 1 as the plastlc potentlal. For thlS case
models 1* and 2* show only minor improvement over the origl-
nal models, and the results are not presented here.

The pure bendlng and pure stretchlng cases presented so far

are not indlcative for the load history in a real structure.
A more realistic test on the accuracy of the material models
376

1,0 ~ ~~
-, =
M ~
ML 0,8 /
I~

II
0,6
Thickn . Integration

L
0,4 ---- Model 1 -
.. .. . .. Mode l 2

V
0, 2
K
K
0 .0
0,0 0, 4 0,8 1,2 1,6 2 ,0 2, 4 2 ,8 3, 2 L

Flg.3 - Pure bending of beam sectlon

1,0 -- .. ..
M
ML .-/ ~
b' ,, '

0,8
V ····
0,6
L '
;/

0,d
;: /
. >'
----
Thickn . Integrati on

Model 1
./
.. ... ........
;/ Model 2

V
0, ?
K
K
0,u L
0,0 0 ,4 0,8 1,2 1 ,6 2 ,0 2, 4 2, 8 3, 2

Fig.4 - Bendlng of beam section subiected to c o nstant

aXlal straln € = O.7S- € L

M 1,0 -- .=
M
./ I~
.' ..
q.~
0,8
.....
'
L
~

0,6 /'
/. .'
I

I-
~

0,4
;/ Thickn . Integra tion
-
---- Model l'
V
... .... .. ....
/ Model 2-

V
0,2

0,0
:) , 0 0 ,4 0,8 1, 2 1 ,6 2, a 2,4 2, 8 3,2

Flg.S - Bendlng of beam section subjected to constant

axial straln € = O.7S- € L
 377

"
1,0
-...........~ "'''-'....~

0,8 0,8 ~,
.,
1\,,
~

0,6
~.,-, '\ ',
0,6
\ 1\\,'. ' \\1\\
v.

\; '\
0,4 0, 4

0, 2 0, 2

0 ,0 0, 0
0,0 0, 2 0 ,4 0 ,6 0,8 1, 0 0, 0 0 ,2 0,4 0,6 0,8 1 ,0
N N
Thickn . Integrati "" N Th ic kn . Integrat i on
-L
Mode l Mod e l I"

Model 2 Model 2"

Fig.6 - Interaction diagrams for bendlng of beam section sub-

Jected to constant aXla1 strain E = 0.75 EL

1,0
N
NL //
<,"~
~
0,8

//V
,-

0,6
/
/
V I

1/
.I
Thic kn . Integrati on
0,4 I
---- Model I
,V . . . ... . Mode l 2
1/

V
0, 2

0, 0
0,0 0,4 0,8 1, 2 1, 6 2,0 2, 4 2 ,8 3, 2

Flg.7 - Stretchlng of beam s ec tion subjected to constant

curvature K = 1.4 KL
378

1,0

-
M
ML
0,8
/~~ ..:::- :: :::::: ::::::::::: ::: ::: ::~~:;~.~.~ ..
/ r-FL'I

/
0,6
Thi ckn. Inlegroli on
0,4 Model -
0, 2 / Model 2

0,0 ~1/---+----~--4---~---4----+---4---~
0,0 0 ,4 0,8 1,2 1,6 2,0 2, 4 2,8 3,2

Flg.8 - Moment vs. curvature for proportlonal bendlng

and stretching of beam sectlon (~ = 2 ~)
KL £r.
N 0,8
NL
0,6
•••.___ •• .•.• _.•._L 1---' .'-- -
•.' ..-- >- F •I Thickn . Inleg,olion
0,4
----- .-. .-..-.~.-....
- - - - Mod el

0, 2 / Mode l 2

/
0 , 0 ¥----+----l---I----l--~
0,0 0 ,4 0,8 1,2 1,6 2,0

Flg .9 - Axial force vs. aXlal straln for proportlonal

bendlng and stretchlng of beam sectlon (~ = 2 ~)
KL £L

is the case when £ and K are varied proportionally . Two such

strain paths, K/KL = 2£/£L and K/KL = O.So£/£L' are investi-
gated. The former corresponds to a neutral axis at 1;;3 = -h/6
wlth limiting values M = 0 . 89 ML and N = 0.33 NL , while the
latter corresponds to a position 1;;3 = -2h/3 (outside the beam)
with limit values M =0 and N = NL . The results are present-
ed ln FlgS. 8 - 11. Here, both the results of Blenlek's orl-
ginal model with negative definite strain-stress relation-
ship when FL > 1, and the model obtained when using FL = 1
as a plastic potential, are included. Again the modlfled
models 1* and 2* show only minor improvements compared with
 379

0 ,4

-
0,2

0,0
0, 0 0,4 0,8 1, 2 1, 6 1,b 2, 0
0,0 0,4 0,8 1, 2
K

,.
Thickn _ Integrotion Thickn _ IntegroHon
KL
Model ---- Mod,,1
Model 2 ... .. .. ...... Mod,,1 2"

Flg.10 - Moment vs. curvature for proportional bending and

stretchlng of beam sectlon (~ = 0.5 ~)
KL EL

1 , 0 ~~~r----'-----r----'----.

0,8+-----r-~~~---r----+_--~

Thickn Integroti on

Model
0 , 4+----.r---~~---r--~~--~
Model 2

o,o ~--~~---+-----r----+---~

0,0 0 ,2 0,4 0,6 0,8 1,0

Flg.ll - Loadlng paths for proportlonal bendlng

and stretchlng of beam sectlon

models 1 and 2. The largest improvement is shown in Fig. 10

where the peak value of the moment is correctly represented.
The loadlng paths in the stress space for the two proportion-
al straining historles are glven in Flg. 11.

When subjecting the beam segment to CYCllC bendlng the moment-

curvature relationshlps ln Flg. 12 are obtained. Note that
the lowered yield pOlnt obtained when the bending moment is
reversed has nothlng to do with hardening of the material,
380

as an elast~c-perfectly plast~c material d~splays no

Bauschinger effect. As expected, model 2 reproduces fairly
closely the actual behav~our under cyclic bend~ng. Th~s ~s

due to the translat~onal quant~ty M* def~ning the subsequent

yield surface. On the other hand, model 1 exhib~ts an "iso-
tropic hardening" effect, and is hence less suited to pre-
d~ct the behaviour under reversed y~eld~ng. The results ob-
tained when subject~ng the sect~on to a pre-strain before
the cyclic curvature is imposed are not significantly differ-
ent from those presented ~n Fig. 12.

Work-hardening materials

For the case of work-harden~ng materials the un~axial stress-

stra~n curve ~s taken as

0=0
o

20
o

For the case of pure bend~ng the accuracy of the two models
~s comparable with that of the non-hardening case. The re-
sults of bending or stretch~ng of a prestrained section are
given ~n F~gs. 13 and 14. From F~g. 13 ~t ~s clear that the
correct l~m~t case M = 0 and N = NL ~s not reached for model
2*, which ~s due to the fact that the quant~ty M* ~n the ex-
pression for the equ~valent stress goes asymptot~cally to-
wards the value M* = 0.25 ML ~nstead of the correct value
M* = O. For the proport~onal load~ng path K/KL = 2 E/EL
both models reproduce the correct moment-curvature relation-
sh~p, wh~le the relationsh~p between Nand E shows greater
deviations, F~g. 15. For the case K/KL = 0.5 E/E L , however,
the opposite ~s the case, and the moment-curvature relation-
ships display s~gnif~cant deviation from the correct one,
F~g. 16. F~nally, the interaction diagram for the two cases
are depicted in F~g. 17. The l~m~t surface FL =1 ~s here
exceeded for both models. Th~s is due to the fact that f~nite

stra~n ~ncrements are used in the solution, without proper

scal~ng back to the y~eld surface.
 381

M \ ,.

r--r-'--'--'~~~-'~~-.--r-.--.
.1 , 2 -1 . 0 -0,8 1. 0 1, 2

Thickn . Integrati on
·O, B Model
~ ---- Model 2

Flg.12 - Cycllc bendlng of beam sectlon ( € 0)

1,0
N
NL 0,8

0,6
Thickn. In tegra ti on

0,4 Model "

Mode l 2 -

0, 2

0,0
0,0 0,4 0,8 1, 2 1, 6 2, 0 2 , 4 2,8 3, 2

Fig . 13 - Stretchlng of beam section of hardenlng

materlal subJected to constant curvature K =

1,0 ~~
....
~
ML / ~ r"....-

V.
0, 8

//
0,6
;/" Thickn , In tegration
0, 4 / ' :I ----
V
Model I'
... . ... .. , .. . Model 2-

V
,/
0, 2

0 ,0
! I
0, 0 0,4 0,8 1, 2 1 ,6 2 ,0 2 ,4 2 ,8 3, 2

Fig.14 - Bendlng of beam section of hardening

materlal subJected to constant axial
straln € = € L
382

0 , 8~----.---~-----r----.----.

0,6+----+----~--~--_+--__1

..,/' .----.,.r--... ______._

0 , 4+-----r,~~~__
~*==---r_--_+----~
Th ickn . Integra tion
:::--
/ Mode l 1*

V
0,2+--F~r---_+----~----t---~
Model 2*

O , O f---~r---_+----4-----t---~
0,0 0,4 0,8 1,2 1,6 2,0 £
L

Flq.l5 - Proportional stretchlng and bendlng of beam

sectlon of hardenlng material (~ 2 ..f...)
KL € L

:::~±0~:::
Thickn . Integralion

Model . "

-~-----==--
Mode l 2'

0,0 ~---;-----r----+-----~--_;
K
0,0 0,4 0 ,8 1 ,2 1,6 2,0 K
L

Flg.l6 - Proportlonal stretching and bending of beam

section of hard e ning material 0.5 ..f...)
€ L

M
ML
1,0 '-
-- - 1

0,8

0 , 6+----+-+.~~--_P~~----1

° + ----,f-\-
Th ickn . In te gration
,4 Mode l ,"
Model 2 ·

N
0,0 ~----~---+~~~----+---~ NL
0,0 0, 2 0 ,4 0,6 0,8 1,0

Flg.l7 - Loadlng paths for proportional str e tchlng and

bending of beam section of hardening material
 383

Summary and conclusions

Based on numerlcal studles the behaviour of the elasto-

plastlc material models of Crisfield and Bieniek has been
compared wlth the classical flow theory of plasticlty. For
the degenerate case of a beam sectlon subJected to combina-
tlons of bending and stretching both models give results that
agree well with the classical theory for monotonlC loadlng.
However, negative definiteness of the stress-straln matrlx
was detected in some of the examples, necessitatlng modifi-
cations In the flow potentlal. For cycllc loadlng Bleniek's
model give the best descrlptlon of the behaviour.

The adJustment of the coefflclent of the NM term in the yield

conditlons (model 1* and model 2*) seems to have a favourable
influence on the behaviour of the models, although the improve-
ment in some cases lS not large.

For work-hardenlng materlals only models 1* and 2* have been

considered. The results indicate that the behaviour of these
models lS guite similar to that observed for the elastic-per-
fectly plastic material. The effect of varying the shape of
the uniaxial stress-strain curve has not been studied in the
present investlgation.

The test cases presented here give information regarding the

general behaviour of the models in the two-dimensional stress
space. The lmportance of some of the deficiencies of the
models in a more general stress space can only be ascertalned
by the analyses of some tYPlcal shell structures.

References

/1/ Crlsfleld, M.A.: "On an Approximate Yield Crlterion for

Thin Steel Shells", TRRL Report LR658, Transport and
Road Research Laboratory, Crowthorne, Berkshire, 1974.

/2/ Bieniek, M.P. and Funaro, J.R.: "Elasto-Plastic

Behaviour of Plates and Shells", Techn. Rep. DNA 3954T,
Defense Nuclear Agency, Weidlinger Associates, New York,
1976.
384

/3/ Eggers, H. and Kroplln, B.: "Ylelding of Plates with

Hardenlng and Large Deformations", International Journal
for Numerlcal Methods In Engineerlng, Vol. 12, pp. 739-
750, 1978.

/4/ Horrlgmoe, G.: "Nonllnear Flnite Element Models In SOlld

Mechanlcs", Report No. 76-2, Dlvislon of Structural
Mechanlcs, The Norweglan Instltute of Technology, The
Unlverslty of Trondhelm, 1976.

/5/ Eldshelm, O.M.: "Nonllnear Analysls of Elastlc-Plastlc

Shells USlng ~ybrld Stress Flnite Elements", Dr.lng.
Thesls, Dlvision of Structural Mechanlcs, The Norweglan
Instltute of Technology, The Unlverslty of Trondhelm,
1980.

/6/ Ilyushln, A.A.: "Plastlcit€", Edltlons Eyrolles, 1956.

/7/ Crlsfleld, M.A.: "Some Approximatlons In the Nonllnear

Analysls of Rectangular Plates Using Flnlte Elements",
TRRL Supplementary Report 51 UC, Transport and Road
Research Laboratory, Crowthorne, Berkshlre, 1974.
An Efficient Finite Element Method for Elastic-Plastic Analysis
of Plane Frames

J BANOVEC
Umverslty Edvard Kardel] ofL]ubl]ana, YugoslavIa

Summary
A general rrethod for the elastic - plastic analysis of plane frames in-
cluding nonlinear gec:rcetric effects (large displacarentc;, large lquadra-
ticl strains) is presented. A mixed type element with three degrees of
freedan d. o. f. at earn end, and with n (n = J ,2,3, ... ) internal d.o.f.
aloog the element is proposed. The mixed approach W1th the approximation
of the axial force and the curvature presents better the elastic -
plastic state than the well kna,.m displacement approach.

Introduction

Plane frarres are alrrost the rrnst frequent structures in the civil engi-
neering and the nonlinear analysis of these structures has attained coo-
siderable interest. Many publicatioos treat ooly material nonlinearity
(elasto - plastic frarre) or only geanetrical noolinearity (large dis-
placarents and quadratic strains; large displacerrents and small strains;
the secood order effects). cnly a few papers analyse both ncnlinearities,
e. g. [Ref. 3) to [5). Mas t of these rrethods separate the frarre colunns
and beams into large nunber of elements. Usually element unifonn loads
IlU.lSt be replaced by adequate nndal forces and m:xrents.

The aim of the presented paper is to develq:> a rrore effective method,

which could decrease the necessary number of elements and will at the
same tine give sufficient accurate results.

Description o f the problem

A typical beam element is sha-m. in Fig. 1, where notation for element

loads, end actioos, displacarents, and orthononnal exx>rdinate systans
is also defined.

It is assumed that the element follows Bernoulli - Navier hypothesis,

thus shear defonnation is neglected and the cross - sectional area dces
not dlange during the defonnatioo. The element loads do not change their
386

I
I
I
I
t

Fig. 1. Beam elarent

directioo and the intensity of elemmt load resultant forces is assurred

coostant. 'lhe elerrent material follCMS sc::rte nonlinear stress - strain
diagram.

Kinematics

'ltle derivation starts with the expression for the extensiooal strain £

specialised for negligible transverse shear defonnatioo [1]:

(1)
 387

or

D_
~ =
1/10+2~ex
V·· dU
dx
+ dudu
dx dx
-1.0 (2)

or in the ccrnponent fonn

du
(1 + ~)CX)S <P = cos tjJ + dx
(3)
dw
(1 + ~)sin <P = - sin tjJ - dx

du du ~ dw ~
- = -a + -a (4)
dx dx x dx z

where z is the nonnal ooordinate; ~, DM are me dimensional stretching

and bending defonnations; u, ware translation COTpa'lents with respect
to the reference ooordinate systan (a.x , a.z ), which is defined by the
rotation tjJ (Fig. 1). The Eqs. (3) for ~ and <p are usually too c:arpli-
cated for practical use and they will be simplified. Rearranging them:

du
(1 + ~) (oos <p - 1) = cos tjJ + dx - 1 - ~

(1 + ~)sin <p = - sin tjJ - :

and assuming that the reference systan is chosen so that the local rota-
tion <p is a small order quantity (Fig. 1):

(5)

the left sides of (3*) are expanded into Taylor's series aboot the point
<p = 0.0:
ClF ;/F co2
(6)
F = (1 + ~) (oos (!> - 1) ~ F[ + Cl<P (D + -Cl<p-2 2
(peO.O
<0=0.0 <[)=O.O

G = (1 + ~)sin CD ~ Gj + ClCl<PGI (ll (7)

(D=O.O <[)=O.O

Ccnsidering (3 * ), (6) and (7), sirrolified equations are obtained:

388

D = du +cosljJ-1+ ~<p2
N dx 2
(8)
dw . d,
~ <p = - dx - Sl.n '"

~ = (1 + ~)I (9 )

<1>=0.0

where higher term<:; are neglected. O:msidering (3), ~ becx:nes:

(10)

Kinematic relations (8) are still too cmplicated, so it is supposed

that ~ is oonstant aloog the elem:mt:

+ cos ljJ (11)

If <p is small (5), the rigid local rotatioo ljJL is also a small order
quantity (Fig. 1):

(12)

Assuming that:

L IlL _ L
Lo
oos L. = (13)
LO

~ is given by:

L
(14)
~ = LO

Collecting results fran (1), (5), (8) and (14):

du ~
~= dx +cosljJ-1+ 2" <p2

dw
'N<P=- dx -sinljJ
(15)

L
~ = LO
 389

kinematic relations are d:>tained. '!hey are taken into acca.mt in deriva-
tion of equilibriun equations by the energy principle. '!he influence of
~ will be discussed later.

Almost in all publications I which analyse plane frarrES I the reference

coordinate system is identified wi th the rotated element system (e
X
,eZ) I

e. g. lit = 0.0 and it is assumed that the deformOO length of the element
L is equal to L0 (~= 1. 0) •

Stress resultants

An element cross sectioo is considered in Fig. 2.

dA

y
a
x

Fig. 2. Stress resultants

'!he stress vector t X is written in the a:rnponent form:

(16)

where due to Bemculli - Navier hypothesis a and • present the physical

a:rnponents of the stress tensor. '!he resultant force:

r= A
f tXdA=Na +Qa
x z
(17)

where:

N=
A
f adA normal force

Q= S .dA shear force I

A
390

and the bending m:mmt:

M = J iX
A
ax z dA = AJ a z dA (18)

are defined by the integration of the stress vector.

Constitutive equations

'U1e material of the element is ronsidered to be noolinear (Fig. 3):

a =a (E) (19)

Fig. 3. stress - strain diagram

All necessary equatioos will be derived for nonlinear hyperelastic

material [61:

a=
dU
de: (20)

"mere U is element defonnatioo energy per unit initial vol1.lIre. Defining

the energy per unit initial length as

u= J UdA (21)
A

it can be concluded fran (1) and (16) to (21) that the energy becanes
function of stretching and bending defonnatioos ~ and ~:

(22)
M= au
aDM
 391

In the case of real elasto - plastic material , when an irreversible

process may occur, the increnental equatioos for reversible process can
still be used i f the history of previrus load steps is taken into account
[2]. '!he Eqs. (17) and (18):

N =f a dA =N (~, ~)
A (23)
M =f a z dA =M (~, ~)
A

can be observed as two functioos relating four variables N, M, ~ and DM"

O1oosing defonnatioos as independant variables the incrE!l'lental equations
are ootained, Eqs. (1), (19) and (23):

(24)
da
A =
t f dadA
de: St =f de:
zdA
A A

I
t
= Af da
de:
z2 dA

where e. g. t:N presents the nonnal force inCrE!l'lent. '!he Eqs. (24) are
used in the displacenent awroach.

Selecting the nonnal force N and the bending deformation ~ for inde-
pendant variables a roodified fom of (24) is obtained:

(25)

'!he Eq. (24) will be used in the mixed energy principle. I t is coovenient
to introduce mixed defonnatioo energy per tmit initial length, (22):

(26)

:~ = - ~
* au* = M
aOM

It follows from (25):

aM (27)
aN
392

than the seoond derivates of U* are oormutative.

Equilibrium equations

It is well knc1Nn that equilibrium differential equaticns and statical

oonditions for the end actions can be derived using the stationary value
potential energy principle. When using the "exact" kinematic relations
(1) , which means exact acoording to Bernoulli - Navier hypothesis, the
"exact" equilibrium oonditions together with the "exact" statical con-
ditions can be dJtained; while introducing the sinplified kinematical
relations (15) into potentional energy functional the a!J!)roximate equi-
librium conditicns will be derived.

The elerrent potential energy n consist of two parts i. e. of the defor-

maticn energy and of potential of external forces:
LO LO
n (u,w) = J u (%,DM) + J Q (u,w)dx + Q (vk ) (28)
o 0

where Q is the potential of the elerrent loads and Q presents the poten-
tial of the end actions (Fig. 1):

(29)

an = - Sk; k = 1,2, •.• 6

aVk

'!he adnissible variaticns of the displacerrents ou and fJN cause deforma-

tion manges (15):

0% = o( ~ ) + % <P 6<P
dw
% o<p = - 0 ( dx ) (30)

ffi = 6( dq> ) = 6(~ )

-M dx dx

where neglecting the tenns whim contain ~, additional assurrpticn is

made. The variations of end displacements bea:ne:

OvI = 6u(0) 6V3 = c5<Il(0) Ovs = OW(Lo) (31)

Ov2 = f:w(0) 6v,. = 6u(Lo) Ov6 = &%l(Lo)
 393

Using (22), (28) and (29) the first variation of II can be stated in the
form:
LO LO
c5II = J (N 6D.~ + ~1 6DM) dx - J (p-x 6u + p z C'M)dx-
o I
o
6
-l'&6v=0 (32)
k=l -k k

Considering (30) and (31), and using the integration by parts, approxi-
mate differential equations:

(33)

d (.!..... dH _ N (0) 0
dx ~ dx + Pz =

and statical boundary ccnditions

S1 = - N(O) Sit = N(Lo)

S3 = - M(O) S6 = M(Lo)
(34)
S2 = - ( .!..... dM - N <O)x = 0
'1~ dx

S5 = (~:: - N (0) x = L°
are obtained. '!he Eqs. (33) and (34) are quite simple in canparison with
the "exact" ones [1]. I t will be shown that relatively gCXJd results can
be obtained.

In the case of small (linear) strains, the quadratic term O. 5 ~ <p2 in

(15) is neglected and correspcnding equilibrium equations and statical
oonditions are recognised by anitting the terms N <0 in (33) and (34). '!he
results strongly depend on the choise of the reference system (ax , az ),
(Fig. 1). If the reference system coincides with the rotated element
system (ex , ez), the large displacement analysis is considered. In the
case when the rotaticn \jJ is zero, the small displacement analysis is
treated. CCiTbining the rrentioned possibilities four different analyses
can be used:
- large displacement - large (quadratic) strains (Z. - Z.)
- large displacement - small strains (Z. - 8.)
- small displacement - large (quadratic) strains (s. O. t.); usually i t
is narred the seccnd order theozy
- small displacement - small strains (f. o. t.); it is the well known
first order theozy.
394

In the case of CX)nstant linear material along the element siITplified

equations can be solved analytically.

Fig. 4 shGls the analytical results obtained in the analysis of cantile-

ver subjected to an end rrorrent. '!he c:crtparison between tlolo diffen>.nt
rrethods, i. e. 2.. - 2.. and 2.. - s. shews that using the integration in-
terval L0 , very good results are obtained by the 2.. - 2.. rrethod for the
rotation up to 90 degrees. Using two equal integraticn intervals the
relative error for the end displacement w is negligible even in the case
when the cantilever defonns into a half circle.Both rrethods, i.e. 2.. - z..
and 2.. - s. give exact value for end rotaticn <». '!he error in results

U -l EI

n
Bending rigidity

Number of integral
intervals

w M LO
w=
ET

!f-
o- - - - - L °_____...,.1
One integral interval
"Exact" rrethcx:1 Simplified rrethcx:1s
~,l'\., - % error for u,w

w u/Lo w/Lo <!> z.. - 2.. z.. - s.

R R R R
u w u w
0.1 0.0017 0.0500 0.1 - 0.01 0,00 - 24.98 0.04
0.5 0.0411 0.2448 0.5 - 0.18 0.01 - 24.45 1.05
1.0 0.1585 0.4597 1.0 - 0.64 0.12 - 22.78 4.29
II/2 0.3634 0.6366 II/2 - 1.26 1.26 - 19.40 11.07
II 1.0000 0.6366 II 0.00 11.31 0.00 57.08

Several integral intervals

2.. - 2.. rrethod 2.. - s. rrethod
w n =2 n = 2 n - 4 n =8
R R R R R R R
u w u w u w w

0.1 0.00 0.00 - 6.24 0.01 - 1.56 0.00 0.00

0.5 - 0.01 0.00 - 6.08 0.26 - 1.53 0.06 0.02
1.0 - 0.04 0.01 - 5.57 1.05 - 1.38 0.26 0.07
II/2 - 0.08 0.05 - 4.56 2.63 - 1.11 0.65 0.16
II 0.00 0.07 0.00 11.06 0.00 2.60 0.65

Fig. 4. Corrparison of analytical results

 395

obtained by the Z. - s. rrethod using 8 intervals is greater than the

error in results obtained by Z. - Z. rrethod with two intervals. '!he
given exanple and other tests, which are not presented here, derocnstrate
that Z: - Z. rrethod is very efficient and suitable nt.D1'erical rrethod
should be clx>sen in the case of the elastoplastic beam elerrents.

Mixed varionational principle

In defining the elerrent p:>tential functional (28) it was asst.D1'ed that

defonnaticns were related to the displacenents (15). '!he expression for
~ seems to be quite a c:arplex ccnstraint. To relax ita new variational
principle is stated in the form:
La
I r du cp2
11 = 11 - J A. (~ - dx - cos ljJ + 1 - ~ "2 )dx (35)
o
by introducing one independant Lagrangian multiplier A.. Thus u, w, A. and
DN can be varied independently. '!he variation of DN gives:
La
OIlI = 0 = J (N - A.)O~ dx (36)
o
where U - N relaticn is given in (22). '!he Lagrangian multiplier can be
identified as the nonnal force N. Ccnsidering Egs. (26), (35) and (36)
the roodified functional ~s defined as
La La
llII = S U* dx + f 1f dx + Q +
o 0
La d ~
+ J N( ~ + COS lV - 1 + "2 cp2)dx (37)
o
Taking into account (29) and the variaticn of u:
La a
II du L
OII =0 =
o
J N o( -
dx
)dx - J PX 00 dx
0
-

a (38)
- 51 00(0) - 54 oU(L )

and using integration by parts, equilibrium equatims in the directim

(ii) are obtained:
x

(39)
51 = - N(O)

Solving thE!l1:
396

(40)

the above equatioos can be used as the coostraints; thus the elenent
mixed defonnation energy fl.IDctiooal is defined:
LO
11* (w, No) = J [U* (N,~) + Q(u,w)]dx + Q(vk ) +
aL °
+
a
J N(N0 ) (du
dx
+ oos ljJ - 1 + ~
2
c(2)dx (41)

~dent variables ~, cp, and N are expressed in (15) and (40). I t is

suitable that new independent variable No is not the fl.IDctim of x. II*
will be treated in derivation of the numerical solution of the prc:blem.

Finite element approximation

It is new a simple matter to ootain general equatioos for a typical fin-

i te elemant.
The dispZaaement lJ approximation. Returning to (41) it is ooserved that
ooly the approximation for w can be chosen. Considering the elertent as
the part of the frarre the behaviour of w is described with the shape
fl.IDctioo ws ' which rigorously satisfies nodal conditions and with the
bubble fl.IDction wb ' which defines the defonnation of rotated clarrped -
clamped elemant. Coosidering kinematic relatioos (15) and introducing
the non di.Irensiooal coordinate 1; the follewing fl.IDctions are proposed:

Ws = 12 v2 + Is Vs + c N L (1 3 v3 + Is Vs) -°
- 0.25 LO(Cu ljJ - sin ljJ) (1; - E;3) (42)
n
wb = k~7 1k vk n ~ 7

~ cp = - dx
dw
- sin ljJ D = dcp
M dx

where
x = 0.5 LO(l + E;)

12 = 0.25 (2 - 3 1; + 1;3)
(43)
IS = 0.25 (2 + 3 1; - 1;3)
 397

I3 = 0.125 ( - 1 + ~ + E2 - E3)
(43)
I6 = 0.125 (1 + E - E2 - ~3)

and for k) 7:

(44)
dIk dIk
~ (- 1) = Ik (1) = dr (- 1) = dr (1) =0

50 w, CP, depend 00 vk (k = 2,3,5,6, ••• ). '!he third order polynanials

~

(43) are recognised allro3t in all oublications, which analyse beam ele-
ment. '!he above polynanials satisfy the ccnditioo for the shape fl.U1ctioo.
The eZement equations. n* is nClftl dependent 00 vk (k:l 1,4) and 00 NO.
Equating to zero the first variatioo of n* , oonsidering Eqs. (26), (29),
(40), (41) and (42) follClftl1ng expressioos are ootained:

L ° a~
s. =
-lc
J0 (M -aV dx + c._N N cP ~
aV
- p
z
aw)dx
aVk
k k

k = 2,3,5,6,7, ••• (45)

LO LO
51 =- 0.5 J pX dx -
o
No 5
4
= - 0J px dx + No
LO
~ =.L J (- Q- + du + cos I\J - 1 + ~ cp2)dx

1
-u LO 0 ~ dx 2
LO
=.L
LO
J (- ~ + 0 G)dx
0
Q-
(46)

J
where:

~= 0
(47)
~ =0 for7~k~n

vazying the shape fl.U1ctioo Ws the expressioos for end actions 52' 53'
5s, and S6 are ootained. '!he bubble fl.U1ctioo wb does not displace the
ends of the elenent and the correspooding generalized force ~ nust be
zero (47). Similarly NO' Fq. (40), does not influence the glOOal elanent
equilibrium and the correspooding generalized stretching defonnatioo %
must be zen> (47). '!he integrand in (46) can be treated as a difference
between tlNo stretching defonnatioos, the first ooe ~ is defined by
398

oonstitutive relations and the seoond cne 0G by ga:metrical relaticns

(46).
Incremental equations. '!he incranents of ~ and % are caused by incre-
ments t.vk and mo' Fqs. (45), (46):

lISk
n
= i:2
f aM a~
t.Vi ~ ( a~ aVi + N ~ aVi aVk )dx
a<p a<p

i;&4
LO aM a~
+ llNo J(
o
aN av.k + ~ (j) ~
k
) dx

(48)

LO a~
+ t.v,+ - t.vl - m
D 0 aN
dx J-
'!he increrrents (48) can be expressed in more cam:>act fom, i f (24) and
(26) are taken into acoount:

(49)

i,k = 1,2,3, .•• ,n

and the surrmation ccnvention is intrcx:1uced. Using Net/ton - ~phson i te-

ration, Fqs. (47) take the form:

lI% (s) = _ %(s) (50)

lIS (s) =-s (s) for7~k~n (51)

k k
where (s) is the iteration m.miJer, lI% (s) and A\: (s) are oorrectioos
to %(s) and to ~ (s). In canparison with the "classical" beam elenent
(wb = 0) additional unknONl'lS (vk ' k ~ 7) and correspondinq equations
(51) are introduced. Assembling the elements the nodal point equaticns
are obtained using the knCMIl proredure [7], which is not presented here.
'Ib get symretric tangent matrix nodal forces are supposed to be ccn-
 399

servative. '!he Egs. (50) and (51) are used for the statical ooodensa-
tion of oorresponding increments NO and v k (k ~ 7); thus the total
frane matrix presents Sate mcxlified stiffness matrix and it is of the
sane order as in the case of the classical beam analysis.
Comparison of different order bubbZe funations. '!he problan can be
solved analytically only in Sate sinple cases whim can be used for
testing the el.ements with different order bubble functions.

In Table 1 the symbol P j (j = 1,2,3, ... ) is used to present the elerrent

in whim the bending deformation DM is the j-th order polynanial, it
Ireans that the oorrespcnding bubble function is of order j+2, so that the
notation for the classical beam element is Pl' The classical buckling of
the c1arrped - c1.arrqJed beam with linear oonstant rigidity is very suitable
for testing the elerrents.

~
PI P3 P"
Number of % Error % Error % Error
elements
1 6.40 0.06
2 1.32 O. O.
4 0.75
6 0.16
8 0.05

Table 1. Classical buckling of the clanped - clarrped beam

'!he results srow that the relative error for critical normal force is
negligible even in the case when cne P" element presents the beam. To
get nearly the SaIre results 8 equal length PI elerrents have to be used.

Figure 5 shCMS the elasto - plastic cantilever which was analised using
f. o. t. Nurrerical results are calculated by nurrerical integration alonq
the elerrent. '!he j-th order polynanial can be represented by j+l values
of ~, which is approximated as the j-th order polynanial; thus the IlOSt
suitable number of subdivision points for P. elerrent is j+1. Using great-
J
er muTter of points a strong divergence of results may occur. '!he results
in Fig. 5 shaN that me elerrent P" gives a quite good approximation of
cantilever up to 97 % full plastic force P fp'

By different positim of the points different results are cbtained.

Using IDbatto integraticn [10] the Pfp is calculated exactly, while con-
sidering Legendre integration the results for w are rrore accurate, but
the relative error for Pfp is 5 %. '!he ccrrpariscn between PI and PIt
elements shaNS that one P "elerrent gives better results than 4 PI ele-
rrents.
400

~p}
-----------r----------__ E
L
"I

W=

Pf~ full plastic force

W section rroduli of the cross secticn

Fir s t order theory

% Error for w
Different numerical
W Number of Pl elem::mts integration
cne elerrent P Lj
2 3 4 8 IdJatto Legendre
1.1 - 0.1 - 0.1 - 0.1 0.0 0.1 - 0.0
1.2 - 0.6 - 0.1 - 0.0 - 0.0 0.1 0.1
1.3 - 0.9 - 0.2 - 0.1 - 0.0 - 0.0 - 0.1
1.4 - 2.2 - 0.9 - 0.5 - 0.1 0.8 - 0.3
1.45 - 4.5 - 2.3 - 1. 4 - 0.3 3.5 - 0.9
1.47 - 6.5 - 3.8 - 2.5 - 0.7 7.2 - 1.8
%
Error
for 11.2 7.6 5.6 2.7 0.0 4.9
wfp

% Error for w
W Ole elerrent - Lcbatto integration
PLj Ps P6 Pa

1.1 0.104 0.025 - 0.019 o.

1.2 0.055 0.081 0.008 - 0.009
1.3 - 0.043 0.034 0.072 - 0.005
1.4 0.841 0.414 - 0.051 0.042
1.45 3.465 1.524 0.497 0.120
1.47 7.200 3.377 1.464 0.382

Fig. 5. Elasto - plastic cantilever analysis (f. o. t.)

 401

Discussion on aN. Using sirrplified kinenatic relaticns (15) the factor

~ was introduced. Its rooaning can be seen i f global element equilibrium
is cx:nsidered, Fig. 1 (p = 0, ~ = ~):
FI = 51 ex + 52 ez ; p2 = 54 ex + 55 eZ (52)

pi + p2 = -0 ; L (p2 e ) = 53 + 56 (53)
z
Transforrred form (ex = a)
x of the simplified equations (33), (34):
-
pi + pi = 0; ~ LO(p2 9Z) = 53 + 56 (54)

is identical to (53) i f eN is equal to L/L° (15). So the simplified

equaticns exactly satisfy glooal elerent equilibrium. Using the incre-
rrental form of (53):

ti'l + &,2 = 0 (55)

t.L (p2 ez ) + L (AP 2 eZ) + L (p2 ~ ) = L1S3

z
+ L\S6

where (Fig. 1 )

(AP2 eX ) = 6(F2 eX ) - (p2 ~ )

Z
(56)
~ =-e 6ljJ;oo =e 6ljJ
x z z x
the expressions for the end action incranents can be stated in the fom:

API =- AP2 (57)

(58)

When using (57) and (58) the correct tangent matrix is obtained. Many
authors use rrodified Egs. (54) and (58) with ~ ~al to 1 ; this way a
ncnsyrretric tangent matrix is derived and usually the term 5 5 l'.lj! is ne-
glected. Another way of obtaining an approximate tangent matrix is to
introduce L = LO into Fqs. (53) and (58). In the method proposed here
the global element ~librium is satisfied exactlv, while neglecting
6~ leads to an approximate tangent matrix. In the case of zero bending
deformation (truss) this matrix berorres the rigorous tangent matrix.
402

Conclusion

A oonsistent fonnulation for the gecrootrically nmlinear behaviour of

elasto - plastic plane frane; has been presented. TIle nurerical exarrples
show that in IOCSt cases the whole beam (oolUIm) can be treated as a sin-
gle P 4 eletrent, when it is loaded mly at the ends. '!be canputer program
NCNFRAN was develq::.tled using the P 4 element. '!hanks to the presented for-
nulatim it works with all four roothods nentioned above, i.e. Z. - Z••
Z. - 8. • 8. o. t. and f. o. t. '!he program has two versims; the first is
based m the displacetent method [11], while the sea>nd uses the mixed
approam. It is often used in structural analysis of frane;. '!his was the
reasm why the CDrplter costs had to be lowered. It was dale by using the
P 4 elenelt whim is sufficiently accurate for practical use, but less ti-
me ronsuning than P 6 ' though the latter gives better results. Tests shCM
that mixed approam te:JUires less iterations than the displacement appro-
ach.

References

1 Pfliiger, A.: Stabilitatsprd:>lem:! der Elastostatic. Springer 1964.

2 Kachanov, L. M.: Fomldatims of the 'lheory of Plasticity. North -
Holland - Amsterdam - Lcndm 1971.
3 Frey, F.; Lemaire, E.: PM - PBB eletrent. FINEL G - User's manual,
Universite de ~, Laboratoire de mecanique des materiaux et
statique des ronstructims 1978.
4 Vinnakota, S.: '!he Elasto - Plastic Stability of Frane;. 1. C. E.
MOnthly / Vol. III No.1. 1974.
5 Backlund, J.: Large Deflectim Analysis of Elasto - Plastic Beams
and Frane;. INT. J. Mechanical Science, 1976, Vol. 1, 8.
6. Olen, J. T.: Finite Element of Nmlinear Continua. McGraw - Hill
1972.
7 Zienldewicz, o. C.: '!he Finite Eletrent Method. McGraw - Hill 1977.
8 Argyris, J. H.; Dunne, P. C.; Sc:haq>f, D. W.: Ql Large Displacement-
Snall Strain Analysis of Structures with Ihtatimal Degrees of
Freedan. CcItp. Meths. AWL Mem. Eng. 14 (1978), 401 - 451.
9 Fliigge, W.: Handbook of Engineering Mechanichs. McGraw - Hill 1962.
10 Abranr::Mitz, M.; Stegun, I. A.: Handbook of Mathsnatical Functims.
Dover Publ. 1966.
11 BanOlTeC, J.; Marincek, M.: sane oanputer prograns for inelastic
buckling and for instability of planar frcmas. Liege Cbllcquium
m Stability of steel struct.w:es (1977), Preliminary Report.
Elasto-Plastic Boundary Element Analysis

J. C. F. TELLES, C. A. BREBBIA
Southampton UnIVersIty, U. K., UnIversIty of CalIfornIa, Irvme, USA

Swnmary

This paper presents the complete and correct formulation

for the boundary element method applied to plasticity.
Three alternative formulations - initial strain, initial
stress and fictitious traction and body forces approach -
are discussed. The paper explains how different yield
criteria can be lmplemented and gives, in matrix form,
the required expressions for the stepwise plasticity
solutlon. This formulation relates stresses to the initial
elastic solution and the plastlc strains, being very simple
to apply. The procedure is only possible due to the reduced
number of unknowns originated by the boundary element method.

Several examples are presented to illustrate the advantages

of using boundary elements to solve plastlcity problems in
preference to finite elements.

1. Introductlon

In recent years researchers have become increasingly

interested in solving complex stress analysis problems
using boundary methods rather than domain techniques. This
has resulted in a reassessment of classical boundary
integral equation techniques and in their interpretation
as an approximate method of analysis [1]. The concept of
boundary elements - i.e. elements only on the surface of
the dOMain but with interpolation functions similar to
those used in finlte elements - was introduced at the
beginning of the 1970's, the combination of integral
equations accuracy with finite elements versatility
resulted in what is now called the boundary element method.
This technlque offers several advantages over finite elem-
ents such as, the possibility of working with boundaries
extending to infinity; having smaller systems of
404

equations and more accurate results for stress

concentration regions and requiring much less data than
f~nite elements to run the same problem. This last
characteristic is, without doubt, the most interesting
feature of boundary elements for the practicing engineer
or user.
Several important papers have been published on the
application of integral equations for solving plasticity
problems [2, 3, 4J but none of them presented the complete
expressions for strains or stresses at internal points.
These expressions were given by Bui in 1978 [51 for three
dimensional bodies and then Telles and Brebbia [6] in 1979
extended them to two dimensional problems. The latter
authors presented the complete formulation for the boundary
element method applied to two and three dimensional
plasticity problems.
In paper, the application of boundary elements to
th~s

plast~city is presented, employing linear interpolation

tunct~ons for the boundary elements as well as for the

~nternal cells. The problem of accurately integrating

the domain integrals has been solved using a sem~­
analyt~cal approach, which can also be applied to compute

principal values. Three alternative boundary element

formulations are discussed; i) the initial strain, ii) the
~nit~al stress and iii) the fictitious traction and body

forces approach. The paper explains how different yield

cr~teria can be introduceu into the formulation in con-

Junct~on with the normality principle.

The required stepwise solution procedure is formulated in

matrix form and a simple recursive relation is presented
relat~ng the stresses to the initial elastic solution and

the plastic strains. This approach is only possible due

to the reduced number of unknowns originated by the
boundary element method, but would not be suitable for
finite element programs for instance.

The results found using boundary elements for a series of

examples are compared w~th analytical ~,

element results. Some of Lhe appl~Ldl~Cl

advantages of us~ng boundary elements uv~c

techniques to solve infinite mediuJYl piobit:.,

show how one can take advantage of synLil''''
s~gnif~cantly reduce the required numL~L I

all the cases the agreement between boui ",.

other solut~ons is satisfactory.
the considerable reductions achieved ~n c'
requ~red to run a problem us~ng b0U!ld.~. l' t.. _ . l

cells, for ~nstance, are only needed ~r. .:,

plastic~ ty, but they do not ~nCred&t:: L1 '" ," I

wh~ch are still those def~ned on the .." .l".I< ,

2. Governing Equat~ons

The total stra~n rate is assumed to be '.' '''',

~ (u. ~,J
+ U. . ) = ~':"
J,~ ~J
+ ~J:.;
,I

where u~ are the displacement 1 dt", ':01ll1' •• ,

~p are respect~vely the elast~c al,d l:'': c&, '

~J
the stra~n rate tensor. T~me <ieil Vc<LL , , ~. _ ,1

a dot and space der~vat~ves by a co~~a.

The equ~l~br~um cond~t~ons are,

a
~J,~
+ b] =0
~n the ~nter~or ~ of the body, and

p~ - a~JnJ =0
on the boundary r, being b J the body fv I: . . , '
per un~t volume, a~J the st.ress rate CLdl' )l1t::
tract~on rate components per un~ t a.re:i a,_
ward normal to the boundary of t.he be ctr '

If we cons~der the plast~c stra~n raL=~ ~~ ~.

and apply Hooke's law to the elastlc ~d~C

d~menslonal strain rate tensor, the fJl.~

ar~ses,
406

G: shear modulus and v: Poisson's ratio

The above expression can be rewritten in terms of initial

stresses,

(5)

h
were a• P
ij represen t s th e componen t s 0 f th e "1.'n1.'t1.'al
stresses" given by,
.p _ .p + 2Gv .p
a ij - 2G £ij 1-2v £kk °ij (6)

The substitution of equation (4) into (2) and (3), togather

with equation (1) gives [7J,
~j,~~ + l-~v U~,~j = 2(~lj'i + 1~2v ~kk,j) - ~(7)
and
.p 2Gv
£kk n1.) = 1-2v u~,~ n i +
(8)
G(u, ,+~, ,) n,
1.,J J,1. J

Equation (7) is an extended form of Navier's equation and

(8) represents its boundary conditions. The above expressions
can be written in the following form,
. 1· -
Uj,~~ + 1-2v U~,~j = - bj/G (9)

and

(10)

where b j and Pi are pseudobody forces and pseudotractions

given by,

.p .p
b,
J
b,
J
- 2G(£ .. , + _v_ £kk, j)
1.J,1. 1-2v
(11)
and
. .p v .p
I\ Pi + 2G(£ij n,J + 1-2v £kk nil (12)

One can notice that equation (9) represents a set of three

quasi-linear partial differential equations for the
displacements (plasticity terms appear on the right hand
side). This enables us to apply as a fundamental solution
for the boundary element formulation the singular solution
 407

of the Navier equations due to Kelvin [1].

Expressions (1) to (12) are valid for plane strain
(i, j, ~ = 1, 2 ; k = 1, 2, 3) and also for plane stress
(i, ], k, ~ = 1, 2) if v is replaced by v = v/(l+v).

Fundamental Solutions The singular fundamental solutions

corresponding to unit point loads can be written as,
*
u ij =
1
16'lT(1-v)Gr {(3-4v) 0ij + r,i r,j} (13)

for 3-D problems

*
u ij =
-1
8'lT(l-vlG {(3-4v) ~n(r) 0ij - r ,~. r .} (14)
,]

for 2-D plane strain problems

* -1 [ ] ar
Pij = 4CY.'lT (l-v)rCY. { (1-2v) 0ij + Br,i r,j an

- (1-2v) (r . n. - r . n.)} (15)

,~] ,] ~

CY. = 2, 1 B 3, 2 for 3-D and 2-D plane strain respec-

tive1y.
u~. and p~. represent the displacements and tractions in
~] ~]
the j direction due to a unit force acting in i direction
in which r is the distance from the point of application of
the load to the pOint under consideration. The deriva-
tives of r are taken with reference to the coordinates of
the latter point and 0ij is the Kronecker delta.
The strains (E]k) at any pOint within the body are given
by,

- r ,~. o]'k + Br ,~. r . r , k}

,]
(16)

and the stresses,

* -- c jkrs E*rsi - 1
°jki 4CY.'lT(1-v)rU {(1-2v) (r,k 0ij

(17)
 I:
~ .q~~r0ptc tensor of elastic constants
. ~, I Y wr 1 tten as,

< -, I: f. + G (c: Q + Q Q) (18)

" J.l rs l.r js is jr

., ....{J;'res .. l0ns are valid for plane stress lf

.. , n'.Jun9ar.,Y Element Formulatlons

"
:,,~ 5 tart thlS sectl0n by lntroducing an
-·,l",tl0n which lS partlcularly simpl1f1ed
- ~ pl,,~tJC stra1ns to be 1ncompress1ble.
'"- ----, ~ti.-.'8 but is Just1f1ed by the fact that
, ". 1 0f ;..he formulation lS only 1ntended for
r" ,,+ the yon M1ses yield cr1ter1on.

l111tlC11 stra1n formulat1on for three

" l t y p~cblems leads to the follow1ng

'LJ
" PJ' dr - dr + i u* b J d>l
>l 1J

r
~ I (19)
, >l

.~ ,~l Ld for boundary p01nts (C

1J
= Q
1J
/2
, ",", se" [19] otherw1se) and also for
"l':h C = Q If metal plast1c1ty lS
1J 1J
,L P1.2st1.C stralns are dev1atorlc, 1.e.
I='
(20)
i J
: : " -- , '., C'l, , stress rates at lnterlor p01nts
, - ".,' . ~,- "V 'c:e of expresslons (1) and (4). The
(19) yields,

i
,F " '''''11-1 ')'1

-I, ax
*
ap iJ
m
u. dr +
J >l
au *
~ b
ax
m
J
d>l

(21)
I II
 409

Where the derivatives are taken with reference to the load

point (these derivatives are written explicitly to
differentiate from the previous ones indicated by comma).
Notice that we need to perform the derivative of the
plastic strain rate integral, which is strongly singular.
The proper expression for this derivative has been
presented elsewhere [6J and is repeated here for complete-
ness,

·L
a a<1 *jki
'p dfl } 'p
aXm
{
L * £jk
<1jki aX m £jk dfl

(8-10v) 'p
+ 15 (l-v) £im (22)

where the integral on the right hand side is to be inter-

preted in the Cauchy principal value sense and the last
term comes from a boundary integral over the surface of
a sphere of unit radius centred at the singular pOint.
From now on the reader is referred to appendix A for all
components of the tensors related to the fundamental solu-
tions.
The above expressions together with (1) and (4) allow for
the determination of the internal stresses,

<1 ..
1J
L (-<1~jk) Pk dr -
L * uk dr
Pijk

+
L * ) b. dfl +
(-<1 ijk k
L':jk' 'p
£kQ, drl

2G(7-Sv) 'p (23)

- 15 (l-v) £ij

where the last two terms represent the influence of the

plastic strains.
For plane strains the procedure is analogous, the only
difference being the fact that the plastic strain rate
integrals still have to take into consideration the work
* £33)'
performed in the third direction (<1 33i 'p This effect
410

is easily incorporated through the assumption of incom-

pressible plastic strains [4] leading to what follows,

Cij Uj ~ L *
u ij Pj dr -
L *
Pij u j dr +
L * b, dO
u ij
J

+L A*
0' jki
'p
Ejk do (24)

where

A* * + 2vo ]'kr ,1, (25)

O'jki = O'jki 4n(l-v)r

and the internal stress rates,

. .
O'ij
L *
(- O'ijk) Pk dr
-ir *
Pijk uk dr

+ i
0(- *
O'ijk) b• k dO + iA*
oO'ijkR.
'p
EkR. dO

(26)
in which

*
O'ijkR. G r2 [ 4v r ,i r ,j 0kR. - 2\1 0ij 0kR. ]
+ 2 7I(l-v)
(27)

and the last integral is to be taken in the Cauchy principal

value sense.
Plane stress problems can also be solved by using equations
. " O'jki
(24) and (26) w1th * *
= O'jki' "'.
O'ijkR. *
= O'ijkR.' v replaced
by v in all ( )* tensors and the last term in (26) being
substituted by,
G
(28)
4 (I-v)

Initial Stress: In order to introduce the initial stress

formulat1on, let us merely study the plastic strain rate
integral presented in (19). Recalling expression (17) we
see that,
 411

(29)

by simple inspection of (18) we can make,

C'k
J rs CrSJ'k (30)

moreover

CrsJ'k ~pjk = ~prs (31)

where aPrs was given in (6).

Thus,

dO = L* .p
Ejki (Jjk dn (32)

Hence the initial stress formulation can be applied with

equal ease. For plane strain problems the formulation
follows the same pattern and the equivalent expression is,

(33)

and internal stresses can be computed by,

(Jij
L *
(- (JiJk) Pk dr -
L*
PiJk uk dr

+
L *
(- (Jijk) b k dO + L *
EijkR.
.p
(JkR. dn

(34)

where the initial stress integral is to be interpreted in

the principal value sense.

It is worth noting that here the initial stress integrals

do not require the contribution of the work performed in
the third direction, nor the assumption of incompressibility
of the plastic strains needs to be made. This is because
* .p
E33i = 0 and the effect of E33 is already included into the
412

components of ~lj. As a consequence, plane stress problems

can be handled by the above expressions with the replacement
of v by V being the only modification.
Fictitious tractions and body forces: The last integral
presented in equation (19) can be written in terms of the
derivatives of u~, as follows,
~J

-L {G(u~,

2Gv
~J,
k + u~k
~ ,J
,) +

.p
1-2v uiR..,R.. 0jk} E Jk dn (35)
which after integrating by parts gives the identity,

Jr 1<
u ij
2G (.p
&jk n k
+ _v_·p
l-2v &.Q,.Q, nj
) d
r

Jo u ij1< 2 (.p
G Ejk,k
+ _v_·p
1-2v E.Q,.Q"j
) d
n (36)

The substitution of (36) in (19) gives as a result,

Cij Uj = Ir U~j P j dr - Ir P~j uj dr + Ie Uij 5j de

(37)
where OJ and Pj were given in (11) and (12) respectively.
Therefore, we have arrived at a plasticity formulation in
which tractions and body force rates are fictitious (depend
on the plastic strains), but the displacements are correct.
In order to apply equation (37) one has to be aware that
although it looks like the elastic application of the
boundary element method, the internal stresses still have
to be computed by use of equations (1) and (4) , i.e.

L
..:.
°ij
1<
(- °ijk) Pk dr - Jr Pijk Uk dr
1< •

.
+L
.p
(- °ijk) bk dO - Cijk.Q, Ek.Q, (38)
 413

Another feature of this formulation is that in contrast with

the two prev~ous approaches, it needs computation of space
derivatives of the plastic strains (see eq. (11)). This may
be considered as a disadvantage for numerical implementa-
t~on, but, nevertheless, it is a valid procedure for formula-
ting the B.E.M. to plastici~y and still remains to be pro-
perly attempted.

4. Boundary Element Discretizat~on

The numer~cal implementation of the initial strain formula-

tion presented in (24) and (26) was given in [9J. It is the
aim of th~s section to illustrate the initial stress formu-
lation by an entirely similar procedure.

Let us suppose that the boundary of the body is represented

by surface elements while the part (or parts) of its interior
where plast~c~ty ~s l~kely to occur is discretized into a
number of ~nternal cells. Under these assumptions, linear
piecew~se functions are chosen to interpolate tractions,
d~splacements and "initial stresses", the two former over
boundary elements and the latter over internal cells (body
forces are not cons~dered in this text). The unknowns
produced by this discretization are interrelated by a series
of coefficients found by integrating over boundary elements
and internal cells.

From the substitut~on of (33) ~nto (24) comes the following

matrix relat~onsh~p,

H u = ~ p + Q aP (39)

where matrices Hand G are the same as obtained for elastic

analysis and matrix Q is due to the ~nitial stress integral.

Similarly, equation (34) gives,

a G' ~ - H' u + (Q' + E')~P (40)

in which E' is a well defined matrix that represents the

independent terms and Q' stands for the initial stress rate
integral. Matrices H' and G' are the same as those obtained
414

for elastic analysis.

In order to extend the validity of equation (40) to stresses
at boundary nodes, expressions computed by means of strain
displacement relationships and traction rate values along
each boundary element have to be employed. These expressions
do not require any integration and in the local coordinate
system (n l in 2 ) of a linear boundary element located between
nodes i and j are given by,

(41)

(42)

(43)

i· j.
where u l and u l are the displacement rate values at nodes
i and J and t is the element size. For plane stress v is
replaced by v
5. Numerical Integration
The process of integrat10n over boundary elements is well-
known and the procedure presented in [lJ was followed.
For the initial stress rate integrals the semi-analytical
1ntegration scheme fully described in [9] was adopted. The
technique is formulated with reference to a cylindr1cal
coordinate system (r,~) based at the singular point. Due
to the very nature of the fundamental solutio~ analytical
integration with respect to r turns to be very simple and
removes the singularities. Integration with respect to the
angle ~ is then performed using one dimensional Gaussian
quadrature formulae, five integration pOints being suffic-
1ent.
Notice that the procedure is also app11ed to compute the
principal value of the domain 1ntegral presented in (34).
But if higher order interpolation functions as well as
different cell shapes were chosen, the princ1pal value could
 415

still be computed by applying equation (40) to represent a

state of constant init1al stresses, in much the same way as
it was shown in [6J.
6. Elasto-plastic Stress Strain Relations

For the formulation of a theory which models elasto-plastic

material deformation, three requirements have to be met,
these are:

a) Exp11cit elastic relationsh1p between stress and strain

before the onset of plast1c deformation.
b) A yield criterion indicating the stress level at which
plastic flow commences.
c) Relationship between stress and strain for post Y1eld
behaviour.

The yield criter10n for isotrop1c hardening can be written

in general form as,

F(a.
1J
,k) = 0 (44 )

where k 1S a hardening parameter that gives the 1nstantan-

eous position of the yield surface in the n-dimens1onal
stress space. Here, the work hardening hypothesis has been
adopted, hence

(45)

On physical grounds, one can notice that the yield criterion

to be independent of the orientation of the coordinate
system employed, should be a function of the three stress
invariants only. It 1S common to represent two of these
invariants as functions of the deviatonic stresses,

Jl a ..
11
1 (46 )
J2 2' Sij S 1J
..
1
J3
3" S.1J Sjk Ski
where
416

S.
1J 1J 3 J 1 us:
a .. - 1:.
ij
(47)

In the present work, instead of J 3 the alternative stress

invariant 8, known as the Lode angle [lOJ was used.

) -< ~
'IT
- "6 :5.. 8 6
(48)

By using these stress invariants, different yield criteria

can be applied, such as

Tresca:

(49)

where a (k) is the uniaxial yield stress.

o

Von Mises:

~ - a (k)
2 0
o (50)

Mohr-Coulomb:

Jl
:r sin¢ + 1J2 (COS8 - -=1
/3
sin8 sln¢) - c cos¢ o (51)

in which ¢ is the angle of internal friction and c is the

conesion of the material.

Drucker-Prager:

(52)

where,
2 sinp 6c cosp
a. = K (53)
13 (3-sin¢) 13 (3-sin¢)

Mohr-Coulomb hypothesis can be simulated by the Drucker-

Prager cr1ter1on in plane strains if a. and K are written as
[llJ,
 417

a tan<jl 3c
K (54)

For our practical purposes, equation (44) can then be written

as,

(55)

where one can not1ce that f(o .) is a scalar function of o ..

1J 1J
which plays the role of an equivalent stress here designated
by 0e. As a consequence, we can def1ne an equivalent plastic
strain EP whose increment produces an increment in the plas-
e
t1C strain energy as follows,

o .. dE~ = dk (56)
°e 1J 1J

In order to obtain the stress-strain relations for post yield

behaviour, let us first rewrite equation (4),

o . (57)
1J

Within the context of associated plasticity, the flow rule

also known as normality principle, can be described by,

.p
E. (58)
1J

where dA is a proportionality factor, termed the plastic

multipl1er.

The substitution of (58) into (57) gives,

o .. C ijU (EU - a k £ dA) (59)

1J

a}' af
in which ak £ (60)
ao k £ ao k £

When plast1c yielding is occurring, the stresses satisfy

equation (55) which by differentiating gives,
418

F a .. a . - d1jJ k
dk
o (61)
1J 1J

Or, according to (45),

o (62)

From the application of the normality principle to equation

(62) results

a .. a .. ~ a .. a .. dA o (63)
1J 1J dk 1J 1J

If we substitute (59) into (63) and solve for dA comes,

dA (64)

where

y a
ij
C
ijk~
a ko +
N
~ddk a ij a ij (65)

Before we go further the last term in (65) can be examined.

It can be shown that f(o .. ) 1S homogeneous of degree one and
1J
this allows for the app11cation of Euler's theorem as follows,

(66)

The substitution of (66) and (56) into (65) gives,

y a C a +H' (67)
1j 1Jk~ k~

where H' = d~/d£P can be interpreted as the slope of the uni-

e
axial curve plotted as true stress versus true plastic strain.

Equation (64) can now be used to substitute d A in (59) pro-

viding the requ1red 1ncremental stress-strain relations

a . (68)
1J
 419

in wh~ch

1
CiJk~ - y C1Jrnn a mn a op Copk~ (69)

For the appl~cat~on of the above relat10ns to the initial

stress formulation, a further modif1cat1on has proved to be
conven1ent. Let us adopt the following notation,

·e
0, , (70)
~J

·e
where stands for the components of the elastic stress
0,
1J
rate tensor (1.e. these represent the stress values as if a
pure elastic problem were being solved).

Equation (68) can then be rewritten in the following form

·e 1 ·e
0, , 0, , C, , a a ok 0 (71)
~J ~J y ~Jmn mn k ~ Iv

which means that the true stresses can be computed from the
correspond~ng elast~c stresses. In addition to this, the
~n~t~al stress rates presented in (6) can also be calculated
by the relat~on,

.p ·e
o , o 0, (72)
~J ~J 1J

By s~mply exarn~ning equation (5) we not~ce that equation

·e
(40) can be appl~ed for the computat1on of 0" if matrix E'
~J
1S replaced by

E E' + I (73)

where I is the ident~ty matrix.

This g~ves,

·e
0 G' p - H' u + Q* aP (74)

where,

Q* Q' + E (75)
420

Before the application of the above expressions to solve

plasticity problems, it should be pointed out that although
solution algorithms are incremental, always finite sized load
increments are prescribed and this may create some drifts of
the stress level beyond the yield surface. If load incre-
ments are kept sufficiently small this problem is practic-
ally eliminated, but if relatively large load increment sizes
are to be permitted, special techniques of the type presented
in [121 have been found necessary to maintain the stresses
on the yield surface.

7. Outline of Solution Technique

In order to minimise the computer effort for the initial
stress formulation, let us reexamine equations (39) and (74).
For a well-posed problem,a sufficient number of tractions
and boundary displacements needs to be prescribed. The
unknowns are then reordered leading to

Ax = f + Q oP (76)

and slml1arly,
·e
a A' x+ i' + Q* aP (77)

.
where vector x contain the unknown tractions and boundary
displacements.
From the multiplication of (76) by A-I we get

x R aP + m (78)

where
R A-I Q (79)

and

m A-I f (80)

Substituting (78) lnto (77) gives,

a·e S oP + n (81)

in which
 421

s Q* - A' R (82)

and

n £1 - A' m (83)

Note that vectors m and ~ represent the elastic solution to

the ~ncremental problem (actual solut~on in absence of plas-
ticity). Furthermore, equations (78) and (81) rema~n valid
if instead of incremental loading, the total load is applied.
The only reason to proceed incrementally being the constitut-
ive relations presented in (68). This enables us to compute
load at f~rst y~eld by simply scaling down the total elastic
solution by a load factor Ao. The incremental process starts
at this load level and further values of the load factor are
given by,

(84)

where S = Aow; W is the given value of the load increment

with reference to load at first yield.
For elasto-plastic solutions, equations (78) and (81) can
therefore be appl~ed as,

x = R(a P + ~p) + A. m
~ - (85)

and
e (86)
a

or alternatively for pure incremental relat~ons,

x =R aP + S m (87)
and
a·e = S aP + Sn (88)

where in both cases vectors m and n correspond to the appli-

cation of the total load and-a P st;nds for the current init-
~al stress increment.
For a typ~cal load increment (i.e., a g~ven value of Ai)'
422

the in1tial stress increment can be determined iterat1vely

at each boundary node and internal point exhibiting elasto-
plastic behaviour by two d1fferent processes. The former
1S in fact a pure incremental procedure. Once the load
increment Sn is applied, the initial stress increment corr-
espond1ng to the solution of the elastic problem is computed
and has to be applied back into the body, providing an
elastic stress redistribution. This operation, again gener-
ates a new in1t1al stress field to be red1stributed elastic-
ally and so on. Iteration is halted when the contribution
of the last initial stress increment can be neglected.

The process 1S in essence comparable to what was presented

in [13J for the finite element method and is summarized as
follows,

a) compute elastic stress increment by,

eq. (88) if f1rst iteration is being performed or
;e s;P otherwise
b) find true stress increment 0 .. (eq. (71))
1J
c) verify convergence, i.e.
compare dE P calculated with its accumulated value
e
obtained during the current load increment to see
if it can be neglected.
d) calculate initial stress increment by,
.p _·e
0iJ - °ij - °ij
e) accumulate values of initial stress and true
stress,
oJ?1J. oJ?1J + ~~ .
l.J
0.. O. + 0 ..
1J 1J 1J

f) continue with next node or p01nt and start with

b) until all nodes and pOints have been consider-
ed.
g) go to a) for a new iterat10n.

Iterations are performed unt11 convergence is obtained

(with1n prescribed tolerance) at every node or point.

It is 1nteresting to note that in order to avoid cumulative

 423

errors, ~p obtained at the end of 1terations 1S applied to-

gether with S~ 1n eq. (88) for the first iteration of the
next load 1ncrement.

The second process, which proved to be less dependent on

the load 1ncrement Slze but not always more economical,
deals with accumulated values of the elastic stress in a
sim1lar fash10n to the procedure adopted for the in1tial
strain implementation [9J.
The in1t1al stress 1ncrement is kept separate from its accum-
ulated values until convergence 1S obtained as follows,
a) compute elast1c stress (eq. (86»
b) calculate elastic stress 1ncrement by,
·e e _ a , - o~,
°ij = °iJ 1J 1J
c) find true stress increment 0" (eq. (71»
1J
d) verify convergence 1.e.
compare dE P calculated w1th its prev10us value.
e
e) calculate new estimate of in1t1al stress incre-
ment by,
oPlJ = oe,
1J
- 0, ,
1J
f) cont1nue with next node or point and start with
b) untll all nodes and points have been consid-
ered.
g) go to a) for a new iterat1on.

Once that convergence 1S obtained for all nodes and p01nts,

the 1n1t1al stress increment 1S accumulated and its value
is also used as an in1t1al guess for the next load 1ncre-
ment.

Note that ne1ther procedure requ1res computation of the

bounaary unknowns. Consequently equation (85) need only be
used to print the boundary unknowns once convergence is
achieved. In addition we should emphasize that matrices R
and S as well as vectors m and n remain unchanged throughout
the 1ncremental process. These features, also common with
the initial straln procedure, represent a great saving in
computer t1me.
424

Finally a question is posed,

which procedure is more efficient?
The answer can only be obtained in the light of experience.
Regarding precision, the examples solved show that if load
increment is kept small, the initial strain and both initial
stress processes lead to entirely equivalent results (differ-
ence within the prescribed tolerance). Nevertheless with
relation to computer time, the initial strain version has
proved to be faster than the initial stress. This difference
however, seems to be due to the fact that in contrast with
the initial strain program which is restricted to von Mises
yield criterion, both versions of the initial stress routine
are more general and can handle four different yield criteria.
As load increment was increased, the initial strain and the
second initial stress versions exhibited more stable results
than the former initial stress procedure, despite special
techniques to avoid stress drifts beyond the yield surface.
But in some of the examples run with relatively large load
1ncrements, the pure incremental form of the initial stress
program produced equally reliable results, requiring less
computer t1me than its second verS10n.
In conclusion, the above discussion demonstrates the equiva-
lence of the three processes and furthermore shows that the
efficiency of the procedure is problem dependent. Ideally,
all three options should be ava1lable for a general case.

8. Examples
Following the discussion presented in the end of the last
section, the results for a series of examples solved by the
boundary element technique are now compared with analytical
solutions where such solutions are available and with finite
element results.
Notched Tensile Specimen:
This first example is one of the very early plasticity prob-
lems solved by using the finite element technique. Plane
stress and plane strain results have been presented in
several papers, creating a good opportunity to compare the
 425

boundary element computations.

Material parameters are as follows,
E = 7000 kg/mm 2
cr 24.3 kg/mm 2
o
v 0.2
H' 0 (von Mises yield criterion)
Plane stress analysis was carried out using the discretiza-
tion shown in f~gure 1. Note that symmetry was taken into
consideration without boundary discret~zationof the symmetry
axes, this is due to a direct condensation process which
integrates automatically over reflected elements and cells
in such a way that the size of the final matrices correspond
to the reduced number of boundary nodes and internal points
presented.
Figure 2 gives the load displacement curve for this case.
It is seen that the curve remains nearly straight until very
close to the limit load, when a sharp bend then occurs. Such
behaviour was also observed by Yamada et al [14J in an ent-
irely similar problem. The limit load achieved by the bound-
ary elements (20 /0 = 1.21) coincides with the results
a 0
presented by Nayak and Zienkiewicz [12] using four different
finite elements to analyse the same problem. Their limit
load was found to vary between 20 a /o o = 1.19 - 1.23 and
simple triangular, isoparametric linear, quadratic and cubic
elements were used, all four meshes with approximately 97
nodes.
For the plane strain case,becduse of a large spread of plas-
tic zones before lim~t load is achieved, the number of
internal pOints and cells was increased from 33 and 51 to
59 and 97 respectively. Load displacement curve is shown
in figure 3 where equivalent finite element results presen-
ted by Chen [15J are also given. The limit load obtained by
BEM (20 /0 = 1.64) is below the value given by the finite
a 0
element method (20 /0 = 1.85). But, as stated by Chen,
a 0
bound theorems demonstrate that the maximum load should lie
between 1.52 and 1.73, which supports the boundary element
results.
426

Fig 1 Notched tensile specimen Boundary element

and Internal cell discretization ( plane stress case)

2C7.
14 eTo
(8 EM) limit load achieved 1 21

-------123
12
----119 FE limit loads

10

08

06

04

02

00005 00015 00025 00035

Fig 2 Load - displacement curve for notched specimen

Inplane stress
 427

20.
00"

""
2 .0 F. E.col_pse 1.85

B.E.colapse 1.6 4

1. 5

1.0 IV

~c
1---1
0 .5
- - F.E.M
- - - B.E.M .
Uc
1
0.005 0 .0 1 0 .015

Fig 3 Load - displacement curves for notched specimen

In plane strain

~ --
I ,
I "

~--..;'+--.. - I',
~.:__ ~ __~~L __~,1 __ _

Fig 4 Plastic zones obtained by B E M for different

load levels ( plane stram)
428

Spread of plastic zones at lower load levels presented in

figure 4 exhibits good agreement with finite element computa-
tions [13J, [lsJ for the same problem.
Deep Circular Tunnel:
The second example was selected to emphasize the advantages
of boundary elements over "domain" type techniques to solve
infinite medium problems.
A circular excavation studied by Reyes [16J and later by
Baker et a1 [17J with linear displacement triangular and
quadrilateral finite elements respectively is here compared
with boundary element results.
The plane strain problem was analysed under the Drucker-
Prager simulation of Mohr-Coulomb yield criterion (a and K
given by eq. (54» and by assuming the infinite domain to be
initially subjected to a uniform stress field of 1 ksi vert-
ical and 0.4 ksi horizontal (K = 0.4). For the present
o
study, external loads corresponding to the relaxation of
this free stress field were applied along the surface of the
opening.
The material (Rock) was assumed to be perfectly plastic with,
E 500 ksi
c = 0.28 ksi
\i = 0.2
<p 30 0

Boundary element and internal cell discretization is presen-

ted in figure 5 where the plastic zone on complete removal
of the free stress from the boundary of the cavity is also
given.
Stresses along the horizontal section computed at the end of
the relaxation process are compared with the corresponding
results presented by Reyes and Baker in figure 6. Here,
internal stresses outside the discretized region were calcu-
lated at simple internal pOints not connected to any internal
cells.
It is important to note that the refinement of the two finite
 429

Initial uniform stress field

o I 1 ksi

.2.4 ks i

,,
,,
• ,,
/ ._.J._.__ C£

Fig 5 Deep circular tunnel problem Discretization used

for B E results and spread of plastic zone at the
end of relaxation process

a
C
- 7. 5
/aker(F.E.M .)

B EM
/'
, Reyes ( F.EMJ

\./,".
v·,".
-5 .0 ,".
':;,....
0v :"0.';.: •••••••

Bake r (F.E.M.J
- 2.5
1BEM
~ Reyes (FEM)

!Lx 3 4
x
r
5

Fig. 6 Final stresses along the hOrizontal section through

the Infinite medium
430

element meshes (about 253 nodes) should not lead to the

differences in the °y values in f~g. 6. Although no refer-
ence was made by the authors, this discrepancy is probably
due to the outer boundary conditions considered in the two
analyses. The boundary element technique does not require
any outer boundary discretization, but in order to study its
influence ~n the results a quarter of a circle with radius
equal to nine times the radius of the cavity was discretized
~nto s~x boundary elements, this is approximately the extent
of the finite element meshes. The outer circle was then
considered to be free to displace, giving as a result a
better agreement with Baker computations. The second alter-
native was carried out by prescribing zero displacements
over the outer boundary, leading now to improved agreement
with Reyes results.

Rough Punch:
In the third example the elastic-plastic behaviour of a
square block compressed by two opposite perfectly rough
rig~d punches is stud~ed. The problem is analysed under
plane strain condition and material is considered to be per-
fectly plast~c obeying von Mises yield criterion.

By using a very refined mesh of 274 linear displacement

triangular fin~te elements and 173 nodal pOints (fig. 7a),
results to this problem were presented by Chen [15J [181. The
boundary element analysis was performed w~th the discretiza-
tion shown ~n figure 7b, requiring less than one th~rd of
the FE data to run the problem.

The indentation process was developed by prescribing the

flat punch displacements leading to the average pressure-
applied displacement curve presented in figure 8. As can be
seen, agreement between the two analyses has been thoroughly
obtained, both methods slightly exceed~ng (4%) the theoret-
ical limit load 13 p/2o o = 2.5.

9. Conclusions

In this paper the correct and complete formulation for the

boundary element method applied to plasticity has been pres-
 431

p =mea n pressure

W= 0 .5 on

·tt
® Fml te Element mesh @ OiscreUz8t ion used for B.E results
(no bounda ry elemenlS along
symmetry axes)

!It

Fig 7 Geometry of rough punch problem

v3p
200 • fEM .
• B.E.M.

1 1'"4-- - -----,
6 T,~-.,
,,
,,
,
,,
E = 10 7 pS'
00 = 13000 psi ,
L______ ____ _
v = 0 .33
H' = 0

0 .0 1 0 .02 0 .03

Fig 8 Mean pressure - applied displacement curve for

rough punch problem
432

ented. Three alternative procedures - initial stress,

initial strain and fictitious forces - were discussed to-
gether with the possibility of applying different yield
criteria. The required stepwise procedure has been formu-
lated in matrix form using a simple recursive relation re-
lating stresses to initial elastic solution and plastic
strains.

The examples presented show the advantages of using boundary

elements in plasticity in preference to finite elements.
Considerable reductions in the data required to run a prob-
lem can be achieved for the same degree of accuracy. In
addition, problems with boundaries at infinity can be prop-
erly modelled using boundary elements and the resulting
matrices are smaller than those corresponding to finite
element or finite difference techniques.

References

1 Brebbia, C.A.; Walker, S.: The Boundary Element Tech-

niques in Engineering. Butterworths, London (1979).

2 Swedlow, J.L.; Cruse, T.A.: Formulation of Boundary

Integral Equations for Three-Dimensional Elasto-Plastic
Flow. Int. J. Solids Structures 7 (1971) 1673-1683.

3 Mendelson, A.: Boundary Integral Methods in Elasticity

and Plasticity. Report No. NASA-TN-D-7418 (1973).

4 Mukherjee, S.: Corrected Boundary Integral Equation in

Planar Thermoelastoplasticity. Int. J. Solids Struct-
ures 13 (1977) 331-335.

5 Bui, H.D.: Some Remarks about the Formulation of Three-

Dimensional Thermoelastoplastic Problems by Integral
Equations. Int. J. Solids Structures 14 (1978) 935-939.

6 Telles, J.C.F.; Brebbia, C.A.: On the Application of

the Boundary Element Method to Plasticity. Appl. Math.
Modelling 3 (1979) 466-470,

7 Lin, T.H.: Theory of Inelastic Structures. Wiley, New

York (1968).

8 Riccardella, P.C.: An Implementat~on of the Boundary-

Integral Technique for Planar Problems in Elast~c~ty and
Elasto-Plasticity. Report No. SM-73-10, Dept. Mech. Eng.
Carnegie Mellon University, Pittsburgh (1973).
 433

9 Telles, J.C.F.; Brebbia, C.A.: The Boundary Element

Method 1n Plasticity. Proc. 2nd Int. Seminar on Recent
Advances 1n Boundary Element Methods, (edited by C.A.
Brebbia), University of Southampton (1980) 295-317.

10 Nayak, G.C.; Zienkiewicz, O.C.: Convenient Form of stress

Invar1ants for Plasticity. Proc. Am. Soc. Civ. Engrs.,
J. Struc. div. 98 (1972) 949-954.

11 Drucker, D.C.; Prager, W.: Soil Mechanics and Plastic

Analysis or Limit Design. Q. Appl. Math. 10 (1952) 157-
165.

12 Nayak, G.C.; Zienkiewicz, O.C.: Elasto-plastic Stress

Analysis. A Generalization for Various Constitutive
Relations including Strain Softening. Int. J. Num. Meth.
Engng. 5 (1972) 113-135.

13 Zienkiew1cz, O.C.; Valliappan, S.; King, I.P.: Elasto-

plastic Solut1ons of Engineering Problems; Initial Stress
F1nite Element Approach. Int. J. Num. Meth. Engng. 1
(1969) 75-100.

14 Yamada, Y.; Yoshimura, N.; Sakurai, T.: Plastic Stress-

stra1n Matrix and its Application for the Solution of
Elast1c-plastic Problems by the Finite Element Method.
Int. J. Mech. SC1. 10 (1968) 343-354.

15 Chen, W.F.: Lim1t Analysis and Soil Plasticity. Elsevier

Scientific Publishing Co., Amsterdam, The Netherlands
(1975).

16 Reyes, S.F.; Deere, D.U.: Elastic-plastic Analysis of

Underground Openings by the Finite Element Method. Proc.
1st. Int. Congr. Rock Mechanics, Lisbon (1966) 477-483.

17 Baker, L.E.; Sandhu, R.S.; Sh1eh, W.Y.: Application of

Elasto-plastic Analysis in Rock Mechan1cs by Finite
Element Method. Proc. 11th Symp. Rock Mechanics,
(edited by W.H. Somerton), University of California,
Berkeley (1969) 237-251.

18 Chen, A.C.T.; Chen, W.F.: Const1tutive Equations and

Punch-indentation of Concrete. Proc. Am. Soc. Civ.
Engrs., J. Engng. Mech. div. 101 (1975) 889-906.

19 Lachat, J.C.; Watson, J.O.: Effective Numerical Treat-

ment of Boundary Integral Equations, a Formulation for
Three-dimensional Elastostatics. Int. J. Num. Meth.
Engng. 10 (1976) 991-1005.
434

Appendix A

Tensors related to the fundamental solutions that appear in

the text are of the follow1ng form,

G
{ SClr
" [ (1-2v)0 .. r k +
2a1T(1-v)r S on 1J ,

+ v(o'k r . + o'k r .) - yr .r .r k J
1,J J,l ,1 ,J ,

+ Sv(n.r .r k + n.r .r k) +
1 ,J, J ,1 ,

(1-2v) (Snkr .r . + n.o' k + n.o. k )

,1,J J 1 1 J

(Al)

--(;--Q
2a1T (l-v) rf.'
{s (1-2v) (0 .. r k r
1J"
R, + 0kQ,r . r
,1 ,J
)

+ SV(Oa. r r or . + O'k r or
N1 , Jr , k + O'k
J,N ,l 1,N,J

+ O'a r .r k) - Syr .r .r k r a +
IN ,1 , ,1 ,J, ,N

where a = 2,1; S = 3,2; y 5,4 for 3-D and plane strain

respectively.
Also,

+ 20 1..Jr , k r , NJ+ 2v[oa.

N1
r .r k + o'k r or .
, J, J , N ,1

+ 01 k r ,N,J
or . + o'or
I N ,.r
l, k J
+ 20 ko r .r . - 8r .r .r k r 0 } (A3)
N ,1 ,J ,1 , J , ,N

for the plane strain initial stress formulation.

Simplified Calculation Models Applied to Postbuckling Analysis
of Thin Plates

B. H. KROPLIN, H. DUDDECK
Technische UnivefSltat BraunschweIg, Germany

Sununary
A Y1eld funct1on, which descr1bes the plastic flow of a
plane cross sect10n of a thin plate approx1mately is
d1scussed and compared to the layer approach. Further an
art1fic1al (viscous) damp1ng 1S suggested to overcome
1nstab1lit1es due to snap throughs w1th or w1thout b1fur-
cations.

Introduct1on

In the f1rst part of th1S paper some aspects of a yield

funct10n model for th1n plates are d1scussed and compared
to the well known layer approach, /1,2,3/.
In elast1c-plast1c analys1s the theory of thin members
loses much of 1tS advantages for the stress resultants,
aX1al forces and bend1ng moments can only be determ1ned
by a d1scretisat1on over the depth of the cross sect10n
or by a layer approach. To slmpl1fy the calculat10n pro-
cess, 1n order to safe comput1ng t1me, Y1eld funct10n
models are developed, 1n Wh1Ch yield1ng of a cross sec-
t10n 1S descr1bed as a function of quant1ties of the
reference surface only. The Y1eld1ng process 1S reduced
to 1tS most 1mportant features. The Solut1on obta1ned
is not generally valid, but often sufficient accurate,
1f the structural behav10ur is sought, e.g. the llmit
load.
The cons1dered yield funct10n 1S developed by Eggers 1n
/4/. Yield1ng depends on the aX1al forces, the bending
moments and a plast1c stra1n parameter.
The Y1eld function 1S applied to the post buckling ana-
lYS1S of th1n steel plates under aX1al compression and
compared to results obta1ned by a layer approach.
436

The second part deals with a method to overcome snap

through problems with or without bifurcations. The
structure is treated as a viscous one with an artif~cial

damping, such that equilibrium states before and behind

the snap through can be reached in a stable manner. Exam-
plesare carr~ed out on a plate strip.

1. Yield function and the layer model

Thin plate theory

One of the most advantageous features of the th~n plate

theory is that the internal work of the structure can be
expressed in quantities of the reference surface only.
With the Bernoulli hypotheses taken to be valid, the
strains are expressed in strains and curvatures of the
middle surface,

(1)

~ and $ are symmetric and antimetric shape functions,

fig. 1. The subscript M indicates quantities of the
m~ddle surface.

iii = 1
ijj = z

fig. 1: Shape functions for stra~ns and elastic stresses

Analogue to (1) the stress can be expressed as function

of the axial forces and bending moments

(2)
 437

if for g and m the following integral conditions

hold /5/,

jli>dZ =1 fz Ii> zdz =0 , ( 3)

z

J 4J
z
dz =0 !4J ZdZ =1.
z
(4)

According to the orthogonality cond1tion (3,4) g and m

are uncoupled.

The layer approach

In the cornmon layer approach the stress 1ncrements are

calculated from the stra1ns of every layer (5) and an in-
cremental constitutive law (6).

(5)

(6)

The compatib1l1ty 1S sat1sfied a priori by (1). An itera-

tive correction is necessary to satisfy the constitut1ve
law (6) for finite increments. The equilibr1um of the
stresses for a slngle particle in the cross section 1S
not considered. To avo1d undue complexity G'3 is often
neglected and every layer is treated 1ndependently /1,2/.
Integrat10n of the layer stresses over the depth of the
cross sect10n leads to axial forces and moments (7).
Equation (2) 1S not needed.

f!2=J9. zdz (7)

:z

Improved strains ~M ' 1!-M are calculated by an iteration

with unbalanced forces on structural level.
438

The yield function model

The derivation of the yield function model starts from

the alternative assumption that yielding of the cross sec-
tion can be described by an elastic limit state, an lnter-
mediate state and a plastlc limit state. The elastic 11m-
it state is defined by the first fiber yielding; the plas-
tic limit state is reached, if all fibers yield.

With Q (2) substituted lnto the Yleld condition for the

continuum

( Hl11 151 ) (8)

a yield condition for the cross section (9) is obtained,

where
An = na.~ Ja.~gA n 9A

Al2 = 4- na.~ Ja.~g A m~A

In general ~ and ~ depend on the constltutlve relatlon-
Shlp. Only for the limit states simple form occur.

The limit states

For the elastic limit state the elastic stress distribu-

tion is given by \{> 1/t and ~ = (12/t 3 ) Z , fig. 1.
Substitution of (2) into (8) leads to the elastic limit
state

( 1oa)

Sign (A 12 ) provides yielding of the top and the bottom

£lber.
For the plastic limit state biconstant functions \{> and~
flg. 2, are chosen, which are developed from a stress
 439

asz :st
f1g. 2: Shape funct10ns for stresses 1n the plastic
llm1t state

funct10n and sat1sfy the orthogonal1ty condition (3) in

advance. After Subst1tuting (2) 1n (8) and el1m1nat1ng c
the plast1c llm1t state Yl (lOb) 1S glven by

( lob)

The intermediate state

In order to avoid compl1cated functions in the partial

plastic state the yielding is considered as gradual change
from the elastic to the plastic limit state controlled by
means of a plastic strain parameter EP • As yield func-
tion for the cross sect10n
o< a <,
0< b ( 11)

1S chosen. Eq. (11) satisf1es the boundary conditions for

the elastic and the plastic limit state, /4/. Due to the
cho1ce of (11) the yield h1story of the intermediate state
has to be expressed by the parameters a and b only.
They are 1dent1f1ed such that E P calculated from the
yield function (11) fits as close as poss1ble to the pla-
St1C stra1ns and curvatures of a plane reference cross
section under monotonous yielding, fig. 3.
440

m
- reference cross section
"'---'---'_L--'---I_ n=1 fut
Sy2 o yield function (11)
n 1

0,75 ~-~~--i1----t-Y
o
a
11
9-
O,5L-----~----~-----L
I-----L----~------L-----L-~e:-·p-
I-O,5~

fig. 3: Accuracy of the approx1mation for a .9,

b = .0666, A12/ YA 11 ·A 22 '= 1.

Clearly part1al unloadings of the cross section are ne-

glected, which are included in the layer approach. Rota-
tions of the axial forces and bend1ng moments are possible
for the path on the yield function 1S not predetermined.
Depending on A12 the yield function changes its shape,
fig. 4. The neighbourhood of a actual stress state 1S
described by the tangential plane of the yield function
~n stress and strain space,

1 ( av .,..,Il av ·af). av £-p) - 0

V = 2S an0t3 n'-'+' + om5j3 m I"' + a£P -. ( 12 )

In opposite to other yield function approaches /6/ a s-E P

diagram is not needed. The fict10uS force S depends on
the integration path,

S = Soy H d £ p. (13)

eP
 441

S is coupled Wl th Ep by the relationshlp

(14 )

m::fyi A 22
2
10"T-~_ A: A12 :: n m Sy
B: A12 : 0

2
3"

"-.
Alz= t nQ./3J a /3QA mQA
1..-_ _ _ _ _ _ _ _ _ _ _ _ _---..:...____ n =i-~
to Y

J1111 :: 1,0 J1122 = 0,5

-p
fig. 4: Change of the Yleld surface wlth A12 and E

In the incremental functional the forces, ~, ~ are con-

strained to the tangential plane by a Lagrangian mul-
tiplier A, which is interpreted from the Euler equations
-p
as the plastic stram parameter E / 4,7/. The con-
strained funtional then reads

. .~ £P ay· ay .!. P
WM = WM - 25 ( as 5 +2at P E ) = stat .. (15 )
442

Layer approach versus yield function model

The derivation of the yield function model restricts the

application clearly to problems, where
- the unloadings in parts of the cross sections and
- the rotations of the forces and moments
are of small influence.
The ultimate state Y1 is sufficient exact (about 4.5 %)
/8/, the plastic strains and curvatures are components
of the plastic strain parameter. They depend on the inte-
gration path, while the plast1c strain parameter itself
is determined by the yield condition based on monotonous
yield1ng. Thus errors in compatibility may occur, Wh1Ch
affect the distr1bution of the stress resultants in the
intermediate state.
Since the layer model satisfies compatibility f1rst (1)
there such errors don't occur, but the equilibrium of the
stresses is only controlled in an 1ntegrated manner (7).
In engineering applications the loss in generality of the
Y1eld function has to be often considered against the gain
of numer1cal effectivity.
In the layer model the stresses have to be calculated in
every layer for every strains gM , ~M and have to be
stored. In the yield function model the yield function
is determined in advance. Only the iteration on structu-
ral level is needed. This means correction of the devia-
tion from a yield surface one time while it 1S necessary
for every layer 1n the layer approach. For the same
reason the element matr1x can be calculated in the yield
function approach in one step while 1t is calculated in
the layer approach by the sum of the layer matrices.
After own estimations the calculation is about ten times
as fast as the layer model with five layers.

Application
The Y1eld funct10n model was applied to post buckl1ng
 443

analysls of thin steel plates under axial compression.

Since the rotations are small, a total Lagrangian descrlp-
tion was used wlth the strain ln the middle surface

The incremental functlonal is glven ln (17)

W=/(ST~L_ H{ES +rn~~~J dA -jQUdA+b =stat.

A A
where for moderate deflections

= [ no.
a.
n t3
~l

and for plasticity

• N 1
~2 = 25
[~E~.a.p~]
~ =[o,mJ
tL =U;~~]
F Flexibllity of the cross sectlon

b boundary integrals.

Examples

The load deflectlon curve of a slender plate, bit = 80

calculated by various approaches are compared in fig. 5.

The approaches are:

1. A layer model with flve layers and 36 GauE integration

points per element for the material propertles, i.e. 144
integration points in the mlddle surface of a square
plate. An element wlth a complete cubic shape functlon
was used /9/.

2. The Yleld functlon model applied in a displacement

444

approach with a triangular element. Cubic shape functions

out of plane and quadratic shape functions in plane are
used /10/. The element variables are dlsplacements and
midslde rotations.

3. The yield function model was applied in a mixed FEM

formulation with linear shape functions for all unknowns.
/4/.

G G,U
Gy

1.0 ~__~?/Yt=3'175mm
m" _ N/mrrf
w S' n
'0 Gr2SO
E-206500 N/mm2
t ~"0,3
w~ b/1000
alb-O,875

",'=Without yielding

-~-:~-=:=-)(
.5
_ ,~~I
,..x~
first yielding
- .: 7 /
4:;6~Z ,. displ.
5 layers

_,f/ ' , b/t .. SO displ.

2. yield f.
)'
,,( '

mixed
3. yield f.
.1
Wm
1.0 2.0 t

fig. 5: Load deflection curve of axially compressed plates

- Comparison of different approaches -

The load deflection curves of the layer approach (1.) and

the yield function model (2.) show excellent agreement.
With the mixed model the limit load is predicted a little
less for the llnear shape functions overestimate the plas-
tic zones of the plate.
 445

In comparlson to the mixed formulatlon a refined mesh

was necessary to arrive at the same accuracy in axial
forces and bending moments wlth the dlsplacement approach.
Due to the slmpler structured nonlinear element matrlx ln
the mixed approach the computation was considerably faster
than in the displacement approach, despite indefinite
finite matrlzes and more unknowns.

G
Gy
1.
.t alb =1

t
I :
2x

a IG y i

0 100 t 24
G
Gy
1. .t . 2x
-:
t
a/b=1I2
Gb
D Ib
.......
t . a
SLyGY
0 100 t 24
G

[b_
Gy a/b=O
1. ,
.!., . -Gb

t a
t, ~VJhi
0 100 t 24
• experiments (11)
+ mixed FEMJyield function

fig. 6: Comparison of experimental and FEM results

446

In fig. 6 the results calculated by the m~xed formula-

tion and the yield function are compared to the results
of experiments /11/. The influence of the shape, alb was
~nvestigated for various slenderness ratios. The calcu-
lated and experimental results coincide in the full range
of the investigation.

Conclusion 1:

A comprehens~ve comparison of the results leads to the

engineering judgement, that the accurate description of
the structural behaviour, i.e. various yield zones and
plastic mechanism is more important for predicting a cor-
rect limit load in the postbuckling range of thin plates
than a complete consideration of the cross section yield-
ing with an ~mproved description of the stress history
in the intermediate state. Especially in problems, which
require consideration of geometric and material nonline-
arity the yield function model reduces the numerical
effort considerably.
However further investigations are needed to define the
application range of the yield function model ~n general.

2. A viscous approach to structural instabilities

In post buckling of plates snap throughs may occur, even

when imperfect structures are considered. At the point of
instability the Newton Raphson iteration diverges. An
alternative to various techniques suggested, /12, 13, 14/
~s to treat the structure as a viscous one. The static
equations are augmented by damping force Q~, which
carries a part of the external load or the unbalanced
forces until a steady state is reached /15/,

~(x)~ + 9 ~ = I! . (18)
 447

~ means the vector of unknowns, ~(x) describes the non-

linear structural matrlx, E contalns the applied loads,
D is an artlflclal damping and x the derivation of ~

wlth respect to the tlme. With a backward tlme lntegration

scheme and

(19 )

(18) can be wrltten as

D_*n:: °
£M- (20)

E~ are the unbalanced forces. Eq. (20) represents a

Newton Raphson type iteration, where D* is chosen as a
function of the unbalanced forces, thus speeding conver-
gence,
r
onx -- a:: 10 f3::08 Y= 05. (21)
..J

Examples

The method is demonstrated on a snap through of a slmply

supported beam, calculated as plate strlp wlth and without
geometrical lmperfections, WhlCh is loaded by the dis-
placement u 1 ' fig. 7.
Wlth an lmperfection, composed of the first and the sec-
o
ond buckllng mode w =a 1 Sln ITx/l+ a 2 s1n 2 ITx/1 and the daIY\p-
ing a stable equilibrium state after the snap through
was reached within 12 steps, (curve B* ). The saIDe prob-
leID was calculated Wl th the modified Newton Raphson proce-
dure. After careful adaptlon of the lncrements the same
equllibrium state was reached after 11 lncrements and ln
the whole 86 iterations.
o
Wi th a true antimetric imperfection, w = a 2 sin 2 IT x/I,
curve C, no instabillty occured at the critical load of
the first buckling mode. With the mlxed method the first
buckling mode with a lower level of energy was missed.
448

To avoid an accompanYlng eigen value calculation in every

step a dlsturbed unbalanced force

flQ +alflQI
0.=02 (22 )
1+0

was applied together with the damping D~ flQ indicated

a snap through from the second to the first buckllng mode
curve c" . Curve D shows the iteration without damping.

QJ
t - 7 mm
<.l 1.0 I =-1000 mm
....
.....0 E = 2.06 10 5 N/mm2
.S? ~= 0 , 0

il
c
E
L.
0
y
.A.
.m y§ ~,n

Wm
U1 =0 /021 mm

W" a sm"!!.!.-+ a sm ~
1 I 2 I

c t al a2
.1 .2
A 0 ,1 0
BO 0,01 0,0707
C 0 0,0707
number of iteratIons
A*, B", c*: with damping
A,B,C incremental
c 0 without damping
....u
0

.!!!
Qj 20 K l
"0

C
....
L.

C
III
u

.1 Wm
1°
t

fig. 7: Snap through wlth and without damping

 449

Conclusion 2

With an artificial damplng elastic snap throughs can be

calculated in a stable manner. The increments can be en-
larged or the load can be applied ln one step. No adulter-
atlons of the results occur.

The stability of the reached equilibrium state can be

tested by a disturbance of the unbalanced forces or the
damping orthogonal to the reached deflection mode.

Optimisation of the damping with respect to the efficien-

cy of the calculation are subjects for further lnvestiga-
tions as well as the definition of a signlficant distur~

ance for the stability test.

Acknowledgement

The author would llke to acknowledge the support provlded

by the Helsenberg Stlpendlum der Deutschen Forschungsge-
melnschaft and the work of Mr. D. Dlnkler on the numerlcal
examples.
450

Literatur:

Wegmuller, A.W.: Full Range Analysis of Eccentrically

Stiffened Plates, ASCE J. Struct. Div. 100, 143-159
(1974)

2 Baums, B.: tiber das Tragverhalten langsgedruckter

quadratlscher elastisch-plastischer Platten bei gro-
Ben Deformationen. Diss. TU Braunschweig 1977

3 Eggers, H.: Stlffness and Force-Moment Interaction

for Plates with Plastic Flow, J. Struct. Mech. 4,
123-139 (1976)

4 Eggers, H., Kroplin, B.: Yieldlng of Plates wlth

Hardening and I,arge Deformations. Int. J. num. Meth.
Engng. 12, 1978 (737-750)

5 Hlll, R.: The Mathematlcal Theory of Plasticlty.

Oxford Universlty Press London 1950, 318 ff.

6 Crisfield, M.A.: Full Range Analysis of Steel Plates

and Stiffened Platlng under Uniaxial Compresslon,
Proc. Instn. Civ. Engrs. Part 2, 59 (1975) 595-624

7 Eggers, H.: Variatlonal Princlples for Elastiplastic

Contlnua, J. Struct. Mech. 3, (1974-1975) 345-358

8 Robinson, M.: A Comparlson for Yleld Surfaces for

Thin Shells, Int. J. mech. Sci. 13, ; (1971) 345-354

9 Dumont, N.,A.: Traglastberechnung beulgefahrdeter

Rechteckplatten 1m elastisch-plastlschen Bereich nach
der geometrisch nichtlinearen Theorie unter Beruck-
slchtigung geometrischer und werkstoffllcher Imper-
fektionen. Dlssertation 017, Darmstadt 1978

10 Meler, F.: Grenzlastberechnung ausgestelfter Platten

mit Imperfektlonen. Bericht Nr. 80-33 aus dem Insti-
tut fur Statlk, TU Braunschweig

11 Flscher, M., Harre, W.: Ermlttlung der Traglastkur-

ven von einachsig gedruckten Rechteckplatten aus Bau-
stahl der Seltenverhaltnisse u ~ 1 von Versuchen,
Der Stahlbau 8, 1978 (239-247

12 Haisler, W.E., Strlcklin, J.A., Key, J.E.: Displace-

ment Incrementatlon In Non-Linear Structural Ana-
lysis by the Self Correctlng Method, Int. J. num.
Meth. Engng. 11, (1977) 3-10
 451

13 Bergan, P.G.: Solutlon Algorithms for Nonlinear

Structural Problems, Proceedlngs Int. Conf. on en-
glneering Application of the Flnite Element Method,
A.S. Computas, Hovik, 9-11 May 1979

14 Argyris, I.H., Hilpert, 0., Hlndenlang, K., Male-

Jannakls, G.A., Schelke, E.: Flachentragwerke 1m
Konstruktiven Ingenieurbau, ISD-Berlcht Nr. 263
(1979)

15 Krbplin, B.-H.: A V1SCOUS Approach to Post Buckling

Analysls, to appear in: Engng. Struct.
 Finite Element Analysis of Reinforced Concrete Slabs and
Panels

G. MEHLHORN, D. KLEIN
Technische Hochschule Darmstadt, Germany

1 . Introduction

The real behaviour of reinforced concrete structures differs

widely from the results of a linear elastic computation.
Because of cracking of concrete, the nonlinear stress-strain
relations of concrete and reinforcement, and the bond-slip
between concrete and reinforcement, the deformation behaviour
of reinforced concrete structures is extremely nonlinear even
under working l oads.
-("'I-
~/,
!St uopt,O
SSt .10/500
C.ncrtle :

4
{I _J()/AII/m 1
'l.- J()OOOMII/m t
v.-o.z Fig . 1 Crack distribution
Ti. rod. of Shear Wall WT7 of the
1 ~SJ·l,ocm: tests by Leonhardt/Walther
~ :..A$l·~'cm, [1]
-A~, • 1.0 em
I--~ SSt .10/500
64()kN - -
CfGOO up 10 610 kN - -
1qro k N --

* I

1
, " I J '"

Fig. 2 Crack distribution

~~~~~~~~~~~~~: of a simply supported rec-
~ tangular slab tested by
J' § Franz [2)
 453

In Figs. 1 and 2 the crack distribution of two characteristic

planar structures is shown. They are the shear wall WT7 of
the tests of Leonhardt/Walther [1] and a simply supported,
rectangular slab that was tested by Franz [~ • Through the
load-midpoint deflection curve of the slab ~n Fig. 3 the
characteristic behaviour of a reinforced concrete structure
is demonstrated. The curve can be divided into three parts.
After nearly linear elastic behaviour, the first change of the
stiffness is caused by the beg~nning of cracking. The further
softening of the curve depends first of all upon the spreading
of the cracks and the nonlinear deformation behaviour of
concrete. When the reinforcement begins to yield, a second
reduction of the stiffness is evident, which finally leads to
the failure by crush~ng of concrete. The test results were
measured only up to a load of about 52 kN/m 2 in order to save
the test equipment in case of sudden failure.

J
lood
(kNfin2)
50
{ lin. elastic ~fallure load
(uncracked)
./ {Yield lme Theotg}

/ /
V I
~ I'
30 ~working lood (DIN 101fS) Fig. 3. Load-mid-
j pOint deflection

L A{
20 curve of a simply
dis - and re,dodlng supported rectan-
1.1 of the slob gular slab tested
//;
10
by Franz [2]
1
0,5 ~o l,S 2,0 .2,5 40 Defledlon [em}

Of all the calculation methods that were developed to consider

the deviation of the behaviour of reinforced concrete struc-
tures from the linear elastic range, the Finite Element Method
proved to be one of the most powerful tools. This method was
used by many authors, for example [10] to [29J

The purpose of this report is to present the main problems of

computation of the real behaviour of planar reinforced concrete
structures using the Finite Element Method and to present
solut~ons obtained by the Darmstadt Group. More general rev~ews
454

are given by Eibl/Ivanyi [3] , Wegner [4], Schnobrich [5] and

others.

2. Computat~on of Planar Reinforced Concrete Structures by

the F~nite Element Method

2.1 Load Carrying Behaviour of Reinforced Concrete Structures
Re~nforced concrete is a composite material consisting of the
two components, concrete and steel. Its load carry~ng be-
haviour is characterized by the fact that the concrete is
able only to res~st small tensile forces. After cracking of
the concrete, the tensile forces are transmitted by bond to
the reinforcement consisting of steel bars. The d~rect~onand
the diameter of the reinforcement bars are determined by the
designer, usually follow~ng the results of an elastic calcu-
lation of the structure, assuming homogeneous behaviour.

After cracking, however, re~nforced concrete is inhomogeneous

and nonlinear. Its deformation behaviour depends on the
direction and history of the loading and ~s represented by
the material properties of the components concrete and steel
and the quality of the bond between concrete and reinforce-
ment.

2.1.1 Mater~al Behaviour of Concrete in the Biaxial Stress

State
For the computation of planar structures, the behaviour of
the concrete must be considered in the biaxial domain. The
material properties are influenced by many different factors,
among them are the composition of the concrete (kind of
aggregate, cement and water matrix), rate of hardening,
consolidat~on, temperature, loading history (age of concrete

at loading, loading rate, loading duration, cyclic loading,

direction and size of the previous load by reloading) and
long time behaviour (shrinkage, creep).

Because of the different influences, the form of the failure

cr~terion and the stress-strain-relation of concrete in the

biaxial stress state depends on the results of the tests

 455

that are performed to obtaln these relations. For short tlme

loading and proportionally increas1ng load, the tests of
Kupfer [ 6] proved to be the most rellable. D1fferent authors
used these test results to develop analyt1cal formulations
of the failure and deformatlon behavlour of concrete for
applicat10n in numer1cal calculat10ns such as the f1nite
element method .

.0:n
7'd

1,0

][ I
I- bwnd oti~3 fl)('
0.5 F1g . 4. Fallure criterion of
l-
I-
>-
e CfOeXs in O~
di~ction
crode3 in two
concrete in the biaxial state
-<Tn of stresses by the formulatlon
I- ~ directions
of L1nk [7] based on Kupfers
AI tests (6)
-0.5 -1,0

In Fig. 4 the failure criter10n by the formulat1on of Link

[71 is shown. The stresses are normalized 1n terms of the
un1axial cyllnder strength, so that the formulation can be
used for dlfferent grades of concrete. The whole stress
region is described by one relat10n w1thout considering the
boundaries between the reglons of compresslon-compression,
compression-tension or tension-tens1on . Failure criteria such
as that of Link [7] cannot be used as plast1clty cond1t1on.
They only describe a boundary for the maximum stresses and
do not allow any statements about the plastlc deformatlons.

The stress-straln-relat1on of concrete 1S nonllnear from the

beginning of loading. Because of the nonlinear deformation
behaviour, the concrete ceases to be 1sotrop1c in thebiax1al
stress f1eld. The material properties depend on the existing
state of stresses and strains. Dlfferent ways are possible
for the computational adaptat10n of the nonlinear stress-
strain relat1on. One way is the f1n1te formulat1on of the
stress strain relatlon by analytical functions. In the bi-
aXLal stress state, however, d1ffer e ntial formulations are
456

more suitable for the description of redistribution of the

stresses. Link [7) developed an incremental formulation for
the tangent stiffness of the concrete by considering the an-
isotropy in the biaxial stress state. In Fig. 5 some stress-
strain-relations that follow the formulation of Link [7] are
compared with the test results of Kupfer [~ •

,?/~ - o.SI5

-1,0 I---+--fi~"""'=-- OJ/a, - ~o

~/a, _qo I
-~~ ~-~~-~r--~-~~-_~~=s·t--i Fig. 5. Stress-strain
- - - calc. relations of concrete .
-0.6 I-----Hf-+- Comparison between test
results of Kupfer (6) and
analytical formulation of
Link [7J

-1,0 l.,{%o}

2.1.2 Material Behaviour of Reinforcement

As in reinforced concrete only steel bars are used it is
sufficient to study the material properties of the steel
under uniaxial stresses. In Fig. 6 the stress-strain-
relations of the preferably used two kinds of steel are
shown. After a nearly perfect-elastic behaviour, the mild
steel shows a strongly marked yield strength followed by a
more or less constant stress after yielding. At large strains
a hardening effect occurs. The cold-worked steel does not
have a clearly defined yield strength. The end of the elastic
region is marked by a loss of stiffness with a further
increase of the stress in the plastic range. For computation
the mild steel may be idealized as an elast~c - perfectly
plastic material w~th consideration of the harden~ng at large
stra~ns while the stress-strain relation of the cold-worked

steel can be approached by a bilinear relation.

 457

as rolled steel cold worked steel

Fig. 6. Stress-strain relations of reinforcement

Under unloading the steel obeys the same law as under first
loading. At cyclic loading with many cycles, however, the
strength of the steel is diminished. Temperature does in-
fluence the steel properties very much. So the strength is
reduced at high temperature.

2.1.3 Bond Between Concrete and Reinforcement

The interaction of concrete and reinforcement is first of all
influenced by the quality of bond between the two components.
The tens~le forces must be transmitted by bond after cracking
from the concrete by shear stresses into the steel bars. The
properties of the bond should guarantee a good distribution
of the cracks so that the crack width is small enough to
protect the re~nforcement against corrosion.

The quality of bond depends on the condition of the surface

of reinforcement, the diameter of re~nforcement bars, the
strength of concrete, the load history and load direction and
the state of stresses. So the transferable shear stresses are
increased by transverse compression. In Fig. 7 the mechanism
of bond between concrete and ribbed steel bars is demonstrated.
The Fig. is taken from the tests of Goto [al. Before cracking,
the strain under tensile forces will be the same in concrete
and in steel. When the limit of strain in concrete is reached,
458

ribbed steel bar concrete uncracktd lone

Fig. 7. Crack mechanism

force on Ihe inner cracks (tests of Goto [8])
CQncrefe
forces on the
ribs of Ih.
~inforteme"f

the concrete wlll crack throughout and by contractlng will be

propped at the ribs of the reinforcement so that all tensile
forces now will be transmitted into the steel bars. At the
ribs, smaller, inner cracks will develop.

Dlfferent ways have been developed to test the bond behaviour

of relnforcement. In Flg. 8 some of the generally used test
specimens are shown. The concentric pull-out specimen is used
to test the bond strength under tension while in theexcentric
pull-out specimen the conditions in the end of a cantllever
beam is simulated with lateral compression forces in the
anchorage zone. Wlth the lapped-splice specimen the over-
lapplng of reinforcement lS tested. In the tension specimen
the bond behaviour in the tensile zone of the constructlon
is slmulated. This specimen has preferably been used for the
derivatlon of shear stress-slip relations as representation
of the stiffness of bond. In Fig. 9 some characterlstic bond-
slip relations of the tests of Dorr [9] under different
transverse compression are shown. The slip is defined as
relative displacement between reinforcement bar and sur-
rounding concrete.
 459

I~-
conu.nfric pull- out
f
.
<
specimen

Q ·'SlotN/m'
INN/m'}
~
l ~::
-
excenlric pUlI- out
specimen
10

t
lilil -
Q · Ia/~ral
Compr.

-~
lapp~d splice
f??
speCimen

.d{mm]

- =it=zz=====*= tension specimen

Fig. 8. Some test specimens Fig. 9 . Comparison of the

for the investigation of bond -relationships with
T- .!J
between re1nforcement and lateral pressure by the
concrete tests of Dorr [9J

The bond is usually taken to be perfect, that means the dis-

placement of reinforcement and concrete is the same, the
SllP is zero. Only the shear strength may be limited. In
structures 1n which bond 1S specially 1mportant, however,
for example if the load 1S induced 1nto the structure first
of all by the reinforcement, the Sl l P between concrete and
reinforcement may not be neglected for a realistic computation
of the deformat1on behav10ur of the structure.

2.1.4 Re1nforced Concrete Under In-Plane Forces

F1g. 10 shows a planar element of cracked reinforced concrete
under 1n-plane forces. The situat10n of the element is defined
by three geometr1cal parameters : the d1rection of the re-
inforcement bars, the direction of the cracks and the
direction of the princ1pal stresses . The first cracks will
develop perpendicular to the principal tens1le stress. Under
further increase of the load, the direction of the principal
stresses can change because of the transm1ssion of shear
stresses across the crack, so that a second crack may occur
at any angle to the f1rst crack. The shear stresses are
transmitted by aggregate interlock and by dowel action of the
460

local roughness global roughness

., .,.
?,
C>1-'
Cl
~ :!".

'" '" :::I

'?,
Fig. 10. Planar element of
cracked reinforced concrete Fig. 11. Transmission of
shear forces across a
crack by aggregate inter-
lock

reinforcement. Fig. 11 shows a model for the aggregate inter-

lock at the crack. Shear forces are transmitted by global
roughness that results from the meshing of the grains in the
aggregate so that a tangential displacement At of the two
edges of the crack is possible only by a widening An of the
crack. The inclination of the direction of the forces in the
crack is caused by local roughness of the mortar. The dowel
action of the reinforcement is demonstrated in Fig. 12. Any
relative displacement of the edges of the crack causes a de-
formation of the reinforcement bar and a transmission of shear
forces across the crack.

fti3f crock

lolerol compression
DirecJion of
.~.-----

'" main fensile stress

f't. ~
." ~inforcemenl Fig. 12. Dowel action of
,,'\" bar
~$.) reinforcement
 461

A B A Sec/ion A Section B
I I I 1
~ 0:

I
I
J.--_ ___.-1
l
I ',,--r ..,

I
I
II Flg. 13. Cross-
section through a
I cracks I reinforced concrete
I lateral I element under
I reinforctm~nt I bending moment

I I
A B

2.1.5 Flexural Behaviour of Reinforced Concrete

In Fig. 13 a section through a cracked element of reinforced
concrete, loaded by bending moments, and the distribution of
stresses in the cross-section is represented. In section
A-A, the tensile forces are transmitted only by the re-
inforcement, besides a small resultant of possible tensile
stresses in the concrete. Between the cracks, however,
section B-B, the concrete is not free of stresses in the
tensile zone as well as in the compression zone. The tensile
stresses are caused by shear forces (bond) between the re-
inforcement and the concrete and by the effect of the com-
presslon stresses lnto the concrete tooth between the cracks.
The neutral axis is not plane anymore after cracking.

2.2 Modelling of Planar Reinforced Concrete Structures by

the Finite Element Method
By adopting the finite element method to describe the load-
defQrmation behaviour of structures, the system is sub-
divided into finite elements that are interconnected at
selected nodes wi th equal deformations and forces. Linear
stiffness relations between the nodal forces and deformations
of the elements are derived by the use of the displacement
method, the force method, or mixed methods. The assembly
of all elements of the complete structure yields a system of
 462

linear equations between the nodal forces and nodal de-

formations that can be solved for different load stages or
given deformations.

2.2 . 1 Types of Finite Elements

The most versatile but also most expensive method for the
modelling of reinforced concrete is the separate idealization
of concrete and reinforcement by elements (see F1g. 14) .
aJ Example for planat element under inplane farces '
b) Example for la!lered bending element.

concrete logers
anedimensianal linle
element far bond- 51ip contact element
for bond-Slip
onedimensianal eleme'" t.:::===;z::::=~ smeared reinforce -
for re;n(or«menf bo, ment layer

"=====::::»._ - nodal poinls r en-rence surface

Fig. 14. Idealization of concrete and reinforcement by

f1n1te elements

The concrete is described by elements for bending, in-plane

forces, the comb1nation of bending and in-plane forces, or
for shallow shells, dependent upon the problem . The possible
elements differ in the1r shape, the 1nterpolation functions
for the fields of deformations or stresses, the number of the
unknowns and the method of integration about the area of the
element . The material properties of the concrete may be
described by one constitutive law that 1S valid over the whole
thickness of the planar element. If the material propert1es
vary over the thickness of the element, for example after
cracking under bending, a subdivision of the element into
layers with d1fferent thicknesses or partitioning of the
thickness by integration pOints at different distances is
more sU1table . In each element layer or 1ntegration pOint
the actual stiffness relations can be formulated.

The reinforcement 1S idealized by one-dimensional elements

for each steel bar or by planar elements with smeared
material properties if reinforcement meshes are used.
 463

Because of compatibility, the shape functions of the elements

for the re1nforcement depend on the kind of the concrete
elements used.

The bond between concrete and reinforcement is realized by

link elements with one-dimensional or orthogonal spr1ngs
between the nodes of concrete elements and reinforcement
elements or by contact elements. The contact elements are
fictit10us elements with one or two dimensions and no ex-
pansion 1n the third direction. They transmit normal stresses
and shear stresses but allow only a relative tangential
movement of the adjacent surfaces. As compared to the link
elements the contact elements offer not only a more distinct
idealization of bond but allow also a better match of the
concrete and reinforcement elements. The link elements may
result in incompatibilities if higher order finite elements
are used.

If perfect bond 1S assumed, the elements of reinforcement

can be combined with the concrete elements by the assumption
of equal deformations in the nodes of reinforcement elements
and concrete elements into one macro-element. The deformations
1n any plane of the element then are related to the de-
formations 1n one reference plane. This Y1elds a reduct10n
of the number of unknowns and a shorten1ng of the computat10n
t1me. In re1nforced concrete slabs the deformations within
the reference plane must be considered because the position
of the neutral aX1S is not plane and its height differs with
chang1ng load wh1le cracking.

The most econom1cal method is the integrated modelling of

reinforced concrete in one element. The material properties
of the components concrete, steel, and bond are combined to
a fict1tious material whose deformation behaviour is de-
scribed by in-plane force-strain relations or moment-curvature
relations. Because of its economy of computat1on this method
is suitable for the investigat10n of the behaviour of a
structure by a series of iterations.
 464

2.2.2 Propagation of the Crack Development

The development of cracks is considered as single cracks
along the element boundaries, Fig. 15a, or as smeared cracks
(crack fields) within an element, Fig. 15b. The disadvantage
of the first method is that the direction of the cracks is
assumed and that a renumbering of the element mesh is
necessary after the development of the cracks. Within the
element the crack is taken into account by modifying the
material properties without changing the topology. The
cracked concrete element is described by an orthotropic
stiffness matrix, usually derived by the assumption that
tension perpendicular to the crack cannot occur in the

no rension st;f(ening
a) Singlt cracks along the ~emt!nf boundorit3 :

stepped respOl/3e
~r crocking

l\
'\ /' \
nodlll PQinfs eraclc

'b) smeared crack within the element :

uncf'(Jc/led
elemel/t with
_-----,~r:r.."....,r-..,....,~./ one crack direefion

elemel/f wilh
two crockdirecfiol/.r
al arbitrary ol/gle

Fig. 15. Description of the Fig. 16. Some possible

crack devolopment with finite assumptions for strain-
elements softening of concrete
after cracking

concrete. The shear forces parallel to the crack, which

are transfered by aggregate interlock and dowel action of the
reinforcement, are taken into consideration by reduction of
the shear stiffness of concrete.
 465

The tension-stiffening of the concrete, tha t is the carrying

capacity of the concrete between the cracks, may be re-
presented by an average additional charge to the steel
stresses or by an unload i ng section in the stress-strain
relation of concrete, after the maximum tensile stress is
reached . In Fig . 16 some assumptions for the strain-softening
of concrete are compared.

2 . 2.3 Numerical Treatment of the Nonlinearities

By consideration of real material behaviour, the system of
equations describing the relations between the nodal de-
formations and nodal forces is nonlinear and cannot be solved
directly. For the iterative treatment different solution
procedures have been developed. Fig . 17 shows some of them.
In the pure incremental method the load is increased stepwise

p p

L-__________~u U U

incremental me/hod New/on-Rophson modfft~d !lew/on -

iteration Rophson if8ra lion
p p p

incff!menlal melhod+ in&l'rmen/of mil/hod ~
N_Im-Rophson iler. n",,/if. Newton -Raph-
(wn» longenfsnflness) son ltero/ion
(wilh iniliol stiffness)
u
inuemen/af me/hod #
modif. Ntwfon - Roph -
Jon iluo/ion
(new sHffnes$ 0/ any
load step)

Fig. 17. Some solution procedures for the numerical

treatment of the nonlinearities

and the stiffness is recomputed in each load step. However,

the solution drifts away from the real load-deformation
relation, if no equilibrium correction is performed . In the
466

2.2.4 Some Remarks About the Idealization of a Structure by

Finite Elements
A conscientious choice of the right type of element and a
careful discretization of the structure into elements is
necessary to obtain satisfactory results. The best type of
element that should be used depends on the problem to be
studied. The appropriate modelling of the load-carrying
behaviour of the reinforced concrete structure is more
important than the accuracy of the numerical calculations.
The question wether simple elements or higher-order elements
should be preferred cannot be answered satisfactorily. Simple
elements may produce convergence difficulties, if they are
not compatible. On the other hand, by the use of higher-
order elements, the system can be divided only into a few
large elements to reduce the numerical efforts so that the
change of stiffness, for example by cracking, is considered
in too large reg10ns if the element itself is not subdiv1ded
into zones with different material properties. With simple
elements 1t 1S easier to refine the discretization of the
structure in regions with a sharp change of the stress or
strain gradient, for example at a concave corner or unde~
single loads, or at sudden variation of the stiffness.

2.3 Synops1s of the Solution Methods Used by the Working

Group of Darmstadt
In Table 1 the different methods are summarized that are
used by the invest1gators of the working group of the
Institute of Massivbau, TH Darmstadt, and the considered
nonlinearities are stated. All the investigators use the
displacement method. Only the short time behaviour of
concrete is regarded except by Hoshino [11], who cons1ders
the shr1nkage of concrete, and by Ze1 tIer /Dietrich/Mehlhorn
[17] , who investigate the inf luence of creep and relaxation.
Geometr1cal nonl1nearities are included by Kristjansson
[15] into the computation of slabs.
 467

3. Numer1cal Examples
3.1 Shear Wall WT7 of the Tests of Leonhardt/Walther f1J
The system and the discretisation 1nto fin1te elements used
by D6rr [1~ 1S shown in F1g. 18. The wall 1S character1zed
by a concentrated tie rod at the lower side and strong
vert1cal re1nforcement bars. After cracking of the concrete
1n the tension zone, the vertical reinforcement must trans-
mit the load applied to the lower side of the panel upward
into the uncracked compression zone, where it is transmitted
by bond to the concrete and carried by arching action to the
supports. To study the effect of bond-sl1p between the
reinforcement and the concrete, a matter of main importance
1n this example, Dorr [191 idealized the reinforcement by
separated bar elements and represented the bond-sl1p by
realistic shear stress-slip relations including the influence
of lateral compression. F1g. 19 shows the crack-development
of the solution of Dorr [1~.

In Fig. 20 the load-midspan deflection relation of Dorr

accord1ng to two different assumptions about the bond-slip
behaviour is shown and compared w1th the test results and
the results of other authors.
468

J?fO J¢1Z J~12

2~10 15¢n1J J~11 J(J11 U;tz

f\ /\ lif\ Vi ~ V11\
V\1 \V NV1NV
1

1\ /'\ 111"'- V1 I"'J V1 1"'-.

II '\/'\1/ 1 "'-V I~ 1/
1\ /\1 111"'-. V1 I"'J V1 1"'-.'
II '\} ;\'V I""V N V
1\ /~ VI"'-. V1 I"'J V1 1"'-.
'I '\J. ~V !~ /' N V
\ V V 1"'-. V "'J V1 1 "'-.'
'I '\ ,1\ V' 1'\ /' N 1/
,\ V vi"-.. V "'-. V1 1 "'-.'
V 1\ '1\ k'" / N V
['\
"
'!~ "
/'
~lP
"- ./ ".
OzPE

i
-
1-16 I· s+cm-----j
- - - "infofUmenf
!I
Fig. 18. Finite element idealization
of Dorr [19} for the recalculat~on
of the Shear Wall WT7 (see Fig. 1)

Load
p~~~--~~--~~~~~=+----+---~
Miiller [18}
q8~----+---'~~~~~-----+----~----~
Geisfefeld [21}
~6~--~~~--+----r---~---r---~

o Hoshino [11]
t:. Stouder (10] ---!----1
- test results

1 2 J If Oeflection [mm)
F~g . 20. Comparison of calculated load-midspan deflection
relations with the test result of Leonhardt/Walther [1]
 469

3.2 Simply Supported, Rectangular Slab Under Uniform Load of

the Tests of Franz [2]

F1g. 21 shows the reinforcement of the slab and the dis-

cret1sat1on into elements and layers of the calculation of
Kle1n/Kristjansson [ 23J . The f1nite element idealizat10n
depends on the compl1cated distribution of the upper and
lower re1nforcement.

SlM '
SSt +1O/5QQ
MI soo/sso

F1g. 21. Fin1te element ideallzation of Klein et al. [23]

for the recalculation of the rectangular slab of the tests
of Franz [2]

In F1g. 22 the 10ad-m1dpo1nt deflect10n relat10n of the slab

1S compared to the results of var10US calculations. The
calculat10ns d1ffer, besides the use of different element
types and element meshes, f1rst of all 1n the consideration
of the tens10n stiffening of concrete between the cracks.
At the load stages beyond the work1ng load, all the calculated
curves are close together and agree quite well w1th the test
result. After pr1mary crack1ng, however, tens10n st1ffen1ng
is of signif1cant 1mportance, so that 1n the calculat10ns
w1thout consideration of tens10n stiffening effects the
stiffness of the slab is underest1mated.

The distribution of the cracks at the lower and upper side

of the slab by the calculation of Klein/Kr1stjansson [ 23 J
is shown in Fig. 23 together w1th crack distribut10n of the
test.
470

L,;, I
Load Load
{kN;In'l step,
50 V Ii,. :=-- - --
--:;;; ~ I-- --
18
(uncrocked) ~
"0
J ~ 16

/
f+

/L- ~c.resu/~ of Klein et 01 [23J

.1

R ~ 12
JO
'\
/d
1 of.1Schafer
calc. results . 1 [13 J fO

20

~~
calc. resu!f; of Maurer [16J 6

10
t "
2

0,5 1,0 1,5 2,0 2,5 40 Deflection [em}

Fig. 22. Comparison of calculated load-midpoint deflectio n
relat10ns with the test result of Franz [2]

lowpr side
fe3f

," " , ,,""1-'i-

--.r-
...... ./ /
l(/./
/
/
\.' i: )I..,><y.Y- of-f.~.I('")(yx i- / /
Fig. 23 . Comparison of
,'- :t: ¢- ~+ 4" ++-t+-4 ::: . / / IOWl!fside
colculalian
/ . / :t "1-+-l.'++ .t -t-+ -t+ t
/ / ~ XXx)(-+'+-+-ty7-X'- >f: "
the crack development
of the test-slab with
the calculation result
of Kle1n et al. [ 23]
/ //)(A-
/ / - - -i---~'"""
,\.
,
"
/

// j \\... ""-
// \" urper side
ealeulalian

"\ //

'" \
/ /
I /
 471

3.3 Square, Corner-Supported Two-Way Slab Under Central

P01nt Load

In this slab tested by McNe1ce [24] the influence of in-plane

forces on the load bearing capacity of re1nforced concrete
slabs 1S demonstrated. For this purpose, the slab was computed
by Mehlhorn/Krist jansson/Klein [14] w1th and without prevention
of the hor1zontal movement of the corners in the midplane of
the slab. The slab was calculated for both support condit1ons
by includ1ng as well as excluding the nonlinear terms of the
stra1n-deformation relat1ons.

In F1g. 24 the calculated 10ad-m1dpo1nt deflect10n relat10ns

are shown. After cracking of concrete the neutral plane of
the slab moves upwards, so that the midplane 1S stretched
and the corners move outwards. If the expanS10n of the slab
1S prevented by f1xed corners, the spreading of the cracks
is reduced. An 1n-plane compreSS1ve force 1S 1nduced which
leads to a stiffening of the slab and provides a higher load
capacity. By the assumpt10n of geometr1c llnear1ty, this
1nfluence 1S overestimated.

lDod
P{IrN} -

20
Ih"knes.
e m
d- ~u till
eHecIdepih. At =~33 em
ptrr:enlDf (tmt 6 t· 0.85%
1
5
~

--+_T...__geom ,mear
, } •
geom nDn/m fixed corners
I
15~~71:~~~
'failure load
(Yield fine Theory)

- - - frat results

qs 1,0

F1g. 24. Load-m1dpoint deflection relations of the square,

corner supported slab by the tests of McNeice [24].
Calculation results of Mehlhorn/Krist jansson/Klein [1~
472

A comparison of the results of Mehlhorn/Kr~stjansson/Klein

[14] wlth the results of other authors is given in F~g. 25.

Only calculations without consideration of tension stiffen~ng

of concrete as in the calculation of Mehlhorn/Krist jansson/

Klein ~4] are chosen. For comparison the geometrically linear
calculat~ons of Mehlhorn/Krist jansson/Klein are used because
the other mentioned authors neglect the geometric nonlinear
terms. Scanlon [251, Hand/Schnobrich/Pecknold [26], L~n [2~
and Gilbert/Warner [28] use the layered Finite Element Model.
Hand/Schnobr~ch/Pecknold [261 and Lin [27] allow for the ~n­
fluence of in-plane forces. For the higher load levels, the
agreement of the results by the d~fferent calculations is
qu~te suff~cient. The differences after primary cracking
depend on the assumed tensile strength of concrete. Bashur/
Darwin [29} obtain very good agreement w~th the test results.
They use the integrated model for the descr~ption of the
deformation behaviour of the reinforced concrete. In their
moment-curvature relations, cracking of concrete and yiel-
ding of reinforcement are considered.
 473

load
p~~

~~-------r--------+-

¥ r-H-----r------

qz

Flg. 2 5 . Comparison of the calculation results of Mehlhorn/

Krist jansson/Klei n [14J with the results of other a uthors
474

Tab. 1: Solutlon Methods Used by the Darmstadt Group

Ideallsatlon by Flnlte Elements

Author Investlg. Klnd of Model Type of Element

structures

Stauder [10] panels under

In-plane forces
sep. ldeal. of
concrete and
concrete: O .. u; v

relnforcement
relnf. : =-u
bond: onedun. llnk
element
Hoshlno [11] panels under
In-plane forces
sep. ldeal. of
concrete and
relnforcernent
concrete:

relnf. :
L:4/:
=-u
v

bond: onedlIn. contact

element
Schwlng [12] prefabrlcated sep. ldeal. of concrete, relnf. and
panels under concrete and bond SllP llke Stauder
In-plane forces relnforcement bond between prefabr.

il{U:
panels: two-dlm.llnk
element
Schafer [13] slabs, lncludlng comb. plate panel u , u'n, v: v'-dw'n
In-plane forces element (;rnt. pOlnts
V, jWi w, r
\l,x: u, , v: v'x:
W, i
w,~x; W'x;i W'~y
Mehlhorn/Krlst- slabs, lncludlng comb. layered plate
]ansson/Kleln In-plane forces panel element L!{:u, v: w'n
Ui Vi w: W'xi W'y
[14], [15], [23J w'xx' W'xyi W'yy
Maurer [16] folded plate comb. layered plate concrete } [ > { u : v, w,
structures panel element +
relnf. bars ~~~~f.W~~~hs 9 x i9y iBz
relnf. bars. ~"'u

t>t
Zeltler/Dletrlch/ slabs, lncludlng comb. layered plate u'x i U'y:
Mehlhorn [17] In-plane forces panel element v, v'x' V'yi
W: w'x i W'y
Muller [18J panels under
In-plane forces
sep. ldeal. of
concrete and relnf.
concrete +
relnf. mesh
relnf. bars: =
t7 u: v
__ u
bond: contact element
Dorr [19] panels under sep. ldeal. of concrete comb. ~; v
In-plane forces concrete and relnf. wlth relnf. mesh
bars relnf. bars: =<-u
bond: onedlm. cant. element
 475

Formulat~on of the Mater~al Propert~es Descr~pt~on of the Cracks

~- e.-Rel. Fall. Crlt. er- c -ReI. Bond- Develop. Transrn. Irens. St~ff Iteratl.on
of Concrete of Concrete of Relnf. 51 1 9 of Cracks of Shear of Concrete Procedure

unlaxlal hypo of Ml.ses elastlc- elastlc- along fully st~ffen~ng pure lterat.
nanllnear 1.n compr., ~deal- ~deal- element of bond- method w~th
(DIN 1045) max cr 1n plast~c plastlc boundar Ilnks secant
tensl0n stlffness

unlaxlal unrestralned I1near I1near- crack-held fully no gradual cons

11near 1n compr., elastlc elastlc w~th~n the of cracks by
elastl.c max C5 1n element l.nl.tlal
tenslon stresses

unlaxlal hypo of Mlses elastl.c- elastl.c- along the nonllnear --- pure lterat.
elastl.c- (no reduct. of ~deal ~deal- ]ol.nts of normal method
~deal tenslle plast~c plast~c prefabr. shear w~th
plashc strength) panels force secant
lnteract. st~ffness

blaXl.al bl.axl.al nonll.near perfect crack held reduct of stral.n soft l.ncrernental+
nonll.near (L~nk [7] ) w~th~n the shear of concrete mod. Newton/
(L~nk [7] ) element st~ffness Raphson ~ ter

b~ax~al b~ax~al nonl~near perfect crack f~eld reduct~on no ~ncremental+

nonl~near (L~nk [7] ) or elast. w~th~n the of &hear mod. Newton/
(L~nk [7] ) ~deal pI. element st~ffness Raphson ~ ter

b~ax~al b~ax~al nonl~near perfect crack held reduct. stra~n soft ~ncremental+
nonllnear (L~nk [7] ) or w~th~n the of shear of concrete mod. Newton/
(L~nk [7J ) b~l~near element st~ffness Raphson 1 ter

un~ax~al unrestr. comp.J l~near perfect crack f~eld fully no gradual cons
llnear max er ~n elastlc w~th~n the of cracks by
elast~c tensl0n element new st~ffness

blaxlal blax~al b~l~near non- crack held reduct. , no ~ncremental+

nonl~near (L~nk [7] ) l~near w~th~n the depends 0 mod. Newton/
(L~nk [7] ) element the crack Raphson lter
w~dth

blax~al blax~al b~llnear nonllnear, crack held reduct. , st~ffen~ng ~ncremental+

nonl~near (L~nk [7] ) depends on wlthln the depends on of re~nf. Newton/
(Lwk [7] ) lateral element the crack bars Raphson ~ ter
compr. w~dth
476

References

Leonhardt, F. und R. Walther: Wandartige Trager.

Deutscher AusschuB fUr Stahlbeton, Heft 178. Verlag
Wllhelm Ernst & Sohn, Berlin 1966.

2 Franz, G.: tiber die Beanspruchung der Bewehrung von

kreuzweise bewehrten vierseitig frei drehbar gelager-
ten Rechteckplatten aus Stahlbeton. Untersuchungs-
bericht des Instituts fUr Beton und Stahlbeton, Karls-
ruhe, 1970.

3 Eibl, J., G. Ivanyi: Studie zum Trag- und Verformungs-

verhalten von Stahlbeton. Deutscher AusschuB fUr Stahl-
beton, Heft 260, Verlag Wilhelm Ernst & Sohn, Berlin
1976.

4 Wegner, R.: Finite Element-Models for Relnforced

Concrete. US-German Symposium on "Formulation and
Computational Algorithms in Finite Element Analysis",
M.I.T. Boston, Aug. 1976.

5 Schnobrlch, W. C.: Behavl0r of Reinforced Concrete

Structures, Predicted by the Finite Element Method.
Computers and Structures, Vol. 7, pp. 365 - 376,
Pergamon Press 1977, Great Britain.

6 Kupfer, H.: Das Verhalten des Betons unter mehrachsi-

ger Kurzzeitbelastung unter besonderer BerUcksichti-
gung der zweiachsigen Beanspruchung. Deutscher Aus-
schuB fUr Stahlbeton, Heft 229, Verlag Wilhelm Ernst
& Sohn, Berlin 1973.

7 Link, J.: Eine Formulierung des zweiaxialen Verfor-

mungs- und Bruchverhaltens von Beton und deren Anwen-
dung auf die wirklichkeitsnahe Berechnung von Stahl-
betonplatten. Deutscher AusschuB fUr Stahlbeton, Heft
270, Verlag Wilhelm Ernst & Sohn, Berlin 1976.

8 Goto, Y.: Cracks formed in Concrete Around Deformed

Tension Bars. Journal of the American Concrete Inst.,
April 1971.

9 Dorr, K., G. Mehlhorn: Berechnung von Stahlbetonschei-

ben im Zustand II bei Annahme eines wirklichkeitsnahen
Werkstoffverhaltens. Forschungsberichte aus dem
Institut fUr Massivbau der Technischen Hochschule
Darmstadt, Heft Nr. 39, 1979.

10 Stauder, W.: Ein Beitrag zur Untersuchung von Stahl-

betonscheiben mit Hilfe finiter Elemente unter Be-
rUckslchtlgung elnes wlrkllchkeitsnahen Werkstoff-
verhaltens. Dissertation, TH Darmstadt, 1973.
 477

11 Hoshino, M.: Ein Beitrag zur Untersuchung des Span-

nungszustandes an Arbeitsfugen m1t Spanngliedkoppe-
lungen von abschnittsweise in Ortbeton hergestellten
Spannbetonbrlicken. Dissertation, TH Darmstadt, 1974.

12 Schwing, H.: Zur w1rkl1chkeitsnahen Berechnung von Wand-

sche1ben aus Fertigte1len. D1ssertat1on, TH Darmstadt,
1975.

13 Schafer, H.: Zur Berechnung von Stahlbetonplatten.

D1ssertation, TH Darmstadt, 1976.

14 Mehlhorn, G., R. Krist]ansson, D. Klein: Berechnung von

dlinnen Stahlbetonplatten bei Berlicks1chtigung e1nes
wirklichkeitsnahen Werkstoffverhaltens. Forschungs-
ber1cht aus dem Institut flir Mass1vbau der TH Darm-
stadt, Heft Nr. 44, 1980.

15 Krist jansson, R.: Physikal1sch und geometrisch n1cht-

lineare Berechnung von Stahlbetonplatten m1t H1lfe
Finiter Elemente. Dissertation, TH Darmstadt, 1977.

16 Maurer, G.: Zum Tragverhalten des einzelligen Spann-

betonkastentragers. Dissertation, TH Darmstadt, 1979.

17 Ze1tler, W., R. Dietrich, G. Mehlhorn: Erfassung des

raumlichen Spannungszustandes 1n Querschnitten von
Massivbrlicken unter besonderer Berlicksicht1gung des
Spannungszustandes an Bauabschnittsgrenzen mit Spann-
glledkopplungen. Forschungsbericht aus dem Institut
flir Massivbau der TH Darmstadt, Heft 42, 1978.

18 Mliller, T.: Stahlbetonscheiben mit offnungen. Disser-

tation, TH Darmstadt, 1979.

19 Dorr, K.: Ein Be1trag zur Berechnung von Stahlbeton-

scheiben unter besonderer Berlicksichtigung des Ver-
bundverhaltens. Dissertation, TH Darmstadt, 1980.

20 Cornelius, V.: Zum EinfluB der Umlenkkrafte aus Vor-

spannung auf die Tragfahigkeit von Spannbetonbauteilen.
Dem Fachbere1ch Konstruktiver Ingen1eurbau der TH
Darmstadt vorgelegte Dissertat1on, 1980.

21 Ge1stefeld, H.: Stahlbetonscheiben 1m gerissenen Zustand

- Berechnung mit Berlicksichtigung der riBabhangigen
Schubsteifigkeit im Mater1algesetz. D1ssertation, TU
Braunschwe1g, 1976.

22 Stegbauer, A.: Beitrag zur Berechnung von Stahlbeton-

flachentragwerken mit wirklichkeitsnahen Materialge-
setzen unter besonderer Berlicksicht1gung des Platten-
balkens. Dissertation TU Mlinchen, 1975.
478

23 Klein, D., R. Krist jansson, J. Link, G. Mehlhorn,

H. Schafer, K. Schneider: Zur Berechnung von dUnn en
Stahlbetonplatten bei BerUcksichtigung eines wirklich-
keitsnahen Werkstoffverhaltens. Forschungsbericht Nr.
32 aus dem Institut fUr Massivbau der TH Darmstadt,
1976.

24 Jofrlet, J. C., G. M. McNeice: Finite element analysis

of reinforced concrete slabs. Journal of the Struct.
Division, Proc. ASCE, Vol. 97, No. ST3, 1971.

25 Scanlon, A.: Time dependent deflections of reinforced

concrete slabs. Ph. D. Thesis, Department of Clvil
Engineering, University of Alberta, Edmonton, Canada,
1971.

26 Hand, F. R., W. C. Schnobrich, D. A. W. Pecknold: A

Layered Finite Element Nonlinear Analysis of Reinforced
Concrete Plates and Shells. Unlversity of Illinols,
Urbana, Illinois, SRS No. 389, CiVll Engineering
Studles, Aug. 1972.

27 Lln, C. S.: Nonlinear Analysis of Reinforced Concrete

Slabs and Shells. Ph. D. Thesls, Div. of Structural
Engineering and Structural Mechanics, Unlverslty of
California, Berkeley, April 1973.

28 Gilbert, R. J. und R. F. Warner: Tension Stiffening in

Reinforced Concrete Slabs. Journal of the Structural
Division, Proceedings of ASCE, Vol. 104, ST 12, Dez.
1978.

29 Bashur, F. K. und D. DarWln: Nonlinear Model for

Relnforced Concrete Slabs. Journal of the Structural
Divlsion, Proceedings of ASCE, Vol. 104, ST 1, Jan.
1978.
PartN:
Nonlinear Dynamics
Finite Element Analysis of Transient Nonlinear Response of
Reinforced Concrete Structures

M.J. MIKKOLA, H. S. SINISALO

Helsmkl Umverslty of Technology, Fmland

Abstract

Material nonlinearitles of reinforced concrete are described

as follows: tensile cracklng by a criterion deflned in terms
of stresses, inelastic response in compresslon by elastic
plastic relatlonships, and effects of aggregate interlocklng,
dowel actlon and bond failure by approximate means. Governlng
equatlons are derived using the total Lagrangian approach.
In time integratlon, the central dlfference method lS employ-
ed. Numerlcal applicatlons include quaslstatic, cyclic and
lmpulslve loadlngs of slabs and beams. Computed results pro-
vided by the flnite element procedure are compared with
avallable experimental data.

Introductlon

Major sources of nonlinearity in reinforced concrete struc-

tures are the progressive cracklng of concrete in tension
zones and the inelastic response ln compresslon, the in-
elastlc deformatlon of reinforcing steel, and other non-
linearities related to reinforcement and its interactlon
wlth concrete. In long term loadlng, the tlme dependency of
deformation of concrete becomes a characterlstlc part of the
response. In trans lent loading, the effect of high straln
rate on the behaviour of concrete and steel should also be
taken lnto consideratlon.

The finite element method has enabled the solution of the

complicated problems represented by the nonlinear behaviour
of reinforced concrete structures. Several studies on rein-
forced concrete under static loads have been published in
recent years, see the excellent review [1] and the references
given there. Dynamlc behaviour of relnforced concrete struc-
~2

tures has been 1nvest1gated in earthquake engineering. How-

ever, relat1vely few reports deal with the nonlinear trans-
ient response of reinforced concrete structures [2]-[9].
Various aspects of impact and impulsive loadings on buildings
are also discussed in the extensive review by a RILEM commit-
tee [10]. It 1S obvious that in many problems of structural
design, part1cularly in new fields like nuclear and off-shore
engineering, the appl1cat1on and development of powerful and
accurate tools, such as the finite element method, are re-
quired for the assessment of safety and deformational
character1st1cs of structures.

Th1S paper reports on a study of reinforced concrete beams

and slabs subJected to transient impulsive loading. In the
initial state, the composite material formed by concrete and
re1nforcement 1S assumed to behave elastically. Tensile
cracking of concrete is determined by a stress criterion.
Inelastic response 1n compression is described by the theory
of plast1c1ty, Wh1Ch is extended also to the softening
region. As a second model, a yield function and cracking
cr1ter10n depend1ng on the strain rate are applied. The
purpose 1S to allow for the rate effect in tensile and
compressive regions. The reinforcement is described as
smeared and orientated steel layers. The cracks in concrete
are treated as smeared, and the nonlinearity effects in
cracks, such as the aggregate interlocking, the dowel action
and the tension stiffening are given by an approximate
formula.

The equation of motion is derived using the principle of

v1rtual d1splacements in total Lagrangean formulation. The
kinematic equations of plates are taken in accordance with
the theories of von Karman and Mindlin. The problem is dis-
cretized spatially and temporally by use of the finite
element and finite difference methods, respectively. Some
problems of beams and slabs subjected to static, cyclic and
dynamic loads are analyzed and the computed results compared
with available experimental and numerical values.
Modelling of re~nforced concrete

The behaviour of plain concrete in multiaxial states of

stress has been the objective of intensive research for many
years. Const~tut~ve modelling of inelast~c short-term behav-
iour of concrete has been performed on the basis of nonlinear
elastic,hypoelastic, elasticplastic, endochronic and fracture
models [11]-[15]. From the experimental investigations the
references [16] and [17] can be mentioned here. A consider-
able variation of experimental results exists. The primary
reason seams to be in test arrangements. A review of the
proposed plastic yield functions and their interrelations
with experiments is given by Wastiels [18]. Theory of plastic-
ity provides a satisfactory model for the behaviour of
concrete. However, it is apparent that the models which are
capable to describe changes in microstructure of concrete due
to straining (microcracking) and the correspond~ng changes in
response are more real~st~c. Such are the endochronic and the
plast~c-fractur~ng models.

The propert~es of concrete at high strain rates are much less

known than in quasistatic loading. Some experimental results
for uniaxial loading exist [10], [19]-[22]. They indicate
that the compressive strength is increased by 0-80 %, the
tensile strength by 0-50 %, and the init~al modulus of
elasticity by 0-40 %, when the strain rate varies in the
range 10- 4 •.. 10 lis. The effect of two- or three-dimensional
state of stress on the dynamic response corresponds to that
of quasi static case [22]. The strain rate effect on strength
can be incorporated into viscoplastic model or elastic
plastiC model with a yield function depend~ng on strain rate.

In this study, the response of concrete is described by the

elastic plastic model. In compressive region, the strain
hardening and strain softening are included. In the tensile
region, the cracking is determ~ned by a stress criterion. The
strain rate effect is included by rate dependency of the
yield function.
484

Elastic plastic model

The behaviour in one dimensional fashion is illustrated in
Fig. 1. After tensile fracture, the stress should drop to
zero, but in this model it decreases linearly, which is
intended to simulate the tension stiffening effect observed
in reinforced concrete. In compression, the behaviour is
linear up to stress value -0.4 f . Beyond this point, the
c
equation proposed by Sargin [23]

(1)

is employed. fc is the maximum compressive stress and Eo the

corresponding strain. A, B, C, and D are experimental para-
meters. This equation also describes strain softening. Strain
softening does not comply with the stability requirements of
the theory of plasticity, but it is accepted here as well
as the normality rule for plastic strain rates in softening
region (see [24] and [25]).

Fig. 1. Uniaxial stress strain

diagram of concrete

In unloading, the stress reversal obeys the initial modulus

of elasticity. In tensile region, the stress is first reliev-
ed to zero and the cracks are closed before the material can
take any compresSlve stress (Fig. 1).

In general three- or two-dimensional case, the response of the

the elastic plastic material is

.
cr C£ - CAn (2)
 485

where C is the elasticity matrix, n = df/do the gradient of

the yield surface f = 0 and A the plasticity parameter

A = H-1/ IT')
0 = \n C£ /(H+n T Cn)
\n T') (3 )

The notation <x) means <x) = x if x ~ 0, <X) = 0 if x < O.

The modulus of plasticity H(£P) = da/d£P can be found on the
basis of the uniaxial stress-strain relationship. The actual
yield condition, which was applied in numerical calculations,
is

f(o,£P) (4)

The function g(9)

2 2 2 1/2
2 (r -1) cos9+ (2-r) [4 (r -1) cos 9+5-4r]
g(9) , (5)
4(r 2 -1)cos 2 9+(r-2)2

taken in accordance with Ref. [26], defines an elliptic form

for the sixth part of the yield locus in the deviatoric plane.
The angle S, varying between 0° and 60°, is

9 (6 )

The factor r = g(600)/g(00) determines the ratio of the

radius at the "compressive" meridian to the radius at the
"tensile" merid1an. The value r = 1.5 was found to be
appropriate [17]. J l , J 2 , and J 3 are the invariants of the
stress tensor 0ij and deviator Sij

(7 )

£P is the effective plastic strain

(8 )

a and b are parameters to be determined on the basis of

experimental data [16]. Tensile strength f t = 0.1 fc and
biaxial strength fc2 = 1.16 fc yield a = 1.138, b = 0.138
486

for compressive region and a = 5.5, b = 4.5 for tensile

region. The biaxial behaviour of the material is illustrated
by Fig. 2. In Fig. 3, the predicted diagrams of the axial
and lateral strains and the dilatation in uniaxial compress-
ion are shown together with experimental curves [16].

Fig. 2 Biaxial
behaviour of
ELASTIC LIMIT concrete
HARDENING
REGION

_ _ _-SOFTENING

Elastic viscoplastic model

The strain rate dependency of the strength of concrete

could be described by Perzyna's viscoplastic model

(9)

which leads to the dynamic yield condition

1
f - ~-l[ -l(l
y 2
l! l!)-
dada
2 ~vp] = 0 (10)

where ~VP ~E~~E~~/3i is the effective viscoplastic strain

rate Ins~eaa
, OIJ.: J tne
:- J common V1SCOP
. 1 as t lCl
. t y t"heory, ano th er
approach is chosen here. Following the suggestion by Nlls-
son [7], a yield limit, which is dependent on the strain
rate, is lnserted into the yield function (4)

o (11)
 487

E 1%0
-1 -2

o/fc

.8

.6
fc= 32.17 N/mm2
.4
E =32500 N/mm2
.2
I1v %0
-.2 -.4 -.6 -.8 -1 v

Fig J. Axial and lateral strains and dilatation in uniaxial

compression

The effective strain rate is ~ = V2~ . .~.

1J 1J
/3' and the coeffic-
ients are c l = 1.6, c 2 = 0.104, c 3 = 0.0045, obtained by
curve fitting on experimental data. Yield of concrete is
checked w1th respect to the dynamic yield condition (11). The
cracking criterion is changed in similar fashion due to the
strain rate effect.

Cracking

Cracking of concrete occurs when the cracking criterion of

the form (4) is satisfied. The d1rect10n of the cracking
plane 1S perpendicular to the largest principal stress. After
crack1ng, the three-dimens1onal state of stress changes to
~8

two-dimensional, etc. The crack is closed, when the strain

perpendicular to the cracking plane becomes zero or
compressive.

Reinforcement
For reinforcing steel, the uniaxial bilinear elastic plastic
material model is used.

Reinforced concrete
Complete compatibility between concrete and reinforcement
is assumed, so that the deformation is continuous over the
adjoining boundaries of concrete and steel. In slabs and
panels, the reinforcement net is treated as smeared layers
of thin wires. The effect of separate reinforcing bars is
transferred to the nodes of the finite element mesh through
the shape functions. The shear resistance, which stiil exists
in the cracked region due to aggregate interlocking and
dowel action, is taken into account in reduced shear modulus
of concrete [27]

0.1 G (12)

Equations of equilibrium
In total Lagrangian approach, the principle of virtual work
can be presented in the form

I88EdV - If8udVo - It8ud8 + Ip U8udV o (13)

V 0 Vo 8 0 V 0 0
o ot 0

where 8 is the 2nd Piola-Kirchhoff stress, E the Green-

Lagrange strain E .. = (du./dx.+du./dx.+au k /ax.auk /dx.)/2,
~J ~ J J ~ ~ J
f the body force, t the surface traction,po the mass density,
u the d~splacement, li the acceleration, and Vo and 8 0 the
volume and surface in the reference configuration, respect-
ively. Use of the finite element approximation u = Nq
results in a system of nonlinear differential equations with
respect to time

R(q) + M q Q (14)
 ~9

where

(15)

are the internal force vector, the mass matrix, and the
external force vector, respectively. The incremental form of
the equation (14) is

(16)

where the tangent stiffness lK t = lK + lK + lK consists

L G 8
of linear, geometric and stress st1ffnesses. The superscr1pts
1 and 2 refer to two adjacent conf1gurations of the structure
on its deformation path. (For details, see Ref. [28] Ch. 19).
The 1ncrement of stress ~8 = 28 - 18 ~ 8~t is determined by
the constitutive equation

*
8 CE (17)

The Jaumann rate of the Kirchhoff stress (the components of

which are equal with those of the 2nd Piola-Kirchhoff
stress) is related to the convective rate as follows

(18)

Gij is the metric tensor in convective coordinates.

Here the equations of equilibrium have been presented in a

form valid for large deformations and rotations. In appli-
cation to reinforced concrete structures, however, the
observation was made that only in some cases of flexible
beams and slabs the effect of large rotat1ons was necessary
to be taken into account. In all other cases, the accuracy
provided by the geometrically linear theory was satisfact-
ory.
~o

Solution techniques

Several direct time integration schemes have been introduced

and discussed in papers on structural dynamics. Comparative
computations have shown [29] that the central difference (CD)
method and the Newmark ~-method perform accurately and
effectively in various problems of transient loading. In the
solution of transient response of reinforced concrete struc-
tures, the explicit CD-scheme and the implicit trapezoidal
rule or the Newmark ~-method with parameters ~ = 0.25 and
y = 0.5 were applied.

When the CD-scheme is used, the solution qn+l of the equat-

ion (14) at the instant tn+l is computed from formula

(19)

where h = tn+l-t n is the step length. The strain and stress

increments are

~E (20)

The internal force vector at the instant tn+l is evaluated

in accordance with equation (15), using Bn+l and Sn+l = Sn+
+~S. The initial condition qo = (ql-q_l)/2~t is used for
elimination of q-l in the first step. The CD-scheme with
diagonal mass matrix is accurate and simple. As an expllcit
difference method its step length is limited by the largest
natural frequency of the finite element equation system,
h < 2 /w .
max

The incremental equation (20) is solved applying the

trapezoidal rule

(21)

This is an implicit scheme and therefore iteration has to

carried out at each time step. The displacement vectors
i
qn+l an d qn+l
i+l ln
° th e l
°th
a n d (01+ l)th 1°tera tolon cyc 1 es
 491

correspond to conf1gurations 1 and 2 in equation (16). Use

of equat10ns (16) and (21)2 yields

'+1 Hl i
where ~q1 qn+l - qn+l. For the first iteration cycle,
ql+l = q is taken. The iteration is continued until
n -i+l n i+l ,
II Aq II < e: II qn+ 1 - qn II where e: 1S a tolerance parameter.
Consistent mass matrix is used. In order to avoid too drast1c
or artif1cial changes due to cracking, the tangent stiffness
was updated in the first iteration cycle but held constant
thereafter. This also reduces computing time.

The stability of the solution was followed by an energy

balance check [30], [29]. If T, U and W denote the kinetic,
internal and external energies, respectively, the energy
balance criterion is

~U + ~T - ~W < e:(U+T) (23)

where e: is a given tolerance. The energy increments can be

calculated using the trapezoidal integration

~U

~T (24)

In appl1cations, it was found that the CD-method was much

more economical than the trapezoidal rule. Therefore most
of the results were calculated using the CD-procedure.

In quasistatic cases, Solut1ons were achieved by the

incremental equation (16) using similar iteration as in the
implicit scheme.
492

Applications

Statically loaded slab

The corner supported square slab tested under central point

load by Jofriet & McNeice [31] was analyzed in order to
check the performance of the computing procedure. The plate
theories by von Karman and Mindlin were applied. The non-
linear strain-displacement relationships of Mindlin's theory
are

au/ax + (aw/ax)21z+za<p/Clx
Clu/Cly + Clv/ax + (aw/ax) (aw/ay) + z(a<p/
(25 )
lay + a'li/ax)
Yxz = aw/ax + <p

etc. where <p and 'Ii represent the rotations of the normal
with respect to x- and y-axes. The expressions of von
Karman follow when the constraints Y
xz
= Y
yz
= 0 are
~mposed.

In von Karman's plate theory, a 24 DOF rectangular element

was used with 4 nodes and displacement parameters u, v, w,
Clw/Clx, aw/ay, a 2w/axay at each node. Integration was carried
out by 2x2 Gaussian formula. In Mindlin's plate theory, a
20 DOF isoparametric element with 4 nodes and displacement
parameters u, v, w, l!>, 'Ii at each node was employed. 2x2
Gaussian integration rule was used except for the shear
deformation only one integration point. Integration through
thickness was performed by Simpson's rule with 7 integration
pOints.

Load-deflection curves of the point A are shown in Fig. 4.

The influence of tension stiffening was investigated using
three different assumptions for the tensile stress-strain
behaviour. Assumption 2 provides the best fit with the ex-
per~mental curve. Von Karman element with a 3 by 3 mesh for
a plate quadrant was used. The mesh for Mindlin element was
2 by 2.
 493

" I

? ... ffi:
I
I
I
---+--- 0.85%
I
I A z
I P
I I
I I
--j--,--
I I
1 2 3 E
--~-....j.-- 0.2 0.4 0.6 %0
I I
I I
~

36 IN
,/ 7

P[klbfl
3.6

2.8

2.0

1,2,3 KIRCHHOFF M33

1.2
x-x- MINDLIN M22

.4
def
4 8 12 16 20 24 28 32 36 40

Fig. 4. Load-deflection curves of two-way slab tested by

McNelce [31]. The data are f = 5500 pSi, f t = 550 psi,
E = 4.15 10 6 psi, Vc = 0.15; f 60 000 psi, E = 29 10 6
c. s S
PS1,
494

Cyclically loaded beam

Burns & Siess [32] made tests on reinforced concrete beams

subjected to cyclically varying point load. The computatlon
was based on a nonlinear strain-displacement relationship
corresponding to the von Karman plate theory. Six 8-DOF beam
elements with displacement parameters u, du/dx, w, dw/dx at
each node were used for a beam half. Load-deflection curves
for 4 cycles are shown in Fig,S. The loading criterion for
the 1st cycle was to obtain a crack through the thickness in
the middle of the beam, for the 2nd cycle to apply a load
which is half of the load inducing first yield of the rein-
forcement, for the 3rd and 4th cycles to obta i n the same
permanent deflection at zero load as in the test.

24" I t------- -- -- --g A~-;t- ---j

(609.6mm) == - - ---
tp A pt
"2 2"
I I I I I I
12" 66" 6"
(304.8mm) (1676.3mm) (1S2.4mm)

2 8 ~[O]I(203.2mm)
(2 cp2S:mm)", 8" I 12"
(304.8mm)

~ P[kN ]
(203.2mm)
40

CYCLE 1 CYCLE 2

2 v[mm ] 4 6 v [mm)

- 20

-40
- -- EXPERIMENTAL
--0--0-0- CALCULATED

Fig,S. Load-deflection curves of cycllcally loaded beam

tested by Burns & Siess [32].
 495

P[kN] CYCLE 3

CYCL E 4
P[kN]

90

60
v [mm ]

40 v [mm ]

Fig. 5. Continued

Dynamically loaded beam

Nosseir [2J made tests on simply supported reinforced beams
loaded by a freely falling mass. Based on some measurements,
he estimated the contact force between the beam and mass by a
bilinear expression wlth a peak value of 123.9 kN. Midpoint
deflection of the beam versus time is presented in Fig. 6.
The computed curve agrees qualitatively with the experimental
one, but the computed deflection is somewhat smaller and the
phase is lagging behind the experimental . Constant tlme step
h = 0,6 ms was used.
496

v [mml
123.9
295
k:L I
P[kNl

t
,IE
.. 2

i~-----~.....,........J
I
6066 [msl
50" lIffIr
I

1D
(1270mm)
20

(254mm)
10"
2" 00 2 #6
(50.8mm) ~ (2ct>19.05mm)
6"
(152.4mm)
- - EXPERIMENTAL
- - - CALCULATED
10

t[msl
10 20 30 40 50 60

Flg. 6. Mldpolnt deflection vs. time of beam under

dynamic load

Dynamically loaded slabs

A clamped rectangular slab subjected to a Jet force at the

center was analyzed by Stangenberg [3J using a difference
method. The deflection histories of the midpoint are shown
in Fig. 7. The finite element results based on plasticity
and thln plate theory compare favourably with the finite
difference solution, but a considerable deviation occurs
for the results of the thick plate theory. The rate depend-
ent model predicts smaller deflections than the elastic
plastic model.
 497

12000mm PIMNl
I 10
I 8
I
----1- - - -
E I
E I
<:)
<:) I P
:
<:)
<:)
I
I
--"1--T--
I I
----i--t---
: I
I
5ms 100ms

def. [mm)
12
11 0-0 MINDLIN THEORY
2 x 2 el.
10 ~ KIRCHHOFF THEORY
2x2 el.
9 3 x 3 el.
8
7
6
5
4
3
2
1

5 10 15 20 25 30 t [ms)

Fig. 7. Midpoint deflection vs. time of clamped slab under

central jet force. Data fc 22.7 MFa, f t = 2.27 MFa,
Ec = 33 GPa, Vc = 0.15, fs 210 GPa, p = 2400 kg/m 3
498

p 2
1+ =2000 N/m2 A

P = 3.4 MN/m
I = 300 N/m2
+
t = 2.7 ms
t- = 40 ms
t t
Mlndlm's theory

experiment [6 ]
central
deflection
100 [mm] FEM [6]

so P uniformly distributed
over plate area

t [ms]

10 20 30 40
Fig. 8. Central deflection vs. time of one-way slab under uniform
pressure. Data. reinforcement 0.17% and 0.085% fc = 36.6 MPa,
f t = 4.90 MPa, Ec = 28.4 GPa, Vo = 0.15, fs = 450 MPa, Es = 210
GPa, span 2280 rom, width 1230 rom depth 130 rom.
A rectangular slab, simply supported on two opposite edges and
free on the remaining edges, was considered by Nilsson & Johansson
[6] . The slab was subjected to a uniformly dlstributed pressure
load varying with time. In addition to the test, the authors made
finite element calculations which included cracking and plastic
deformations. The central deflection-time history of thlS study
was computed using four elements for a slab quadrant (Fig. 8).
Varying time step h = 0.025 ... 0.5 ms was used for the Kirchhoff
theory with 2 by 2 mesh and h = 0.01 ... 0.2 ms for the Kirchhoff
theory with 3 by 3 mesh and for the Mindlln theory.

Discussion
The numerical results obtained for the problems of beams and slabs
indicate that the present method is capable to predict satisfac-
torily the response of relnforced concrete structures subjected
to transient loading. More numerical results and experimental
data are still needed. Modelling of rate dependent behaviour should
be developed. More realistic description of aggregate interlocking,
dowel action and bond slip is also necessary.
 499

References

1 Bergan, P.G.; Holand, I.: Nonlinear finite element

analysis of concrete structures. Computer Methods
in Applied Mechanics and Engineering 17/18 (1979)
443-467.

2 Nosseir, Mohamed S.E.B.: Static and dynamlc analysis

behavlour of concrete beams failing in shear.
Ph. D. Thesis, Unlverslty of Texas, 1966.

3 Stangenberg, F.: Nonlinear dynamic analysis of rein-

forced concrete structures. Nucl. Eng. and Design
29 (1974) 71-88.

4 Rebora, B.; Zimmermann, Th.1 Wolf, J.P.: Dynamic

rupture analysis of reinforced concrete shells.
Nucl. Eng. and Design 37 (1976) 269-

5 Buyukozturk, 0.1 Connor, J.J.: Nonlinear dynamic

response of reinforced concrete under impulsive
loading: Research status and needs. Nucl. Eng. and
Design 50 (1978) 83-92.

6 Nllsson, L.; Johansson, I.: Analys med finit element-

metod av luftstotvagsbelastade enkelspanda betong-
plattor. FOA Rapport C20204-D4, Stockholm, November
1977.

7 Nllsson, L.: Impact loadlng of concrete structures.

Publication 79:1, Department of Structural Mechanics,
Chalmers Unlversity of Technology, 1979.

8 Sinlsalo, H.S.f Tuomala, H.T.E.; Mikkola, M.J.:

Nonllnear finlte element analysls of reinforced
concrete slabs subJected to transient impulsive
loading. Int. Conference on Concrete Slabs, 3-5
April 1979, Dundee, Scotland. Advances ln Concrete Slab
Technology, Proceedln0s publlshed by Pergamon Press,1980.
9 Sinisalo, H.S.; Tuomala, M.T.E.; Mlkkola, M.J.:
Nonlinear dynamlc analysis of reinforced concrete
slabs. Plastlclty ln Relnforced Concrete, Final
Report, IABSE ColloqUlum, Copenhagen, 1979, PP.
239-246.

10 The effect of lmpact loading on bUlldlngs. State-

of-the-art reports. Materiaux et Constructions
(RILEM) 8 (1975) 44, 77-129.

11 Cedolin, L.; Crutzen, Y.R.J.; Del Poli, S.: Tri~xial

stress-strain relationshlp for concrete. J. Eng.
Mech. Div. ASCE 103 (1977) EM3, 423-439.

12 Coon, M.D.; Evans, R.J.: Incremental constitutive

laws and thelr assoclated fallure criteria with
application to plain concrete. Int. J. Solids
Struct. 8 (1972) 1169-1183.
500

13 Chen, W.F.: Constitutive equations for concrete.

Plasticity in Reinforced Concrete, Introductory
Report, IABSE Colloquium, Copenhagen, 1979, pp.
11-34.

14 Bazant, Z.P.; Bhat, P.D.: Endochronic theory of

inelasticity and failure of concrete. J. Eng. Mech.
Div. ASCE 102 (1976) EM4, 701-722.

15 Bazant, Z.P.; Kim, S.-S.: Plastic-fracturing theory

for concrete. J. Eng. Mech. Div. ASCE 105 (1979)
EM3, 407-428.

16 Kupfer, H.B.; Hilsdorf, H.K.; Rlisch, H.: Behaviour

of concrete under biaxial stresses. J. of ACI 66
(1969) 8, 656-666.

17 Gerstle, K.H. et al.: BehaVlour of concrete under

multlaxial stress states. ASCE Annual Convention,
Chigao, Illinois, October 1978.

18 Wastiels, J.: Failure criteria for concrete under

multiaxial stress states. Plasticity in Reinforced
Concrete, Final Report, IABSE Colloquium, Copenhagen
1979, pp. 3-10.

19 Read, H.E.; Maiden, C.J.: The dynamlc behaviour of

concrete. Systems, Science and Software, Report 3
SR-707, 1971.
20 Goldsmith, W.; Pollvka, M.; Yang, T.: Dynamlc
behaviour of concrete. Experimental Mechanics 6
(1966) 2, 65-79.

21 Grlner, G.R.; Sierakowski, R.L.; Ross, C.A.:

Dynamic properties of concrete under impact loading.
Shock and Vibratlon Bulletln 45 (1974-75) 131-142.

22 Takeda, J.; Tachikawa, H.; FUJimoto, K.: Mechanical

behaviour of concrete under higher rate loadlng than
in static test. Mechanical Behaviour of Materials,
Symposium Procedure, Kyoto, Japan, 2 (1974) 479-486.

23 Wang, P.T.; Shah, S.P.; Naaman, A.E.: Stress strain

curves of normal and lightweight concrete in com-
pression. ACI Journal 75 (1978) November, 603-611.

24 Palmer, A.C; Maier, G.; Drucker, D.C.: Normality

relations and convexity of yield surfaces for un-
stable materials or structural elements. J. Appl.
Mech., Trans. ASME 34 (1967) June, 464-470.

25 Nguyen, Q.S.; Bui, H.D.: Sur les materiaux elasto-

plastiques a ecrouissage positif ou negatif.
J. de Mecanique 13 (1974) 2, 321-342.
 501

26 Willam, K.J.; Warnke, E.P.: Constitutive model for

the triaxial behaviour of concrete. IABSE Seminar on
"Concrete structures subjected to triaxial stresses"
May 17-19, 1974, ISlmS, Bergamo, Italy.

27 Lin, C.S.; Scordells, A.C.: Nonllnear analysis of

RC shells of general form. J. Struct. Div. ASCE 101
(1975) ST3, 523-538.

28 Zienklewicz, O.C.: The flnlte element method (3rd

ed.). McGraw-Hill, London 1977.

29 Mikkola, M.J.; Tuomala, M.T.E.; Slnisalo, H.S.:

Comparison of numerical integration methods in the
analysis of impulsively loaded elasto-plastlc and
vlscoplastic structures (Submitted for publlcation) .

30 Belytschko, T.; Schoeberle, D.F.: On the uncondlt-

ional stability of an impllclt algorithm for non-
llnear structural dynamics. J. Appl. Mech. ASME 42
(1975) 865-869.

31 Jofriet, J.e.; McNeice, G.M.: Finite element analys-

is of reinforced concrete slabs. J. Struct. Div.
ASCE 97 (1971) ST3, 785-806.

32 Burns, N.H.; Siess, C.P.: Repeated and reversed

loading In reinforced concrete. J. Struct. Div.
ASCE 92 (1966) ST5
An Analysis of the Stability and Convergence Properties of a
Crank-Nicholson Algorithm for Nonlinear Elasto-Dynamics
Problems

L. C. WELLFORD, JR., S.M. HAMDAN

University of Southern CalIfornia, Drexel University,
Los Angeles, USA Philadelphia, USA

Summary
The stability of nonlinear step by step computing schemes
for solid mechanics problems can be insured by either re-
quiring that the energy of the approximation be conserved or
by insisting that the energy be bounded for all time. While
techniques exist for analyzing conserving procedures, few
analysis methods exist which are capable of establishing the
boundedness of the energy of nonconserving computing schemes.
In this paper an analysis procedure for demonstrating the
stability and convergence of nonconserving step by step solu-
tion algorithms is introduced. As a typical example this
technique is used to analyze a Crank-Nicholson-Galerkin al-
gorithm for nonlinear elastodynamics problems.

Introduction
Energy analysis procedures have been used by several investi-
gators [1-4] to study the stability of step by step solution
algorithms for nonlinear structural dynamics problems. This
previous work has been principally directed to developing
algorithms which, either implicitly or by constraint, con-
serve some scalar or energy like quantity from time step to
time step. various unconditionally stable algorithms have
been constructed using these procedures. In this paper a
method of analysis for nonconserving algorithms is developed.
This analysis is based on the premise that, in the approxi-
mation, energy should rema1n bounded for all t1me. This meth-
od of analysis should be useful in analyzing various noncon-
serving implicit algorithms, load extrapolation algorithms,
and explicit algorithms.
 503

The Physical Problem

In the Lagrangian formulation of nonlinear elasticity the mo-

tion of a solid body is defined by the following equation:

pu 1· - [tkJ (0 + U. )),k = f. (1)

1J 1, J 1

where p is the density, u i is the displacement vector, t kj is

the stress tensor, and fi is the body force vector. The tech-
niques to be introduced here can be used in analyzing discrete
algorithms for (1). However, in order to concisely present
the method, it is useful to consider an alternate problem.
This problem is a one dimensional analog of (1). In the one
dimensional case the motion of a flat slab of material, I, is
considered. The function u represents the displacement com-
ponent, perpendicular to the slab, of planes in the slab.
The displacement component u is governed by the following
equation of motion:

pu - D[t(l + Du)l = f (2)

where D = a/ax, t is the stress, and f is the body force.

The analys1s procedure to be introduced here is designed for

the case in which the stress t is known in terms of powers
of the strain. Thus it would be particularly useful in con-
sidering hyperelastic materials and the case in which the
strain energy is known in terms of a power series in the strains.
However, in order to analyze a problem for which existence and
regularity propert1es are available, a particular case is con-
sidered. This is the geometrically nonlinear problem. In
this case t = EE
x
and E
x
= Du + 1/2 (Du)2 where E is the modulus
of elasticity and EX 1S the strain. In this problem it can
be shown, using energy arguments that, in the case when homo-
geneous boundary conditions are applied, if fEL2(O,oo;L2(I)),
a solution exists such that

This is the regularity that would be expected in a shock wave

propagation problem. In the transient vibration problems con-
sidered here the Solut10n is much more regular. Thus 1t is
assumed that for some 1nteger m : 2
504

• P.m-l
UEL",,(O,ooiW2 (I)).

Similar regularity assumptions will be introduced for higher

time derivatives when needed subsequently.

Regularization Procedures
Stability and convergence results, of the type discussed in
this paper, are extremely difficult to derive using the second
order equation (2). Thus (2) is represented as a first order
system and certain regularization (artificial viscosity) terms
are added to obtain the following form:
pv - D(t(l + Du)) + £v = f
.
u - v + au o (£,a>O) (3)

The analysis is limited by considering the case in which homo-

geneous boundary conditions on u are applied. In addition,
the nonlinearity in the equation is simplified by neglecting
the third degree term in the displacement gradients. This
restriction could be eliminated but at the expense of an in-
crease in the complexity of the arguments. Then the equation
(3)1 is satisfied in its weak form and (3)2 is sat~sfied in
its strong form to obtain
• 3 2
(pv,w) + (E(Du + 2" Du ),Dw) + £(v,w) (f,w)

it - v + au = 0 (4)

for all w in the original solution space of velocities.

The Semi-Discrete Solution

In order to form an approximate model which is discretized
in the spatial variable, a set of global finite element basis
functions is introduced. These basis functions define a fi-
nite dimensional subspace Sh(I) x Sh(I) of the original solu-
tion space. The elements of Sh(I) are assumed to sat~sfy an
inverse hypothesis [5] of the following form (for positive
constant C*):

I IDU I IL 2 (I) < rII I I U L

2
(I ) (UES h (I)) . (5)
 505

It is assumed that an imbedding relationship [6] of the fol-

lowing form is satisfied by the elements of Sh(I):
1
IIDUII L (I):' ch- p(p+l) IIDUII L (I) . ( 6)
p+l p

In addition, if k denotes the degree of the piecewise poly-

nomials in Sh(I), it is assumed that there exists a constant
K independent of h such that
INF 0
UE:: Sh(I)llu-ull~(I) :: Khk+l-mllull~~+l(I) (7)

Let (U,V)ESh(I) x Sh(I) be the approximation for the dis-

placement u and the velocity v. Then the approxlmate version
of (4) is defined as follows:

(pV,W) + (E(DU + ~ DU 2 ) ,DW) + s(V,W) (f,W)

U - V + aU = 0 (8)

The equation (8)1 is satisfied for all WESh(I) while (8)2 1S

satisfied pointwise at the nodes.

Stability and Convergence of the Semidiscrete Approximation

Let U and V denoted weighted projections of u and v into the

subspace Sh(I) defined by

(E[ 1+3Du]D(u-U) ,DW) 0

(EVD(v-V) ,DW) o ( 9)

where v = 1 + 3/2(Du + DU) . It is assumed that the projec-

tion is optimal in the sense of (7) • Let the approximation
u - U and e
u = =
errors be defined by e v - V. These errors
v
can be decomposed as follows:

eu E
u + &'U
E
u = u - U eu U U
e E +&' E v - V cf V V (10)
v v v v v

The energy error can be defined in certain weighted norms

U(W) !(EVDW,DW) - Strain Energy Error

D(W) !pl Iwi IL 2 (I) - Kinetic Energy Error

2

It is necessary, in the nonlinear problem, to scale these

506

energy errors by defining auxiliary energy error measures

U*(W) = U(W)e- xt ~*(W) = ~(W)e-xt (11)

where
SUP 3· -
X = t ~ [0,00] [2v(DU(t) + DU(t))]
XEI
In addition total energy errors can be defined by P(w,r) =
~(w) + U(r) qnd P*(w,r) = ~*(w) + U*(r). The errors 8 u and
8 v can be established through the following Theorem:
Thm. 1. If the exact solutions u and v are sufficiently regu-
lar and the mesh discretization parameter h is sufficiently
restricted, there exist positive constants gl' g2' and g3'
independent of hand t, such that

(12)

Proof: Setting w = W in (4)1' subtracting (8)1 from (4)1'

making use of (10), and setting W = 8 v ' the following equa-
tion is obtained:

ddt ~ (8v ) + (E'JD.g'

u
,D8v ) + ~
p
rl (8v )

3
ED8u 2
(2 ,D8v ) - '
(pE v , 8 v t

- (E(VDE u - ~ D8u DEu,D8v )

(13)

Similarly differentiating (4)2 and (8)2 with respect to the

spatial coordinate, subtracting the two equations, multiply-
ing by EvD8 , and integrating over the spatial domain, the
u
following equation is obtained:

d
dt U(8u ) - (EvD8v,D~u) + 2aU(8u )

(3E [D~ + DU] D8 ,D8 )

4 u u

- (EV[DE u - DE v + aDE u ] ,D8u ) (14)

Adding (13) and (14), it can be seen that

 507

(3E (O~ + OU)O/,:' ,08 )

4 u u

+ (-2
3E08 2,08 ) + 1> (15)
u v

where 1> is a truncation error term

1> = -(PE + EE ,8) - (E[vOE 3 080E ),08)

--2
v v v u u u v

- (E'J [DE
u
- DE
v
+ aOE ) ,08 )
u u

using (9) certain troublesome terms can be omitted from this

expression to obtain

1> = - (p E 3 E [DE 2 + 08 DE ),08 )

+ E E ,8 ) + (-2
v v v u u u v

- (EV[OE
u
+ aOE u ) ,08 )
u

The terms on the right in (15) can be estimated using the

Holder inequality, (5), and the elementary inequality ab ~
1 2 Y 2
2y a + 2" b (y > 0). In addition to estimate the nonlinear term
the following relationship, obtained from (6) is required:
4 Cl 4
1108)IL (I) ~_1108uIIL (I)
4 h 2
Typically, the first two terms on the right of (15) take the
following form:

(34E (O~ + OU) 08u ,08)

u
~ XU(8 )
u

(-l2 E08 2,08 )

u v

where v* = 1 - ~II OU + Dull L (I). Using similar methods to

bound the terms in 1>, it c~n be shown that, for certain para-
meters Yl' ... ' Y6 (with dependence on h indicated), the right
hand side of (15) denoted R is
2
R ~ XU(8u ) + d l U(8u ) + d2~(8v) + d 3 U(8u ) + d 4 (16)
508

where

v *2 Eh \ 1

3EC*2 y 1
2p + Y2 + 2p
3EIIDEuilL
00
(I)C*\4(~)
+ 2ph

31 I DEu I I Loo (J)

-----:;--- +
2v*y (-!) h v*
4 h
3E I I DE) I Loo (I) I I DE) I ~2 (I)
2~ 2 II E~II ~ 2 (I) +

2
+ 0: I I DEU I I L2 (I) )

where v = 1 + ~I IDU + Dull L (I)·

00

The constants Yl' ... 'Y6 are adjustable. They can be chosen
3
such that d l = dS/h , d 2 = X, and d 3 = d 6 h where d S and d 6
are independent of h. Then by making minimal restrictions
on the regularity of the solutions, it is possible to show
that, because of (7), d 4 ~ d 7 h 2k - l Then from (15) and (16)

~ F(8 8) + ~ ~(8v) + (20: - d 6 h)U(8u )

dt v' u P

Introducing the expressions for the scaled energy error from

(11), it can be seen that

~
dt
F*(8 8) + ~ ~*(8 ) +
v' u p v
(20: - d 6 h)U*(8u )

It is clear from this equation that, if h is sufficiently

restricted, the algorithm has positive damping and the result
(12) is obtained. •
 509

If the initlal values of velocity and displacement are defined

by weighted energy proJections of the type introduced in (9),
the value of F* (~v,8u) at t = 0 lS zero. Thls is the case which
will be considered in thls analysis. The case, when F*(8v ,8u )
at t = 0 is not zero, could be handled by a straightforward
modificatlon of the techniques presented here. The inequality
(12) has the followlng form:

(17)

Suppose, the calculations are to be carried out untll time

t = T. Then for any tE[O,T), the inequality (17) implles that

d F* < f (0) - f F* + f (T) F*2

dt - 1 2 3

When equality occurs In this equation, it can be transformed

to a Bernoulli equatlon. The resulting equation can be solved,
and the solutlon can be used to establish the following bound:

F*(t) S. fS (
1+
[~-.::fr _fAt)
4 2 e -
f4 + f2

where f4 = ~ f22_4flf3 and fS = (f 2 - f 4 )/2f 3 . The algor-

ithm is stable If the numerical damping constant f4 is real
and the term f4 - f2 is small. The algorithm converges (for
a fixed amount of damplng) only If fS decreases wlth h. The
algorlthm converges to the undamped solution only if the numer-
ical damping constant f4 decreases with h. From (12) it can
be seen that

f
4
= J
g2
2
-
4 g XTh2k-4
gl 3e

Clearly if k'::' 3, there always exists a value for h such that

f4 is real and arbltrarlly close to g2. Thus under the con-
dition k:: 3, the algorithm can be stabilized by simply re-
flning the mesh. In additlon if k:: 3, the artificial damping
constant can be made proportional to h~ where ~< 2 while still
retaining the stabillty property. Thus, without causlng In-
stability, the numerical damping constant can be made propor-
tional to h, resulting in convergence to the undamped solu-
tion. The size of fS can be establlshed by emploYlng the
510

binomial theorem to define f 4. Asymptotically as h'" 0, f 5

O(h 2k - l ). Thus
D)
F *( ~v'C!>u _< ch 2k - l

The total energy error of the approximation can be estimated

using the triangle inequality
(18)

The first term on the right in (18) is a truncation error term

which can be estimated from (7) and (9). The second term on
the right hand side of (18) is dominant, and the following
result is obtained:
Thm.~ If the exact solution to (4) is sufficiently regular,
if the subspace Sh (I) is characterized by k ~ 3, and if h is
sufficiently restricted, the approximation (8) is stable and
convergent, and
F(ev,e u ) ~ ch 2k - l

The following comments about the results of this analysis seem

•
appropriate:
1. The energy growth equation (12) contains, on the
right hand side, two competing forces. The interpolation
error term, characterized by gl, acts to cause conver-
gence. The nonlinear term characterized by g3 acts to
cause divergence. The damping term, characterized by g2'
is, in a way, an innocent bystander. Only when the trun-
cation error term is made more powerful (by increasing k)
can the truncation error overcome the nonlinear effects
to produce stability.
2. The restriction that higher degree polynomials (k ~ 3)
be used in the finite element model can be relaxed if we
give up some of the results of the analysis. For lower
degree interpolation procedures short time stability,
with damping constants depending inversely on h, can be
obtained.
3. From Thm. 2 standard w~ error estimates can be ob-
tained by taking the square root of both sides. The re-
sult shows that there is a loss of accuracy of ~ power
in h, as compared to the error of the linear problem.

The Crank-Nicholson-Galerkin Procedure

The semi-discrete algorithm, discussed in the previous sec-
 511

t1on, sat1sfactorily 1ncorporates a spat1al d1scret1zation

Wh1Ch leads to stab1lity. The Ob]ect1ve of th1S section is
to def1ne a temporal discret1zat1on procedure, motivated by
the semi-d1screte model, which has a stable temporal discre-
tizat10n and inher1ts the stable spatial d1scret1zation of
the semi-discrete model. Let, typically, un = u(n~t) where
n 1S the time step number 1n a step by step solution algorithm
w1th t1me step length ~t. Let u n+~ = (n+l
u + u n )/2. F1nally,
let a Crank-Nicholson-Galerkin version of (8) be defined as
follows:

n+l n) k
(p V ~t- V ,W + (EDu n + 2,DW) + 3 (E(DUn+~)2,DW)
2"
+ E(Vn+~,W) = (fn+~,W)
u n + l _ Un n+~ n+~
~t - V + aU o (19)

for VW€5 h (I). The equation (19)2 is satisf1ed at each node

and everywhere in I.

Weighted proJections Un and Vn ~f the exact solution un and

v n are def1ned through the semi-d1screte methods (defined
1n (9)) w1th all time funct10ns be1ng evaluated at t1me step
n. Total errors and decomposed errors are defined as in (10).
. 11 y, e n = u n - U.
Typ1ca n T h e energy error at d
1screte'
t1me
u
p01nts 1S def1ned 1n the follow1ng weighted measures:

Un(W)

~n(W)

Aga1n, 1t 1S necessary to scale the energy error measures by

def1n1ng temporally d1screte auxil1ary energy error measures

-~ln
where

(20)
[: +~
2
and
512

As before fully discrete total energy errors can be defined.

For example, F*n(w,r) = n*n(w) + u*n(r). The fully dlscre-
tized errors ~~ and e~ can be estimated through the following
theorem:

Thm. 3. If the exact solutlons u and v are sufficiently re-

gular and If the discretization parameters h and ~t are suf-
ficiently restricted, there exist positive constant gl' g2'
and g3 independent of h and ~t, such that

l~t {F*n+l (~v' tE)

u
- F*n (~v'~u)} ~ gle -xn~t (h 2k - l + ~t 4)

g2F*n+~(~v'~U) + g3ex(n+l)~th-3(F*n+~(~v'~u))2
• (21)

Theorem 3 can be proved by a straightforward adaptation of

the methods presented In Theorem 1. In order to establish
the stabllity and convergence properties of the algorithm
(19), the fully discretlzed energy growth equation (21) must
be utillzed. The size of the energy error can be estimated
through the followlng Lemma:

Lemma 1. Let w(t) be a non-negative functlon. Let w(t) be

deflned on the discrete set P = {0,~t,2~t, .•. ,N~t} by wn
w(n~t). Suppose that there exist positive constants a l , a 2 ,
and a 3 such that
n+l n wn + l + wn
w - w :: Hal - ~ta2 ( 2 )

wn + l + wn 2
+ ~ta3 ( 2 ) (22)

for n = 0,1,2, ••• ,co. In addltlon, suppose that WO = 0 and that

a4 = ~ a~ - 4a l a 3 > O. Then there exists a constant c such
that i f
~t :: c
the discrete function wn + l satisfies
 513

( 23)

where

Proof: w*n + a in (22) ,

5
a slmpl1fied formula can be obta1ned to define a bound for
the solution

Equivalently, a bound for the solution can be defined by

w*n+l ~ w*n+l where w*n+l 1S the smallest root of the fol-
lowing equation:

It can be easily shown that

-B(l- ~ 1- 4AC )
w*n+l B2
2A ( 24)

where
ilta 3
A - -4-

ilt n
B 1 + T(a 4 - a 3 w* )

ilt a3 n n
C = -(1 - T(a 4 - _w* »w* •
2
By reduc1ng ilt, it 1S poss1ble to make the term 4AC/B 2 in
(24) small. Thus the square root term in (14) can be evalu-
ated using a binomial series. The following expression is
obtained:

C AC 2 (1 + 2AC
2
+ a) . (25)
- B
B3 B

The term a represents the summation of the higher order terms

in the b~nomial series. These terms are proport~onal to powers
of ilt (ln part~cular powers > 1). Thus as t ->- 0, a ->- o. Sim-
514

plifying (25) using (24), it can be seen that

"'t
(l-T(a 4 -a 3 w* n ))w* n
w*n+l ::: + S
1 + "'t (a - a w*n)
2 4 3

where Sis a term which approaches zero as '" t .... O. The cor-
·
respon d 1ng . ,1n terms 0 f wn+ 1 an d wn.1S d e f'1ne d as
express10n
follows:
(26)

where

Let us investigate the case when wn : as (or correspondingly

-as::: w*n ::: 0). This condition is satisfied when n = 0 because
of the hypothesis WO = O. Subsequently in this proof it will
be shown that, in fact, this condition is satisfied for all n.
Because of this limitation on the size of wn , the first two'
terms on the right in (26) involve exponentially decaying
terms. These terms behave nlcely and can not cause insta-
bility. The third term on the right is an error term caused
by the nonlinearity in the equation. At any particular time
step, by shrinking "'t to zero, this term can be eliminated.
However, small errors, in particular time steps, can bU11d
up over many time steps to produce instability. To insure
stability it is, thus,necessary to show that the error term
S, defined as follows, is negative:
",ta 3 (w*n)2
S =
41
(",2 y -
'I'
1)

where
"'t
1 - T(a 4 - T
a3 n
w* )

1 = 1 "'t
+ T(a 4 - a 3 w* n )
1 { ",ta 3 1/!w*n }
y = -2 1 + .2. + ex
1 21

Because of the condit1on that w*n ::: 0, the fact that ex is

proportional to powers of "'t :: 2, and the fact that ex = 0 when
 515

w*n = 0, there always exists a time step ~t which is small

enough so that y < 1. Thus if ~t is sufflciently restr1cted
in size,S < 0 and from (26)
n+l
w ~ as(l - ].1 ) + wn ].1 (27)
n n

By emploY1ng the inequality (27) and the condition that WO = 0,

it can be shown that as long as ~t is sufficiently restricted
and wn :::. as for n = 0, .•. ,N
N+l N
w :::. as(l - IT ].1) (28)
i=O 1

for N=O,l, ... ,oo. The size of the term on the right hand
side of (28) can be fairly easily est1mated. In fact
N
1 - IT ].1i -<1 N+l
-].1 (29)
1=0

where

].1 (30)

In addit10n, for x ~ 0 small enough

x
e
-2x 1 - 2 _< e -x
~ (31)
1 + ~
2

From (30) and (31), it can be seen that for ~t small enough
-2 (a 4+a 3 a S ) (N+l)~t N+l -(a 4+a 3 a S ) (N+l)~t
e :::'].1 :::. e (32)

Finally from (28), (29), and (32), lt can be seen that if ~t

is sufficiently restricted and 1f wn :: as for n = 0, .•. ,N
1 -2 (a 4+a 3 a S ) (N+l)~t
wN + :: as (1 - e ) (33)

The result (23) can now be obtained by an induction proof.

Let N = 0 and note that by hypothesls WO O. Thus, from (33)
1 -2(a4+a3aS)~t
w :: as (1 - e ) < as

Slm1larly lettlng N = 1 and noting that WO

516

equation (33) gives

2 -4(a 4+a 3 a S )6t
w < as (1 - e ) < as

and so on.
•
From the results (21) and (23) the energy of 8 v and 8 u can be
estimated. Initially it is convenient to assume that, as in
n
the semi-discrete case, F* (#v,8u )=O when n=O. The energy
error growth condition (21) contains certain time dependent
functions. To make the equations agree, certain restrictions
must be employed. Suppose that the calculations are to be
carried out until t=(N+l)6t. Then for any time step n=O,···,N,
equation (21) gives

(34)

Comparing (34) and (22), it is clear that as long as k~3,

6t=ch£(£~2), and h is sufficiently restricted, there exists

a posltive numerical damping constant

( 3S)

such that for all time steps n<N

where
 517

Estlmatlng q5 uSlng the blnomlal Theorem (ln the way introduced

ln the seml-discrete analysls), the following estimate is
obtained:

*n+l 2k-l
F (~ ,~ ) < c(h + 6t 4 )
v u

Flnally, emploYlng the triangle lnequallty as in (18), lt can

be shown that the followlng result holds:
Thrn. 4 If the exact Solutlon to (4) is sufflclently regular,
£
lf the subspace Sh(I) lS characterlzed by k~3, lf 6t=ch
for £~2, and lf h lS sufflclently restrlcted, the fully
discretized algorlthm (19) lS stable and convergent,and

•
(36)

The following comments, concerning these results, seem

approprlate:
1. The stablilty property of the algorlthrn, lnherent
ln the numerical damplng term (q4 ln (35)), depends
on the temporal truncation error. From (35) it can be
seen that, lf there lS real physical damplng in the
problem (characterlzed by positive values of a,E, and
g2)' posltive numerlcal damping can be obtalned by
slmply maklng 6t=ch. However, if there lS no physlcal
damplng ln the problem, convergence to an undamped
solution can be obtalned only if g2=f(h). Then, in
order to obtaln posltlve numerical damplng, lt lS
necessary to make 6t=ch£ for £~2.
2. Other temporal dlscretization schemes could be
lntroduced to integrate the equatlons of motlon forward
ln tlme. However, it should be noted that any scheme,
wlth a temporal truncatlon greater than the Crank-
Nlcholson method, may be very dlfficult to stabilize.
In these schemes the loss of temporal convergence rate
may make lt very dlfflcult to produce a posltive value
for the numerlcal damping parameter q4.
3. If the square root lS taken on both sldes of (36),
standard L2 type error estlmates can be obtained.
4. From the error estlmate (36) it can be seen that the
nonllnear problem has no loss of convergence rate, in
the temporal dlscretizatlon, as compared to the linear
problem.
518

References
1. Belytschko, T., and Schoeberle, D.F., "On the Uncond-
ltional Stabillty of an Impllcit Algorithm for Nonllnear
Structural Dynamlcs", Journal of Applied Mechanics,42,1975,
pp. 865-869.
2. Hughes, T.J.R., "Stability, Convergence and Growth and
Decay of Energy of the Average Acceleration Method in
Nonllnear Structural Dynamlcs", Computers and Structures,
6, 1976, pp. 313-324.
3. Hughes, T.J.R., Caughey, T.K., and Liu, W.K., "Pinlte
Element Methods for Elastodynamics which Conserve Energy",
Journal of Applled Mechanics,45,1978, pp. 366-370.
4. Wellford, L.C., and Hamdan, S.M., "An Analysls of an
Impllclt Flnlte Element Algorithm for Geometrically Non-
linear Problems of Structural Dynamics, Part 1: Stablllty",
Computer Methods ln Applled Mechanics and Engineerlng,14,
1978, pp. 377-390.
5. Oden, J.T., and Reddy, J.N., Mathematical Theory of
Flnite Elements, Wiley Interscience, New York, 1976.
6. Wellford, L.C., "On the Theoretlcal Basis of Flnite
Element Methods for Geometrlcally Nonlinear Problems of
Structural Analysis", Formulatlons and Computational
Algorlthms in Flnite Element Analysis, Ed.,K.J. Bathe, et.
al., MIT Press, Cambrldge, Massachusetts, 1977.
Part V:
Solution Methods
Direct Solution of Equations by Frontal and Variable Band,
Active Column Methods

R. L. TAYLOR, E. L. WILSON, S.1. SACKETT

University of California, Berkeley, USA University of CalIfornia, Livermore, USA

ABSTRACT

The variable band, actlve column and the frontal

methods for the solutlon of linear algebraic equatlons are
compared. Areas considered in the comparison of the two
methods are:

1. The number of numerical operatlons required for the

trlangular decompositlon and resolution for dlfferent
load vectors.

2. Logical operations and data transfer requirements asso-

ciated with the formation of equations, decomposltion
and resol ut lon.

3. Use of the methods for nonlinear problems, dynamics and

substructure analyses.

4. Ease of computer implementation on mini-computers,

mlcro-computers and vectorlzed computers.

A major contribution of the paper is the demonstratlon that

the blocked, variable band, active column method requires
exactly the same number of numerlcal operations as the fron-
tal method. Based on advantages ln other areas, lt is con-
cluded that the varlable band, active column method is
preferable for the Solutlon of a broad range of problems in
computational mechanics.
INTRODUCTION

In this paper we consider two techniques which are com-

monly employed to directly solve the algebraic equations

resulting from a structural or a finite element analysis.

The two methods compared are the variable band, act1ve

column method (also commonly referred to as prof1le or sky-

line method) and the frontal method. Each of these methods

has been extensively described 1n the llterature: for the

frontal method see [1 ]-[8], and for variable band, active

column methods see [9]-[14]. Proponents for each of these

methods make claims that one method is superior to the

other, usually w1th little or no analysis to substantiate

the claims. In our paper we examine several of the factors

which influence the cost of solv1ng large sets of algebraic

equat1ons, e.g. those result1ng in three dimensional ana-

lyses. The factors considered include number of operations

to effect triangular decomposition, ordering of the elimina-

tion process, resolution vs. decompos1tion effort, vectori-

zation of algorithms, etc. Based upon our evaluation and

experiences we conclude that variable band, act1ve column

methods are generally preferable to frontal methods for a

broad range of problems in computat1onal mechanics.

DIRECT SOLUTION OF FINITE ELEMENT EQUATIONS

In many structural or f1n1te element methods it 1S

necessary to solve a large set of algebraic equatlons of the

form
 523

Ku f (1 )

where f is a specIfIed nodal vector, u IS the nodal SolutIon

vector to be computed, and ~ IS a symmetric, positIve defin-
ite, sparse matrix. In lInear static analysis (1) occurs
naturally, In lInear dynamic analyses (1) occurs when uSIng
step-by-step integration methods, and In nonlinear analyses
(1) occurs after linearizatIon by, for example, Newton's
method. The sparsIty of ~ results from the basic structure
of fInIte element approximation by locally based functIons.
The arrays are normally formed by direct sums over elements
(direct stiffness) and with ~e and 1e as the appropriate
element arrays for the problem class beIng considered we may
wrIte

K 2: ~e (2 )
e
and

f
2: 1e
e
The connectIon propertIes of each element will define the
non-zero structure of the sparse matrIx K. In variable band
methods the non-zero envelope of K is called the profile of
the matrIx. The profile wIll be defined in terms of equa-
tion numbers assigned to each degree-of-freedom at each node
and the list of nodal numbers attached to each element. In
most implementations of the variable band method the equa-
tion numbers are assIgned independently of the nodal
numbers. Often profIle optImization schemes are employed to
obtaIn "good" ordering of the equation numbers. We shall
524

discuss this further in a subsequent section of the paper.

In the frontal method the connection properties of the

elements and the element ordering will define the maximum
front width. For either approach a different ordering of
equations or elements will create a different sparsity of
the K matrix. This fact is crucial to the efficient solu-
tion of (1) by direct methods. What we seek always is the
ordering to effect a solution with a minimum effort.

In the next section we will summarize the steps

required to solve (1) by variable band and frontal methods.

SOLUTION STEPS FOR VARIABLE BAND AND FRONTAL METHODS

All practical direct solution methods are versions of

elimination or factorization algorithms. The direct solu-
tion to (1) is obtained by constructing the triangular
decomposItion

where D is a diagonal matrix, U is an upper triangular

matrix with unit diagonals and ~t denotes matrix transpose
of U The solution to (1) may now be constructed by solv-
ing

f (5 )

x (6 )

and

U u = 1..
 525

The solutions of (5) to (7) are relatively simple. The

solution to (5) is called a forward solution since we con-

sider the equations in forward order. Equation (6) is a set

of uncoupled equations which 1S solved sequentially by sim-

ple divis1ons. The solution to (7) is obtained by starting

with the last equation and working to the first - - this

step is usually called backsubstitut1on.

It is clear that once the triangular decomposition is

constructed the solution can be obtained for any number of

specified right-hand sides f e1ther simultaneously or

sequentially. In the sequel we shall refer to equations (5)

to (7) as the resolution step. In iterative or step-by-step

time integration methods the same factors of K are often

used many times, consequently, the efficient resolution of

equations can be as important as eff1cient triangular decom-

position of K (which may be performed very few times in an

analysis compared to the number of resolutions).

In large analyses the number of nonzero entries in the

triangular factors greatly exceed the core capacities of

modern computers. In fact, the capacity of the backing

storage may be severly taxed by this class of problems. For

example a problem w1th 20,000 equat10ns may have more than

10,000,000 terms in the triangular factor for U. Therefore,

we cannot consider in-core methods as a general approach to

solving (1)

In the variable band algorithm we will assume that U is

526

blocked into groups of columns approximately equal to one-

half of the available high-speed storage and all blocks are
placed in secondary storage * The factorization is per-
formed in sequential block order by reading previously fac-
tored blocks into high-speed storage for each block to be
factored. The limiting size on the problem to be solved is
that two blocks must reside in-core simultaneously. All
accesses to secondary storage are for large blocks of data.
The number of numerical operations between reads is signifi-
cant since one large block of data is operating on another
large block of data. In this way, the access time is small
ccmpared to in-core numerical effort for most structural and
f nite element problems.

The main steps to solving (1) by a blocked, variable

band algorIthm may be summarized as:
(1 ) Determine equation numbers for nodes, compute profIle
of matrix K and determine block structure for triangu-
lar factors~
(2) Compute element arrays k and f , equation numbers for
each node on the elementeand st5re results on low-speed
storage.
Assemble arrays K and f using previously computed ele-
ment arrays and blocking for the triangular factors.
Store assembled blocks on low-speed storage.
Construct triangular factors of K using assembled
blocks. Store blocks of factors on low-speed storage.

~or the CDc-76oo the blocks may be stored in the

large core memory and single vectors moved to the fast
core for the triangular decomposition and resolution
steps. See section on vectorization for further de-
tails.
 527

(5) Perform resolut1on uS1ng blocks of factors and !.

When a new solution 1S requ1red only step 5 and those steps

involving !e and f are required.

The frontal approach solves the set of algebraic equa-

tions as they are assembled. As p01nted out in the paper of
Irons [1], the only requirement to perform a Gauss elimina-
t10n for any equat10n 1S that the elements of the equation
need to be completely assembled. In the sparse matrix
defined by a f1nite element problem interaction with an
equation being el1minated on the "front" of act1ve nodes 1S
def1ned by the elements processed to date. Furthermore, all
equat10ns already el1minated are removed from subsequent
cons1derat1on unt11 the resolut1on step. Thus, the number
of nodes act1ve 1n the front depends on the order elements
are 1ntroduced. In the frontal method the total matr1x K is
never fully assembled pr10r to factorizat1on, rather the two
steps are 1nterweaved. The only information required in
h1gh speed core during solution is the act1ve part of the
front. In add1tion, the equation(s) to be eliminated after
coeff1cients of each element are assembled may be anywhere
in the front, in contrast to other d1rect procedures where
elimination proceeds in equation order. In the frontal pro-
cedure once an equat10n is el1m1nated it is no longer
required for the rema1nder of the triangular decomposition
and may be removed from the front. The removal of an equa-
t10n leaves a zero row/column 1n the active frontal equa-
528

tions. When additional equations are eliminated the

existence of a zero row/column will increase the apparent
number of operations (even with zero checks in the outer
loops). Furthermore, when the next element is added all
zero row/columns may not be filled. To minimize this detri-
mental effect two procedures have been proposed:
( a) Longevity based frontal methods where equations which
survive for longer duratIons are put in the lower equa-
tion numbers of the active front [4]. This method
requires additional computations to determine the order
of elimination and placing in the front a step
called prefront by Irons.
( b) Shift methods to fill the vacated row/column positions
[3]. An equation to be eliminated is first moved to a
buffer and then used to modify the other terms. As the
other terms are modified those in equation positions
larger than the one eliminated are shifted to the left
and/or up one position to fill the vacated space. The
added indexing to effect the shift is minimal.

Other strategIes have been proposed, however the two cIted

appear to be the most common. The shift strategy is optimal
in that no zero row/columns can exist after elimination of
an equation. In our construction of operation counts for
the frontal method we assume that shifts are employed to
fill zeros.

We can summarize the steps to effect a frontal elimina-

tion as follows:
(1 ) Perform a prefront to determine the order for elimina-
tion of equations [7]. Compute maximum front width.
(2 ) Compute element arrays k and f , corresponding equa-
tion elimination informKtion,-~nd store results on low
speed storage.
 529

(3) For each element retrieve the element arrays, assemble

into frontal arrays, and eliminate equations which are
completely assembled at this step. Store eliminated
equations in buffer and when buffer is full transfer to
low-speed storage.

(4) Perform resolution steps using frontal equations in

low-speed core.

For a single set of equations steps 2 to 4 can be comblned.

Figure 1 illustrates the steps in a frontal solution uSlng a

very simple example.

OPERATION COUNTS - FRONTAL METHODS

In order to assess relative efforts between frontal and

variable band solution methods the operations required to

perform a triangular decomposition and a resolution will be

computed for a set of sample problems. The problems are

slmple three dimensional meshes constructed from 8-node

Lagrangian elements, 20-node Serendipity elements, or 27-

node Lagrangian elements. The meshes are described by a

block of elements with m, m, and n elements in each direc-

tlon (subsequently described as "mxmxn" meshes of elements).

The maximum front widths (in nodes) are described in terms

of the number of elements on an m-edge as:

- m (m+1 ) + 2, for 8-node elements

- (2m+1) (m+2) + (m+1) (m+1) + 7, for 20-node elements

- (2m+1) (2m+3) + 12, for 27-node elements

The results are given ln Table 1 for several values of the

number of nodes, M, on an m-edge where

530

M m+1 , for 8-node elements

- M 2m + 1, for 20-node and 27-node elements.

Table 1 shows that for a given mesh of nodal layouts, use of

20 node elements with quadratIc interpolation give smaller
front-widths in three dimensions than 8-node elements which
use only linear interpolations (this is prImarily due to the
fact that one node in the middle of each face is omitted).
In addition the use of 8-node elements requires processing
of 8 times the number of 20-node elements. To the authors
knowledge this rather surprising result has not been noted
previously, although it has been recognized that mid-side
nodes are treated very efficiently in frontal methods [1 J.
Moreover, use of 27-node Lagrangian elements produces
front-widths which are asymptotically 1 + 11M times those of
the 8-node element meshes. Thus, for a given number of
nodes use of quadratic order elements should not, in gen-
eral, signifIcantly increase the cost of solving equations
over that required for linear order elements (e.g., In Table
3 compare 8-node 8x8x32 mesh with a 27-node 4x4x16 mesh,
since these two have the same numbe~ of nodes). The number
of operations to perform a triangular decomposition by the
frontal method may be estImated by summing the operatIons
performed for each element added to the mesh. If after
addIng an element the resulting front width is I and after
eliminating all the equations appearing for the last time
the front width is J, an estimate for the number o£ opera-
tions performed for each element is
 531

- 1*(1+1 )*(1+2)/6 - J*(J+1 )*(J+2)/6 for tnangular

decomposltion,
- 1*( 1+1 ) /2 - J*(J +1 ) /2 for a forward solution or
backsubstitution.

For each of the meshes considered the total number of opera-

tions for triangular decomposition is given in Table 3 and
the total number of operations for a forward solution or
backsubstitution lS given Table 6. The triangular decompo-
sltion is approximately proportional to maximum front width
squared for the problems considered, whereas, a resolution
is proportional to maximum front wldth (e.g., the 8-node
8x8x32 mesh requires only 25% less effort ln the triangular
decomposition than the 27-node 4x4x16 mesh, and from Table
6, resolution is only 10% less). Similarly we note in the
tables that using 20 node elements requires 43% and 50% less
effort to construct a trlangular decomposltion and resolu-
tion for the resulting algebraic equations than those of the
8-node element, respectively. The case for the 20-node ele-
ment seems conclusive, however, one should recall the study
in [15] where serious deterioration of results were observed
for distorted Serendiplty elements subJected to high gra-
dlent solutions. The quadratic Lagranglan elements do not
degrade significantly when subjected to the same tests
(recently Wachspress has shown that Lagra~glan quadratic
elements maintain complete quadratlc polynomial solutions
provided boundaries of the element are planar and nodes are
uniformly spaced). For paralleleplped element meshing the
532

20-node element is the obvious choice.

In order to indicate possible limitations of the fron-

tal method we have computed core demands for a three-
dimensional problem with three degrees-of-freedom at each
node This is typical of problems in continuum and fluid
mechanics. For a mesh with equal number of nodes in the two
directions defining the front surface only small meshes can
be considered before the high speed core of most modern com-
puters is exhausted (In most engineering applications M
would normally be substantially greater than eleven). In
addition to the core required for the front matrix, addi-
tional memory is required for front vectors and a buffer to
store the eliminated equations. If each eliminated equation
were written to low-speed storage as eliminated the peri-
pheral processing costs on many computers would greatly
exceed the numerical operation costs. Thus, the buffer
should be substantial so that only large blocks of elim-
inated equations are processed. After the high-speed core
demands for the frontal equations is exceeded, the front
matrix must be partitioned in a manner similar to that used
for the variable band algorithm. Once this occurs any
advantages of frontal methods rapidly evaporate since multi-
ple passes through the blocks are required to eliminate each
equation [7]. In the next section we consider the operation
counts for the variable band method. Guided by the order
frontal methods process equations we are able to demonstrate
that the variable band method is as efficient as frontal
 533

methods and for large problems undoubtably superior. Before

proceeding to this section, however, we should note that
frontal methods require extensive moving of information into
and out of the frontal matrix and vectors. In addition
searches for the location of equations in the front are
required. In the resolution processes the moving of infor-
mation is very detrimental to overall efficiency, whereas,
in triangular decomposition it is necessary to keep the
front as small as possible . Furthermore the cost of moving
information is offset by the ease of assembling element
arrays into the front, a step which is not trivial for vari-
able band methods.

OPERATION COUNTS - VARIABLE BAND METHODS

The number of operations necessary to perform a tri-

angular decomposition for equations solved by a variable
band method depends on the ordering of the equations. Vari-
ous methods have been proposed to automatically compute a
"good" numbering order (e.g., see [16]). In Figure 2, we
show schematically how we assume our equations to be parti-
tioned and in Figure 3 we indicate how the columns interact
during the triangular decomposition step. Figure 3 shows
that a mixture of tall columns with short columns is very
desirable since it is the intersection of the columns which
defines the number of operations which will actually be per-
formed during the triangular decomposition of K. For a
given mesh it is extremely difficult to estimate the number
534

of operations required to effect a decomposition, however,

the number of terms within the profile defines the number of
operations required to perform a forward solution or a back-
substitution. In order to obtain the operat1on counts for
triangular decomposition we constructed software to simulate
the solution process and sum the numerical operations that
would be required to perform each reduction. In construct-
ing operation counts we considered two numbering systems:
(a) Minimum bandwidth numbering. A numbering which starts
numbering equations at one corner, increments along the
shortest edge, moves to the next row of nodes in a con-
tinuous process until all rows of nodes on one level
are numbered, and repeats for each level in the mesh
will produce a m1n1mum bandwidth numbering for the
meshes considered (see Figure 4a).
(b) Frontal Numbering. In this method we merely number the
equations in the order the frontal method would perform
eliminations. This glves maximum column heights signi-
ficantly larger than band numbering but produces a mix
of tall and short columns such that the mean column
height 1S significantly less than that for band number-
ing (see Figure 4b).

No operation counts were constructed using an automatic

numbering scheme such as [16], nor do we infer that frontal
numbering is actually optimal for the variable band method.

Simulations were performed for the same meshes con-

sidered for the frontal method and the results are glven in
Tables 4 and 5 for the 27-node Lagrangian element and the
20-node Serendipity element. The operation counts for the
8-node linear element were not constructed, however, we note
that band and frontal numbering coincide in th1S case.
Furthermore, based on results obtained for 20-node and 27-
 535

node elements operation counts are expected to be similar to

those of the frontal method.

Our results show that frontal numberlng of equations

for variable band solution produces substantially the same
number of operations as for a frontal solution (in two
dimensional problems set up and solved by hand we obtained
identical number of operations for the two methods, conse-
quently, any dlfferences noted in the tables may be due to
the estimators used for frontal operation counts). In con-
trast to this, using band numbering the operatlon counts
were significantly increased, often double the frontal for
the triangular decomposition. Comparing number of terms
within the profile of the variable band wlth resolutlon
operatlon estimates for the frontal method we again obtained
agreement for frontal numbering but substantial differences
for band numbering - - especially for the 27-node element
meshes. In resolution using a variable band method very
little indexing or manipulation exists compared to numerical
operations (compared to fairly large amounts of indexing and
manipulation in frontal programs we have seen). In calcula-
tions we have performed, actual timing on various computers
for programs written in ANSI-FORTRAN always show variable
band methods to execute faster than frontal methods provided
equations are numbered in frontal elimination order for the
variable band method.

The one area where variable band methods may be infe-

536

rior to frontal methods (when the front matrix is completely

contained in the high-speed core) is assembly of the blocks
of equations. Front is nearly optimal since all the
required coefficients of K are in the high-speed core,
whereas, in variable band methods the coefficients required
for each element can be in more than one block of K. Use of
frontal numbering of equations will generally give better
ordering of element for assembly, but more care must be
taken to produce efficient assembly by variable band methods
than for frontal methods. Of course, once the frontal
matrix cannot reside in the high-speed core frontal methods
begin to suffer from the same difficulties. The simplicity
of the triangular decomposition by variable band methods
will more than compensate for the added assembly effort in
most large problems, in resolution assembly is not a factor.
The use of sub structuring methods can also be used to minim-
ize both assembly and solution costs by reducing the number
of elements in each part to be assembled, as well as, avoid-
ing operations on unfilled zeroes within the profile (or
front!). We consider this aspect in a subsequent section.
First we will discuss briefly our experiences with some
notions of vectorization for the variable band method.

VECTORIZATIONS FOR THE VARIABLE BAND METHOD

The variable band method can make extensive use of

parallel operations or vectorization features of modern com-
puters. In the triangular decomposition and forward solu-
 537

tion the basic operations in the inner loops of the algo-

rithm are vector dot products. In the backsubstitution step
the basic operation is a subtraction of a vector times a
constant from another vector. For the blocked variable band
algorithm (called FISSLE - a Fast Implicit Solver for Sys-
tems of Linear Equations) used in the NIKE-finite element
programs at the Lawrence Livermore National Laboratories
[17J vectorization notions have been programmed for the
CDC-7600 and CRAY-1 computers. Most of the subroutines used
in FISSLE are written in ANSI-FORTRAN, only three subpro-
grams are vectorized:
(a). A dot product routine RECOL to compute the inner pro-
ducts for the triangular decomposition and forward
solution steps.
(b). A special inner product routine RECOLD to compute the
elements of the diagonal matrix Q in the triangular
decomposition.
(c). A routine COLBAC to compute the vector sum of a scalar
with a vector in the backsubstitution step.

The solution algorithm has been implemented on the CDC-7600

and CRAY-1 computers at LLNL. The FTN compiler on the CDC-
7600 optimizes the loops in the routines while the CFT com-
piler on the CRAY-1 automatically produces code to utilize
the CRAY vector hardware for the computations.

Although the FTN and CFT compilers produce very effi-

cient object code, it is still possible to obtain a signifi-
cant increase in execution speed by coding RECOL, RECOLD,
and COLBAC in assembly language (COMPASS for the CDC-7600
and CAL for the CRAY-1). On the CDC-7600 the blocks are all
538

located in large core memory (LCM). Since the data is

accessed sequentially it is possible to fetch operands con-

siderably faster than if the data were located in the fast

core (SCM). Achieving high access rate, however, requires

that no bank conflicts occur. By coding in in assembly

language it is possible to move a vector operand to SCM at

very little cost. With one vector in LCM and one in SCM no

bank conflicts will occur. In addition it 1S possible to

utilize the machine registers more efficiently and by split-

ting the loop in half, to overlap memory access w1th compu-

tations. It is possible to take advantage of the machine

features on the CRAY-1 to Y1eld similar speed increases.

The assembly language versions of RECOL, RECOLD and COLBAC

are in fact an average of 2.7 times as fast as their Fortran

counterparts on the CDC-7600 and an average of 2.13 times as

fast as their Fortran counterparts on the CRAY-1 (see Table

8. ) .

Figure 5 shows the timing curves obtained with FISSLE

on the CDC-7600 and the CRAY-1. These curves were con-

structed by solving problems similar to the model problems

used in this paper. The advantages resulting from vectori-

zation are readily apparent in the fact that the asymptotic

slope of the factorization timing curve remains less than

the theoretical limit of two. This is due principally to

the operation overlap obtained by using assembly language

and, on the CRAY-1, the exploitation of chaining in vector

operations. It is also interesting to note that a
 539

resolutIon is over two orders of magnitude faster than a

trIangular decomposition for large average column heights.

This has important implications for the solution of linear

step-by-step dynamic problems and for the solution of non-

linear problems by quasi-Newton or other iterative solution

algorithms which re-use triangular factors many times.

SUBSTRUCTURING

In many problems, especially those with branches,

zeroes within the profile may not fill during the triangular

decomposItIon. ThIS is quite detrimental to the overall

effIciency of solving equations. When a frontal numbering

order is used to solve equations these zeros will persIst in

both the variable band and the frontal methods - - neither

is Immune to the problem. A simple example of a problem in

WhICh zeroes persist within the profile is shown in Figure

6. If the equations are numbered as shown then a group of

zeroes persists at the branch line after the triangular fac-

tors are computed. In order to elIminate operations with

zero elements a check on every possible operation must be

performed - - which is ObVIously more inefficient than per-

forming operations with a small number of zero elements. It

is possIble to SkIP these operations by employing a sub-

structure analysis on each of the branches first. The

results of each substructure analysis can then be assembled

with the remaInder of the problem and the solution process

completed.
540

If operation counts are constructed for the substruc-

ture analysis one again reaches the conclusion that the
frontal ordering of the equations is the crucial aspect of
the analysis both frontal and variable band methods will
require about the same solution effort. Furthermore, both
methods are equally adaptable to generating the substructure
matrices with only minor coding changes to the algorithms.

The use of sub structuring also will make the list of

elements significantly shorter in each part. This will
further enhance the assembly step in the variable band
method by reducing the possible number of searches through
the element file to assemble each block.

CLOSURE

In this paper we have demonstrated that both the fron-

tal method and the variable band, active column method
require the same number of operations to perform a triangu-
lar decomposition or a resolution. This is contrary to the
prevailing opinion and even Irons has communicated directly
to us his belief that variable band techniques are about
midway between the band with full triangle and the front
methods for numerical operations [18]. It is our belief
that the case is now made in favor of variable band methods
over frontal methods especially for very large problems.
The algorithms are intrinsically simpler to code and in
resolution require much less indexing and no relocation of
information. In this study we have compared only the
 541

variable band and frontal methods for the solution of large

problems. There is every reason to believe that neither of

these methods is optimal for three-dimensional problems.

Either sparse matrix methods or iterative methods may ulti-

mately prevail. Nevertheless, many existing programs are

using variable band methods and it should now be apparent

that knowledge gained from the frontal method can enhance

the efficiency of solving problems which contain elements

with mid-side nodes.

REFERENCES

1. B. M. Irons, "A Frontal Solution Program for Finite Element

Analysis," Int. J. Nwn. Meth. Engr., 2, 5-32, 19JO.

2. R. J. Melosh and R. M. Bamford, "Efficient Solution of Load-

Deflection Equations," J. Struct. Div. , ASCE, g), ffJ-76 , 1969.
3. E. Hinton and D. R. J. Owen, Finite Element Programming, AcadEmic
Press, London, 1gJ7.

4. P. Hood, "Frontal Solution Program for Unsymmetric Matrices,

Int. J. Nwn. Meth. Engr., 10, 379-399, 1976.
5. R. J. Collins, "Dynamic Destinations: A Method for Removing Redun-
dant Operations in the Front Algorithm," Int. J. Nwn. Meth. mgr.,
12, 1042-1043, 19J7.
6. S. F. Abbas, "Some Novel Applications of the Frontal Concept,"
Int. J. Nwn. Meth. Engr., 15, 519-536, 1980.

7. G. Beer, "A Second Generation Frontal Solution Program," (to

appear) .

8. R. J. Coll ins, "A Mod ified Prefrontal Routine,"

Int. J. Nwn. Meth. mgr., 11, 765-766, 19J7.

9. A. Jennings and A. D. Tuff, "A Duect Method for the Solution of

large Sparse Symmetric Simultaneous Equations,"
large Sparse Sets of Linear Equations, Ed. J . K. Reid, AcadEmic
Press, London, 1971.

10. D. P. Mondkar and G. H. Powell, "Large Capacity Equation Solver for

Structural Ana.lysis," Compo and struct. , 4, 699-728, 19J4.
542

11. C. A. Fellppa, "SolutIOn of Lmear Equations W1 th Skylme-Stored

Symmetric Matrix," Canp. and struct. , 5, 13-29, 1975.
12. E. L. Wilson and H. H. Dovey, "Solution or Reduction of Equilibrium
Equations for large Complex structural Systems, "
Mv. Engr. Software , 1, 19-25, 1978.

13. E. L. Wilson, "Solution of Sparse Stiffness Matrices for structural

Systems," Sparse Matrices, Ed. 1. S. Duff, SIAM, Philadelphia, 1979.

14. R. L. Taylor, "Canputer Procedures for Finite Element Analysis,"

Ch.24,The Finite Element Method, 3rd. ed., by O.C.Zienkiewlcz
McGraw-Hill, London, FJ17.

15. J.A. stricklin, W. S. Ro., E. Q. Richardson and W. E. Halsler, "On

Isoparametric vs Linear strain Triangular Elements,"
Int. J. Num. Meth. Engr., 11,1041-1043, 1977.
16. N. E. Glbbs, W. G. Poole, Jr., and P. K. Stockmeyer, "An Algonthm
for Reducing the bandwidth an Profile of a Sparse Matrix,"
SIAM J. Num. Anal. ,13, 236-250, 1976.
17. J. O. Hallquist, "NIKE2D: An Implicit, Finite-Deformation, Finite-
Element Code For Analyzing the static and Dynamic Response of Two-
Dimensional Solids," lawrence Livermore laboratory Report, UCRL-
52678, March 1979.
18. B. M. Irons, personal communication.
 543

TABLES AND FIGURES

I
1 Table 1.
1
1 1
: Mruumum Front Width for 3-d Meshes :
1 1
1 1
: wi th M-nodes on fuch Edge :
1 1
1 1
+--1
1
Element Type 11
1 1
1 1
M : &-node 2O-node 27-node I
I 1

3 i
1
14 22 27 j
1
1 1
5:
1
32 36 47 I1
7:
1 1

1
~ ~ ~ I1
1 1 1
I 9:1 92 e6 111 I1
1
1 1 1
: 11 : 134 120 155 :
I
1 I i

1
Table 2. 1
1
1
1
3-d Element Types Wl th 3-dof/ node 1
1
1
1
: &-node (1) 20-node (S) 1
1 27-node (1) 1
1
1 1 1
1
M 1iFront Core Front Core
1
1
1 Front core I1
1 1 1
1 1 1 1
1 1 1
1 3 1
1 42 9J3 66 2211 1
1 81 3402 1
1
1 1 1
1 1 1
1
1 5 1
1 96 4656 108 5886 I 153 11781
1 1 1
1 1 1
1
1 7 1
1 174 15225 174 15225 I 237 28203
1 1 1
I I I
I
I 9 I
I 276 38226 I 258 33411 I
I 345 59685
I I
I I I
I
: 11 1
I 402 81003 : 360 64~ : 477 114003
I I I I
_I_L I 1
544

i Table 3. i
i i
iI Triangular DecomIXlsi tion Operations (in millions) iI
i II
Element Type
I
I
Mesh 8-node I 2O-node 27-node I
I I
~----~--------~---------+--------~I

2x2x8 .01 .11 .29

3x3x12 .04- .91 2.52

4x4x16 .19 4.54 13.15

5x5x20 .65 16.83 50.21

6x6x24 1.86 50.76 154.84-

8x8x32 10.50 304.94 959.65

I
Table 4. I
I
I
I
27 node lBgrangian Element I
I
I
I
Band Numbering Frontal Numbering I
I

-
I

Mesh No. 1000 Mean Million 1000 Mean Million!

I
I
Size :Ells. Terms Column Ops. Terms Column Ops. I
I
I
I

2x2x8 425 18 43 .42 15 36 .28 I

3x3x12 1225 100 82 4.32 77 63 2.44

4x4x16 2673 353 132 24.50 260 97 12.89

5x5x20 4961 964 194 93.20 694 140 49.52

6x6x24 8281 2225 267 312.61 1578 191 153.27
8x8x32 1m85 8514 453 2010. 5937 316 953.72
 545

Table 5.
20 node Serendipity Element

Band Numbering Frontal Numbering

- - - _ ._-
Mesh No. 1000 Mean Million 1000 Mean Million
Size Eqs. Terms Column Ops. Terms Column Ops.

2x2x8 261 8.2 31 .13 7.5 29 . 11

3x3x12 712 40 56 1.12 35 50 .f57
4x4x16 1505 132 88 5.32 11 5 77 4.43
5x5x20 2736 345 126 22.00 300 110 16.54
6x6x24 4501 773 172 67.00 669 149 50.10
8x8x32 10017 2842 284 408.00 2449 245 302.51

Table 6.
Forward Solution or Backsubstitution
Operations (in thousands)

Element Type
Mesh 8-node 2O-node 27-node

2x2x8 --- --; 9 7.3 15.2

3x3x12 4.0 32.8 76.5
4x4x16 12.2 114.2 259·7
5x5x20 30.4 297.7 693.9
6x6x24 66.0 665.0 1578.3
8x8x32 232.9 2440.6 5937.0
546

Table 7. 1
1 Vectorizatlon Timings in Fortran and Assembly language :
: for CRAY-1 and COC-7600 Computers :
: Time In nano-seconds to produce a single element for an n-vector. :
: REGOL REGOLD COLBAC:
1 Time I CAL Time I CAL Time ICAL 1
1
: n=1
: CAL 3Xl.0 1.00 3000.0 1.00 600.0 1.00
! CFT(ON=V) 2200.0 2.75 9000.0 3.00 1400.0 2.33
: CFT(OFF=V) 2100.0 2.63 4000.0 1.33 1300.0 2.17
: COMPASS 2300.0 2.&3 2000.0 0.67 1000.0 1.67
: FTN(OPT=2) 3500.0 4.38 6000.0 2.00 1300.0 2.17
1
:n=B
: CAL 362.5 1.00 375·0 1.00 175.0 1.00
: CFT(ON=V) 937.5 2.59 1125.0 3.00 337.5 1.93
! CFT(OFF=V) 850.0 2.34 1000.0 2.67 662.5 3.79
: COMPASS 1012.5 2.79 1000.0 2.67 1050.0 6.00
: FTN(OPT=2) 1487.5 4.10 1500.0 4.00 1425.0 8.14
1

: n=64
: CAL 70.3 1.00 1 109·4 1.00 56.3 1.00
: CFT(ON=V) 167.2 2.38 203·2 1.86 75.0 1.33
! CFT(OFF=V) 576.6 8.20 734.4 6.71 531.3 9.44
: COMPASS 446.9 6.36 562.5 5.14 478.1 8.50
: FTN(OPT=2) 806.3 11.47 5015.6* 45.86 817.2 14.53 :
1 1
1
: n=1000 1
1
:1 CAL 30.7 1.00 57.0 1.00 46.6 1.00 :
: CFT(ON=V) 69.8 2.27 97.0 1.70 49.2 1.06 :
: CFT (OFF=V ) 540.0 17.59 666.0 11.68 513· 7 11.02 :
1 COMPASS 365.3 11.90 442.0 7.75 386.4 8.29 1
1 FTN(OPT=2) 1113.9* 36.28 872.0 15.30 1482.9* 31.82 1

* Due to large number of bank COnflIctS.

 547

[-4 5 2 ;-7 8 ] = LG
Kll K12 K13 K14 K15

K22 K23 K24 K25

K33 K34 K35 FRONT Coefficient Array

before Elimination.
K44 K45
K55

[ K14 K24 K34 K44 K45J= EO Equation being Eliminated.

K= 4 Pivot location
JJ= 7 Global equation number

[-4 5 2 8 J= LG

Kn K12 K13 K15

K22 K23 K25 FRONT Coefficient Array

after Elimination. Note
K33 K35 treatment of row/column 5.
Shift prevents zeroes ln
KS5 frontal rows and columns.

Figure 1. Elimination of an Equation During Frontal Elimination

548

Block # 1 2 3
3 cals 2 cals 2 eals
2 6
3 4 7
5 8 10 13
9 11 14
12 15
16 17
18

1 2 3 4 5
Block 1 llJ11314151 K-Array
11 13 15 I Pointer to Diagonals

1 2 3 4 5 6 7
Block 2 1 6 17 1s19110 111 112 I K -Array
[ill] Pointer

1 234 5 6
B1 aek 3 113 114 115 116 117 Ils I f-Array
I 4 I 6 Pointer

Figure 2. Blocking Format for Variable Band Solution.

 549

Figure 3. Interaction Between Tall and Short Columns

for Variable Band Methods during Trianqular
Decomposltion.
550

21
i7
46 941 36 .31 26
0

71 66 61 56 51
..
81•
96
•
91 86
76

Figure 4a. Band Numbering for 27-node Lagrangian Element.

(2x2xn "1esh)

11

17
29
61

3 27 # 23 fJ 7 64

7 75 71 55
111

81 7) 73 tE7 114

1 5 12 105
129

131 12) 12 0107

Figure 4b. Frontal Numbering for 27-node Element.

(2x2xn Mesh)
 551

Figure 5. FISSLE CPU COST

"
, .. :-,
! ' FISSlE:7800
! FACTORIZATION
i
l00~--~----~~~~_----
I
1 ________~-+----~~~~----~
"'-li--~;-'-'--':-I
-~.~- -:"~l,~l , .~ '. ' i
, ! :I: ' '1,
I I

: 'r ,"- ~-
-" : I : ::
I I

10r-~=-~~~--~~----r-~----~~
., LINE SOR: 7800 - +
(456 ITER., 10~ ' : •

§... . - COlSOl : 7800 ' -

. FACTORIZATION
.....

lr-------~-7~--~~.-
_4~:::~~_
--_-_--~~----~.-----~
_-,_..._-J-_______
.+-_. ---
_____.. --_ ... _- ---+--:;:-...:.:_-: _.
.
-:::::-7. ~~: ~~ ,
.-
, .~-=-: . t
- - - 1---
. --.----1---....L-_____ __._

; i
1--7""---="":-, --:-;- ,
: !! t ! !:
COLSOl : 7800 ....:
R.H.S. REDUCTION
, + BACKSUB. :
: j! I ~ •
I:
1~~~~
• . -. ,. -~
-, ~
-:~-.-_~
--- - - r-...L- f -7" - ~- FISSlE : CRAY
. .... .. . . .• -

,. ,' +
--..,-......;-y--+-~:+ R.H.S. REDUCTION
BACKSUBSTITUTION
-'-;~ 'TI -----'---"-~-"' -
: ' !I :f I: ;
i
; !' 4, 1!
1T"Tf
I : I

'I i i i!1
.Ol~
! I . __~__~~~~
__~~~~~~____~~~~~~ : I I

...o 8 ...§
AVG. COL. HGT. -
552

2 5

12 13

Substructure 1.
1 2 3 4 5 6 7 8 9--11--13 1 2 3 410 9
XXX X XX XX
XX X XXX
XX XX XXXX
X XX XXX
X X X X0 0 III III
X X X0 0 III
X X X0 X Substructure 2.
X X0 X 5 6 7 8 911
XXX
XXXX xxxx
XX X xxx
XX xxxx
X xxx
QQ
iii
Substructure 3.
x - Non-zero in Triangular 9--11--13
Factor. QQQ
o - Zero ln Triangular Factor. Qxx x
Qx x
a - Substructure Coefficients. xx
x

Figure 6. hample with Persistent Zeroes within Profile of K.

Solution by Iteration in Displacement and Load Spaces

P.G.BERGAN
The Norwegian Institute of Techno!ogy, Trondhelm, Norway

Summary
A method of adjustlng the load lntensity in order to minimize
the residual (unbalanced) forces from the equilibrlum equa-
tions is suggested. This technique may be used as a means
of controlling the steps for load or displacement pattern
incrementatlon schemes. It may also be used for obtaining
improved convergence during equllibrium iterations of initi-
al stress (strain), quasi-Newton or true Newton-Raphson type.

Introduction

The solution algorlthms for analysis of nonlinear static ana-

lysis are normally based on stepwise incrementation of loads
or of displacements. The load stepping schemes usually com-
bines a slmple Euler-Cauchy incrementation with equilibrium
iterations of the Newton-Raphson type. For reasons of effl-
clency and rellablllty the Slze of the steps should be adapt-
ed to the degree of nonlinearlty of the system during the
course of solution. Special measures must be taken In order
to deal with limit pOlnts, bifurcations and unstable branch-
es. Solutlon algorlthms of thls type are discussed in Refs.
/1/ - /4/.

Stepwise solution by incrementing a characteristic displace-

ment component /5/ has also been a popular method. The load
intensity is adjusted during the equilibrlum iteratlons so
that the force reaction at the preselected component vanish-
es. The main advantage by this method is that the modified
stiffness matrix may remaln posltive deflnite even for un-
stable systems. Even though this approach may be very
554

successful for small-scale and lucid problems it 1S often

not suited for large-scale, practical problems.

This paper discusses a solution technique that has similari-

ties with the two methods d1scussed above. It allows for in-
crementation of either loads or a suitable displacement pat-
tern. Further, both the load intensity and the displacements
are continuously corrected dur1ng the equilibrium iterations
/6/.

The basic 1dea

In conventional incremental-iterative solution schemes the

obJect1ve of the equilibrium iterations is to find the con-
figuration of the structure for which the associated internal
forces exactly balance the applied external forces. In other
words, the pOint in load space is being kept unchanged while
iterat10ns are carried out in the displacement space. In
most situat10ns the analyst is not particularly 1nterested
in a solution for specific values of the loading vector R;
he rather wants to be able to def1ne the over-all solution
path by some typical discrete p01nts. Intermediate values
are eas1ly obta1ned by interpolat1on. Since it 1S not essen-
t1al to fix the loads at prescr1bed values it 1S tempting
continuously to adJust the load 1ntensity p along with the
1terative adjustments of the displacements r.

The principle for the method is shown in Fig. 1. Let pOints

"a" 1n the two spaces denote a prev10usly found equilibrium
configuration for which the resulting internal forces R. t
1n
exactly balance the applied forces R . A new load increment
a
~Rab is now applied and the corresponding incremental dis-

placements are found using the Euler-Cauchy method or some

slmilar scheme.

( 1)

and
( 2)
 555

where KI is the incremental or tangentlal stlffness. The

new conflguratlon r l produces lnternal forces Rint,l' see
points marked "1" In Fig. 1. The difference between the
applled forces Rb and the internal forces Rlnt,l lS usually
interpreted as unbalanced or residual forces and applled In
a Newton-Raphson dlsplacement correctlon. Instead of dOlng
thls the residual force lS here minlmlzed by adJusting the
lntensity of the applled forces to another point "bl" on the
loading path. The corresponding minlmized unbalanced force
Run b a 1 , 1 lS then used In a Newton-Raphson correctlon of the
displacements. In general, these corrections may be written
-I
r.J+l = r J + KI,J Runb al,J ( 3)

where

R
unbal,j
= R - R
int,J
(4 )

Index j lndlcates the number of the lteration cycle and KI .

,J
lS the stlffness matrlx that may represent the true gradient
or be an approximatlon to it.

R
1
r
I

a) load space b) displacement space

Fig. 1 - Iteration with adaptatlon of load intensity

556

The minimization of unbalanced forces and corresponding ad-

justment of load intensity should normally be carried out
for every lteration cycle. Such a sequence of operations is
illustrated in Fig. 1. It is also possible to carry out the
force minimlzation only directly after a new increment has
been applied and keep the loading level fixed during the
following equilibrium iterations. Constant load increments
could then be used throughout. The adjustments of load in-
tenslties would then produce an automatic loading algorithm
in which the degree of nonlinearity governs the actual load-
ing steps. This is illustrated in Fig. 2 for a one degree-
of-freedom snap-through problem. The initial, elastic stiff-
ness has been used ln the incrementation, cf. Eq. (1) and
the trial load increment is constant throughout. The mini-
mization of unbalanced forces brings the loading level di-
rectly back on the curve in this case, and no further itera-
tions have been necessary. It is seen that the actual load
increments are nicely adjusted to the nonlinearity of the
system. There has been no need for a special unloadlng cri-
terion. A slmllar case wlth several degrees of freedom

Flg. 2 - Solutlon of a one degree-of-freedom system wlth

constant load steps and load intensity adaptation
 557

would requlre equillbrium lteratl0ns, but the properties

of adapted load steps and automatic unloading essentially
remaln. The special case shown in Flg. 2 corresponds to In-
crementatlon of the displacement by constant amounts.

Solution for proportional loadlng

In case of proportional loading applied to the structural

system the loads may be related directly to a reference load
vector R f ln terms of a loading parameter p
re

R= p Rref (5)

ThlS linear relationshlp implles that the load path is a

straight line in load space, see also Fig. 3. Assuming that
a new step has been taken from point "a" or that equlllbrium
iterations are belng carried out, the condition of minlmizing
the unbalanced forces may be stated

aap I I Runball I aap II pR ref - Rint I I o (6)

prescrlbed
load path

minimum
dlrection

Fig. 3 - Minimization of unbalanced forces for proportional

loading
558

This is the condition for finding the best suited load level
The "s~ze" of the unbalanced force should be measured
in terms of some norm, e.g. a weighted Euclidean norm defin-
ed by
N
II X II (E (7)
k=l

where Wk are the weights and N is the dimension of the vector

x. Us~ng this ~n Eq. (6) the best adapted load level is

N
E W2 R R
k=l k ref,k int,k
(8)
N
L
k=l

The direction of the minimized unbalanced force is perpen-

dicular to the load path in the scaled load space, see Fig.
3. Note again that these operat~ons of load adaption may be
carried out after each new load step as well as after each
cycle during equilibrium iterations.

The we~ghts may ~n many cases be chosen as unity. However,

when both translat~onal and rotational degrees of freedom
with correspond~ng loads are ~nvolved, the weights may be
used to counteract that the forces have d~fferent un~ts. As
w~ll be mentioned later, the possibil~ty of using weight may
also be advantageous when convergence problems occur.

Solution for non-proportional loading

Eq. (5) is not valid for a non-proportional load h~story,

and the load ~ntens~ty must be expressed in a more general

form

R = R(p,r) (9)

It ~s ~mpl~ed that dependency of forces on the load~ng para-

meter must be known in advance or can be establ~shed on the
bases of the deformed configuration (e.g. pressure loads) .
 559

The process of m1n1m1z1ng the unbalanced forces is now based

on an auxil1ary local Ilnear1zat1on of the loading path. In1-
t1ally, 1t 1S assumed that R depends only on p. The process
starts by taking a load step from the egu1librium state "a"
to a new load level R(Pb)' see F1g. 4. The new displacements
are found from Egs. (1) and (2) and the internal forces Rint
are establ1shed accord1ngly. It is then assumed that the
load path may be approx1mated by a secant between "a" and
"b", and the unbalanced force 1S min1m1zed w1th respect to
this Ilne through the cond1t1on

Cl P-Pa
Cl~ I I Runball I -;;-11
op
R
a
+ - - ( R -R)
Pb-Pa b a -
R
1nt
II o (10)

Adopting the we1ghted norm 1n Eg. (7) the m1nimizing load

level 1S

N 2
Z Wk2 (R b ,k - R k)
k=l a,

secant path

R
n
true
load
path

Fig. 4 - Min1m1zat1on of unbalanced forces for non-propor-

t10nal loading
560

The unbalanced force with respect to the corresponding point

on the true load path (marked "bl" in Fig. 4) is now

(12)

These operations may now be repeated in connection with

equilibrium iterations. Rb should then be taken as the adapt-
ed loading level from the last iteration cycle while Ra is
kept constant.

It can be expected that the method of auxiliary linearizations

works well for relatively smooth loading paths, however, it
can also be applied in cases of Ikinks" on the path. It may
also be used w1thout modifications for deformation-dependent
load1ng. Calculation of Rb is then based on the last found
(approx1mate) configuration with associated load 1ntensity.

Incrementation of displacements

The fact that the present technique automatically finds a

suitable load level opens the possibility for a simple way of
incrementing displacements rather than loads. Of course, also
the load incrementation of Eq. (1) produces an incremental
displacement pattern. The qua11ty of these displacements pri-
marily depends on the gradient that is used; the true tangent
st1ffness usually gives a good pattern while approximate gradi-
ents or the elastic stiffness may give a very poor set of in-
cremental displacements. For instance, a symmetric structure
with symmetric loading may buckle into a asymmetric mode where-
as use of the elastic stiffness would always give a symmetric
displacement pattern.

A good guess for a new set of incremental displacements is

that it should have the same shape as the incremental dis-
placements between the last two equilibrium configurat1ons

(13)
 561

The scaling factor y assures that ~rab assumes a suitable

size. For 1nstance, the norm of the incremental step may
be required to be the same as for some reference step ~r f
re
by choosing

II ~r ref II
y (14)
IIMil1

~rref which should have an ideal S1ze, may for instance be

taken as the 1ncremental d1splacements at the first linear
step (the only step that is calculated from incremental
loads) .

Discussion of performance

As mentioned earlier, load level adaptat10n may be app11ed

directly after each new 1ncrement as a means of step control.
The method then implies automatic unload1ng after l1mit
p01nts, even w1thout introducing any special unloading cri-
terion. Exper1ence so far 1nd1cates that the method 1S
well suited for this purpose, part1cularly for systems with
a smooth behaviour.

Even more interesting is the aspect of using the method as

a way of improv1ng convergence dur1ng equilibrium iterations.
This 1S part1cularly true 1n connection with initial stress
(strain) type iteration and iteration with approx1mate gra-
dients; the prospective ga1ns are greatest for these techniq-
ues. It can be expected that the convergence 1n connection
w1th true grad1ent 1terat10n (Newton-Raphson) will be at
least as good w1th load level adaptation as without. In
any case, the load level corrections do not cost much com-
putationally.

In general terms, convergence is hardest to obtain when

using only the elastic stiffness throughout the entire cal-
culat10ns; this case will now be discussed. The unbalanced
force min1mizat10n always gives an unbalanced force that is
perpend1cular to the load path, which 1S a straight l1ne
562

for proportional loading. Thus, these unbalanced force

components always have a linear dependency between them.
-1
When these forces are multiplied by K , which is kept con-
stant, a corresponding linear dependency also is introduced
between the components of the displacement corrections. In
an N-dimensional case the displacement corrections thus fol-
low a hyperplane, which in an unfortunate case may have no
intersection with the true solution path. In such a situ-
ation the iteration will not converge.

The question of convergence is illustrated for a system with

only two degrees-of-freedom in Fig. 5. The path for propor-
tional loading is a straight llne In the Rl -R 2 plane and the
mlnimized unbalanced forces are always perpendicular to this
dlrection. When these forces are sequentially multiplied by
the same lnverse gradient they produce displacement correc-
tions that all follow the same stralght line. The final so-
lution is s~mply the point of intersection between this line
and the true solution path. Fig. 5 also shows that, if the
iterations start from a point in the displacement plane
that lies on the convex slde of the path, It may happen that

diverging direction
r
2

converging direction

t ideal direction

displacement space

R21~1 3 pr
1

b2 bl b
~:pace
Rl
Flg. 5 - Convergence propertles when using constant gradlent
 563

there is no intersect10n between the llne of displacement

corrections and the Solut1on path. Thus, there will be no
convergence. The figure also indicates that the best direc-
t10n for the d1splacement correct1ons would be a line perpen-
dicular to the true d1splacement path.

The simplest way of assuring converging iteration is to

update the grad1ent at 1ntervals. Another possibil1ty is
to start up the 1terations from a d1fferent point in solution
space whenever convergence is poor. Th1S could be done by
gOlng back to the last equ1l1br1um configuration and using
only half of the prev10usly used 1ncrement. In Fig. 5 th1S
corresponds to 1terat1ng from the midpo1nt between points
"a" and "1", Wh1Ch would assure convergence for that part1cu-
lar sltuat1on.

A third remedy that could be used 1S to rotate the direction

of the d1splacement correct1ons. This may be don~ through
matr1x transformat1ons or by use of acceleration factors.
There 1S no point 1n uS1ng the same factor for all components
since that would not change the d1rect1on of displacement
correct1ons. One way of achieving accelerat10n is through
use the following factor for component k 1n Eq. (3)

~rk,j-l + ~rk,j
(15)
~rk,J-l

where 1nd1ces j-l and J refer to the two previous iteration

cycles. Eq. (15) should be overruled by requirements for
min1mum and maX1mum values, e.g. 0.5 and 2.0, respect1vely.

Scal1ng of the components 1n load space may have an effect

similar to using accelerat10n factors.

Numerical studies

The algorithm is tested for a spr1ng system suggested 1n

Ref. /6/ and shown 1n Fig. 6. This case 1S a var1ation of
564

the classical pinned bars snap-through problem with the ex-

tension that this system also may buckle out of the A-B-C
plane. In th1S two degrees-of-freedom system vi is the snap-
through component while v 2 is associated with bifurcation 1n-
stability. The third displacement component in space 1S re-
strained. By varying spring constants, geometry, load combi-
nation and imperfections it is possible to simulate a w1de
range of coupled nonlinear phenomena involving llmit p01nts,
instable branches, symmetric and asymmetric bifurcations etc.

Further, it is relatively slmple to present the results on

graphical form for a system with only two freedoms. The po-
tential equilibrium and 1ncremental forms for th1S problem
are given in an Appendix.

The current stiffness parameter /3/,/4/,/6/ has been util1zed

extens1vely in the numerical studies. For a case of propor-
tional loading and incremental solution with true gradient
(tangent stiffness KT ) this quantity is defined by for step i

sp (16)

where index 1 refers to the in1tial state and the first step.

ideal system side view of

imperfection

F1g. 6 - Nonlinear model system w1th two degress-of-freedom

 565

When only an approximate gradient or the elastlc stiffness is

used S may rather be found from
p

lIPi II MIl I
S (17 )
P lIP l ll llr i l1

where I Illr.1 I I is an Euclidean norm taken of the incremental

displacements between the last two equilibrium configuratlons.

It should only be noted that the scalar quantity Sp gives a

good characterization of the over-all stiffness of a struc-
tural system. Sp starts out with value 1 for the initial
state, it is less than unity when the system becomes softer,
and it is larger than unity when the system becomes stiffer
than in the initial state. An important property of S is
p
also that it is exactly zero when passlng limit points and
is negative for unstable branches of the solution path.
Moreover, the current stiffness parameter may be used for
automatic step size control be prescrlbing a desired change
LIS for S durlng each new step. Sp may also be used to
p p
guide the algorlthm at limlt points. These properties are
discussed in more depth in Refs. /3/, /4/ and /6/.

The first study of the model problem involves a case of

slight geometric imperfection with initial displacement
v2 = 0.01 L. Further data are given in Flg. 7. Curve A
shows the results for pure Euler-Cauchy incrementation with
liSp 0.01, curve is Euler-Cauchy combined with true Newton-
Raphson iteration and liSp = 0.05, and, finally curve C is
the new method wlth only elastic stlffness, load increment-
ation and LIS = 0.05 for the trlal load steps. The results
p
for the last two methods coincide in the figure while curve
A for pure Euler-Cauchy deviates somewhat because of trun-
cation errors. The jump of the stiffness parameter because
of transition to a bifurcation path is shown in Fig. 7c.

The method of adJusting load intensity during increment and

lterations is demonstrated for a case of a geometrically
566

0.5
a.

O.O+-____~______~__~~----~~~--~_

-0.5

o2

1.0

b. 0.5

3.5
o1
s
P
1.0

A
c. 0.5

-0.5

Fig. 7 - Study of model problem with geometr1c imper-

fection. Ho =1,5, vo=O.Ol, mo =0.4, m1=0.0,
P2=0.0. Curve A: Euler-Cauchy with ~Sp=O.Ol,
Curve B: Incrementation with Newton-Raphson,
Curve C: Elast1c stiffness matrix and load
adaptation.
 567

perfect system (v 2 0) with a small load perturbation

P 2 = 0.01 P l ln Fig. 8. The very sharp deformatlon mode
transition after the completion of the transverse bifurcation
mode should be noted. Even with many attempts it was impos-
sible to make the usual Euler-Cauchy incrementation wlth
Newton-Raphson iteration pass this critical point of the
path. Therefore, it was very satisfying that the new method
with only linear elastlc stiffness, but with acceleratlon
factors, dld work. The extreme slngularity at the critical
point is accentuated by the plot of the current stlffness
parameter ln Flg. 8.b.

V2

a.

sp

-0.5

-1.0
!t

Flg. 8 - Study of model problem with load pertubatlon.

Ho =1.5, vo=O.O, mo =0.4, ml=O.O, P2=Pl/100.
568

The model problem also lends itself to a study of asymmetric

buckling. This is done by making the force-displacement re-
lation of the transverse spring nonlinear, see Appendix.

1.0

A
B
O.S
'
•••••
.
/.
...... / C

'.

a.

'. '.
" . ....

°2
loS

1.0
~ ..............................

O.S .....~ ......

:'O.S 1.0 loS 2.0 3.0
b. 0.0 .'

O.S

1.0

Fig. 9 - Study of model problem with nonlinear spring.

Ho=l,S, vo=O.O, mo =O.4, P2=O. Curve A: ml =O.2,
vo=-O.Ol, Curve B: ml=O.O, vo=O.Ol, Curve C: ml =O.2,
v o =O.01.
 569

In this connection the cases of a small negative, and a posi-

tive geometrical imperfection were considered, see curves A
and C in Fig. 9. For reference, also a constant sprlng co-
efflc18nt was tested along wlth positive imperfection, see
curve B. An lnteresting aspect here is that the lateral
buckling follows the lnitlally unstable assymetric branch
in case of a negative imperfection while an imperfectlon in
the opposite directlon guides the solution into the initlally
stable, asyrnrnetrlc branch. Curve B follows the lateral path
for syrnrnetrlc buckling. Also these results have been pro-
duced by the technlque with load adaptatlon during itera-
tions.

Results from some more test cases are described in Ref. /8/.

References

/1/ Oden, J.T.: Finite elements of nonllnear continua,

McGraw-Hlll (1972).

/2/ Tillerson, J.R.; Stricklin, J.A.; and Halsler, W.E.:

Numerical methods for the Solutlon of nonlinear pro-
blems ln structural analysis, ln Nonllnear structural
problems, (editor Hartung, R.F.), AMD Vol. 6, (1973)
67-101.

/3/ Bergan, P.G.; Horrlgmoe, G.; Krakeland, B.; and

S¢reide, T.H.: Solution technlques for nonlinear
finite element problems. Int. J. of Num. Meth. ln
Engng., Vol. 12 (1978) 1677-1696.

/4/ Bergan, P.G.: Automated lncremental-lteratlve solution

schemes. First Int. Conf. Numerical Methods for Non-
linear Problems, Swansea, U.K., Sept. 2-5, 1980.

/5/ Argyrls, J.H.: Contlnua and discontinua, Proc. Flrst

Conf. Matrix Meth. in Struct. Mech., Wright-Patterson
A.F.B. Ohio (1965).

/6/ Bergan, P.G.: Solutlon algorthms for nonllnear struct-

ural problems. Proc. Int. Conf. on Engineering Appli-
catlon of the Flnite Element Method. Computas, Norway,
(1979), also to be publ. ln Computers and Structures,
(1980) .
570

/7/ Bergan, P.G. and S¢reide, T.H.: Solution of Large Dis-

placement and Instability Problems Using the Current
Stiffness Parameter, Proc. Int. Conf. on Finlte Elements
in Nonlinear Solid and Structural Mechanics, Gei10,
Norway, 1977. Also in Finite Elements in Nonlinear
Mechanics, Vol. 2, Ed. Bergan P.G. et a1., Tapir
Publishers, Trondheim, (1978) 647-669.

/8/ Marum, S.E.: Testing of a New Method for Solution of

Nonlinear Problems. (In Norwegian). Student Project.
Division of Structural Mechanics, The Norwegian Institute
of Technology, Trondheim, (1980).

A:e:eendix - Eguations for the model :erob1em

The initial geometry is defined by

v2
H ,
Ho v
L 0 L

Spring coefficients

a+bv 2 , a , bL
k1 = const. , k2 m
0 m1
k1 k1

Nondimenslona1 forces and displacements

Auxi11iary quantities

f
f 1 + H2 + v 2 f2 (....2)~ - 1
0 0 0 f1
2 2 1+f2
f1 1 + (H o -0 1 ) + (v + 2)
0 f3
~
Nondimensiona1 potential energy

Equlllbrlum equatlons

P1 2f 2 (H o - 01)

P2 mo 0 2 + m10~ - 2f 2 (v o +0 2 )
 571

Incremental equil~brium equat~ons

Partitioned and Adaptive Algorithms for Explicit Time
Integration

1: BELYTSCHKO
Northwestern University,Evanston, USA

Swnmary

Two computational algorithms for the explicit t1me integra-

t10n of the equations of mot1on with different time steps
within a mesh are presented. The first is a mesh partition,
which has been general1zed so that an arbitrary number of
t1me steps can be used. The second method chooses the time
step for each node and element, and this time step depends
on the current material properties, so that the method 1S
adapt1ve. Results are presented for some simple one d1men-
sional nonl1near problems.

Introduction

In explicit time integration of transient problems, the

t1me step is limited by nwnerical stability. Th1S llmit is
inversely proportional to the maximum frequency of the mesh,
Wh1Ch 1n turn, varies inversely with the traversal t1me for
a wave across the element. Hence, 1f a mesh conta1ns very
short or st1ff elements, the time step required for sta-
bility becomes very small.

In f1nite difference treatments of cont1nua, th1S charac-

terist1c of explicit methods often 1mposes only small pen-
alties, for finite difference gr1ds are usually qU1te regu-
lar and only use a slngle type of "zone" or element. On
the other hand, in finite element applicat1ons, large var1-
atlons ln element size are commonplace. Furthermore, dif-
ferent types of elements, such as beams, constant strain,
and linear straln elements are often combined 1n a mesh,
and the maxlmwn frequency of these elements varles substan-
tlally. Thus the computational penalties incurred by Sln-
gle time step explicit integration of finlte element meshes
 573

are often substantial.

Several remedies have been proposed for this d~ff~culty.

Belytschko and Mullen [1,2] have developed expl~cit-impl~cit

methods where the mesh is part~tioned ~nto doma~ns by nodes

and the partit~ons are s~multaneously integrated by explic~t

and implic~t methods. They have shown that the max~mum fre-
quency in the explic~t partition governs the t~e step, so
that if the stiff portions are ~ntegrated ~mplic~tly, the
t~me step can be much larger. Hughes and Liu [3] have pro-
posed an expl~c~t-~mpl~c~t part~t~on by elements w~th s~mi­
lar stab~l~ty properties; the element bas~s of the~r part~­

tion simplif~es the algorithm and enhances ~ts compat~bi­

lity w~th general purpose finite element software.

However, explic~t-impl~c~t methods suffer from the disadvan-

tage that the addition of the implicit algor~thms ~ncreases

the s~ze and complexity of the program. Moreover, in non-

linear appl~cat~ons, the reliabil~ty of ~mplicit algor~thms

does not equal that of explicit algorithms.

For these reasons,~n [4] and [5] part~t~oned algor~thms were

developed and exam~ned ~n wh~ch explicit integration ~s
used throughout the mesh but different time steps are used
in the partitions. The algorithms used l~near ~nterpola­

t~ons in time of the displacement at the ~nterfaces between

part~t~ons, which ~n [5] were shown to y~eld stab~l~ty for
a model problem. Wright [6] has ~ndependently proposed a
sim~lar method.
In th~s paper, two further developments are rev~ewed. The
first, for which a more deta~led descr~ption ~s in prepara-

tion [7], is based on a constant veloc~ty ~nterpolation on

interfaces. It ~ncreases the versat~l~ty of the part~t~on

immensely, and by us~ng th~s approach, a single time step

code has been modified to treat an arbitrary number of par-
titions within a mesh by a few FORTRAN statements. Th~s

algor~thm ~s presented ~n the th~rd section.

The second development, descr~bed ~n the fourth sect~on, ~s

a generalization and extension of this scheme, in which

 574

each node and element is associated w1th an independent

"clock" and time step. The time steps are automatically
computed and altered throughout the simulation. Thus the
method 1S in many senses adaptive: as the elements change in
sizes or change in st1ffness because of the state of stress,
the local time steps are changed so as to minimize computa-
tions. The scheme is not w1thout flaws or shortcom1ngs: 1t
requires more storage for each element and node, and it does
not equal the reliability of the partit10ned technique. How-
ever, th1S is only a first, tentative step in thlS direction.
This method may have sign1ficant potential, but 1tS full re-
alization will require further development and a better under-
standing of 1tS stabil1ty characteristics and time step re-
str1ctions.

Governing Equations

We w1ll here restr1ct our treatment to the semidiscret1za-

tion of the momentum equations which w1ll be written in the
form

f (1)

where

f (2)

Here a superposed dot denotes a time derivat1ve, M 1S the

mass matrix, u a column matrix of nodal displacem;nts, fext
the external ~odal forces and fint the internal nodal f~r­
ces. For a nonlinear problem, the internal nodal forces are
obtained by

( 3)

where ~E is the Boolean connectiv1ty matrix, VE the volume

 575

of element E, and B the straln rate-dlsplacement matrlx

whlch gives the element strain rates in terms of the nodal
veloclties by

.
E B ~ = ~ ~E u
. (4)

For a linear, undamped structure

K u (5)

so that Eq. 1 can be written as

Mu + K ~ (6)

The central difference formulas give

un+~ un-~ "n

+ I1t n u ( 7)

n+l n
u u + I1tn+~ un+~ (8)

1 (I1t n-~2 I1tn+~)

I1t n + (9)
2"

where superscrlpts deslgnate the tlme step and fractlonal

superscripts designate a mldstep value.

Comblnlng Eq. (1) wlth Eq. (7), we obtaln

+ (10)

For the purpose of computing the new internal forces, we

also need to update the stresses by

n+l
o (11)

where ~reflects any frame invarlant stress rate that may

be needed. Equatlons (8-10), in conjunctlon wlth (3), (4)
and QU, provlde the explicit algorithm.
576

The stabil~ty limit for the central difference method with a

constant time step is given by

lit < (12)

where w is the maximum frequency of the mesh, wh~ch cor-

max
responds to the maximum eigenvalue of

K u ( 13)

For a constant stra~n element, an upper bound on the maximum

frequency is g~ven by

2c
w (14)
max T

where c is the wavespeed and 1 m~nimum length across an ele-

ment. Hence

(15)

wh~ch ~s the smallest wave traversal time across an element

~n the mesh.

Th~s stab~l~ty condit~on has been proven only for linear

problems; see, for example Fujii[8]. For nonl~near systems,
no str~ct stability cond~t~ons are available but using a
t~me step of 70% to 90% usually leads to stable computations.
Stability in time integrat~on of nonlinear systems should be
checked post facto by an energy balance.

For this purpose, the following formulas can be used

wn+l,ext

(17a)
 577

wn,~nt
E
+
I
2" l'It J ~n+~{~n + ~n+l) dV (17b)

VE

w~nt
L w~nt
E
(17c)
E

I .n+~ T
(~ ) M un+~ (18)
2"

where w~nt and wext are the ~nternal energy and external
work, and T is the k~net~c energy. Energy balance then re-
quires that

~,ext + ~,int + ! (Tn-~ + Tn+~) < TOL

2

(19)

where TOL ~s usually less than 0.05.

If the time step varies during the computation, Eq.(12) does

not hold. In a spec~al~zed analys~s g~ven ~n [91, where the
rat~o of success~ve time steps and wl'lt was cons~dered con-
stant, ~t was shown that vary~ng the time step ~ntroduces
an art~f~c~al v~scos~ty wh~ch ~s positive when the t~me step
decreases, negat~ve when the t~e step ~ncreases. The lat-
ter ~s never stable according to a Neumann
analys~s. The
analys~s [91 ~mpl~ed a constantly ~ncreas~ng time
g~ven ~n

step; numer~cal experiments have shown that l~m~ted ~ncreas­

es ~n time step over restr~cted t~me ~ntervals are not de-
stab~l~z~ng. However, ~ncreas~ng t~me steps should be
treated w~th care unt~l this aspect is understood better.

Part~tioned Algorithm

In th~s scheme we subd~v~de the mesh into NG groups of ele-

ments. The element group G ~s updated with a t~me step l'ItG
578

and the group time is denoted by tG. In addition, we use a

master time tM which is incremented by a time step 6t M. It
is required that all 6tG be integer multiples of 6tH , and
that for adjacent groups, the time steps be integer multi-
ples of each other.

Only the element al1gnment within the groups needs to be

specif1ed by the data. The nodal time steps are automati-
cally generated w1th1n the algorithm. For th1S purpose,
each node is allocated two words of storage, the nodal t1me
of node N, tN' and its last time step value 6tW

The essence of the procedure is as follows. Whenever the

master clock tM is 1ncremented, all element groups are
checked. Any group Wh1Ch has fallen beh1nd the master time
1S updated w1th its own t1me step 6tG and the stresses and
the clock t1me tG are updated for this group. At the same
t1me, any nodes connected to element group G are assigned
the t1me step 6tG if 1t is larger than the prev10usly as-
slgned t1me step. Therefore, a node at an 1nterface w1ll
always be updated with the largest time step among the ele-
ment groups connected to 1t.

After the element groups are completed, the nodes are

checked. Any nodes for which the clock is not ahead of the
master t1me are updated.

A flow chart of the scheme 1S given 1n Table 1. Other than

the data for inputting element groups, only about 15 FORTRAN
statements are requ1red to 1mplement this scheme in a slngle
t1me step scheme.

The scheme is most appropriate for elements with a velocity

stra1n formulation. Otherw1se, the displacements at the
element group time must be obtained by linear 1nterpolate
of the displacement prior to comput1ng the strains.

Table 1
Partitioned Explic1t Algor1thm

o
o. in1tial conditions: ~ ,
 579

for all groups; tM=O

1. tM -+- tM + Li tM
2. compute !ext, set !int = 0, LitN = 0

3. loop over element groups: G = 1 to NG

4. if tG>t M, go to 7

5. update stresses in all elements in group G us~ng time

step LitG using nodal veloc~ties ~ and Eqs. (4) and (11);
compute element internal forces and add into f~nt;
increment group time tG -+- tG + LitG -

6. for all nodes connected to group G, if LitG>LitN let LitN=LitG

(~.e. pick time step to be the largest of any element
group connected to it)

7. end of loop over elements

8. loop over nodes: N + 1 to NNODE

9. ~f tN > t M, go to 11

10. update veloc~t~es and displacements of node N using

LitN , tN -+- tN + LitN
11. end of loop over nodes

12. go to 1

Adapt~ve Algor~thm

The second algor~thm to be discussed here ~s ~ntended to

take advantage of changes ~n local stable time steps that
occur as the solut~on evolves. An example is the elastic-
plastic response of structures, where the stable time step
~ncreases substantially whenever the mater~al becomes plas-
t~c and cont~nues to be loaded. The prev~ous scheme ~s not
well su~ted to taking advantage of such situations because
the element groups are ~dentif~ed prior to the run as ele-
ment data and because of the restrictions on the ratios of
the t~me steps between part~t~ons.

In this algorithm, each element is advanced by its own t~me

step litE' and each node is advanced by its own t~me step
Li~. The nodal time step ~s the min~murn of the element
580

t~me steps of all elements connected to the node.

The basic feature of this scheme is that both the ~nternal

forces and the time der~vatives of the internal forces,

!"~nt ,
are stored for all nodes. Whenever a node is to be
updated, we f~rst update fint by

(20)

where fint ~s the latest value of the t~me der~vative.

Elements require the storage of the stresses at the previous
to time steps. Whenever an element is updated, its new
stresses are computed and the change in ~fint
-E
~s computed
by

~fint BT [ _ 1 _ (an+l _ £n)

-E ~tn+~ -
J E
VE

+ 1 (£n - £n-l)] dV (21)

~tn-~
E

The purpose of the second term in the above equat~on is to

Oint
el~minate the element's old contribution to f . The con-
tribution of the element ~s then added to the nodal forces
by

+ fint + LT ~fint (22)

~E -E

To drive the algorithm, a master time, tM is needed. This

master time ~s incremented by ~tM' which is always the smal-
lest increment ~tE and ~tN among all elements and nodes,
respect~vely.

A flowchart of the algor~thm ~s g~ven in Table 2. Note that

the time step for each element 1S recomputed cont~nuously;

however, ~n order to reduce truncation error and the desta-

 581

bilizing effects of increasing time steps, llmlts are set on

the maxlmum lncrease.

The threshold In step 2 lS essentlal to the method. If

the element tlme lags the node tlme by only a small amount,
failure to update the element can result In the development
of rather large errors.

This adaptlve scheme requires substantlally more storage

than a slngle time-step,expllcit scheme and is also more
susceptlble to roundoff errors. However, on a computer wlth
12 dlgits, thls roundoff error was not found to be slgnlfl-
cant over several thousand time steps in tests that have
been made.

Table 2
Adaptive Explicit Algorithm
o ._!:z ·lnt
o. condltlons: u , u , f = 0
o for all elements and nodes, tM
1. loop over elements: E = 1 to NELE

2. If tE > tM - threshold, go to 7

3. compute new element tlme step 6tE while computlng new

stresses by Eqs. (4) and (11)

4. modlfy total nodal force derlvatlve matrlx by Eqs.(2l)

and (22)

5. If 6t M > 6tE , set 6tM = 6tE

6. update element tlme: t + tE + 6tEi modify nodal tlme

steps for nodes connec~ed to element E

7. end of loop over elements

8. loop over nodes: N = 1 to NNODE

9. If tN > t M, go to 14

10. update !int for node N by Eq.(20)

11. compute !ext for node N

12. update nodal velocities and dlsplacements for node N

by Eqs.(7) and (8).
582

13. update nodal time tN + tN + t,~

14. end of loop over nodes

15. tM + tM + t,tM
16. go to 1

The time step for adjacent nodes and elements cannot be com-
pletely arbitrary. One ~mportant restr~ction is that the
element time step must be smaller than the nodal time steps
of its nodes. Th~s restrict~on detracts from the method
substant~ally, for ~t would be des~rable to be able to
increment the element at the largest possible t~me step,
s~nce updating an element is far more costly than updating
a node.

Example

An example to which the two methods were applied is shown

~n Figure 1. It ~s a bar in a uniax~al state of stress. A
solution by an explic~t-expl~cit part~tion with l~near d~s­

placement ~nterpolation was previously reported in [5].

A compress~ve stress of magn~tude 0.1 was appl~ed at the
left-hand end as a step function in t~me. F~gure 2 shows the
stress in the f~fth element from the left (x=4.5) obta~ned

by the part~oned technique described here. The results,

as those in [5], show the usual dispers~on and aliasing,
wh~ch are exacerbated by the large discont~nu~ty ~n element
lengths. F~gure 3 shows the stress in the same element for
the adapt~ve solut~on, which is qu~te s~milar ~n character.
In this problem, the adapt~ve method provides no advantage,
since the stab~l~ty limits do not change w~th t~me. Both
computations are clearly stable, which was conf~rmed by an
energy balance check.

The adapt~ve method was also tr~ed for other problems of the
same type. At t~mes the solution was observed to drift
from the correct solution. It was not a violent Neumann
type of instab~l~ty, although energy balance was v~olated.

These appear to be due to certain combinations of t~me steps

 583

in adjacent nodes and elements and the cholce of the ele-

ment threshold (step 2 in Table 2); adequate restrlctlons
and gUldelines remaln to be developed.

10 Elements (~=O.l)

I 2 3 4 5 6 7 8 9 1011ll 21I 222324252627282930 31r

(T(tl=!>11111 III
I
I I, I I I I I I I I r x
10 Elements (~=l) I 10 Elements (~=l)

Flgure 1. Mesh for sample problem; wavespeed c 1.0

o

-0.1

Ul
Ul
Q)

'"
~ -0.2

10 20 30 40 50
Tlme

Flgure 2. Comparlson of solution by partltioned algorlthm

wlth analytlc solutlon.

Analytlc

Ul
Ul
Q)

'"
.j.J
rJl
-0.2

x = 4.5

10 20 30 40 50
Tlme

Flgure 3. Comparison of solutlon by adaptive algorlthm

wlth analytlc solutlon
584

Acknowledgement

We would gratefully like to ackowledge the support of the

Electr~c Power Research Corporation and the National Sc~ence

Foundation.

References

1. Belytschko, T.B. and Mullen, R.: Mesh Partitions of

Expl~cit-Implicit Time Integration. Formulations and
Computational Algorithms in Finite Element Analysis,
ed~ted by J. Bathe, et. al., M.I.T. Press, (1977)
673-690.

2. Belytschko, T. and Mullen, R.: Stab~lity of Explicit-

Implicit Mesh Partitions ~n Time Integration. Interna-
tional Journal for Numer~cal Methods in Engineering,
Volume 12, (1978) 1575-1586.

3. Hughes, T.J.R. and L~u, W.K.: Impl~cit-Explic~t F~n~te

Elements in Transient Analysis: Stability Theory. ASME
Journal of Applied Mechan~cs, Volume 45, (1978) 371-374.

4. Belytschko T. and Mullen, R.: "WHAMS: A Program for

Transient Analysis of Structures and Continua. Struc-
tural Mechan~cs Software Series, Volume 2, edited by
N. Perrone, et. al., (1978) 151-212.
5. Belytschko, T.; Yen, J. and Mullen, R.: M~xed Methods
for T~me Integration. Computer Methods and Appl~ed
Mechanics ~n Engineering, Volume 17/18, (1979) 259-275.

6. Wr~ght, J.P.: M~xed Time Integrat~on Schemes. Computers

and Structures, Volume 10, (1979) 235-238.

7. Belytschko, T. and Robinson, R.: An Algorithm for Mixed

Time Step Integrat~on. Subm~tted for publicat~on.

8. Fuj~i, F.: Fin~te Element Schemes: Stab~l~ty and Conver-

gence. Advances ~n Computational Methods in Structural
Mechan~cs and Des~gn. Edited by J.T. Oden, et. al.,
Un~vers~ty of Alabama Press, Alabama (1972) 201-218.

9. Belytschko, T. and Mullen, R.: Explicit Integration of

Structural Problems. F~n~te Elements in Nonlinear
Mechanics. Ed~ted by P. Bergan, et. al., Volume 2,
(1977) 697-720.
Dynamic Relaxation Applied to the Quasi-Static, Large
Deformation, Inelastic Response of Axisymmetric Solids ,.

S. W KEY, C. M. STONE, R D KRIEG

Sandia LaboratOrIes ¥ " , Albuquerque, NM, USA

ABSTRACT

The use of dynamic relaxation as a solution strategy for the

quasi-static, large deformation, inelastic response of sol-
ids is examined. The underlying mechanics, the constitutlve
theories of interest, the incremental form of the equations,
the spatial discretization, and the implementation of dynamic
relaxation for path and/or time dependent material response
are each discussed. The mechanics are carried out in the
current configuration of the body described by a fixed spa-
tial coordinate system and using the Cau~hy stress. Finite
strain constitutive theories for elastic, elastoplastic, and
creep behavior are introduced. An incremental form of the
problem allowing a sequence of equilibrium solutions to be
found is presented. A constant bulk strain, bilinear dis-
placement isoparametric quadrilateral finite element is em-
ployed for the spatial discretization. The solution strategy
used to generate the sequence of equilibrium solutions is
dynamic relaxation which in the form adopted is based on
explicit central difference pseudo-time integration and arti-
ficial damping. It is used to find the next solution as a
result of an increment in time and/or load. Each solution
must satisfy equilibrium to within a prescribed tolerance
before proceedlng to the next increment. Several example
calculations are presented.

*This article sponsored by the U. S. Department of Energy

under Contract DE-AC04-76-DP00789.

k;A U. S. Department of Energy Facility.

586

1. INTRODUCTION

The use of dynamic relaxation as a solution strategy for the

quasi-static, large deformation, inelastic response of sol-
ids is examined. Dynamic relaxation as a means for generat-
ing solutions to statics problems goes back at least to Otter,
Cassell and Hobbs [lJ. The basic idea is to convert a static
problem into a dynamic problem by adding acceleration and
damping terms to the equations and then with "optimal" damp-
ing integrate the transient response out until the quiescent
or static solution is obtained. The idea is eminently adapt-
able to the problem of large deformations and inelastic
materials. The result is an explicit, iterative strategy
which has a large radius of convergence as noted by Under-
wood [2J. A considerable literature concerned with the
transient dynamic response of solids and structures can be
brought to bear on the problem [3-11J.

The purpose here is to present an analysis procedure where

large deformations, finite strains, inelastic time indepen-
dent and dependent constitutive modeling, and arbitrary load
incrementation are all combined for problem solving.

Of particular significance is the disassociation of the in-

cremental constitutive evaluation scheme from the load incre-
mentation time step and related equilibrium iteration proce-
dure. The approach is to integrate the constitutive equaticns
accurately with subcycling if necessary, independent of the
time step selected for load incrementation. This is an ap-
proach urged by Bushnell [12J and a practice of long standing
in transient dynamic structural response calculations [13J.
 587

2. MECHANICS

While the purpose is to consider axisymmetric and planar

problems, the treatment in this section is general for the
sake of brevity and completeness. The treatment of continu-
um mechanics found in Truesdell and Toupin [14J is used.

Equations of Equilibrium. A body 1V is given which occupies

a finite region of Euclidean space. Subjected to prescribed
body forces and surface tractions, the body 1Vundergoes the
motion xi = Xi(Xa,t). The particles of the body are identi-
fied by the coordinates xa. They are referred to as material
coordinates, and the relation of the particles to the coordi-
nates xa does not change in time. The places in space which
the particles occupy during the motion are identified by the
coordinates xi. The function xi describes the motion of the
particles Xa through space as a function of time t. It is
the motion xi which is sought.

The place occupied by the body at t = 0 is taken as the ref-

erence configuration. In this configuration the body is
assumed to be strain free, though not necessarily stress
free. Only material coordinates if which coincide with the
spatial coordinates xi in the reference configuration are
considered. Thus, in the reference configuration, Xi (if ,0) =
xa.

The problem is stated in terms of the principle of virtual

work. The differential form

k
an = pE aX k dv (1)
1V ~l

is to vanish at all points along the path of motion for all

variations aXk satisfying the displacement boundary condi-
tions on ~2. The integration is performed over the current
configuration of the body 1V, where p is the mass density in
that configuration, t km is the Cauchy stress--the stress in
588

the current configuration, and sk is the surface traction

which is acting on 7 4. The comma in x k
,m denotes covariant
differentiation.

The divergence theorem is employed to display the equilibrium

equations. In anticipation of using the finite element method
to generate approximate solutions, the case where oX k is
,m
only piecewise continuous is considered. Interior surfaces
where the discontinuities of oX k occur are denoted by 7°.
Only surfaces 7° which are sta ti~~ary with respect to the
material are considered. The situation is pictured in Fig-
ure 1 where n k is the normal to 7° and the symbols + and -
denote the respective sides of the surface. The result is

- f (tkm,m k
+ pf ) ,~xk dv + f km
(t km _ t_
+ )nmox k da
r 7°

_ s k )ox da
+ ,f (tkmn
m k 0 (2)

7 1

The differential form will vanish if and only if the respec-

tive integrands vanish. The resulting expressions are equili-
brium

t km fm o in r ( 3)

the jump condition at a contact discontinuity

o on g> ° (4 )

and the traction boundary conditions

(5)
 589

The displacement boundary conditions are

FII 1 The body 'If WIth surface tractloru I' on the boundary ~ I and a prescnbed motion on the boundary Jl An Intenor
boundary cI 0 with a umt nonnal vector ,./r IS PictUred

a
on y> (6)

It is important to realize that these equations are completely

general and applicable for arbitrarily large deformations.

Strain and Strain Rate. In finite deformations there are

many strain measures which are useful. The majority of them
can be computed from the deformation gradient Fk defined by
a

ayk Q
.:.A- (x"',t) (7)
axel

The left Cauchy-Green tensor is computed from the deformation

gradient as

(8)

The quantity Ga~ is the metric tensor in the coordinates of

the reference configuration. The velocity v k is defined as
590

vk(Xa,t) - ~ (Xa,t) (9)

at

The stretching is given by

d km 1
2 (vk,m + vm,k) (10)

with the spin by

1
wkm 2 (vk,m - vm,k) (11)

In the constitutive models which follow, only the left

Cauchy-Green tensor Bkm and the stretching d km are used as
strain and strain rate measures, respectively. There are a
wealth of additional strain and strain rate variables which
could be introduced but they are not needed here. Further
reference to them can be fcund in Truesdell and Toupin [14].

3. CONSTITUTIVE THEORIES

The stress relations considered are those of elementary

character. They represent three classical forms of material
behavior and are remarkably adaptable in performing useful
engineering analyses. They are rubber elasticity which exhi-
bits path independent and time independent behavior, classical
plasticity which exhibits path dependent but time independent
behavior, and classical creep which has both path dependent
and time dependent behavior. These attributes have important
consequences when incremental strategies are considered.

Rubber Elasticity. Rubber is a material used because of its

ability to remain elastic when subjected to large deforma-
tions. Blatz and Ko ~5] have made an extensive study of its
stress-strain behavior when subjected to large strains. They
 591

consider the behavior of both foam and continuum rubber.

Here only the continuum rubber is treated. The left
Cauchy-Green strain (8) is used. Using the third invariant
of the strain which is given by

Idet (F~)I (12)

the cauchy stress is computed as

(13)

Here, grs is metric tensor in the spatial coordinates, ~ is

the shear modulus and v is their generalization of Poisson's
ratio to finite strains. Based on data sited by them, V is
taken equal to 0.463. This gives J 3 an exponent of 13.3 and
a marked hardening behavior under compression. This descrip-
tion for bulk compression matches the data to 700,000 pounds
per square inch. This is a hyper-elastic relation.

Plasticity. Plasticity is the behavior characteristic of

ductile metals. Figure 2 shows results which are typical of
the behavior of a metal bar loaded first in uniaxial tension
followed by uniaxial compression. The straight line repre-
sentation in Figure 2 is an idealization of this behavior.
This is the approximation which results from the plasticity
re1atlons employed here, Goe1 and Malvern [161. It is based
on the notion of a universal hardening curve from which gen-
eral triaxial behavior is predicted. The finite strain
treatment of Key, Krieg, and Biffle [17] is used. It is
stated in a rate form where the invariant co-rotational or
Jaumann stress rate is related through a tangent modulus to
the stretching. The result is equation (14).
592

Fig. 2. The typical behavior of a ductile metal bar loaded

first in uniaxial tension followed by uniaxial compression.
The straight line approximation is characterized as an
elastic modulus E, a yield stress t , a strain hardening
modulus Et , and a hardening parametgr a, where kinematic
hardening is obtained with a = 0, isotropic hardening is
obtained with a 1, and a linear combination of the two
is obtained for a between zero and one.

V'rs
t (14)

' l d '~ng ~s
Wh en no y~e ' crsmn ~s
, occurr~ng, , the ~sotrop~c
, , tensor
Agrsg mn + 2~grmgsn, where A and ~ are the Lame parameters.
When plasti; yielding is occurring, that is, when! ~~,~~ij _
2 c~ ~,r~ 2 lJ
k = 0 and ~rs ~ > 0, the moduli are replaced by

(15)
 593

where

rs I;,rs /[2k 2 (1 + H/3~)] 1/2

n

I;;s t ' - a;s

rs

/2k S ~ P
3 Hld rs I , k (O) (2/3) 1/2 t
0

\/
a 2 P
(1 - S)
rs 3" H d rs ars(O) 0

Isotropic hardening is described by S = 1. Kinematic harden-

ing is included in a rather obvious way by letting S = 0 and
prescribing the center of the yield surface a. .. to move
\/
according to a. ij = (1 - S) H d~s 1 ~J
The prime denotes
deviatoric components, the superscript p denotes the plastic
part of the stretching, and H = EEt/(E - Et } .

Creep. Creep occurs in man-made and natural materials, and

in ductile materials becomes significant when the temperature
on an absolute scale reaches one third to one half the
material's melt temperature. It is a time dependent material
594

mechanism in that the stress depends principally on the rate

at which the material is deformed. Elastic terms are also
necessary in this temperature range and are included here.
In what follows the common Norton power law for secondary
creep will be used, Penny and Marriott [18J. It is parti-
cularly descriptive of dislocation glide and climb deforma-
tion mechanisms in metals, Gittus [19J.
It is stated in a rate form where the invariant co-rotational
stress rate is related linearly to the difference between
the total stretching minus the inelastic stretching d C .
rs
Thus

~rs c rsmn (d _ dC ) (16)

mn mn

where

n - 1
-2-
l2 0 (l
,2
t:1J,t "i j ) exp(-Q/R8)t"
rnn

o and n are constants determined from the material, 8 is

temperature, Q is the effective activation energy, and R
is the universal gas constant.

4. INCREMENTAL FORM OF THE PROBLEM

The notion of an incremental solution is fundamental to the

bulk of the methods for finding a motion Xi(Xa,t) which
generates a stress history in equilibrium with the applied
loads. It is assumed that at time t , the stress t n satis-
n rs
fies the equilibrium equations (3), (4), and (5) i and the
stress t n is the result of integrating the constitutive
rs
models with the strain histories derived from the known
motion up to tn' The prescr1bed loads are incremented to
time tn+l and a predictor/corrector method is introduced to
find the new configuration x~+l = Xi(Xa,t n + l ) which has all
 595

the equilibrium properties which were deemed necessary at

tn and accuracies acceptable to the constitutive model evalu-
ation. To indicate the incremental quantities a ~ is used.
For example, ~t = tn+l - tn' ~trs = t~:l - t~s' etc.

A basic assumption which underlies most incremental treat-

ments, including the one here, but which is rarely stated is
that the motion between xi and xi+l is linear. As a conse-
n n .
quence, the incremental velocity given by v~+1/2 = ~xi/~t is
constant over the time increment. In what follows equil1-
brium is tested in the conf1guration at tn+l which is a trial
configuration until equilibrium is established. To do this
the stresses at tn+l must be evaluated. Following Hughes
and Winget [2 OJ , the one parameter family of configurations
is introduced

i i
(1 - a)x n + aX n + l (17)

The gradient h ij of u i ~xi with respect to x n+C't

j is given by

h .. u .. (18)
1J 1,J

From this gradient the strain increment e ij is given by

e .. (a) (19)
1J

Thus, e .. (0) is the Green-St. Venant strain increment, e .. (1)

1J 1J
is the Signorini strain increment, and

1
e .. sym (20 )
1J 2"
596

Without the need for further linearization the configuration

halfway between nand n+l is selected for evaluating the
stretching, spin, and computing ~trs. The midpoint configu-
ration is optimal in the sense that no quadrat1c terms are
i i i i
needed to accurately evaluate (dx dx )n+l - (dx dx In. The
terms in the co-rotational derivative involvinq the spin wrk
used in the constitutive equations (14) and (16) are for the
purposes of taking rigid body rotations of a material point
relative to the spatial coordinates xi into account. In
incremental form they are an orthogonal rotation through an
incremental angle. Hughes and Winget [20J have provided a
modern account of this process and have provided a direct
way to evaluate the orthogonal rotation matrix R .. from the
1J
spin w ..• Thus,
1J

1 i -1 1
R ..
1J
(oi
k
- l1t 2' w k) (gkj + ~t 2' wkj ) (21)

Half angle trigonometric formulas are used to get the square

root of [R;
1 Rij = Ri~k Rkj
~
. With these constructions the
constitutive models (14) and (16) can be integrated over
the increment from n to n+l. First the stress t n and the
rs
applicable state variables an are advanced to n + ~ by
rs

-n+~ R~i R~j t~.

t (22)
rs r s 1J

-n+~
a R~i R~j a n.. (23)
rs r s 1J

Using d n + l / 2 and l1t, the constitutive equations are inte-

rs n+l/2 . n+l/2
grated and new stresses t and state var1ables a
rs rs
are obtained. These are then rotated from n + ~ to n + 1
by the same process as in equations (22) and (23). Mid-
interval constitutive evaluation is used by Hallquist [21J
and Biffle [22J.
 597

Of part1cular importance is the integration of the con-

stitutive equations from time n to n + 1. The rubber
elasticity is straightforward in that a new left Cauchy-
Green strain is computed every time from which the Cauchy
stress is computed from (13). Since this is a hyperelastic
model, 1t is independent of the path followed to get to the
configuration at n + 1. The plasticity model is integrated
from n to n + 1 assuming the stretching is constant which is
consistent with (17). For a constant stretching path the
integration is very nearly exact and is extremely reliable,
and independent of the specific value for the time interval
tit, [23-25J.

The creep model is also integrated from n to n + 1 assuming

the stretching is constant. The creep equations, however,
are not nearly as easy to numerically integrate. They are
referred to in the mathematics community as "stiff". Only
with considerable effort and great expense can they be nu-
merically integrated with conventional methods from time n
to n + 1. In this case, a semianalytic integration is used.
Domains in stress and strain rate space are identified where
various nearby differential equations with exact solutions
are applicable. The solution path over a time step may re-
main within a single domain or may pass through two or more
doma1ns requ1ring the solutions to two or more of the differ-
ential equations to be applied, one after the other over the
time step. In this way arbitrarily large strain or time
intervals can be accurately and reliably used. An absolute
maximum of seven subincrements are required so that computa-
tional t1me is not excessive. Th1s approach, while conceptu-
ally straightforward is highly tailored to the constitutive
model at hand.

This approach to constitutive equations uncouples stability

and accuracy in their evaluation from the time step size
used in load incrementing schemes w1th their attendant equil-
ibrium iteration, cf., Bushnell [12J • This approach contrasts
598

with that where ~t in the constitutive integration scheme

corresponds to the ~t used in incrementing the load.

If this coupling is pursued, many numerical schemes can be

developed. Argyris, Vaz and William [26] have documented a
number of these and have provided the guidance necessary for
their successful use.

While seemingly elaborate, these are in practice simple and

quick calculations. This treatment has the property of being
objective and uses a stra~n increment which while appearing
"linear" is the same order of accuracy as using the nonlinear
strain increments at either n or n + 1, respectively. It
remains only to introduce the spatial discretization in order
to discuss the dynamic relaxation solution strategy.

5. SPATIAL DISCRETIZATION

The treatment up to this point has been in terms of an arbi-

trary body undergoing an arbitrary motion. Of interest here
are the axisymmetric and plane strain problems. Because the
planar case may be obtained from the axisymmetric case by
moving the symmetry axis arbitrarily far off, only the axi-
symmetric problem is discussed. Both of these problems are
spatially two-dimensional. No attempt is made to do the
axisymmetric problem in general but rather a specific ele-
ment with specific interpolation functions is introduced
from the outset.

If cylindrical coordinates, r, 8, z are introduced, an axi-

symmetric body can be characterized by a cross-section in the
r-z plane which when revolved about the z-axis generates the
body. Such a cross-section with a quadrilateral element is
pictured in Fig. 3 (such an element is frequently called a
"ring element" because when it is revolved about the z-axis
a ring is generated). As indicated in Fig. 3, the element
is a four node isoparametric element [27]. Thus, if ya re-
presents the value of the function y at the node with index
 599

L-------------------------~_r

Fig. 3. The cross-section of an axisymmetric body with a four node iso-

parametric quadrilateral element. The isoparametric coordinates
range between minus one and plus one.

a, and ~a represents the interpolation function at the node

with index a, the function y is described within the element
by

y (24)

where the bllinear isoparametric interpolation functions are

given by

(1 - a)(l - b)/4 (1 + a)(l - b)/4

(25)
4
(l + a)(l + b)/4 ¢ (1 - a) 11 + b)/4

When all the elements around node 1 are considered, the basis
function obtained from the interpolation functions in each
element can be pictured. Flgure 4 shows the result for this
case. As can be seen in Fig. 4, the basis functions obtalned
 600

Fig. 4. The basis function wI obtained from considering the contributions

of the respective interpolation functions from each element surround-
ing nodal point 1.

in this manner have only piecewise continuous gradients along

the element boundaries. As a result, implicit in the finite
element equations which result from this choice of basis func-
tions is an approximation to the jump conditions (4) along
each element interface.

When the basis functions are introduced into the principle of

virtual work and the variations taken, a discrete form of the
differential equations (3), (4), and (5) is obtained as

(26)

where, with no sums on the index k,

a vector
 601

k~a
s ~k a vector.

This is a discrete statement which requires the divergence of

the stress field [T] to be in equilibrium with the prescribed
body forces and surface tractions [F].

The integrals for [T] over each element are performed numeri-
cally using Gaussian quadrature rules. A 2x2 integration is
used to evaluate the area integral. In the process a con-
stant bulk strain is used. This may be viewed as a 1 point
integration of the volumetric strain energy and a 2x2 inte-
gration of the deviatoric strain energy, [28-30].

6. DYNAMIC RELAXATION

As a solution strategy for statics problems, dynamic relaxa-

tion involves first converting the equilibrium equations into
equations of motion by adding an acceleration term, second
introducing an artificial damping, and finally 1ntegrating
forward in time from initial conditions until the transient
dynamic response has damped out to the static result with
equilibrium satisfied. An early introduction of the idea 1S
given by Otter, et al. [~, but a more recent work which
summarizes all of the significant contributions on the topic
since then is Underwood [2]. In reference [2] cons1derable
detail is presented about dynamic relaxation and its use in
nonlinear problems.

Dynamic relaxation is attractive for three reasons: it is

vectorizable, it is versatile, and it is reliable. Because
it can be made explicit, it is highly vectorizable for modern
digital calculations. In an explicit form it is ideal for
dealing w1th large deformations, finite strains and inelastic
material behavior. It is incredibly reliabie and indefatigable
602

in seeking an equilibrium state. The method does suffer from

a certain lack of elegance.

To produce a dynamic problem, an acceleration term is added

to the equilibrium equation (3). Thus,

(27)

where r is a spatially varying density selected to m~nimize

the numer of steps needed to reach equilibrium and T is a

pseudo-time scale connected w~th the dynamic relaxation and
separate from real time t. The acceleration term is discretized
the same way that it would be in a true dynamics calculation,
[31]. This leads to the discrete system

(28 )

where

q a vector

[M] [l rcp~ dVJ , a diagonal matrix

At time t equilibrium is satisfied so that [T] = [F] • A

n n n
new solution is initiated by incrementing the load to time
t n +l • In general, equilibrium will not be satisfied, so that

(29)

Central difference expressions are introduced first for the

acceleration in terms of a velocity p and then for the velocity
in terms of the displacement q. Thus, the equations of motion
 603

(29) become

(30)

Here, 0 is the damping, the optimum value of which leads to

the minimum number of steps i to reach equilibrium.

Every step i leads ~o a new trial configuration (x~+l)i and

n+l)1 . The path 1n
trial stress ( trs . space traced out by the

steps is artificial; it is a by-product of the dynamic re-

laxation, as is the advance in time T. The trial states i
represent equilibrium iterations. Figure 5 depicts the
process in a multidimensional configuration space of the
nodal point positions. The point n is an equilibrium solu-
tion and the point n + 1 is the equilibrium state being
sought. The curved path between nand n + 1 traces out the
true solution. The spiral path marked with tics and para-
meterized by steps in T is the sequence of trial states
generated by the dynamic relaxation method. The straight
line from n to the last step calculated from dynamic relaxa-
tion is the 1nterval over which the stress t n + l is evaluated
rs
using the real time step 6t. The integrals in (26) are re-
evaluated at each step i using the trial geometry and when
equilibrium is achieved a straight line approximation to the
true path between nand n + 1 is obtained. This scheme
uncouples the path dependence and real-time dependence of
the constitutive behavior from the arbitrary sequence of
trial states generated by the dynamic relaxation method.

Convergence is based on achieving an acceptably small equili-

brium unbalance. Because the converged solution is a straight
line approximation, the true state n + 1 will not be found,
but a nearby equilibrium state will be found nonetheless.
 604

Fig. 5. A model equilibrium iteration sequence in a multidimensional configura-

tion space of nodal point positions developed with dynamic relaxation
showing convergence at load step n+l. The straight line path from n to
the last step calCUl!ted from dynamic relaxation is the interval over
which the stress t~~ is evaluated using the real time step ~t.

The only question remaining is how to select the variable

density r, the pseudo-time step ~T, and the damping 8 in
order to find a converged solution in the minimum number of
steps i. The performance of dynamic relaxation is tied to
the minimum natural frequency Wo and maximum natural fre-
quency wI of the discrete equations.

The damping per cycle which is demonstrated by Eqn. (30) is

frequency dependent. For a given damping factor 0 the de-
crease in amplitude per cycle is greatest for the lowest
frequency component. The damping is then chosen to provide
critical damping for the lowest frequency. This expression
is

(31)
 605

Note that the range on 8 is (0,1). A stability analysis on

this set of explicit equations produces a critical pseudo-
time step given by

tn (32)
c

The damping 1S stabil1zing although in a practical sense

this is negligible. If the problem were linear so Wo and
wl were fixed, then the numer of time steps required to
reduce the amplitude by a factor of ten would be

(33)

From Equation (33) it is seen that any action to reduce the

ratio wl/wo speeds convergence.

For the linear problem and a uniform mesh of dimension h, the

maximum frequency wl is glven by

2 c/h (34)

Here c is the uniaxial strain sound speed given by

[(A + 2~)/rJ~. While the problem is nonlinear in the large,
as equilibrium is approached at the end of each load step
the behavior is linear. With (34) the condition on 6T is
essentially the well known Courant-Friedrichs-Levy condition
which says that 6T can be no larger than the time it takes
for information to travel between the "closest" two nodes in
the mesh.

On the basis of these observations, the concept of mesh

homogenizat1on is introduced. The densities r are selected
element-by-element to give the same transit time across each
606

element, regardless of the size of the element. Thus, in-

formation about equilibrium unbalance is transmitted uni-
formly over the mesh. This process gives the lowest possible
value to the maximum frequency wl ' the goal being to mini-
mize the ratio wl/wo'

Even with changing sound speed and mesh distortions 6T is

easily controlled. If necessary, the density r can be re-
defined between load steps to recover a uniform element
transit time.

The fundamental frequency Wo is continuously estimated using

an approximate Rayleigh quotient, [2J. At each step i in
the dynamic relaxation, a new estimate (w ). is computed from
o ~
(35) •

(W ) (35)
,0

where [KJ loc is a diagonal stiffness matrix whose jth com-

ponent is computed from

(36)

With each estimate of the fundamental frequency a new value

of the damping 0 is computed. This has the virtue that the
lowest active mode will be found in the event that the fun-
damental mode is not participating, [2J.

7. APPLICATIONS

The first problem while elementary is a relaxation or stress

redistribution calculation. A changing stress field provides
the maximum feed back in a creep constitutive model, equation
(16). The material is naturally occurring bedded halite (NaCl)
found some 650 meters below the surface in the southwestern
 607

United States. The properties for use in equation (16) are

given in Figure 6 with several of the coefficients combined
into a single leading constant for the Norton power law.

The calculation is for the hole closure which takes place as

the result of a suddenly created cylindrical hole in a hydro-
statically stressed "infinite" body of halite. The surface
of the cylindrical hole is traction free. This is a one-
dimensional problem in which only a single row of elements
proceeding radially outward from the cavity is needed, Fig-
ure 6. At time t = 0, the solution is that for an elastic
material which gives high deviatoric stresses at the edge of
the hole and dying out within 20 to 30 hole radii. Of in-
terest is the hole closure which is influenced by the trans-
ient stress redistr1bution. The goal here is to show the
arbitrary size of the load increment in terms of ~t. Based
upon a time step selection procedure taken from [26], a
critical time step was computed for this problem. The hole
closure associated with ~tcr1tical' 10 x ~tcritical' and
100 x ~t 0tO 1 is given in Figure 7. It is clear from
cr1 1ca
this figure that the choice of ~t influences the accuracy
of the calculation but the amount of computational effort
associated with each solut1on is found to be identical.
This illustrates the uncoupling of the constitutive model
integration from the dynamic relaxation solution algor1thm.
Both of the smaller ~t solutions show ident1cal hole closure
rates at 10 years while the largest ~t has Just begun to
approach steady creep at 10 years.

The second problem is a two-dimensional calculation con-

cerned with the response around an underground opening.
Figure 8 shows the dimensions and boundary conditions. Fig-
ure 9 shows the stratigraphy incorporated in the calculation
with the mechanical properties given in Table I. Figure 10
shows the finite element mesh used. Zero friction was used
on the sliding interfaces representing the clay seams. Again,
the opening is assumed to occur instantaneously at t = O.
The calculation is characterized by stress redistribution as
600

a result of the disturbances introduced in the in situ hydro-

static stress field by the drift opening. While there is a
great deal of detail in the calculated results which could
be displayed, only the floor to ceiling convergence at the
room centerline is shown in Figure 11. Shown as well in
Figure 11 is the cumulative number of iterations needed to
generate the solut~on as time accumulates. During the early
stress redistribution phase, the majority of the calculational
effort takes place. Once this period is passed, the problem
takes on the character of a creep calculation and very few
iterations are needed to obtain the solution.

The third problem considers the necking of the plane strain,

steel tensile bar shown in Figure 12. Deformation controlled
loading was used with the left end displaced in 100 equal
steps to an increase in the axial length of the bar of 78%.
The material is modeled as an isotropic hardening, elastic-
plastic metal w~th a bilinear representation for the stress-
strain curve. Material property data is shown in Figure 12.
To insure that the bar would neck, a geometric imperfect~on

was built into the model. The problem is character~zed by

large plastic strains both axially and rad~ally as seen ~n

the deformed mesh of Figure 12. The amount of neck~ng,

characterized by the non-dimens~onal width change of the bar,

is also plotted in F~gure 12. Significant necking begins
to occur after the ax~al strain has exceeded 40%. The solu-
tion was found to be sensitive to the convergence tolerance
applied to the analysis. Informal comparisons of results
with other codes and element choices have shown a convergence
tolerance of 0.1% to give acceptable results. Severe dis-
tortion of the elements near the neck would normally cause
some concern but favorable comparisons with other codes
seemed to give no justification for repeating the analysis
with a finer mesh.

The last example considers an elastic-perfectly plast~c in-

finite cylinder subjected to an internal pressure. There
is an exact solution for this problem [3~ so that a comparison
 609

can be made with theory. Flgure 13 gives the material prop-

erties and geometry used ln the analysis. The numerical
solution matches the theory qUlte well. The algorithm handles
elastic-plastic problems wlth no difficulty.

TABLE I. MATERIAL PROPERTIES USED IN DRIFT ANALYSIS

DEGRADED ANHYDRITE OR
SALT SALT POLYHALITE ~g~~~~~i~'!SALT
Young's Modulus (Pa) 2.5 x 10 10 2.5 x 10 10 7.3 x 10 10 5.28 x 10 10

Po~sson' s Rat1Q 25 .25 .30 .28

Elast1c Bulk Modulus (Pa) 1.7 x 10 10 1. 7 x 10 10 6.1 x 10 10 4 02 x 10 10

Elast1c Shear Modulus (Pa) 1.0 x 10 10 1. 0 x 10 10 2.8 x 10 10 2.06 x 10 10

Creep Constants

A (Pa -5 S-l) 3 16 x 10- 45 1. 012 x 10- 44 ------- 6.32 x 10- 46

n 5. 5. ------- 5.

Coeffl.cl.ent of Frlctl.On)J O.
610

T
11.

~~r
1.~~ .j
P = 1.41 x 10 7 Pa

Material Proj2erties
9
G = 9.93 x 10 Pa
K = 1. 65 x 10 10 Pa

EC = Aa
n

3.47 x 10- 44 Pa- 4 • 9 -1

A = s
n = 4.9

Figure 6. Problem Description for the Externally

Pressurized Infinite Cylinder.
 611

.036

.032

.028

.024
100. x IItcr

.
~
"'-
.020

..~
-
,
Q)

.016
-..-:..
--........
o
.-I
U
~~ ....
Q)
lit 10. x lit _~
-~
.-I
o cr _
..,.,.".'" ..,;"
~

::r:: .012
__ -.:-':;:""-;; = II t
__ ~.... cr
/ ......
.008
/:/
.,//

.004
V-;/'
/
/

o.
o. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
Time (Years)

Figure 7. Radial Creep Closure of the Cylindrical

Hole for Several Choices of lit.
612

l !III

19.32

I
g
5.03

104.62

Lx
Figure 8. Dimensions and Boundary Conditions
for the Two-Dimensional Analysis of
an Underground Opening
 613

POL YHALITEI DEGRADED

ANHYDRITE ANHYDRITEI
SALT SALT

POL YHALITEI
ANHYDRITE CLAY SEAM

-
=
~
601.7
606.6
609.9
611.4

622.1
(MODEL BOUNDARY)

=
VVVVV\

~ 626.7 ~
..i:ivViM
~
634.6
639.5
-- --
--
VI/VVII\

647.1
- 649.5
~

663.6
664.5
670.6

683.4
~ 685.8 ~

693.7
-
=-==-
-:.-:..
~ 698.0

708.0
~ 710.5 (MODEL BOUNDARY) 'VVINV\
711.7
-==~
~-=- 717.8

Figure 9. Geologic Profile and Stratigraphy

Used in Model of Underground Opening
614

~~

Figure 10 . Finite Element Mesh Used for

Analysis of Underground Opening
 615

2000
.14
1800
/
.....-
....-
-- -
....-
....- ....-
.12 1600

s
-- --- - --
Iterations

'"--
1400

,,,
~

Q)
1-1
.10
::I
rn
0
I 1200 rn
r-i
I::
0
U •.-1
.08 .j..I
.j..I
'1-1 1000 III
1-1
•.-1 Q)
1-1 .j..I
C H
r-i .06 800 r-i
III III
U .j..I
•.-1
0
.j..I
1-1 600 Eo!
Q)
:> .04
400
.02
200

o. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
Time (Years)

Figure 11. Vertical Drift Closure and Total Number of Solution

Iterations Versus Creep T~me for a Single Drift in
Salt
 616

GEOMETRY
WA = 3. 0 in. ( . 0762

IIf! !!! !!! ! !f!t!!!!!

m)
A B e D
WB = 3.0 in. ( . 0762 m)

We = 2.96 in. (. 0752 m)

WD = 2.96 in. (.0752 m)
L = 9.0 in . (.2286 m)
I, .1

-----------l
6L = 7 . 0 in. (.1778 m) L

IIIIIIIIII!I~ ~
I" L + 6L

PROPERTIES 11
E = 30. x 10 6 psi (2.07 x 10 pal H =5 x 10 4 psi (3.447 x 10 8 pal
0y = 4. x 10 4 psi (2.758 x 10 8 pal v = .3

- O.OS

Cooverqanc.
- 0.10 Tolerance ttl

- 0 . 15

- 0.20

- 0.25 L--.........---"_~--'-~'---o---'---I
O. . 1 .2 .3 •• .5 ., .7 ••

AX.IAL STRADl, .Vl.

Figure 12. Problem Definition and Results for the Necking

of a Plane Strain Steel Tensile Bar .
 E 2.07 x lOll Pa

\! .3
Theory
t ---] a=l. b=2.
8 6 Finite Element Results
cry = 3.1 x 10 Pa
Pb j k = cr
y 113

o. ;p 1.0

0.9

0.8

0.7
.><:
N
0.6 ~
N'><: ~
81'" b "-<D
b
0.5

0.4
1.
-0.7 0.3

O.~I ____ ~ __ ~~ __- L_ _ _ _ ~_ _~

-0.81 10.2
1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.2 1.4 1.6 1.8 2.0
pia ria
~
Figure 13. Comparison of Theory and Finite Element Results .....
for an Internally Pressurized Infinite Cylinder
618

REFERENCES

1. J. R. H. Otter, A. C. Cassell, and R. E. Hobbs, "Dynamic

Relaxation," Proc. Institution of Civil Engineers, 35
(1966) pp. 633-656.

2. P. Underwood, "An Adaptive Dynamic Relaxation Technique

for Nonlinear Structural Analysis," Lockheed Palo Alto
Research Laboratory, Report No. LMSC-D678265, Palo Alto,
California, USA, in preparation.

3. R. W. H. Wu and E. A. Witmer, "Finite-Element Analysis

of Large Elastic-Plastic Transient Deformations of Simple
Structures," AlAA Journal, Vol. 9, No.9, September 1971.

4. J. F. McNamara and P. V. Marcal, "Incremental Stiffness

Method for Finite Element Analysis of the Nonlinear Dyn-
amic Problem," ONR Symposium on Numerical and Computer
Methods in Structural Mechanics, Urbana, Illinois,
September 1971.

5. M. Hartzman and J. R. Hutchinson, "Nonlinear Dynamics

of Solids by the Finite Element Method," J. of Computers
and Structures, Vol. 2, pp. 47-77, 1972.

6. J. H. Heifitz and C. J. Costantino, "Dynamic Response of

Nonlinear Media at Large Strains," Journal of the Engin-
eering Mechanics Division, ASCE, Vol. 98, No. EM6,
December 1972. ----

7. T. Belytschko and B. J. Hsieh, "Non-linear Transient

Finite Element Analysis with Convected Coordinates,"
IJNME, Vol. 7, 1973, pp. 255-271.

8. J. o. Hallquist, "DYNA2D - An Explicit Finite Element

and Finite Difference Code for Axisymmetric and Plane
Strain Calculations (User's Guide) ," UCRL-52429, Lawrence
Livermore Laboratory, University of California, Livermore,
California (March 1978).

9. G. Maenchen and S. Sack, "The Tensor Code," Methods in

Computational Physics, Alden, Fernback and Rotenberg,
eds., Vol. 3, Academic Press, New York, 1964.

10. M. L. Wilkins, "Calculation of Elastic-Plastic Flow,"

Methods in Computational Physics, Alden, Fernback and
Rotenberg, eds., Vol. 3, Academic Press, New York, 1964.

11. L. D. Bertholf and S. E. Benzley, "TOaDY II, A Computer

Program for Two-Dimensional Wave Propagation," SC-RR-68-
41, Sandia Laboratories, Albuquerque, New Mexico,
November 1968.
 619

12. D. Bushnell, "A Strateqy for the Solution of Problems

Involving Large Deflections, Plasticity and Creep,"
Int. J. Num. Meth's Engr., Vol. 11 (1977) pp. 683-708.

13. D. L. Hicks, "Hydrocode Subcycling Stability," Mathe-

matics of Computation, submitted for publication-.----

14. C. Truesdell and R. A. TouPln, "The Classical Field

Theories," Encyclopedia of Physics, Vol. III/I, Springer-
Verlag, Berlin, 1960.

15. P. J. Blatz and W. L. Ko, Application of Finite Elastic

Theory to the Deformation of Rubbery Materials," Trans-
actions of the Society of Rheology, Vol. 6, 1962, pp.
223-251.

16. R. P. Gael and L. E. Malvern, "Biaxial Plastic Simple

Waves with Combined Kinematic and Isotropic Hardening,"
JAM, Vol. 37, 1970, pp. 1100-1106.

17. S. W. Key, J. H. Biffle, and R. D. Krieg, "A Study of

the Computational and Theoretical Differences of Two
Finite Strain Elastic-Plastic Constitutive Models," in
Formulas and Computational Algorithms in Finite Element
Analysis, edited by Bathe, Oden, and Wunderlich, MIT,
Cambridge (1977).

18. R. K. Penny and D. L. Marrlott, Design for Creep, McGraw-

Hill, London, (1971).

19. J. H. Gittus, Creep, Viscoelasticity and Creep Fracture

in Solids, John Wiley and Sons, New York (1975).

20. T. J. R. Hughes and J. Winget, "Finite Rotation Effects

in Numerical Integration of Rate Constitutive Equations
Arising ln Large-Deformation Analysls," Int. J. Num.
Meths. Engr., in press.

21. J. o. Hallquist, "NIKE2D: An Implicit, Finite Deforma-

tion, Finlte-Element Code for Analyzing the Static and
DynaIl'ic Response of Two-Dimensional Solids," Lawrence
Livermore National Laboratory Report UCRL-52678, Univer-
sity of California, Livermore, March 1979.

22. J. H. Biffle, "JAC: A Two Dlmensional Finite Element

Computer Program Which Uses a Nonlinear Conjugate Gradl-
ent Technique for the Solution of the Nonlinear,
Quasistatic Response of Solids," Sandia National Labora-
tories, Report ln preparation, Albuquerque, New Mexico.

23. R. D. Krieg and S. W. Key, "Implementation of a Time

Indepennent Plastlcity Theory into Structural Computer
Programs," in: J. A. Stricklin and K. J. Saczalski
(eds:), Constitutive E uations in Viscoplasticity:
Computational and Engineerlng Aspects, AMD Va. 0,
(The Amerlcan Society of Mechanlcal Engineers, New York,
1976), pp. 125-138.
620

24. R. D. Krieg, "An Efficient Numerical Method for Time

Independent Plasticity," SAND77-0943, Sandia Laboratories,
Albuquerque, New Mexico (November 1977).

25. R. D. Krieg and D. B. Krieg, "Accuracies of Numerical

Solution Methods for the Elastic-Perfectly Plastic Model,"
J. Press. Vess. Tech. 99 (November 1977) pp. 510-515.
26. J. H. Arqyris, L. E. Vaz and K. J. William, "Improved
Solution Methods for Inelastic Rate Problems," Compo
Meth's App. Mech's. and Engr. 16 (1978) pp. 231-277.

27. B. M. Irons, "Engineering Applications of Numerical

Integration in Stiffness Methods," AIM Journal, Vol. 4,
No. 11, November 1968.

28. J. C. Nagtegaal, D. M. Parks, and J. R. Rice, "On

Numerically Accurate Finite Element Solutions in the
Fully Plastic Range," Computer Methods in Applied
Mechanics and Engineering, Vol. 4 (1974) pp. 153-177.

29. T. J. R. Hughes, "Equivalence of Finite Elements for

Nearly-Incompressible Elasticity," LBL-5237, Lawrence
Berkeley Lahoratory, University of California, Berkeley,
1976.

30. T. H. H. Pian and S. W. Lee, "Notes on Finite Elements

for Nearly-Incompressible Materials," AIM Journal, Vol. 14,
No.6, June 1976, pp. 824-826.

31. S. W. Key, "A F1nite Element Procedure for the Large

Deformation Dynamic Response of Axisymmetric Solids,"
Computer Methods in Applied Mechanics and Engineering,
4 (1974) pp. 195-218.

32. W. Prager and P. G. Hodge, Theory of Perfectly Plastic

Solids, John Wiley and Sons, New York (1963).
On the Efficient Solution of Nonlinear Finite Element Systems

H. D MITTELMANN
Ulllversltat Dortmund, Germany

Summary
We cons1der var1ational problems resp. var1at1onal in-
equal1tles which correspond to quasilinear elliptic boun-
dary value problems without resp. with llnear constraints.
Finite element discret1zat1on leads to nonlinear algebraic
systems resp. llnearly constrained nonlinear programm1ng
problems. Eff1c1ent algor1thms are given for the solution
of these problems. Convergence 1S shown and some numer1cal
results are presented. Applications are for the unrestric-
ted case all p,-oblems Wh1Ch are glven by a m1n1mizat1on
problem and for problems w1th constraints e. g. elast1c-
plast1c torsion or elast1c contact problems.

1. Introduct1on
In the following we will consider optimization problems of
the form

f(x) = m1n
S (1. 1)
M
S = {x e.JR , g(x) G'l'x + g < O}
o -

where f : JRM -+ JR .
1S a str1ct 1 y convex not necessar1'1 y

quadratic functional and G £ JRM, P, g £ JRP. We assume that

o
(1.1) 1S obtalned by d1scretiz1ng an infinite-d1mensional
var1ational problem of the form

jh(x, u, vu, ... ,Vmu)dx = m1n, m > (1 • 2)

R K

where h 1S not necessar1ly quadratic in vmu, m > 1 and K

622

is a closed convex subset of a sUltable function space V

of vector functions u=(u 1 ' ... ,u£) defined on the domain
R <; lR n .

It lS therefore natural that (1.1) is not a general linear-

ly constralned nonllnear optimization problem but it has
certain special properties. The Hessian matrix FE: lRM,M of
f e. g. is a sparse matrlx when (1.2) is discretlzed by
finlte-difference or flnite element methods. Furthermore
G is usually a sparse matrlx, too. In typical appllcations
from structure mechanlcs as e. g. the contact problem
every column of G has only a few nonzero elements. If the
obstacle or the contact problem for membranes is consldered
p has the same order of magnitude as M but for contact pro-
blems of elastic bodles (cf. e. g. 1 7 1 and the papers Cl ted
there) p is considerably smaller than M. Further occurren-
ces of (1.2) in applications are well-known and we need
not give a list of them here. It is also not necessary
to explain that (1.2) corresponds to a nonlinear variatlo-
nal inequality respectlvely to a free boundary value pro-
blem for ltS Euler equatlon.

Problem (1.1) (resp. (1.2)) lS often solved by Uzawa's method

(cf. e. g. 4 I). We recall that the dual problem of (1.1) lS

~ (A *) max ~ ( A) , AEO U (A ) *' £ lRP

A>O
~ (A) *'
£A (x )
min £A (x) * (1 .3)
XGU(x )
T
£A (x) = f (x) + A g(x)

where £A (x) is the Lagrange function of (1.1) and ~(A) the

dual functlon. Uzawa's method now applies the gradient pro-
Jectlon method to (1.3) yieldlng the algorlthm:

Let Ak ~ 0, k > 0, be glven. Then compute ~(Ak)' e. g.

by solvlng
T
'V £, (x)
x Ak
'Vf (x) + Ak'Vg (x)
° (1 .4)
 623

if this 1S also a suffic1ent condit1on for the m1nim1zation

of ~A (x). Denote the solution by x k . Then improve Ak by

(1 .5)

which 1S not d1ff1cult to compute Slnce VA~(Ak) = g(x k ).

For the case of quadrat1c f convergence of the sequence
{xk } can be shown for values of p in a range

p <

(cL I 4 I, p. 94) with suitable constants a o ' a 1 •

In every step of this algorithm an unconstrained nonlinear

optimization problem 1n JRM respect1vely a nonlinear system
of order M has to be solved.

Recently it was suggested in lui to use the Augmented La-

grang1an method or1ginally developed independently by
Hestenes and Powell to problems of the form (1.1) with
convex g. This algor1thm comb1nes the advantage of reducing
the problem via duality to a problem with positivity con-
straints with the 1dea of penalization. Under sU1table as-
sumptlons, e. g. the penalty parameter r > 0 need not tend
to 1nf1nity, Rockafellar proves convergence of this algo-
rlthm Wh1Ch becomes arb1trarily fast for increasing r. In
every step, however, again an unconstra1ned problem has
to be solved.

Slnce numer1cal experience has shown (cLI41, p. 113) that

for many problems of the form (1.1) but with

M
S = {x E JR , c < x < d} (1 .6)

a projected relaxat10n method is more efficient than

Uzawa's method and beyond that solves the restricted pro-
blem faster than a similar unrestricted problem (cf. I 4 I) ,
624

we shall concentrate on relaxation. In its simplest form

this is the algorithm:

(k) M (k+1)
Let x e. S = {x E JR , x .::. o} be given. Compute x accor-
ding to

_ (k+1) (k+1) (k) (k)

where 'Vf (f.),x. k - (x 1 , .•. ,x.~- 1 ' x.~ , •.. ,xM ),
~ ~,

O.~, k are suitable positive numbers and w.~, k relaxation pa-

rameters.

The proJected point-SOR-N method ~s obta~ned for the choice

O.~, k = f ~~
.. (x.~, k) (1 .8)

where F = (f~k) is the Hess~an matrix of f. The convergence

properties of (1.7), (1.8) are well known for quadratic f
(cf. e. g. 14 I). In the nonquadratic case global conver-
gence was proved in 1 9 1 and for a modified relaxation me-
thod ~n 1101. In computations such methods usually quite
rapidly approx~mate the proper subspace of act~ve con-
stra~nts correspond~ng to the (approximat~on of the) unknown
contactregion(cf. e. g. 141, p. 135) while most of the
time is spent in finding good approximations for the remai-
ning var~ables.

Thus it is desirable to have an algorithm which is faster

than relaxation for unrestricted problems and to use that
after a nearly proper subspace has been found. Since one
cannot determine when th~s is the case that algorithm has
to be globally convergent for the restricted case, too.
A two-phase algorithm for (1.1), (1.6) in this spirit was
given in 1111. In the following we shall consider this and
a completely d~fferent two-phase algor~thm.
 625

2. A preconditloned conjugate gradient method

In this section we define an algorithm for the approximate
solution of

f(x) min
S (2.1)
M T
S {xEJR , G x + go > O}

h
were G = ( EM,-E ) EJRM, 2M , EM t h e MxM 1
d entlty matrlx
' and
M
T T T i
go = (-c ,d ) . We denote the columns of G by g ,i=1, ... ,2M,
T
and deflne I(x) {lE{1, ... ,2M}, glx + goi = O}.

Let G = (gi) iEI' Q = E - G GT and for x = x(k) let

I I M I I
I (x (k)) Gk = G1 and Qk analogously. QI lS the projec-
Ik =
I

k
tor on the subspace of the actlve constraints. Finally let

r =
k

Next we assume as ln [10] that the grad lent of f may be

written as

vf(x) = A(x)x-b(x). (2.2)

If (2.1) is obtained by the dlscretizatlon of a problem cor-

respondlng to a quasilinear elllptic equation of second or-
der usually a canonlcal though not unique decompositlon (2.2)
exists. The matrix A lS glven E'xplicltly for two problems
in [10] and compared with the Hessian matrix of f. Usually
A has slmilar propertles as symmetry and structure but is
easier to compute.
Now we glve an algorlthmlc deflnltlon of the projected cg-
SMOR-N method and shall comment on in afterwards.

The projected cg-SMOR-N algorithm

Let x(o)ES and P_1EJRM be arbitrary. Set j k o and

ilo = O.
626

Step 2 I f <IIQkrkll < IrUI and llk

then set Ik = I k - {~}and Ak = 1.

Step 3 Compute zk as the directlon vector taken by one

step of the SSOR-algorlthm for the linear systew
A<x(k)~ = b(x(k)) startlng from x(k) and keeping the com-
ponents corresponding to Ik fixed.

S = {
O, if j =
T T
° or Ak = 1 or llk
k zkrk/zk_1rk_1 otherwise.

Step 4 Determine ~k as the maximal admissible step length

~ T T (k) ~
in direction Pk· Compute ak=zkrk/PkA(x )Pk· If a k ~ a k
then set a k = uk' llk+1 = 1 while a k = ~k' llk+1 = otherwEe. °
Step 5 Set x(k+1) = x(k)+ akTkPk' for the cholce of Tk see
below.
o, if j = m, m °
t
>

j 1, i f Ak = 1
j + 1 otherwise,

k k + 1 and go to Step 1.

This algorithm is in the usual sense a primal feasible

direction method and in generallzation of Rosen's gradient
projection method the direction vectors are the projection
of vectors obtained by a back and forth-sweep of the SOR-
method for a certain linear system and which are modified
by a cg-like rule with a restart after at most m steps.
In order to assure global convergence zig-zagging
has to be avoided by the strategy used in Step 2.
 627

Although the above algorithm seems to be rather compllcated

it in fact needs more storage than projected relaxation
methods but less operatlons (cf.[10]). The essential feamres
are the rather weak dependence of the convergence rate on
the value of the relaxation parameters used ln Step 3 and
the fact that the dependence on m is also not very strong
with optimal values between m = and m = 5 for many pro-
blems.

A thorough study of the choice of the direction taken ln

each step shows that thlS is always a feasible one. Espe-
cially it is not necessary to insert pure gradient-projec-
tion steps. Global convergence is proved in [11] by first
showing that the vectors zk are uniformly related to the
projected gradient Qkrk and that P~zk ~ r~zk for Tk ~ 1/Y k
where at least bounds for Yk may be given. The essential
fact is that zk and Qkrk are related through

(2.3)

where Sk is a positive definite matrix which in this case

is also symmetric. Then the descent properties are stated
and the proof is completed by showing that zig-zagging
cannot occur.

Slnce in one step of this algorithm only one constralnt

may be inactlvated and usually only a few may become active
there is a certain need for a method which relatively fast
approximates the proper subspace and thus yields an acce-
leration at least in the case when many constralnts have
to be activated or inactivated during the computation. For
this purpose a modified relaxation method namely MOR-N:

(k) f. (x. k)
(k+1 ) ( . (d 1 1, ) )
xi max c i ' mln i'x i - wi,k a (x )
ii i,k
(2.4)
i = 1, ••• ,M, a ii diagonal element of A in (2.2)
628

was successfully applied in [11]. For a proper choice of

relaxation parameters w. k this algorithm itself was shown
1,
to be globally convergent.

The resulting two-phase algorithm solved different nonli-

near model problems several times faster than projected
relaxation methods alone. Hence, at least for the case
of bound-constraints we obtained an efficient method. Gene-
ralizations to general linear constraints might be possible,
whereas convex nonlinear constraints are more diff1cult to
treat with feasible direction algorithms. Here, however,
the augmented Lagrang1an method may be applied advantageous-
ly.

3. Numerical Results

The cg-SMOR-N and the MOR-N algorithm as described in 2. but

without projection were applied in [1o]to several unrestric-
ted problems and compared w1th other related methods as e.g.
the classical SOR-N method or the cg-SSOR-N algorithm as
suggested 1n [ 1]. A cons1derable improvement was obtained
concerning the total amount of work and its dependence on
the choice of the parameters involved.

Related bound constrained problems namely the minimal sur-

face over an obstacle, i. e.

h (u, V'u) (1 + IV'u I 2) 1/ 2 (3 • 1)

in (1.2) were solved in [11] with the two-phase algorithm

of the last section. The second phase was started when the
number of active constraints had the same positive value
for two consecutive steps. In order to illustrate the be-
havior of the method we graphically present typical results
for discretization by linear fin1te elements.
 629

The domain R = {(x,y), x 2+ y2 < 2} was triangulated resul-

ting in a triangulation with 475 interior nodes and
g (x,y) '" 0,

r 2 2 1/2
a(l-r Ib)
.
,1f x = 0, -b ~ Y ~ b
¢ (x,y)
l 0 otherwise

with a = 1.255, b = 1.439 were chosen in (3.1). The star-

ting vector was taken as the pOints on a sphere of radius 2
with center at the orig1n. In Fig. 1 the work measured in
essential algebraic operations per variable is shown which
is needed 1n phase I I to reduce the residual II Qkrk II to
10 -6 for varying values of the (constant) relaxation para-
meter w. In Fig. 2 the decrease of the residual is shown
for the two-phase method in comparison to the projected
MOR-N algorithm with the "optimal" value w = 1.9 given by
Fig. 1. Here the amount of work is measured in the equiva-
lent of that necessary for a relaxat10n step. Finally in
Fjg. 3 the development of the number of active constraints,
it is zero for the start1ng vector, is shown.

Operat.
15000

10000

5000
cg-SMOR-N

1.4 1.6 1.8 2.0 w

Fig. Number of operations in phase I I for different

values of w
630

---
residual
0
10
MOR-N

10
-2 "'----
-...
"""" .......... .......
-4
10

phase I phase I I
-6
10

10 20 0
relaxation steps
Fig:. 2 Residual for relaxation and 2-phase method

act. constr. I
17

/
.,P" cg-SMOR-N

10 I
MOR-N I
"
5
I
10 20 30 40
relaxation s:eps
Fig:. 3 Number of active constraints

As an example from structure mechanics we consider the in-

dentation problem for the nonlinear membrane. Finite-diffe-
rence or finite element discretizations of membranes were
treated by many authors. We mention here only [12, 16].
In [16] a thin flat unstrained membrane covering the region
R in the (x,Y)-plane is considered which is loaded by a
force acting in the z-direction. If u, v, w denote displace-
ments in the x,y,z-direction, respectively, then under cer-
tain simplifying assumptions but with a nonlinear strain-
displacement relation the displacement components of the
deformed membrane satisfy a coupled system of three non-
linear differential equations. Under further assumptions
which are mainly u = v = 0 and after transformation to
nondimensional variables the last and most nonlinear of
 631

these reduces to

(3w 2 + w2 ) w + 4w w w + (w 2 + 3w 2 )w =-0 (3.2)

x y xx x y xy x y y

The boundary conditlon is w = 0 if the membrane is kept

fixed there.

(3.2) is the Euler equation of the functional (1.2) with

h(w, vw) (3.3)

and hence we have to solve the homogeneous Olrichlet pro-

blem for a pseudo-Laplace-operator which was also consi-
dered e. g. in [3, 1 4] .
The complete solution of the nonlinear membrane problem may
be found then as ln [16] using the obtained displacement w
together with u = v = 0 as starting solution for the system
in u, v, w. It is well-known that for values of p near 1
or considerably greater than 2 a direct treatment of (1.2),
(3.3) by classical methods is impossible in practice
(cf. [3 ]). The appllcation of the proposed method in the
case p = 4 yielded results comparable to those for the mi-
nimal surface problem. The main difference was that the
relaxatlon parameters and the damping factor 'k in Step 5
had to be taken smaller. This is caused by the fact
(ct. the inequallties for W k and Tk in [11]) that
T T l ,
y A(x)y ~ y F(x)y (the choice of the matrlx A lS rather
canonlcal) for the membrane problem while the opposite in-
equality is valld for the first example (cf.[10]).

In Fig. 4 the z-displacement along the y-axis of a nonline-

ar membrane over the unit circle is shown which is pressed
by the uniform load 0=2,6,20,60 agalnst the obstacle
2 2 2
.. ( ) = { 1 for x + (y-o,25) < 0,25 ,
~ x,y 1,50therwlse,

The unit circle is subdivided by a triangulation with

alTerage edgelengthh = 1/12 and 475 interior nodes.
632

Fig. 4: z-displacement of nonlinear membrane

4. A multlgrid method for variational inequalities

One of the methods for the approximate solution of boundary
value problems which at least under certain conditions may be
quite efflcient is the so-called multl-grid method. We shal
explain it for an abstract llnear boundary value problem
glven by

Au = f. (4. 1 )
For a sequence of stepsizes {h,q,} with

h
0
> h1 > h2 > ... > h,q, > . .. > 0

(4. 1 ) is to be discretized by a finlte-difference or a fi-

nite element method yielding the finite-dimensional prob~s

(4.2)

i. e. u,q,' f,q, are elements of a finite-dimensional vector

 633

space V£' £ = 0,1, ... and A£ : V£ ~ V

£
is a linear opera-
-1
tor. ~£ = A£ lS assumed to exist and ~£ is an approxi-

mation of <lit Furthermore to connect the grid functlons on

V£' £ = 0,1, ... we need a linear and injective prolongation
operator
P£,£-1: V£_1~ V£' £ = 1,2, ...

and a Jinear and surjective restriction operator

1,2, ..•

Finally a smoothlng procedure

(4.3)

will be used, where the linear operators G£, H£ satisfy

-1
A£ = H£ - G£, I the identity operator. Then in the follo-
wlng quasi-Algol procedure we define a llnear MG-algorithm
as In [6].

prodedure MGM(£,u,f); integer £; array u,f;

comment £ actual level number,
lnput uti) outpur. u(l+1)
u £' _. £ '

f right-hand side of eq. (4.1) i

If £ = 0 then u := ~
o
*f 0else
--
begin integer ji array v,d;
for j := 1 step 1 until v do u := ~£(u,f);

d := r£_1,£* (A£ u-f); v := 0;

MGM ( £ -1 , v , d , ) ;
>I'
U := u - P£,£-1 v
end MGM;

Thls procedure performs one MG-step starting from an appro-

ximatlon U~l) to u£ using the levels £, £-1, ... ,0 and on
each of these except the last one at first v smoothing
steps are executed. ~ may, however, ltself be an applica-
o
tion of a smoothing procedure. This method may be genera-
lized to nonllnear problems (cf.[6]).
634

In order to apply such methods to problems of the form (1.1),

(1.6) one could e. g. use this nonlinear MG-procedure where
~o is the application of v steps ~r according to relaxation
with project1on and also~~ in the smoothing steps is sub-
stituted by~~r • We call this method MG1-algorithm. This
method was applied to problems similar to those considered
in 3. and at least in the computations it did not converge
globally ([5]). A global convergent MG-method may however
be obtained if in the algorithm defined in 2. §tep 3 is
changed and the vector Pk is determined as the direction
taken by one call of the procedure MGM for the linear pro-
blem A (x (k))x = b (x (k) ) starting from x (k) and keeping the
components corresponding to Ik fixed.

Global convergence may then be proved along the lines of

the proof in [11] if Pk and Qkrk are related as in (2.3)
by a matrix S (suppressing the index k) with similar pro-
perties, essentially positive-definiteness. For smoothing
by relaxation G~, H~ in (4.3) are given by

(4.4)
T
if A~ D~ + L~ + L~ . Then a simple computation shows that

for v=2 we have on the level ~

1,2, ••. (4.5)

where S
o

The matrix S~ is not symmetric. The equally unsymmetric ma-

trix H~, however, is pos1tive-definite for 0 < w < 2 if
A~ has this property. Hence for sufficiently small w B~ 1S
positive-definite and hence S~ if the second term on the
right of (4.5) is multiplied by a damping factor.

If this last method is called MG2-algor1thm then we propose

as a second two-phase method that obtained by the applica-
tion of the MG 1-algorithm to approximate in phase I the
 635

proper subspace and of the MG2-algorithm as the second phase.

In computatlons performed by W. Hackbusch for problems of
the type (1.2), (3.1) dlscretized by finite-difference me-
thods on a quadratic mesh, R being the unlt square, this
method also brought a cost-reductlon compared to simple
projected relaxatlon which, however, was comparable to
though not greater than that for the projected cg-SMOR-N
method of 2.

5. Conclusions

We have seen that two-phase methods as proposed in 2. and

4. are efficient algorithms for the solution of variatio-
nal inequallties, when the feasible set has the form (1.6).
Further numerical experience, however, is needed in order
to compare lt wlth other algorithms which could be applied
to these problems. We mention here especially quasi-Newton
methods for problems with a sparse Hessian (cf. e. g.[15,
17] ). These methods for unconstrained minimization pro-
blems could be generalized to the constralned case. In the
convergence proofs for these methods, however, it is as-
sumed that in every step a linear system is solved exactly
and one cannot see an easy way how to get rid of thlS as-
sumption.

References

Concus, P., Golub, G. H. and D. P. O'Leary,

Numerical Solutlon of nonllnear elliptic partial dif-
ferential equations by a generalized conjugate gra-
dient method, Computing 19 (1970), 321-339

2 Cottle, R. W., Giannessi, F. and J.-L. Lions (eds.),

Variatlonal Inequalities and Complementarity Problems,
John Wiley, Chichester 1980

3 Glowlnski, R. and A. Marrocco, Sur l'approximation par

elements finis d'ordre un, et la reolution par penali-
satlon-dualite, d'une classe de problemes de Dirichlet
nonlinealres, Rev. Francaise Automat. Informat. Re-
cherche Operationnelle ser. Rouge Anal. Num. R-2 (1975),
41-76
636

4 Glowinski, R., J. L. Lions and R. Tremolieres,

Analyse Numerique Des Inequations Variationelles,
Dunod, Paris 1976

5 Hackbusch, W., private communication

6 Hackbusch, W., On the convergence of multi-grld ite-

ratlons, Report 79-4, Universitat zu Koln, to appear
In Beitrage Num. Math. 9

7 Klkuchi, N. and J. T. Oden, Contact Problems in Elasti-

city, TICOM Report 79-8, The University of Texas at
Austin (1979)

8 O'Leary, D. P., Conjugate gradient algorithms in the

solution of optimization problems for nonlinear ellip-
tlC partial dlfferential equations, Computing 22(1979',
59-77

9 Mlttelmann, H. D., On the approximate solution of non-

linear variational inequalities, Num. Math. 29 (1978),
451 - 462

10 Mittelmann, H. D., On the efficient solutlon of nonli-

near flnite element equations I, to appear in Num. Mafu

11 Mittelmann, H. D., On the efficient solution of nonli-

near finite element equations II, submitted to Num.Mafu.

12 Oden, J. T. and T. Sato, Finite strains and displace-

ments of elastlc membranes by the flnlte element method,
Int. J. Solids Structures 3 (1967), 471 - 488

13 Rockafellar, R. T., Lagrange multipliers and variatio-

nal inequalities, in [2]

14 Scarplni, F., A numerlcal approach to the solution of

some variational inequalities, in [2]

15 Shanno, D. F., On variable-metric methods for sparse

Hessians, Math. Compo 34 (1980), 499 - 514

16 Shaw, F. S. and N. Perrone, A numerical solution for

the nonllnear deflection of membranes, J. Appl. Mech.21
(1954), 117 - 128

17 Toint, Ph., On the superlinear convergence of an algo-

rithm for solving a sparse minimizatlon problem. SIAM
J. Num. Anal. 16 (1979), 1036 - 1045
The Numerical Calculation of the Contact Problem in the
Theory of Elasticity

J. J. KALKER, H. J. C. ALLAER T, J.DEMUL

Delft University of Technology, Delft, SKF-ERC, Nieuwegem,
The Netherlands The Netherlands

Summary
In the present paper, distlnction lS made between frictional
and frictionless contact problems. After their formulation,
the epitome of frlctlonal problems, V1Z. the problem of
frictional roillng contact is considered, and the two most
recent and successful methods for deallng wlth lt are pre-
sented. Then the frictlonless problems lS considered. Two
palrs of methods are presented, each palr conslsting of a
non-varlatlonal method, wlth ltS variational counterpart
authored by us. The flrst palr, V1Z. Paul & Hashemi I Kalker
& Allaert, deal wlth the general half-space frictlonless
contact problem. The second palr, V1Z. Reusner I Kalker & De
Mul, deal with the contact of bodles of revolutlon, WhlCh
may be approximated by half-spaces. The codes implementlng
the frictionless problems suffer from long operatlng tlmes.
Research lS stlll ln progress to speed up the codes.

1. Introductlon

Conslder two deformable bodles, va, a = 1, 2, wlth surfaces

ava, see fig. la. They are pressed together wlth a force N,
so that they are brought into contact. In an undeformed
state they would lntersect, see flg. lb. The dlstance h of
the bodles at a pOlnt of thelr surfaces lS measured along
the common normal on the bodles; lt lS counted negatlve
where the bodies intersect, posltive where they do not. In
order to counteract the penetration, a dlsplacement u a (In
body a), a = 1, 2, comes into belng. The contact condltion
lS then that the deformed dlstance D lS non-negatlve:

D d~f h - u 1 - u 2 __> 0 ; 0 : contact

n n

u~ dlsplacement component ln the dlrection of n a ,

outer norffial on av a (I )
638

In addition we will normally have that the normal component

of the surface traction in the contact patch is compressive:

Pn ~ 0 in contact o outside it (2 )

Outside contact, the displacement ~s prescribed as a in u

Aa c aVa, and on the remainder of the surface, the surface
u
load pa is prescribed as pa

a a
p u (3 )

In the contact region K = {xl x E aVa, D = O}, fr~ctional

forces act. They are commonly modelled in three possible
ways:
1) Fr~ction is neglected: the shear traction in the contact
area is set zero. This is su~table, e.g. in lubr~cated

contacts in which hydrodynamic effects are small; see

also sec. 4.
2) Friction is infinite: the relative tangential velocity
of the bodies in the contact patch (the slip) is set
zero. This ~s suitable, e.g. when the friction forces
are relatively small while the coeffic~ent of friction
(see below) is large.
3) Friction is finite and acts according to Coulomb's law,
which reads:

The tangent~al traction IpTI < - fPn ' f: coeffic~ent

of fr~ction. If the slip w f 0 ~ w =- APT '

A > 0 , and IpTI =- fPn

where the tangential component PT d~f P - Pn n (4 )

2. Variational formulations in a special case

The above problems may be formulated as variational

problems [2, 3, 19J.
In the case of one of the bodies being rigid, and the other
an elastic half-space z > 0, the variational formulation of
problem 1) reads:
 639

min!
u,p together
II z=o
Pn(X'Y){~Un(X'Y) - h(x,y)}dxdy

sub Pn(x,y) < 0

sub PT(x,y) 0

u + 0 as (x,y,z) + 00 (5)

By minimizing over "u,p together" we mean that u and p

correspond to the same elastostatic field. For instance, in
the case of problem 1),

u
n
(x, y) - 1 - cr
- 21i'G
II z=o
Pn (x' , y' )
I
dx'dy'
2
-
2
(x-x') + (y-y')
(x,y) : Cartes~an coordinates on the surface of the

half space (6)

In the case of problems 2) and 3), an incremental argument

must be pursued.
Let T be an ~ncrement of time. Thus, problem 2) becomes,
when t is a generic instant:

min!
u,p together
II z=O
Pn(x,y,t){~un(x,y,t)-

- h(x,y,t)}dxdy + II K(t-T)
PT(x,y,t){~UT(x,y,t) -

- UT(X,y,t-T) - p(x,y,t;T)}dxdy sub Pn ~ 0 (7)

where p(X,y,t;T) is the shift, dur~ng the t~me-~nterval T,

of the r~gid body w.r.t. the half-space at (x,y) at the
time t. Note that the contact area K(t - T), and the elastic
field, are known at the time (t - T). These quantities at
t~me tare proveded by the m~nim~zat~on step (7).
In the case of problem 3), the traction bound g d~f - fp n
must be known beforehand. Aga~n we must work incrementally,
and the min~mum pr~nciple becomes:

min!
u,p together
II z=O
Pn(x,y,t){~un(x,y,t)-

- h(x,y,t)}dxdy + fI PT(x,y,t){~uT(x,y,t)-
J z=O
640

- UT(X,y,t-T) - p(x,y,t;T)}dxdy
2 2
sub Pn ~ 0 , IpTI ~ g (g = - fPn ' known) (8)

We prove the validity of these principles in a heuristic

manner; rigorous proofs may be found in 2, 3 .
P r o o f : The contact conditions (1), (2), (4) w1ll be
shown to be the Kuhn-Tucker conditions of the above vari-
at10nal principles. To that end we observe:

PS(x,y,t)g(x,y,t)dxdy =

S = 1, 2; g: known function; uS' Ps belong to the same elas-

tostatic field.
Let D ~ 0, DPn =
0 be the optimal Lagrange multiplier of
Pn ~ O. Let A ~ 0, A(g2 - Ip T I 2 ) = 0 be the optimal Lagrange
multiplier of g2 - Ip T I 2 ~ O.
Then:

o= II z=O
OPn{un(t) - h(t) + D}dxdy + II z=O
OPT{uT(t)-

- uT(t - T) - p + 2APT}dxdy ; VOPn ' OPT

~ 1. h(t) - un(t) D > 0 , deformed distance;
if Pn < 0 ~ D 0 ~> contact,
if Pn 0 ~ D > 0 ~ non-penetration.
~ 2. uT(t) - uT(t - T) - p = - 2APT (= sh1ft w.r.t. rigid
body = slip);
if 1PT 1 < g ~ A 0 slip = 0,
1f 1PT 1 = g 9 A > 0 , slip = - 2APT' see (4).
This completes the proof for problem 3) .
As to problem 2), it suffices to observe that the "infinite
friction" means that g = 0 outside contact, so that A is,
in fact, free and hence slip = {uT(t) - uT(t - T) - p} is
unrestricted, while 1nside contact, IpTI < g, so that slip
= {uT(t) - uT(t - T) - p} = O.
As to problem 1), the extra condition PT 0 induces the
freedom of the slip = {uT(t) - uT(t - T) - p}, and only the
contact format10n remains. QED
 641

3. The rolling contact problem

As an example of a frict10nal problem, we consider the

problem of roll1ng contact [4J, V1Z. "Two bod1es rollover
each other wh1le friction is present in the interface. Their
r1gid body speeds are not prec1sely matched, so that a rig1d
slip occurs in the interface, Wh1Ch calls up an elast1c
field. It 1S requ1red to f1nd the tangential traction PT 1n
the 1nterface, and in part1cular 1tS resultant."
This problem 1S of importance 1n railway slmulat1ons, where
the problem needs to be solved very many times, so that
speed of caluclat10n is of the essence.
In the railway application, the contact area and the normal
pressure (- Pn) carr1ed by it are known, they follow from
the theory of Hertz (1881; see e.g. [lJ).

For the sake of solv1ng the equat10ns of elasticity, we

approx1mate the bodies by the half-spaces {z ~ O}, so that
the relations of Bouss1nesq & Cerrutti (see [lJ) are val1d.
In a symbol1c form, they read:

u. (x,y) =
1
ff K
p (x' ,y' )H .. (x-x' , y-y' )dx'dy',
J 1J

where, e.g. H33 (x,y) = 12~G0 / x 2 + y2 -1

summation convention 1S observed (9)

It will be clear, that an 1mplementat1on [5J of the rolling

contact problem on th1S bas1s will be very slow. In fact,
on an IBM 370/158, calculat10n t1mes of around 10 s/case
seem the best that can be achieved. In th1S Solut1on method,
the tangent1al traction is d1scretized as constant over each
mesh of a rectangular net in the contact area. The elastic
energy due to the tangential tract10n becomes then:

E u (x,y)p. (x,y)dxdy
1 1

N
~ I I
I,J=l 1,j=1,2
M2I + 1 ,2J+]' P2I+i P2J+]'
642

i,j = 1,2 : indicate components. I,J = 1, ... N

indicate meshes, N in total. P2I+i : value of the

rectangular traction component Pi (x,y) at the center

of the mesh I. (10)

The variat~onal pr~nciple (8) becomes:

min!
{Pi}
(Y L
I,J=l i,j=1,2
(~M2I+~,2J+J P2I+i P2J+j) +

Y
I=l i=1,2
L (u 2I + i (t - T) + P2I+i(t;T))P2I+~ QI)

2 2 2
sub P2I+i + P2I+2 gI ' I = 1, ... ,N ; QI: area of

mesh I ; u 2I + i (t - T): component of displacement at

center of mesh I at t~me (t - T). P2I+i: analogous

(11)

Here the contribution of the (known) normal traction dis-

tribution has been omitted. The problem is solved by a
mathemat~cal programming algorithm based on Newton's method.
Solution of problem (11) is the traction distribution at
t~me t when that at time (t - T) ~s known, which latter
enters the problem through u 2I + i (t - T).
In order to f~nd the steady state solution, ~n which the
traction d~stribut~on over the contact is independent of
time, several elementary time intervals must be traversed.
After about 1 - 1.5 contact w~dths have been traversed by
rolling, convergence is normally achieved.
From the description it is readily understood that this is a
slow affair, even though:
1. only the total force is required, which reduces N;
2. the matrix (M 2I + i ,2J+j) need be calculated only once;
3. the problem is convex;
4. in Newton's method modified for bounds we possess a fast
optimization routine;
5. generally, only one optimization iteration need be per-
formed per time step.
At present, the bottle neck of th~s method are the "curved"
 643

2 2 < 2
restrictions P2I+l + P2I+2 gI· In the present program, p
is represented by the polar representation (P2I+l,P2I+2)
= (r I cos ¢I ' r I sin ¢I) when r I = gI' so that the restric-
tion reduces to the bound r I ~ gI.
The rectangular representat10n 1S mainta1ned when r I < gI'
because the polar representation becomes slngular when
r I = 0, and this introduces spur10us Kuhn-Tucker points. As
a consequence, we shift back and forth between represent-
ations, whereby the adaptation of the gradient and Hessian
is most consuming.

It w1II therefore not cause surprise, that approximate

solution methods have been looked for. At present, the most
successful and general of these 1S the so-called "simple
theory" [4, 6, 7, 8, 9, 10J. It 1S based on the replacement
of (9) by:

(12)

where L is constant to be determined. How this is done, is

shown in [4, 6J. It may be shown [4J that the steady state
rolling contact cond1tions are:

ap'l'(x,y)
w(x,y) = slip p(x,y) - L , x: rolling
ax
direction (a)

I PT I ~ g ; w t 0 =l> W =- APT ' A > 0 , I PT I = g (13)

The follow1ng Solut1on method is descr1bed 1n [10J. Let T be

an increment of x, T > 0; working from high values of x
towards low, we rewrite (13a):
WT = PT - LPT(x + T;Y) + LPT(x,y).
Call LPH =- PT + LPT(x + T;Y), and set:
PT = PH if IpHI ~ g(x,y) , PT = PHg/lpHI if IpHI > g(x,y).
Indeed, IpTI ~ g, and
WT = PT - LPT(X + TiY) + LPT(X,y) =
=- LP H + LP H = 0 if IpTI < g , and
M4

IpH I
L(l - -g---)PT = - APT
IpH I
A (-g---) - l)L ~ 0 if IpT I g
which is Coulomb's law.
The algorithm must be provided with a boundary condition,
for wh~ch reference is made to [10J. It is extremely fast, a
reduction factor of 100 in speed of calculat~on is achieved
with respect to the algorithm of sec. 3.1, with errors of an
estimated maximum of 15% of the maximum value of the total
tangential force. Such high errors are acceptable in rail
vehicle dynamics, where owing to pollution of wheels and
ra~ls, unpredictable deviations of up to 40% from the theory
have been observed.

4. The frictionless contact problem

As was observed in sec. 1, the frictionless contact problem

obtains when the contact ~s lubricated. It also occurs when
the contacting bodies may be, for the sake of elasticity,
considered as half-spaces with the same elastic constants.
Under these circumstances, the problem can be decomposed
[llJ into a frictionless problem, and a frictional problem
of the type considered ~n sec. 3. This decompos~tion is of
great importance technically; it obtains e.g. in the contact
between wheel and rail.
The first contact problem to be solved was a frictionless
one (Hertz, 1881, see [lJ). In it, the undeformed distance h
(see fig. 1 and eq. (1», is a quadratic function of x and
y, Cartesian coordinates on the surface of the half-space.
Many problems may be solved with the Hertz theory; we refer
to the previous section for an example.
Notable exceptions occur in the railway application when the
flange of the wheel comes into contact with the rail, and in
bearings, where under high loads the rounded ends of the
rollers come ~nto contact with the rings. In both cases, the
speed of operation of the algorithms is of the essence, in
the case of railway application because of the railway
vehicle simulation, in the case of the bearing because there
also motion simulations are desired.
 645

Unlike the rolling contact problem, no "simple" truly fast

theory has been proposed apart from the Hertz theory.
We will describe four methods of solution ~n some detail.
Two of them, viz. Paul & Hashem~ ([12J, sec. 4.1) and
Reusner ([13J, sec. 4.3) are so-called class~cal methods.
Based upon the Bouss~nesq & Cerrutti integral (6), a number
N of displacement quantit~es are expressed ~n N independent
variables, such as the pressures (- Pn(x,y)) and the contact
width, under the assumption that contact ~s established
(D = h - un 0), wh~ch fixes the displacement quantities.
This yields N (poss~bly non linear) equations for the in-
dependent variables, which are solved. Then it ~s verified
whether the non-penetration condit~on D ~ 0 and the compres-
siveness condition Pn ~ 0 are everywhere satisfied. If they
are, the solution has been found; ~f they are not, the
contact area ~s changed, and a new iteration is performed.
The other two methods, viz. Kalker & Allaert and Kalker &
De Mul, are variational methods that are described here for
the first time. They are not the f~rst var~ational methods
to appear in the literature: pr~ority lies with the method
of Fridman & Chern~na [14J, dating from 1967.
Independently, Conry & Seireg [15J, publ~shed a sl~ghtly

deviant variational method, and they were the first to

introduce the not~on of design into contact theory.
The theory was further developed by Paczelt (e.g. [16J), who
nowadays ~s interested in des~gn questions for contact
problems. Kalker [17J publ~shed a relatively fast method for
the calculation of contacts which are slender and fa~rly

sharply pointed. This work is based on line contact theory

[17, 18J; in the bearing appl~cat~on, the method can unfor-
tunately be used only when the rad~us of the rounding at the
ends is more than three times larger than the radius of the
roller. In pract~ce, these radii may be ten times smaller
than the radius of the roller, which renders the line
contact method useless for the bearing applicat~ons; instead
Reusner's method ([13J, sec. 4.3) ~s currently used.
646

The contact area is discreti~ed by rectangles, see fig. 2.

The displacement u I in the centroid of QI is expressed in
the P J by Boussinesq & Cerrutti, see (6):

(14)

In accordance with the contact condition u I = hI' the linear

equations (14) are solved for PJ (collocation method) .
The edge of the contact area 1S given by Pn = 0, wh1le the
behaviour of Pn near the edge of the contact is parabolic,
see fig. 2b. Based on these new {PJ}' the points A , B. on
1 1
the edge of the contact (see fig. 2b) are determ1ned so that
the behav10ur of fig. 2b is best approximated; between the
new Ai and Bi , the same number of elements are conceived as
before, and the equations (14) are again set up, etc. The
process stops when Ai and Bi have converged, which happens
after about four to ten iterations. Note that the coef-
fic1ents KIJ have to be recomputed for every iterat10n. As
th1S is a major job, it const1tutes the bottleneck for this
method.

The contact area is discret1zed by rectangles with sides

parallel to the (x,y) axes. First we w111 consider the case
that all rectangles are congruent. Then we calculate the K IJ
of (14) as follows:

a2F
Let f(x,y) be such that dxdy = f(x,y) ; then

f(x,y)dy = [[F(X,y))X 2 )Y2 (15)

x=xl y=Yl

-1
Now, if f(x,y) = I x 2 + y2 9 F(x,y) =
-1
= x sh- 1 (Y/ixi) + y Sh- 1 (x/iyi) ; sh z = inverse

hyperbolic sine In (z + / z 2 + 1) (16)

 647

so that

(( F( ) I -x J -1
Jxx=x JY I -Y J -k "close" (17a)
x,y -x +1 y=y -y +k
I J I J

When (xI - x J )2 + (YI - YJ)2 » 12 + k 2 we may simpllfy this

to:

4lk
"far"
/ (XI-X J ) 2 + (YI- YJ)2

2 2
if (xI - x J ) + (Y I - YJ )

e.g. 4 (17b)

The optimization problem becomes:

( 18)

This is a simple quadratic programming problem posed long

ago (1967) by Fridman & Chernina. As a result, the PJ
- representative of the pressure in the center of QJ - are
glven, and it is found which QJ belong to the contact and
which do not. In other terms, the boundary of the contact
area is given very roughly indeed. The need for improvement
of this feature led Paul & Hashemi to their method descrlbed
in sec. 4.1; in the results, the points Ai' Bl of fig. 2a
presumably lie on the contact boundary. We could do the same
thing, but we preferred to calculate the boundary of the
contact as follows.
In the contact problem with D(x,y) = h(x,y) - un(x,y) we
shift the bodies over a dlstance T w.r.t. each other in the
x-dlrection. The un and Pn shift a distance T too: In fact,
the entire contact field shifts over a dlstance T. It is then
ah I 'i' ap
easy to see that - - = {, K J from
aPJ ax I€contact,JEcontact IJ ax- ' -
which ax- follows.
M8

Similarly, !PJ may be found.

ay
These may be used to find a boundary point, see fig. 3; we
are of the opinion that this method is as accurate as Paul &
Hashemi's to find the boundary. We tend to opine that our
method is somwhat better than Paul & Hashemi's on the follow-
ing grounds:
1. The results are equally accurate.
2. The optimization idea is more general than the P.&H.
method, since any combination of rectangles may con-
stitute a contact area. This is not so with P.&H., since
their contact must be convex ~n one direction at least.
Very efficient methods are available for the quadratic
programming problem, which are equivalent to only a few
matrix inversions.
3. The matrix of influence numbers need be calculated only
once. When all rectangles are equal, the invar~ance of
the discretization w.r.t. the translation over a mesh
implies that the number of influence numbers to be cal-
culated is only O(N) rather than O(N 2 ), where N is the
number of elements. In comb~nat~on with the simplified
element (17b), this renders the calculat~on of the matr~x

truly fast.
Whereas ~n the railway application, congruent rectangles
seem appropriate, this is not so in the bearing application,
where stress concentrations occur at the transition to the
sharp rounding of the roller. The large gradients of the
stress necessitate a locally very fine mesh, so f~ne that
the number of equal elements would run into many hundreds or
even thousands, which would lead to a very slow algorithm.
When the discretization is no longer uniform, the ~nfluence

numbers of the collocat~on (14) are no longer symmetric,

KIJ KJI • This means that there no longer exists a quadra-
~
aQ(PJ)
tic form Q(pJ) such that aPI =
~
5 .
KIJ PJ - hI' that ~s,
there is no optimization problem fully equivalent to the
collocat~on problem. The most precise way to ~mplement

problem 1) is now to solve the following problem:

 649

_ 1/6 / x 2 + y2 3

a 4 p(X-X',y-¥') = / (x _ x,)2 + (y _ y,)2 -1

axax'ayay

XI+1I )XJ+1J )YI+kI )YJ+k J

M = [[[[ P(x-x' y-y') )
IJ ' x = x -1 x'=x -1 y=y -k y'=y-k
I I J J I I J J

fJ Q h(x,y)dxdy
I

min! ~ I MIJ PI P J - I HI PI sub PI ; 0 VI

I,J I
2 2
when (XI - x J ) + (YI - YJ)

2
- kJ ) } ; (19)

with C » 1 , e.g. 2:

MIJ ~ 16
I (xI - x J )
2
+ (YI - YJ)
2 -1
lK1Jklk J (19b)

In the present matrix (M IJ ) we must calculate 32 logarithms

per element which is not covered by (19b), while in the
collocation matrix (K IJ ) of (17), we need only eight loga-
rithms under the same circumstances. This renders the
calculat10n of (M IJ ) roughly 4 times as slow as that of
(K IJ ). Present research is d1rected towards diminishing th1S
factor.

Reusner's method concerns itself with the calculation of the

frictionless contact between two bodies of revolution. Then,
the undeformed distance h 1S given by:

h(x,y) = A x 2 + B(y) (20)

where A 1S in most cases constant, but may depend on y.

Under the assumptions of line contact theory [17, 18J, see
sec. 4, it may be shown that the normal pressure is semi-
elliptical, Pn = - C /a(y)2 - x 2 where a(y) is the half-
width of the contact area at y.
Also, the Hertz theory [1] predicts a normal pressure of the
650

form Pn =- C /1 - (x/a)2 - (y/b)2, where a and b are the

semi-axes of the (elliptical) contact area; i.e. Pn is also
semi-elliptical for fixed y.
Reusner then assumes that for fixed y Pn is semi-elliptical
whenever (20) holds. His method of calculation assumes a
contact area of a form sketched in fig. 4, where each rect-
angle carr~es a semi-ell~ptical pressure distribution. The
calculation proceeds roughly as follows.
At the centroid of each rectangle the normal d~splacement un
and its second derivative u nxx are calculated. They are
equated to B(y) and 2A, respectively (cf. (20)). This y~elds
2N non-linear equations with the 2N unknowns PJ and a J (a J :
half-width of contact area at element J), which are solved
iteratively. Proper measures are taken to enforce the valid-
ity of non-penetration (1) and compressiveness (2) con-
ditions. A correction is made for the finite depth of the
1
body, which is a roller of rad~us (2A)' rather than a half-
space. The speed of calculation on an IBM/370/158 ~s several
minutes.
The advantage over Paul & Hashemi is that with Reusner the
number of elements is roughly 1/5th to 1/10th, but the
influence numbers are harder to calculate.

Just as Kalker & Allaert was designed as an improvement of

the method of Paul & Hashemi, the method of Kalker & De Mul,
which is at present not yet finalized, is designed to
improve upon Reusner. The most important bottleneck of
Reusner seemed to be the setting up of the non-linear
equations. Th~s seemed to us to be due to the fact that the
influence numbers of Reusner are not expressible as element-
ary functions the way the K1J (see (17)) and the M1J (see
(19)) are. Consequently, we started to approximate the pres-
sure distribution of f~g. 4 by one which is constant over
each Reusner mesh. This turned out to be too inaccurate,
whereupon we started to work with an element of the shape
given in f~g. S. As variables we took the area of the
element, 1.S pa, and the height-width ration (p/a), and we
 651

m1n1mized the familiar object funct10n (see (19)):

m1n! C = ~ ILJ MIJ PI P J - ~ HI PI' sub (1.5 pa)I ~ 0 I.
Note now thaf the gradient and Hessian of this C consists of
MIJ and the1r first and second derivat1ves with respect to
a I and a J , wh1ch can, however, easily be expressed 1n
elementary functions. The function C is this no longer qua-
dratic, and the problem must be solved w1th the methods of
non-11near programming, for wh1ch we chose Newton's method,
modified for bounds. At present the program takes about 15
sec. on an IBM 370/158 computer for a problem which is com-
parable to the one ment10ned at the end of sec. 4.3.
It is hoped that more will be gained when the computation of
the Hessian is improved, see sec. 4.2, where 1t is ment10ned
that present research 1S d1rected towards that end.

5. Conclusions

In the case of frict10nal rolling contact, it has been shown

that var1at1onal methods are the only methods available for
solving the problem as accurately as one wishes. If a small
accuracy is taken into the bargain of an extremely fast
operation, recourse may be had to a class1cal "approximate"
method.
In the case of frict10nless contact, classical and var1-
ational methods are foughly equally fast. That under these
C1rcumstances the small edge observed 1n favor of the
variat10nal methods, may very well be due to the "speed-
consiousness" of the present authors.

L1terature

[1] A.E.H. Love, A treatise on the mathemat1cal theory of

elasticity, 4th ed. (1926), Cambridge U.P.
[2] G. Fichera, Problemi elastistatici con vincoli uni-
laterali: il problema di Signorin1 con amb1gue con-
d1zioni al contorno, Mem. Accad. Nat. Lincei, ser. 8,
I (1964) p. 116-140.
[3] G. Duvaut, J.-L. Lions, Les 1nequations en mecanique
et en phys1que, Dunod, Paris (1972).
 [4] J.J. Kalker, Survey of wheel-rail rolling contact
theory, Veh. Syst. Dyn. ~ (1979) p. 317-358.
[5] J.J. Kalker, The computation of three-dimensional
rolling contact with dry friction, Int. J. Num. Meth.
Engng, !! (1979) p. 1293-1307.
[6] J.J. Kalker, Simpl~fied theory of rolling contact,
Delft Progr. Rept! (1973) p. 1-10.
[7] K. Knothe, D. Moelle, H. Ste~nborn, ROLCON, e~n

schnelles, vielseitiges Digitalprogramm zum rollenden

Kontakt, ILR Mitt. 55 (1978) (Report Berlin T.U.).
[8] L. Bansagi, Ein analoges Modell und die Simulat~onser­

gebnisse zur Nachb~ldung des rollenden Kontaktes nach

der Theorie von Kalker, Report MAN Augsburg (BRD),
K 096 556-EDS-08 5.0000.01 (1979).
[9] J.G. Goree, E.H. Law, User's manual for Kalker's
simplified non-l~near creep theory, Rept US Dept.
Transport, FRA/ORD-78/06 (1977).
[10] J.J. Kalker, A fast algorithm for the simplified theory
of rolling contact, (1980). Available from the author.
[11] A.D. de Pater, On the reciprocal pressure between two
bodies, Proc. Symp. Rolling Contact Phenomena, J.B.
Bidwell (Ed.), Elsevier (1962) p. 29-75.
[12] B. Paul, J. Hashemi, An improved numerical method for
counter formal contact stress problems, Rept US Dept.
Transport, FRA/ORD-78/26,
[13] H. Reusner, Druckflachenbelastung und Oberflachenver-
sch~ebung im Walzkontakt von Rotationskorpern, SKF
Schweinfurt (1978).
[14] V.M. Fridman, V.W. Chernina, Iteration methods applied
to the solution of contact problems in bodies,
(Russ~an), Mekh. Tverdogo Tela AN SSSR 1 (1967)
p. 116-120.
[15] T.F. Conry, A. Se~reg, A mathematical programming
method for design of elastic bodies in contact, J.
Appl. Mech. 38 (1971) p. 387-392.
[16] I. Paczelt, Iteration method applied to the solution of
contact problems of elastic systems with elements in
unilateral relation, (Russian), Act. Techn. Acad. Sc~.
 653

Hung. ~ (1974) 1, 2, p. 217-241.

[17J J.J. Kalker, The surface displacement of an elastic
half-space loaded in a slender, bounded, curved sur-
face region with application to the calculation of the
contact pressure under a roller, J. Inst. Maths Applics
~ (1977) p. 127-144.
[18J J.J. Kalker, On elastic line contact. J. Appl. Mech. 39
(1972) p. 1125-1132.
[19J J.J. Kalker, Variational principles of contact elasto-
statics, J. Inst. Maths Applics 20 (1977) p. 100-219.

Captions of figures

Fig. 1 Two bodies in contact.

Fig. 2a Discretization according to Paul & Hashemi.
Fig. 2b The behaviour of Pn near the edge of contact.
Fig. 3 Finding a boundary point according to Kalker &
Allaert.
Fig. 4 Contact area assumed by Reusner.
Fig. 5 The element of Kalker & De Mul. The factors ~ may
be changed.
654

Common
normal
(b)

negative distance h

(b)

n~o~c~o~n~ta~c~t~~c~o~n~ta~c~t_______ x

Boundary
pOint
Exterior Penalty Methods for Contact Problems in Elasticity

J.1:0DEN
The Umverslty of Texas, Austill, USA

Summary
This paper reviews recent results of aden, Kikuchi, and
Song [4] on the use of exterior penalty methods as a basis
for finite element approximations of contact problems in
linear elasticity.

Introduction
Th1s note reviews several results recently obtained by the
author and his colleagues, N. Kikuchi and Y. J. Song, on
the use of exterior penalty methods for unilateral problems
in elasticity theory. The physical problem considered is
that of a l1near elastic body deflected by the applicat10n
of external forces so that it comes in contact with a rigid
fr1ct10nless foundation. An appropr1ate var1ational state-
ment of this problem is as follows:
Find (u,o) E V x M such that
B{u,v) - <o,y n (v»
_

_ - s> > 0
<T - o,y n (u)
= f{v) V

V
v EV

TE M
} (l)

Here,

on r D C r} = space of admissible displacements.

u,v ( V = displacement vectors

M h E H- 1/2 {r c ) I T < O}

o = contact pressure
656

D
B(u,v) 2fJ Jrl E:.(u) E: D•• (v) dx , where
1J - 1J-

B : V x V ->- lR is the virtual work, fJ the

shear modulus, the deviatoric strains;

1
(u) = -2 (u. . + u. .)
E: ••
1J - 1,J J,l
<.,.> denotes duality pairing on W' x W , where

W = Hl/2(f ) and fc is the contact boundary:

C
Clrl = f = I'D U r F U rc ; I'D n rC = ~ ,

Yn : V ->- W is the normal trace operator;

y (v) v • n v E: (Co (IT) ) 3 , n being

n -
a unit vector normal to the boundary f

s = initial gap between the body and the founda-

tion and is given data in W.

f E: V' f (v) = frl f · v dx +

- -
f f
t . v ds
F

f the body force and t the prescribed surface

traction, both assumed to be L 2 -functions.

The spaces V and Ware endowed with norms,

v. .v. . dx ;
1,J 1,J

and are Hilbert spaces (meas fD > 0) We assume that rl

is a "sufficiently smooth" domain. Additional information
on this class of problems and properties of these spaces
can be found in the monograph of Kikuchi and Oden [1].
The physical situation is illustrated in Fig. 1.
The not1on of exter10r penalty methods for con-
strained minimization problems provides an alternative way
to formulate problems of this type. If E: is an arbitrary
positive number, we may seek u_E: E V such that
 657

B(~E'~) + ~ <j(Yn(~E)-s)+'Yn(~» = f(~)

Ii v E V (2)

Fig. 1 Elastic body in contact with rigid

fr1ction less foundation.

where J: W + W' 1S the Riesz map defined on W and (~)+

denotes the positive part of ~ (i.e. ~+ = sup (~,O) ,

"sup" being the supremum relative to the order relation
< on W) •
The fact (which is proved in [4]) that B(o,o) is
cont1nuous and V-elliptic makes it possible to show that (2)
possesses a unique solution ~E for every E > 0 More-
over, we also have the condition

}
There exists B > 0 such that
(3)
I <T,y n (v)
_ >I
B IITllw I < sup Ii T € W'
v€V

Moreover, set

a
E
- !E j (y (u ) -s)
n -E +
(4)

Then, under these cond1tions it 1S possible to show that

i) the sequence (u,a)
_E E
converges strongly
in V x W' to the solution (~,a) of (1)
and
1i) there exist constants Cl ,C 2 > 0 , independent
of E, such that
658

Approx~mations

The properties of the penalty formulation outlined

above make (2) an attractive basis for finite element ap-
proximat~ons: I} instead of two unknown f~elds as ~n (I),
there is only the unknown regularized displacement f~eld

u 2} problem (2) is posed on the whole space V rather

-£
than in the constra~nt set - E V I y n (v)
K = {v _ -< s}. 3} the
perturbed solution u can be made arbitrar~ly close to
-£
the solution u by taking £ sufficiently small; and
4} an approximat~on G£ of the contact pressure can be
computed using (4) after the problem (2) has been solved,
with little extra cost and effort.
It is known, however, from experience with penalty
methods for other types of constraints that it may be
necessary to integrate the penalty terms in (2) ~nexactly

in order to obtain a satisfactory finite element approxima-

tion. In ant~c~pat~on of s~m~lar diff~culties here, ,ye
prepare for reduced integration schemes by introducing the
quadrature rule

L
E G
J(f} =
e=l
Je(f} , J
e
(f)
L
j=l

(S)

wherein is the number of hnite elements in a finite

E
element mesh approximating rC , We are the quadrature
e J
weights, and n· are the quadrature points for element
-J
re c rC , I < e < E .
We assume here that G can be
taken suff~ciently large that J( } defines an exact ~n-
tegrat~on of polynomials f of a g~ven degree: J(f}

'" frC f ds •
 659

Next, let us suppose that a fam11y of f1nite-d1men-

slona1 subspaces Vh of V has been constructed uS1ng
regular fam11ies of conforming fin1te elements. Then our
finite element approximation of problem (2) (or (1)) assumes
the fo11ow1ng form:

F1nd ~~ E Vh such that

Behavior of the Approximation

While it may appear that the reduced-integrat1on,
exterior penalty, f1nite element method characterized by
(6) involves only the space Vh of approx1mate d1sp1ace-
ment fields, there 1S 1ntrinsic to every such formulation
an approx1mation Wh of W'. To see this, note that 1f
we introduce the perturbed Lagrange functional

h h 1 h h h
L (v ,T ) 2" B (~ ,~ ) - f (v )
£ -

(7)

h h
then its saddle p01nts (u ,0 ) are characterized by the
-£ -£
system

h h J[o h£ y n (v_ h ) j V ~ h E: Vh
B(u ,v )
-£ -
- f(v h )
} (8)
h h h £Ohj} > 0 h
J{ (T -0£) [(Y n (~£) -s) + £ - V T € Mh

Here Wh is a fin1te dimensional subspace of W' and Mh

1S the set of Th € Wh such that Th < o. Thus,

from which 1t follows that we can take

(9)
660

where n7 are the quadrature points in the definition of

-J
J{.} • Substituting {9} into {all gives {6}.
It follows that the penalty method {6} is equivalent
to the perturbed Lagrangian {or perturbed mixed method} {a}
and that there is connected with {6} a space Wh defined
by each choice of J and Vh . With these observat~ons in
mind, we seek to use such spaces Wh which satisfy the
following three conditions:
A. The quadrature rule J{.} is such that
J{ThTh} {Th'Th} =
V Th'Th f Wh
h
B. There exists cr£ f Wh which is uniquely
defined by {9} for each solution u h of {2}.
-£
c. There exists a constant 8h > 0 such that

_ h }]
IJ[Thy n {v
h

}
8hllTllo< sup
h
~ E Vh
lI~hlll {lO}

for every Th € Wh

Under conditions A, B, and C, it is shown in [4] that

i} there exists a unique solution u h to {6} for
-£
every £ > 0 , meas ro > 0, n = nh }
{assuming
ii} ~f cr h£ is defined by {9}, then {uh,cr
-£
h } converges
£
to a mixed finite element approximation {uh,cr h }
as £ ->- 0 , where

B{uh,vh } - J[cr h ,y {v}]

h f{v h } V vh E: Vh
n -
h h h V Th E Mh
J[{T -0 } {y n {u_ }-s}] > 0

Here and elsewhere s is assumed to be the

trace of a function s € Vh on rC .
iii) Under mild additional assumptions on J , there
exsits a constant C' independent of £ and
I
h , such that
 661

1
+ "D
iJh
IIyn (u-v
- -
h ) 110 r ]
, C

h
+ inf [II a_1 h Ilw' + <1 -a,y (u)
n -
1hE~
_ s>11/2] + inf I<T-a h ,y n (u)
-
-s> I
1/2
1eM
1/2
+ ~ II a h II 0 r + EJ } (11)
iJh ' C

where EJ is a measure of quadrature error. A sinu1ar esti-

mate of II a-a h£ 110 , r C is given in [4].

Discussion of Results

The key to the success of the reduced-integration penalty

method is the use of an integratlon rule J which will
yield a constant 8h independent of h ln the discrete
stability condition (10). In the following, we list spaces
Vh ' integration rules J , and computed stability parameters
8h for some typlca1 two-dimensional elements. Complete
proofs of these results are given in [5].
We use the notation:
9-node biquadratic elements
a-node isoparametric
4-node bl1inear
6-node quadratic trlang1e
3-node linear triangle

G3 3-point Gaussian quadrature

G2 2-point Gaussian quadrature
G1 1-point Gaussian quadrature for J
S Simpson's rule
T Trapezoid rule
-1
P2 discontinuous piecewise quadratic
polynomials
discontinuous piecewise linear
polynomials
discontinuous piecewise constants
662

P~ continuous piecewise polynomials

of degree k,(k = 2,1) J
Fifteen examples follow:

1. G3 FAILS

2. G2 FAILS

3. Gl FAILS

4. S CONSTANT
(stable)
5. T CONSTANT
(stable)
6. G3 FAILS

7. S CONSTANT
(stable)
8. T CONSTANT
(stable)
9. G2 FAILS

10. T CONSTANT
(stable)
ll. G3 FAILS

12,. Q8 G2 FAILS

13. Q8 Gl FAILS

14. Q8 S CONSTANT
(stable)
15. Q8 T CONSTANT
(stable)

Of course, even though one of the above methods may

lead to a constant Bh lndependent of h and, therefore,
a stable approximation of the contact pressures, there may
result a loss in the asymptotic rates of convergence of
these methods. Accordlng to the above table, a conformlng
approximation is apparently needed to produce a convergent
scheme. In the above list, "FAILS" lndicates elther Bh=O
or Bh is dependent upon the mesh parameter h (e.g. Bh=O(h)).
Extenslve numerlcal experlments have been performed
using the elements listed above [5]. In all cases, the
experiments confirm the predicted lnstabilltles (or stablll-
ties) and, for stable schemes, the rates of convergence.
 663

The problem considered here is the indentation of a

rigid cylindrical punch into a rectangular block of iso-
tropic, linearly elastic material with modulus E = 3000
and Poisson's ratio v = 0.49999. Half of the body is
modeled with the non-uniform mesh of Q2-elements shown in
Fig. 2. Other dimensions are indicated in the figure. The
cylinder is pushed into the body an amount 8 = 0.8. Results
of a typical numerlcal experiment are reproduced in Figs. 3
and 4.
Three different integration rules were used to obtain
three different solutions of this problem, corresponding to
cases 1, 2, and 4 above. Computed distributions of con-
tact pressures for these cases are shown in Fig. 4. As pre-
dicted, cases 1 and 2 are unstable; widly oscillating

I
-Llhearly Elastlc Founda tlon-
I
1000
1 4

J
0.3

I. 8 .1
Flg. 2 Model for numerical example of a
Signorini problem.
664

pressure distributions are obtained and further computa-

tions indicate that these oscillations become more irratic
as the mesh is refined. Use of Simpson's rule, however,
leads to a stable scheme and the smooth continuous approxi-
mation of the contact pressure indicated in the figure.

100

200

300 - - Hertz' Solution

o By 2-point Gaussian
400 • By 3-point Gaussian

Fig. 3 Contact pressure by Guassian

quadrature rules.

100

200
Hertz' F.E.M.
Solution Solution
300
a 3.873 3.6
p 258 275
400

Contact a: contact area

Pressure p: max. contact pressure

Fig. 4 Contact pressure by Simpson's Rule for

numerical integration.
 665

Acknowledgement. The support of the U.S. Air Force Office

of Scient~fic Research under grant F-49620-78-C-0083 is
gratefully acknowledged. The work reported here was part
of a broad project on f~nite element methods for constrained
problems done ~n collaboration with Professor N. Kikuchi
and Col. Y. J. Song.

References

1. Kikuchi, N.; aden, J.T.: Contact Problems in Elast~city.

SIAM, Philadelphia, (to appear).

2. aden, J.T.; Kikuchi, N.: F~nite Element Methods for

Constra~ned Problems in Elasticity. International
Journal for Numerical Methods ~n Engineer~ng, (to
appear).

3. aden, J.T.; K~kuch~, N.: A Study of Penalty Methods

for Problems in Nonl~near Elasticity, IUTAM Symposium
Proceedings, Pergamon Press, Oxford, (to appear).

4. aden, J.T.; K~kuch~, N.; Song, Y.J.: Reduced Integra-

tion and Exterior Penalty Methods for Finite Element
Approx~mat~ons of Contact Problems in Incompressible
Elast~c~ty, TICOM Report 80-2, Austin, Texas, 1980

5. Song, Y. J.; aden, J.T.; K~kuchi, N.: R.I.P. Methods

for Problems ~n Elast~c~ty, TICOM Report 80-7, Austin,
Texas, 1980.
Approximate Solution of Nonlinear Problems in Incompressible
Finite Elasticity

R. GLOWINSKI, P. LE TALLEC, V RUAS DE BARROS

Umverslte Pans VI, University of Texas, Austm, INRIA, Le Chesnay,
France USA France

1. INTRODUCTION
The main goal of this paper is to describe some numerical
methods for solving nonlinear variational problems in incom-
pressible finite elasticity.
In Sec. 2 we discuss a decomposition princlple for a large
class of variational problems and then derive several itera-
tive methods of solution from this principle. We give in
Sec. 3, the formulation of elastostatics two-dimensional and
axisymmetric problems for incompressible Mooney-Rlvlln mate-
rials.
In Sec. 4 we descrlbe the application of the algorithms of
Sec. 2 to the iterative solution of the mechanlcal problems
of Sec. 3. In Sec. 5 we give some brief indicatlon on the fi-
nite element approximation of the problems of Sec. 3
Finally some numerical results obtained applying the methods
of Sec. 4,5 are presented and discussed in Sec. 6

2. DECOMPOSITION OF VARIATIONAL PROBLEMS AND ASSOCIATED ALGO-

RITHMS.
2.1. A family of Variational Problems.
Restricting our attention to real Hilbert spaces, we conslder
two such Hilbert spaces V and H equipped with II· II and I· I ,
and inner products ((.,.)) and (.,.), respectively.
Let B E £(V, H) and F, G be two proper, convex, lower semicon-
tinuous functionals from H,V to IR, respectlvely, such that

(2.1) dom(G) ndom(FoB) t ~ .

 667

We associate wlth the above V,H,B,F,G the following minimiza-

tion problem

(P) Find u E V such that J (u) :<; J (v) V v E V,

where J : V + IR is defined by

(2.2) J(v) F (Bv) + G (v) .

It appears from (2.~) that J(.) and therefore (P) have very
special structures and taking this into account it is then
quite natural to design special methods of solution for (P).

Remark 2.1 : Most of the considerations which follow can be

applied to variational problems like

(2.3)

where f EV ' (dual space of V) and where Al (resp. A2 ) is a

monotone operator (possibly multivalued) from H to H' (dual
space of H) (resp. V to V') i the operator A = B* oAloB+A2,
from V + V'is not in general the differential (or subdif-
ferential) of a functional J. For many results on these
generalizations we refer to P.L. LIONS-B. MERCIER [1] and
GABAY [2]. If we suppose that in addltlon to (2.1) we also
have

(2.4) lim J(v) = +00

\\v\\++oo
then (P) has a solution which is unique if J is strictly
convex.

Remark 2.2 : The applications in finite elasticity that we

have in view are actually related to nonconvex minimization
problems.

2.2 A decomposition principle.

Let us define W cV xH by
668

(2.5) w {{v,q} €VxH, Bv-q = O} •

Problem (P) is then clearly equivalent to

Find {u,p} €W such that

(TI )
j (u,p) ~ j (v,q) V {v,q} € W,
where

(2. 6) j (v,q) = F{q)+ G{v).

Remark 2.3 : The new problem (TI) has clearly a mixed formula-
tion "flavor" since the linear relation Bv-q 0 suggests the
introduction of a Lagrange multiplier.

Remark 2.4 : Problems (P) and (TI) are equivalent, but consi-
dering (TI) we have in some sense simplified the nonlinear
structure of (P), at the expense however of the new variable
q and of the relation

(2.7) Bv-q o .

Actually since (2.7) is a linear relation, very efficient

techniques may be used to treat it ; in the following this
will be done by using, simultaneously, penalty and Lagrange
multipliers, via a convenient augmented lagrangian functio-
nal.

2.3. An augmented lagrangian functional associated with (TI).

Let r >0 ; define ~r :VxHxH + IR by

(2.8)

It can be proved that if {U,p,A} is a saddle-point of~

r
over VXHXH, i.e., if

{2. 9}
 669

then {u,p} solves ('IT), Le., u solves (P) (with p BU).

2.4. A first algorithm for solving (P).

To solve (P) and ('IT) we shall compute the saddle-points of
£r by an algorithm derived from the duality algorithms con-
sidered, for example, in GLOWINSKI-LIONS-TREMOLIERES [3,4],
Chap. 2. Such an algorithm applied to the solution of (2.9)
is

(2.10) >,.0 EH, given

n n n ,n+1
then if n ~O, A given, we compute u ,p ,A by
~ Find {un ,pn} E VxH such that
(2.11)
t.er(Un,pn,A n } $£r(v,q,>,n} TV {v,q} EVXH,

(2.12) An +1

Concerning the convergence of algorithm (2.10}-(2.12) it

can be proved (see [5, Chap. J] and [6, Chap. 5]) that under
very reasonable assuwptions on F,G,B and if

(2.13) 0 <p <2r,

then we have as n .... +00

n .... u
(2.14) u strongly in V,

n ....
(2.15) p P = Bu strongly in H,

(2.16) An .... A weakly in H,

where u is the solution of (P), and where A is such that

{U,p,A} is a saddle-point of £r over VXHXH.

Remark 2.5 : The only nontrivial step in the above algorithm

is clearly the solution of the minimization problem (2.11).
Actually to solve (2.11), taking into account its special
structure, ~t is very convenient to use a functional block
relaxation method (like those discussed in CEA-GLOWINSKI [7];
670

see [5],[6] for more details on the block relaxation solu-

tion of (2.11)). If this relaxation method is used, and if in
the calculation of {un,pn} we only do one inner relaxation,
starting from { u n-l ,p n-l} , we obtain the algorithm described
in Sec. 2.5.

2.5. A second algorithm for solving (P).

The new algorithm is defined by

-1 0
(2.17) {u ,A} E VxH, given,

then for n ~ 0, u
n-l ,A n given, we compute p n , u,/\
n, n+l £y

Find pn E H such that

(2.18)

Find un E V such that

(2.19)

Remark 2.6 : Several variants of (2.17)-(2.20) are ava~lable.

We can for example

(i) Exchange the role of q and v (see also Remark 2.7)
(ii) Update also An between the steps (2.18), (2.19) ;
doing so we obtain the following variant (due to
GABAY [2]) of (2.17)-(2.20)

(2.21) {u -1 ,A o} E VXH, given,

(2.22)
 671

rl r (n
u ,pn ,/\,n+]/2) -<rlr ( v,p n ,II.,n+l/2) Vv E V,
(2.24)

q and v playa more symmetrical role in (2.21)-(2.25) than

in (2. 17) - (2 . 20) •

Remark 2.7 : If one uses algorithm (2.17)-(2.20), it is re-

commended to solve, in the second step, the problem which
has the best properties of ellipticity (for more details,
see [5,6]). If one uses (2.21)-(2.25), the character of the
problems ellipticity does not matter since q and v play a
symmetrical role. -

Concerning the convergence of (2.17)-(2.20) it is proved in

[5,6] that under vary reasonable assumptions on F,G,B we
still have (2.14)-(2.15) i f

1+ 15
(2.26) o < p <--2- r.

If G is linear it follows from GABAY-MERCIER [8] that we can

again take 0 <p <2r.

2.6. Remarks on the choice of p and r.

For a given r, the optimal choice of p is very close to p=r,
as shown by the various numerical experiments that we have
done on the algorithms of Sees. 2.5 and 2.6. The choice of r
is a more delicate matter. Theoretically, the larger r is,
the faster is the convergence of (2.10)-(2.12). Actually, for
large values of r the problem (2.11) will not be well-condi-
tioned and its accurate solution w111 be a costly operation
moreover, for very large values of r, round-off errors play
a significant(negative) role. As we can see, we have there-
fore two contradictory behaviors as r increases. The global
672

effect of these phenomena on (2.10)-(2.12) is to produce an

algorithm which is not very sensitive to the choice of rand
which is very robust.

2.7. Relations with Alternating Directions Methods.

2.7.1. Relations between algorithms (2.17)-(2.20), (2.21)-
(2.25) and A.D.I.
We suppose for simplicity that V=H and B=I. We suppose also
that F and G have as differentials (or subdifferentials) Al
and A2 , respectively. Note that Al and A2 are necessarily
monotone (possibly multivalued) operators. Then (F) is equi-
valent to

where the equal sign has to be replaced by 3 if Al and/or

A2 are multivalued. Suppose that p=r ; then by elimination
of An in (2.17)-(2.20) we obtain

-1
(2.28) u given,

then for n ~ 0,

n-l n-l
ru A2 (u ),

(2.30)

Setting u n + l / 2 = pn+l, we finally have

n n
ru - A2 (U ),

We recognize in (2.31), (2.32) a Douglas-Rachford Alternating

Direction Implicit (ADI) method (see [9]). Similarly, by
elim1nation of An and An+l/2 we obtain from (2.21)-(2.25)
(still supposing p=r)
 673

n+l/2 A ( n+l/2)
ru - 1 u

which is (see [10]) a Peaceman-Rachford ADI method. Such ADI

methods involving fairly general monotone operators Al and
A2 have been studied in [1,2].

2.7.2. An initial value problem interpretation of (2.17)-

(2 . 20) and ( 2 . 21) - (2 . 25) .
It follows from Sec. 2.7.1 that if p=r and V=H, B=I, then
(2.17)-(2.20) and (2.21)-(2.25) can be seen as implicit sche-
mes, based on fractional steps, for the discrete time inte-
gration of the initial value problem

(2.35)
~~ + A(u) 0

where A

From that ~nterpretation r appears as the inverse of a time

step ~t (i.e. r=l/~t). As shown in [11], this interpretation
of the avove algorithms, using the initial value problem
(2.J5), may be very helpful in obta~ning insight concerning
the behav~or of these algorithms ; for example, the larger
r is, the safer the algorithms.
The numerical integration of initial value problems by ADI
methods is discussed in [1] under fairly general assumptions
on Al and A2 ·

2.7.3. Further comments.

The solution of problems like (P) by decomposition-coordina-
tion methods via augmented lagrangians seems to be due to
GLOWINSKI-MARROCCO [12,13, H] (see also POLJAK [15]). For
more details and various applications see also [1,2,5,6,8,
11,16-22]. Ref. [11],[22] in particular describe the appli-
cation of the above algorithms to large displacement calcu-
lat~ons of flexible, inextensible, bending pipe lines.
To the best of our knowledge the relations between the al-
674

gor~thms of Sec. 2.5 and ADI methods have been observed for
the first time by CHAN-GLOWINSKI [23]. In Sec. 2 we have
basically followed the presentation of FORTIN-GLOWINSKI [5]
and GLOWINSKI [6].

3. - FORMULATIONS OF ELASTOSTATIC PROBLEMS FOR INCOMPRISSI-

BLE MOONEY-RIVLIN MATERIALS.
3.1. Notation and mechanical assumptions.
A fundamental problem in nonlinear e1ast~city is the calcu-
lation of the deformations of a solid body made of an homo-
geneous, incompressible, hypere1astic material, subjected
to volumetric forces ~~ (po: density in the reference con-
figuration) and superficial forces ~o' In a lagrangian for-
mulation, the related energy functional, corresponding to a
displacement field ~ is

- = Jnp0 Jan
(3.1) n(v) (o(v)-f'v)dx - S·V df ,
- - - - 0-
2
where n is a bounded domain of lRN corresponding to the re-
ference configuration ; an(= an1 u an 2 ) is the boundary of n,
the body being fixed on anI' We have denoted by o(~) the
stored energy functional (per mass unit). For a Mooney-Riv-
lin material we have

(3.2) a (v)

(3.3)

with, in (3.2), (3.3), I. the i-th invariant of the FFt ten-

~ "''''
sor, where

(3.4) F Ilv+I

and E 1 ,E 2 are coefficients which are material dependent.

The displacement also has to satisfy the incompressibi11ty
condition which here has the following form

(3.5) det F(v) 1 a.e. on n.

 675

Remark 3.1 : We have supposed 1n (3.1) that S and ~ are lndepen-

- 0-
nent of v. 'TIllS corresponds to the standard simplifying assumption
known as the dead loading hypothesis for Wh1Ch ! and §o do
not vary during the motion ; if we deal with pressure type
forces, then §o is a function of u (the actual displacement),
given by

(3.6) S
-0

where q(~+~(~)) 1S the pressure at the point ~+~(~) of the

actual configuration and ~ is the unit outward normal at dQ,
in the reference configuration.

3.2. Mathematical formulations.

We shall give in this section several possible formulations
of the elasto-stat1c problems ; it is still an open problem
to prove the1r equivalence in general (see [24]-[26] for a
discussion of these equivalence, and also the comments of
Sec. 3. 2 . 4) .

3.2.1. Formulation by minimization.

It is reasonable to suppose that those displacements u cor-
responding to stable equ1librium position obey

u is a local minim1zer over K of the functional

(3.7)
v -+ TT(V),

with for an incompressible Mooney-Rivlin material

K 1
(3.8)
a. e. ,

The existence of solutions for (3.7), (3.8) is proved in [27].

3.2.2. Formulation by equilibrium equations.

The equilibrium positions correspond to the solutions of
676

the following system of nonlinear partial differential equa-

tions
U E: K,
(3.9)
(Dn(u) ,v)+ Jrl p[u,v]
--
dx o TV v E: X,

where Dn is the differential of n (on (H1(rl»N) and where

d
(3.10) [~,~] = ~ (det F(u»v . . ,
~,J --- ~ - ~,J

(3.11)

(in (3.]0) we have used the standard summation and derivative

notations). The above function p, which is clearly a Lagrange
multiplier associated with the incompress~bi1ity condition
(3.5), appears as a pressure.

3.2.3. Formulation by augmented lagrangian.

We proceed as in Sec. 2 by "relaxing" the linear relation
(3.4) using an augmented lagrangian. We obtain thus the fol-
lowing formulation :

2 NxN .
Find {~,;, ~} E: W = XxYx (L W» , stat~onary point
over W of the augmented lagrangian
(3.12)
.eR(~'~'~) = n(v) + ~11~~+;-~1I~2

- Jrl~· (~~-;-~) d~ ,

where, in (3.12), we have R> 0 and

3.2.4. Some relations between formulations (3.7), (3.9) and

(3.12) .
The following results are proved in [24] :

a) Formulations (3.9) and (3.12) are equivalent.

b) Every "smooth" solution of (3.7) is a solution of (3.9),
(3.12),
 677

c) If the functional IT is convex (case of a Mooney-Rivlin

materlal if N=2), then every solution of (3.12) is such
that u is a mlnimlzer of

similarly, for R sufficlently large every solution of

(3.12) is such that F lS a minimizer of ~ +.eR(!;'~'~).

Remark 3.2 If {u,F,:\} is a solution of (3.12), then the

condltlon

implles

(3.13) -d (_d _ _ (p G)-A. ) = p f..

J dU.. 0 lJ 0 l
l,J

From (3.13), A appears as this part of the first Piola-Kir-

choff stress tensor, corresponding to the incompressibility.
We observe also that an algorithm solving (3.12) gives the
stress fleld directly.

3.3. Axisymmetric problems.

3.3.1. Formulatlon of the problems.
We consider the case of an aXlsymmetric incompressible hyper-
elastlc body subJected to an aXlsymmetrlc system of forces.
The problem is to flnd the aXlsymmetrlc positions of equili-
brium. We denote by u = {u l ,u 2 } the dlsplacement, with u l the
radial displacement and u 2 the axial one. Using an {r,z}
system of coordlnates we have

)
o
(3.14) I+Vu o
e (u)

u
where e(u) = 1+ ~
r
is the extenslon ratio in the circumfe-
rential direction, and where
dU.
_J
dZ
678

Using the notation of Sec. 3.1, we have (with e=e(u))

2
e + (0 .. +u. )
2
1J 1, J

det(~+?~) = e(1+ul,lu2,2-ul,2u2,1+ul,1+u2,2).
For incompressible materials we have

det(~+?~) = 1 a.e.

implying in turn

2 2 2
I 2 (tl) = e (o .. +u . . ) + lie.
- 1J 1, J

3.3.2. Lagrangian formulation.

Let ~ be the half meridian section of the reference domain
we associate to ~ the following spaces (with 1 s p < +(0)

ol P = {¢I f I¢(x) I P rdr dz < +oo} ,

~3¢
3r ' az
1 _ I 3¢ 2
If - {¢ ¢, E ol }

we shall use in the sequel the notation dx = rdrdz.

With respect to the general case we do the following modifi-
cations :

a) Spaces
X {VE (lfl)2, v=O on 3~1}'

and with 1 s i, j s 2

124
z {ll
o
, ll· . } Eol
1J
x (.e) }

b) Augmented lagrangian

olR(V,G,ll) R
= n(v) + -2 f (0 .. +v . . -g .. ) 2 dx
- ::: '" - ~ 1J 1, J 1J -
(3.15)
-J ( R
ll .. (o .. +v .. -g .. )dx + -2 f (g -e(v)) 2 dx-
~ 1J 1J 1J 1J - ~ 0 - -
 679

-f S"l
~o(e(v)-g
-
)dx.
0-

Remark 3.3 : A possible variant of.e R in (3.15) can be obtai-

ned by replacing e(~) by go in TI(~) (leading to a functional
TI (~,go))'

4. - ITERATIVE SOLUTION OF THE EQUILIBRIUM PROBLEMS.

We apply to the solution of the equilibrium problem (3.12),
the iterative methods described in Sec. 2, keeping in mind
that (3.12) is a nonlinear, nonquadratic, non convex problem.

4.1. A first algorithm for solvlng (3.12).

We follow Sec. 2.~ ; using the notation of Sec. 3.2.3 the
algorithm is

(4.1) \0 given in (L 2 un )NxN,

"
then for n ~ 0, \ n known we obtain Un,F n and \n+l
"
:::::
(\ n E (L 2 (S"l) ) NXN') by
::::::

'"
{Un,F n } XxY and V {v,G} Xxy we have
-
E E
'"
(4.2)
JJR(~n'F'n,:n)
L__ .~ S
JJ
L
( G ,n) ,
R ~, '" ' ~

(4.3) \n_ p (l7u n +I-F n ) , p > O.

::::: "'''':::::::::::

Remark 4.1 :Problem (4.2) is equivalent to the nonlinear

system

nnn, n n n
(4.4) .eR(~ ,!
,~ )
s .... R(~ '~'0) VG E Y, ;' E Y,

n
(4.5) d~R (~ ,! n ' ~ n ).~ = 0 "Iv E X, U
n
E X,

whose block relaxation solution leads to the algorithm des-

cribed in Sec. 4.2.
680

4.2. A second algorithm for solving (3.12).

4.2.1. Description of the algorithm.
The second algorithm is given by

(4.6)

then for n;?: 0, un-1,~n given, we obtain ;n,~t,~n+1 from

(4.7)

(4.8)

(4.9)

4.2.2. On the solution of problems (4.7), (4.8).

Problem (4.8) 1S equ1va1ent to

Find un to X such that

4.10)
Vv to X

which is an unconstrained minimizat10n problem whose solution

is quite easy, particularly if R is sufficiently large. If N=2
then the functional to minimize 1S quadratic, and therefore
solving (4.8), (4.10) is equivalent to solv1ng a linear problem
related to a second order partial differential operator which
is independent of n and whose discrete variants are linear sys-
tems associated with positive definite matrices independent
of n (we shall use therefore a prefactorization of these ma-
trices). If N=3 or in the axisymmetric case problem (4.8),
(4.10) is no longer linear jit can be, however, efficiently
solved by a preconditioned conjugate gradient algorithm.
Problem (4.7) is a more delicate one (apparently, at least)
if N=2, (4.7) is reduced to (omitting indice n) :

Find F to Y and minimizing over Y the functional

(4.11)

G ....
'"
J!G 2
[RG. . - 2 (R (u .
1J 1,J
. +0
1J
) - A .. ) G.
1J 1J
] dx
-
 681

w~th

1 a.e. on ~}.

Since derivatives of G (and F) are not ~nvolved in (4.11),

~ ~

we can solve this last problem pointwise. We have therefore

to solve an infinity (in theory) of four-dimensional problems
of the following class

Find {F .. }
-- ~J
E c:. and minimizing over C) the functional
(4.13)

where

(4.14)

We diagonalize the above quadratic relation using as new

variables

(4.15)

Using b defined by (4.15), problem (4.13), (4.14) becomes

Find ~ E C and minimizing over C the functional

(4.16)

where

(4.17)

The extremizers of (4.16), (4.17) are given by

(4.18)
682

where the scalar p (Lagrange multiplier of £.b 2 2) is a

~ ~

solution of

(4.19)

2 2
We suppose that zl+z4 ~ 0 ; then it can be shown that (4.19)
has a unique solution in ]-R,+R[ ; moreover, using the im-
plicit function theorem (see [24], [25] for more details) one
can show that this solution p of (4.19) between -R and R is
prec~se1y the one associated with the global minimum of
c + RC~-2z.c. over £.c~
~ ~ ~ ~ ~
= 2, to which, therefore, corresponds
a unique global minimizer given from p in (4.16), (4.17) (ac-
tually there is no other minimizer (local or global) than the
above global minimizer).

Solving (4.19) on ]-R,+R[ is a trivial problem; we have

then b from p and (4.18), and then F from band (4.15).

Remark 4.2 : In the actual computer experiments that we have

done, we never encountered the situation zf+z~ = 0 ; in fact
we have the feeling that for R sufficiently large this cannot
happen in the context of problem (3.12) if N=2.-

The solution of problem (4.7), in the axisymmetric case and

for genuinely three-dimensional problems, leads to a far more
comp1~cated discuss~on, however the same general ideas st~ll

apply. we refer to [24], [25] (resp. [28]) for the axisymme-

tric (resp. three-dimensional) case.

5. - FINITE ELEMENT APPROXIMATION OF THE EQUILIBRIUM PROBLEMS.

5.1. Synopsis.
A most important step to the computer use of the iterative
methods of Sec. 4, for solving the elasticity problems of
Sec. 3, is the reduction of these problems to finite d~men­

sional one via convenient approximation methods. As i t can

be guessed the main d~fficulty to overcome is the incompres-
sibility condition (3.5), coupled to the other non1inearities
of the problems under consideration ; a guideline to the
 683

numerical treatment of the incompressibility condition is

given by those results concerning the numerical analysis of
incompressible flows governed by Stokes and Navier-Stokes
equations (owing to the very abundant litterature concerning
this subject we refer to the corresponding bibliographical
references 1n [24], [30] and also to the papers of Malkus
and Oden in these proceedings) .

In this paper, restricting our attention to two-dimensional

problems we shall describe br1efly two families of finite
element approximations, preserving the decomposition proper-
ties of the continuous problems :

a) A fairly class1cal one, previously used in [24], [25] , [29] .

b) A very recent one introduced by the third author in [30].

Correspond1ng numerical results w111 be shown in Sec. 6.

5.2. Approximation by quadrilateral finite elements.

The numerical solution of the lagrangian problem (3.12)
using the algorithms described in Sec. 4 requires the intro-
duction of finite dimensional spaces Xh,Yh,Zh approximating
X,Y,Z. An approxiillate gradient operator sh from Xh to Zh
must also be def1ned.

In this section 5.2 the spaces Xh,Yh,Zh are constructed

using quadrilateral f1nite elements, with the displacements
interpolated at the vertices, the pressure and the tensor F
at the center of each element.

5.2.1. Quadrangulation of ~.

We suppose that our domain D is a polygonal open set in ~2

which can be decomposed into quadrilaterals

(5.1)
k
UIh
E
~k·

Here Ih 1S a finite set of numbers and Dk is the image by

a mapping J k of a reference rectangle ~ such as the one
shown in F1g. 5.1 i this mapping J k is defined by
684

"
X2 A4
AI

AS "Xl

AZ A3

F~gure 5.1 : Reference Rectangle

(5.2)

where, in (5.2), x ak is this vector whose coordinates are

- th
those of the node a of the k element and where ¢a is the
shape function associated to the node a of the bil~near f~­

nite element {~,{Aa}!=l' Ql(n)}, Ql(~) be~ng the space of

the b~linear polynomials defined on ~, ~.e.

(5.3)

5.2.2. Discrete spaces and discrete operators.

We can now define the following spaces (with Yh!k = Yh!~k)

(5.4)
 685

(5.5) Ph {qh\qh\k const.} ,

1
(5.6) Xh Hh n X,

(5.7) Zh (P ) 4
h

(5.8) Yh Y n Zh'

Remark 5.1 : It follows from (5.2), (5.4) that the above fi-
nite element approximat1on is of the isoparametric type . •

Concerning the approximate gradient operator sh' two qU1te

natural choices can be made :

Either

or

( 5 . 9) 2 sh (
~h) \k = L 2 -proJec
. t '10n 0 f "~~h on Zh'

5.2.3. Formulation of the approximate problems.

We consider only the plane strain problem S1nce the exten-
S10n to the axisymmetr1c case 1S qU1te obvious ; taking
1nto account the results of Sec. 3.2.3, 3.2.4 the discrete
approximation of (3.7), (3.8) that we consider is def1ned by

;h minimizes over Yh the funct10nal

(5.10)2
~h + ~\\I+Sh(uh) - ~h\\2+Jrl~h '~h dx,
~6

It is important to observe that from the construction of Yh

and if we suppose that ~h and ~h are known in (5.10)2 then
this last minimization problem can be solved elementwise ;
therefore the algorithms of Sec. 4 can also be applied to
the solution of problem (5.10).

Concerning the convergence of the discrete solutions to the

continuous one we refer to LE TALLEC [24].

5.3. Approximation by incomplete quadratic simplicial fini-

te elements.
We shall briefly discuss in this section a new type of fini-
te element approximation, using triangles ; these finite
elements have been recently introduced by the third author
in [30] (see also [31]) and have been denominated AQL (for
Asymmetric Quasi-Linear) .
We follow [30] ; this new approximation is defined as
follows :

Let ~~ be a triangulation of ~ and let T E ~h with vertices

{oiT}i=l; let P 4 / 3 be the space of the polynomials in two
variables, of degree ~2, defined over T and such that their
restriction to two given edges, e lT and e 2T for example,
are linear functions (e iT is the edge of T, opposite to
01T) ; we clearly have dim P4/3 = 4. We i~troduce now the
following set of degrees of freedom {aiT}i=l' where:

a iT is for i=1,2,J the functional value at ciT'

a 4T is the functional value at the midpoint of edge e 3T .

A fundamental property of the AQL element is given by the

following

Proposition 5.1 : Let ~ = {v l ,v 2 } be a vector valued func-

tion defined over a triangle T and whose components belong
to P 4 / J , then

det (I+Vv) is linear over T.

~ ~~
 687

We refer to [30) for the proof of the above proposition.

It follows then from Prop. 5.1 that

(where GT = centroid of T), and (5.11) clearly suggests to

require the incompressibility condition

det (I+\lv) 1

at GT only.
Notice that because the nonsymmetrlc structure of the AQL
element we have to make some restriction on the triangula-
tion upon which the space of the approximate displacements
is to be deflned ; one suggests in [30) the following pro-
cess that generates nearly as general a mesh as any other
finite element triangulation :
We first partltion Q into convex quadrilaterals arbltrarily
then each quadrilateral is subdivided into two triangles by
one of its two diagonals (any of them can be used). The
edges over which the restrictlons of the second order poly-
nomials of P 4 / 3 are allowed to be quadratlc are preclsely
those dlagonals.

Using the above AQL finlte elements we approximate X,Y,Z

by Xh,Yh,Zh as follows:

Hl o -
(5.12) {Yh E C W), YhIT EP 4j3 'IT E 'eh } ,
h

(5.13) Ph {qh1qhlT const. 'IT E 'eh },

1
(5.14) Xh X n Hh

4
(5.15) Zh (Ph) ,

Concerning the approximate gradient operato~ sh we use

688

Using the above spaces and operator the approximate problem

is then defined exactly as in Sec. 5.2.3 by (5.10), and the
local properties of (5.10)2 are preserved by the new element.

Remark 5.2 : We refer to [30], [31] for a detailed discussion

of the AQL element, and also for various two and three dimen-
sional generalizations.

Remark 5.3: It will be interesting to use the AQL element

(and its genera1izat1ons) for the finite element solution of
Stokes and Navier-Stokes problems.

6. - NUMERICAL EXPERIMENTS.
All the numerical experiments displayed in this sect10n deal
w1th Mooney-Riv1in materials and either a plane strain or
aX1symmetric situation ; moreover the numer1ca1 results pre-
sented here have been obtained using the quadrilateral fini-
te element approxim~tion discussed in Sec. 5.2. Other numeri-
cal experiments are done in [24] ,[25] (including the numeri-
cal treatment of plane stress problems) ; numerical exper1-
ments using th1s Ruas' element of Sec. 5.3 (and some of its
generalizations are discussed 1n [30] ,[31].

6.1. Stretch1ng of a thick cracked rectangular bar.

We consider a thick rectangular slab of Mooney-Riv1in mate-
rial, with a non-propagating crack in its m1dd1e, submitted
to vertical stretch1ng forces applied at 1tS extremities.
The initial configuration of the lower part of the bar and
of the crack is shown on Fig. 6.1 (a). Under the action of
the external forces this bar is stretched and its computed
equilibrium position (plane strain assumption, with a(~)=

E1(I1(~)-2» is shown on Fig. 6.2(b) (the various data con-

cern1ng this problem are 1ndicatcd on F1g. 6.2).
US1ng p=R=8 we have convergence of algorithm (4.1)-(4.3) in
20 1terations, corresponding to a computat10na1 time of 3.2
seconds on CDC 6400. The computed stresses at the boundary
 689

-4
match the applied tension with a 10 precis1on.

INITIAL (01 FINAL (bl

CAACK u,. 0

THE BOUNDARY CONDITIONS ARE

INDICATED ON THE FIGURE

C1 =1. PSI

TRACT I ON =6. OPSI

HEIGTH =1.75IM
WIDTH =1.95IM
CRACK LENGTH =0.50IM
STRAIN ENERGY =2. 212FT L8

UHAX =3. 77IH

Figure 6.1

6 . 2. Combined Inflation and Extension of a Circular Cylin-

drical Tube.
We cons1der here a circular cylindr1cal tube, made of an
incompress1ble isotropic elastic material, whose strain func-
tion 1S of Mooney-Rivlin type. This tube is inflated by im-
pos1ng a fixed radial displacement to the inner surface d~l'
the outer surface be1ng free of tractions. An analytical
690

solution of this problem is given in CHADWICK-HADDON [32]

under the assumption that both extremities are stress free
and remain horizontal. We have approximated these conditions
by restricting the axial displacement to be zero at the mid
cross-section an 3 and leaving the upper section traction free.
This physical configuration is a close approximation of the
case treated in [32] and is described in Fig. 6.2, where we
have represented the upper half part of the tube in its ini-
tial configuration.

I,
I,
I,
I,
I,
anI an2
I,
,I
I,
I,
I,
I,
r
I 1'-.
V/////////////// '/ '/'//// '/'/ '/////"

F:'gure 6.2

USlng the notation of [32], the parameters ln this problem

are

2a(~) = .875(I(~)-3)+ .125(I2(~)-3),

N = (outer radius/inner radius) in the reference

configuration
Q final inner radius/initial inner radius.

The numerlcal values glven in Table 6.1, both for the ana-
lytical and numerical solution correspond to

EXTV = final height/initial height,

 691

r:XTII final outer radius/initial outer radius.

The computed Cauchy stresses 0ee and orr in the mid cross-
section are indicated (as functions of r) on Fig. 6.3, for
different values of the parameters Nand Q ; these computed
values (indicated by crosses on Fig. 6.3) are exactly loca-
ted on the curves obtained analytically in [32] and repro-
duced on Fig. 6.3 ; this confirms the validity of our compu-
tations.

N 1.4 1.4 1.8 1.8 2.2 2.2 2.2

Q 1.2 1.6 1.6 2.0 1.6 2.0 2.2

EXTV analytical .9460 .8583 .8991 .8432 .9252 .8794 .8578

EXTV computed .9460 .8582 .8995 .8434 .9261 .8801 .8584

EXT. anal¥dCarl.ll91 1.3700 1.2486 1.4334 1.1774 1.3146 1.3879

EXTH computed 1.1192 1.3701 1.2489 1.4339 1.1778 1.3154 1.3882

Table 6.1

Comparison between analytical results (from

[32]) and computed results.
692

Cl =0 . 8750pSI
C2 =0. 1250pSI

HEIGT H =1.2IN
3. 0 a ee

0.5 ____
. .. . . .
~_~ _ _ _ _ _ __
~

O. 1
0.2
0.3
O. 4
O. 5
0.6
0.7
0.8 0
rr

Figure 6.3
 693

REFERENCES

1 Lions P.L., Mercler B., Splitting algorithms for the sum

of two nonlinear operators, SIAM J. Num. Anal., 16, (1979),
pp. 964-979.

2 Gabay D., Methodes Numeriques pour l'Optimisation Non

Lineaire, These d'Etat, Universite Pierre et Marie Curie,
Paris, November 1979.

3 Glowinski R., Lions J.L., Tremolieres R., Analyse numeri-

que des Inequations Variationnelles, Vol. 1, Dunod-Bordas,
Paris, 1976.

4 Glowinski R., Lions J.L., Tremolieres R., Numerical Ana-

lysis of Variational Inequalities, North-Holland, Amster-
dam (to appear).

5 Fortin M., Glowinski R., Resolution numerlque de problemes

aux limites par des methodes de Lagrangiens augmentes (in
preparation) .

6 Glowinski R., Numerical Methods for Nonlinear Variational

Problems, Lecture Notes, Tata Institute, Bombay and Sprin-
ger-Verlag, Berlin, 1980.

7 Cea J., Glowinski R., Sur des methodes d'optimisation par

relaxation, Revue Francaise Automatique, Informat., Re-
cherche Operationnelle, R-3, (1973), pp. 5-32.

8 Gabay D., Mercier B., A dual algorithm for the solution

of nonlinear variational problems via finite element
approximation, Compo Math. Appl., Vol. 2, N° 1 (1976),
pp. 17-40.

9 Douglas J., Rachford H.H., On the numerical solutlon of

heat conduction problems in two or three space variables
Trans. Amer. Math. Soc., 82 (1956), pp. 421-439.

10 Peaceman D.H., Rachford H.H., The numerical solution of

parabollc and elliptic differential equations, SIM10.
App1. Hath., l (1955), pp. 24-41.

11 Bourgat J.F., Dumay J.M., Glowinski R., Large displace-

ment calculations of flexible pipelines by finite ele-
ment and nonlinear programming method, SIAM J. Sc. Stat.
Comp., ! (1980), pp. 34-81.

12 Glowinski R., Marrocco A., Sur l'approximation par ele-

ments finis d'ordre un et la resolution par penalisation
dualite d'une classe de problemes de Dirichlet non li-
neaire, Compt. Rend. Acad. Sc. Paris, t. 278A (1974),
pp. 1649-1652.

13 Glowinski R., Marrocco A., On the solution of a class

of nonllnear Dirichlet problems by a penalty-duality
method and finite elements of order one, in Optimization
Technique : IFIP Technical Conference, G.I. Marchouk, ed.
Lecture Notes in Compo Sciences, Vol. 27, Springer-Verlag,
Berlin, 1975, pp. 327-333.
694

14 Glowinski R., Marrocco A., Sur l'approximation par ~l~­

ments finis d'ordre un et la resolution par penalisa-
tion-dualit~ d'une classe de probl~mes de Dirichlet non
lin~aires, Revue Francaise d'Autom. Inf. Rech. Operation-
nelle, R-2 (1975), pp. 41-76.

15 Poljak B.T., On the Bertsekas' method for minimization

of composite functions, in International Symposium on
Systems Optimization and Analysis, A. Bensoussan and
J.L. Lions eds., Lecture Notes in Control and Information
Sciences, Vol. 14, Springer-Verlag, Berlin, 1979, pp.
179-186.

16 Glowinski R., Marrocco A., Sur l'approximation par ele-

ments finis d'ordre un et la resolution par penalisa-
tion dualit~ d'une claGse de probl~mes de Dirichlet non
lin~aires. Rapport Laboria 115, 1975 (extended version
of Ref. 14}.

17 Glowinski R., Marrocco A., Numerical solution of two-

dimensional magneto-static problems by augmented lagran-
gian methods, Compo Meth. Appl. Mech. Eng., 12 (1977),
pp. 33-46.

18 Marrocco A., Experiences numeriques sur des problemes

non lineaires resolus par elements finis et lagrangien
augmente, Rapport Laboria 309, 1978.

19 Mercier B., Sur la Theorie et l'Analyse Numerique de

probl~mes de plasticite, Th~se d'Etat, Universite Pierre
et Marie Curie, Paris, 1977.

20 Chan T., Glowinski R., Finite element approximation and

iterative solution of a class of mildly nonlinear ellip-
tic equations,STAN-CS-78-674, Compupter Sc. Department,
Stanford University, 1978.

21 Begis D., Analyse Numerique de l'ecoulement d'un fluide

visco-plastique de Bingham par une methode de lagrangien
augmente, Rapport Laboria 355, 1979.

22 Bourgat J.F., Glowinski R., Le Tallec P., Decomposition

of Variational Problems and Applications in Finite Elas-
ticity, in Partial Differential Equations in Engineering
and Applied Science, R.L. Sternberg, ed., Marcel Dekker,
New-York, 1980, pp. 445-480.

23 Chan T., Glowinski R., Numerical methods for a clasR of

mildly nonlinear elliptic equations, Atos do Decime Pri-
meiro Coloquio Brasileiro de Matematica, Vol. I, IMPA,
Rio de Janeiro, 1978, pp. 279 318.

24 Le Tallec P., Numerical Analysls of Equilibrium

Problems in Incompressible Nonlinear Elasticity, Ph. D.
dissertation, The University of Texas at Austin, 1980.
 695

25 Glowinski R., Le Tallec P., Numerical solution of problems

in incompressible finite elasticity by augmented lagran-
gianmethods (I) Two-dimensional and axisymmetric problems,
submitted to SIAM J. Appl. Math.

26 Le Tallec P., Oden J.T., Existence and characterization

of hydrostatic pressure in finite deformation of incom-~
pres sible elastic bodies, TICOM Report, University of
Texas at Austin, 1979.

27 Ball J.M., Convexity conditions and existence theorems

in nonlinear elasticity, Arch. Rat. Mech. Anal., 63 (1977),
pp. 337-403.

28 Glowinski R., Le Tallec P., Numerical solution of problems

in incompressible finite elasticity by augmented lagran-
gian methods (II) Three-dimensional problems (to appear).

29 Glowinski R., Le Tallec P., Une methode numerique en

elasticite non lineaire incompressible, Comptes Rendus
Acad. Sc. Paris, t.290B (1980), pp. 23-26.

30 Ruas V., A class ofaxymmetric simplicial finite element

methods for solving finite incompressible elasticity pro-
blems, to appear in Compo Meth. Appl. Mech. Eng.

31 Ruas V., Sur l'application de quelques methodes d'elements

finis a la resolution d'un probleme d'elasticity incom-
pressible non lineaire, INRIA, Rapport de Recherche 24,
May 1980. -----

32 Chadwick P., Haddon E.W., Inflation-Extension and Ever-

sion of a Tube of Incompressible Isotropic Elastic Mate-
rial, J. Inst. Maths. Applics., 10, (1972), pp. 258-278.
Incompressible Finite Elements: The LBB Condition and the
Discrete Eigenstructure

D.S.MALKUS
IllmOis Institute of Technology, ChIcago, USA

Summary
The basic stability condition for mixed/Lagrange multiplier
variational principles in incompressible media problems is
the LBB condition. It plays an essential role in determin-
ing whether or not the problem is well-posed and governs the
choice of finite elements in the discretization. This paper
examines the discrete eigenstructure of a well-known La-
grange multiplier formulation for linear elasticity or
Stokes-flow. It shows how the weak incompressibility con-
straint is reflected in the elementary-divisor structure of
the eigenproblem whose solution determines the finite ele-
ment approximation to the natural modes. In this context
the discrete LBB condition can be seen to be a condition
determining the limiting disposition, as the mesh parameter
decreases, of a matrix pencil with infinite eigenvalues.
Pure pressure modes and the load vectors required to trans-
mit them are paired in cyclic subspaces of the infinite
eigenspace. This pairing can be related to a well-known
heuristic interpretation of the LBB condition. The rela-
tionship between the natural mode eigenproblem and the
eigenproblems which determine the norm of the inverse of
the discrete operator of a static or steady-flow problem is
described. Finally, because of the equivalence between
classes of mixed and penalty formulations, it is shown that
these results apply to penalty/reduced integration finite
element methods.

Introduction
This paper discusses certain aspects of the discrete
stability condition for finite element models in incompress-
ible elasticity or fluid flow. The term stability is used
to denote the uniform invertibility of the discrete operator
in the variational statement of a static or steady-flow
problem [1-4J. The approach taken here is to view the dis-
crete problem from an algebraic point of view. This approach
does have its limitations because, by nature, a stability
 697

condition is an analytic rather than algebraic concept. On

the other hand, stability conditions often have obvious alge-
braic consequences, such as the requirement that certain
matrices of coefficients in the discrete equations be posi-
tive-definite, or that others be of full rank [1-4]. The
purpose of this investigation, then, is to frame further
analysis of the discrete stability condition for incompress-
ible finite elements in as complete as possible an under-
standing of its algebraic consequences. As a result, it is
hoped that this will contribute to the continuing investi-
gations into determining which incompressible element formu-
lations are computationally reliable and which are not.

Lagrange Multiplier Method

Consider the following familiar variational equation char-
acterizing solutions to problems in incompressible elasticity
or steady Stokes-flow:

{ Find U E H';" ¥ V E: H (1 )
B(U,V) = F(V)
F E: H* is a standard load integral. H is the Hilbert space
HIo X HO• U = (B,p) ~ H is a displacement or velocity/
pressure pair, and the subscript "0" denotes that B may
satisfy appropriate essential boundary conditions. For the
purpose at hand, there is no loss of generality in assuming
that these conditions are homogeneous, so that H is a linear
space. H* is the dual of H. The form of B(U,V) is assumed
to be:
B[(B,p),(~,q)]= l[2YE: ij f ij -E-uq- fup]dD. (2)

where E: ••
1J
=~ (~Ui
uX.
+ ~Uj)
uX.
, f ..
1J
=~ (~Vi
uX.
+ :v j )
uX.
and repeated
J 1 J 1
indices imply tensorial contraction; p is shear modulus, and
-pip is the hydrostatic pressure function for elasticity;
p is a viscosity with p the pressure in Stokes-flow.
It can be shown that B(U,V) satisfies three important
conditions which guarantee that the problem of (1) is well-
posed, given appropriate conditions on the domain [3J and
698

assuming that the essential boundary conditions are such

that the Korn inequality [5] is satisfied. It also may be
required that the pressures p ~ HO satisfy some appropriate
constraint [3,4J. The three conditions are:

ii.

iii. B(U,V) = B(V,U)

\I • \I
H is the usual product norm on H, and Cl and C2 are
constants. The need for (i-iii) is established in [1-3J
and a proof that these conditions are satisfied by (2) is
gi ven in [4 J.
Here consideration is given to the discrete problem aris-
ing when (1) is discretized using a finite element trial
space Shx Th C H, with trial functions in Sh satisfying
appropriate essential boundary conditions and pressures in
Th satisfying any necessary constraints. The mesh paramete~
h, tends to zero as the mesh is refined. Conditions i and
iii of (3) carryover to the discrete model and condition
ii becomes

(4)

The stability condition of prime importance to the conver-

gence of the discrete model [1-4J is
lim inf C~ ~ C; > 0 (5)
h-o
which is the uniform invertibility requirement.

The LBB Condition

Under the conditions put forward in the previous section
it has been shown [1,2J that the condition (3ii) will be
satisfied if and only if
(6 )
 699

( . , .) denotes the HO-inner product on ll. This has dis-

crete analogue

sup
h
I ( E: I:.
J.J.
,JIll ~ ch3
Ilu'" II H1=1
Inequality (6) is the continuous LBB condition and inequality
(7) together with the requirement
lim inf c~ ~
h-+o
c; > 0 (8)

is the discrete LBB condition. That these conditions

guarantee the satisfaction of the previously stated conditions
in (3ii) and (4-5) follows from the results of [1 and 2J and
is implicit in the derivations of [3J. That the discrete LBB
condition guarantees uniform invertibility will also follow
from the work presented here.

Matrix Equations of the Finite Element Model

B(U, V) = F(V) leads to matrix equations ~U = F which have
the following partioned form when the appropriate nodal
ordering is chosen:

(9)

Let u,p,etc. denote vectornof nodal values of trial functions

Bh, ph, etc. Then the matrices of (9) are implicitly defined

uT ~2 v = 2(1)
.n.
E-~j~jda ¥Bh ,r E: Sh

T (10)
u ~l P
T
v F=

The following matrix will also be useful:

pT .9q =
.n.
1
phqh dll >fph,qh E ~ (11)

These matrices are evaluated with respect to standard f.e.m.

700

nodal bases of Sh and Th.

The matrix ~2 plays a special role in this investigation.
Because of the assumption that the Korn inequality [5J is
satisfied, it means that there exist constants a- > 0 and
S > 0, independent of h such that
(12)

This uniform equivalence between the strain energy and the

Hl-norm allows substitution of the strain energy norm for the
Hl-norm without loss of generality. Furthermore, the matrix
Q of (11) gives the HO-norm on~. This means that without
loss of generality in 0-4 and 6-7), Iluhll H can be replaced
by
(13)
where

(14)

In what follows the mass matrix ~ of Sh will be needed:

uT ~ v = ~)dh ~h dD Jrf )dh ,~h E: Sh (15)

.fl.
Three Eigenproblems
A. Natural Modes

l[y
Define the weakly constrained Rayleigh quotient

E ij E- ij - EiiPJdIl.
R(u,p) - (16)

It is easy to see that this has discrete equations of

stationarity

~u (17)
or

(18)
 701

B. Convergence
Assuming the consistency of the finite element approxima-
tion, that is assuming that Sh x Th can accurately approximate
arbitrary members of H [3J, what is needed to establish the
convergence of the finite element solutions to (9) towards
the exact solution is verification of condition ii* (equation
4). Using the norm given in (13), the satisfaction of ii*
can be determined from the following eigenproblem:

(19)
The following fact is surely well-known by most, and is
stated here as the first theorem for the sake of completeness.
The proof of this result and the proofs of all other theorems
stated here may be found in an extended version of this paper
[6 J.
THEOREM 1: I f I A 11 .s. I A 21 .s. ... .s. I ). N I are the eigen-
values of (19), then
inf sup IB(Uh,Vh)1 1/..11
II vhll=l Iluhll=l

C. LBB Eigenproblem
The eigenvalues of the following eigenproblem determine
the satisfaction of the discrete LBB condition:
(20)

where
(21)

which eigenvalue determines the inf sup in (7) is established

by
THEOREM 2: Let k = dim ~ and n = dim Sh. Let O'l.s. 0""2.s.
•.• .s. O'k be the k largest eigenvalues of (20). The inf sup
of (7) (w. r. t. II· II) is given by VaI
if k .s.n, and 0 if k> n.

The purpose of the remainder of this paper is to establish

the relationship between eigenproblems A, B, and C. It turns
out that the natural mode problem A and the convergence
eigenproblem B are closely related through the LBB eigen-
problem C. In order to see this it is necessary to take
702

into account the form of eigenproblem A, which is rather

different from the other two. It is required to draw on
knowledge of the canonical form of matrix pencils of arbi-
trary divisor structure [7-9J. This is because of the singu-
lari ty of I:1Ei)0 and the indefiniteness - and even possible
singularity of K. There may fail to be a complete set of
eigenvectors. 'if c~ = 0 for some h in (7), there may not
even be a full set of eigenvalues.

The Eigenstructure of Problem A

If Dl = 0 for some h, the LBB condition is not satisfied
as stated. This can happen for a Dirichlet problem [3,4J,
but can be easily remedied by applying an appropriate con-
straint to the pressure unknowns. This reflects the fact
that in such a problem, in which there are no stress boundary
conditions, the pressure is determined only up to an ambient
pressure which is nowhere specified. If it is not already
apparent, it will be shown later that the lack of such a
constraint will render ~ singular by virtue of a rank de-
ficiency in ~l of order one (a constant pressure). For
practical purposes in obtaining solutions, one might want to
impose the constraint explicitly, but in theory the full
statement of the continuous LBB condition can be restricted
to range over pressures which are in the quotient space of HO
by the nullity of the adjoint of the divergence operator [4J.
This has the effect of removing the constant pressure mode
from consideration in (6-7).
Also 0i = 0 i f k = dim ~ > dim Sh = n. This can happen for
grossly inconsistent (overconstrained) choices of element
type and pattern for Bh and ph [lOJ. This can also happen
for other element types and patterns - if for very large h -
the number of essential boundary conditions "uses up" too
many degrees of freedom on Bh. But these last two instances
are well-recognized and easy to avoid. Perhaps the most
interesting case for which Dl = 0 is the case in which the
discrete LBB condition as stated in (6-7) is purposefully
not satisfied. This is done in order to decrease the con-
straint count by rendering many of the constraints redundant.
 703

The elements of crossed-triangle pattern used by B. Mercier

[llJ are of this type. The idea is to enlarge the null-space
of the discrete divergence operator (represented in weak form
by ~l). In [11 J a consistency proof is given for the quad-
ratic member of this family of macro-elements, but the possi-
ble effect on the stability of such a choice of elements is
-T
not discussed. By enlarging the null-space of ~l beyond the
size of the null-space predicted by the difference between
column and row dimension, one necessarily introduces a rank
deficiency of corresponding order to the adjoint of the dis-
crete divergence operator (represented weakly by ~l)' This
rank deficiency is introduced regardless of whether the ad-
joint of the continuous operator has a null-space, and is
potentially of high order (as opposed to the Dirichlet case
in which the rank deficiency is of order 1). So assume the
following dimensionalities:
n = dim Sh, n >0
{ k = dim ~, 0 <k <n ( 22 )
m dim Ker ~l' 0 .$. ];l < k
From (19) one may easily see that (}l = 0'2 = ••• = crm = O.
Defini tion: Uh e, Sh X Th is "mean isochoric" if Uh =
. T T - T
[u : p J and ~l u = O.
Mean isochoric is used to denote those fields for which the
~h part satisfies the weak form of incompressibility con-
straint. The word 'mean' indicates that these fields may not
be exactly incompressible, though the reverse is true: iso-
choric implies mean isochoric.
THEOREM 3: There are n-m eigenvectors of problem A. There
are n+k-m eigenvalues, n-k+m of which are finite, and
2(k-m) of which are the infinite repeated root. There
are n-k+m mean isochoric modes with finite frequency and
k pure pressure modes, of these k-m are of infinite fre-
quency. The other m pure pressure modes are ill-disposed
and of indeterminate frequency. Problem A is defective
of k eigenmodes: There are k-m 2 x 2 Jordan blocks in the
infinite eigenspace, each with a pure pressure mode as an
eigenvector and a dilatational mode as cyclic generator.
The remaining deficiency in eigenvectors results from the
m ill-disposed pure pressure modes, each of which corre-
sponds to a row and column minimal index of zero.
The results of THEOREM 3 can be put into the following
704

picture of the canonical form of (18) under the equivalence

of singular pencils [7J:
COROLLARY 3.1: There exist non-singular matrices Band §
of real numbers such that B(ts - A [Jj tl) 0 J)§ =

m { 0
_ ~Ol
I
___ +- ----
I

[~ ~J
I
I

1
2(k-m) I • (23)

t~ ~]I ___ _
I

1- ;..
I n-k+m-A
n-k+m
{
This is a block diagonal matrix. Un-
specified entries are zero.
It is useful here to introduce an auxilliary eigenproblem:
*
B. Adjoint LBB eigenproblem
- T -1 -
lSI Jj lSI P = A Qp (24)
Let ql' ... ,qn be the null eigenvectors of (24) and PI'··· ,Pk-m
be the eigenvectors of (24) with nonzero /I.. Let" 1" refer
to orthogonality in the norm of ~ (given via Q):
Lemma 3.2.1: The qi span the kernel of lSI' Ker lSI' and
Pi span the coimage, Colm RI = {Ker ih~ L. For each Pi'
the vector z.l = M- I rvKl p.l is dilatational, i. e., KIT z l. /.L 0,
t"V I"V

and each p. is the dilatation of the corresponding Zl. in

l_l -T
that p. = Q Kl z. / A ..
1 '" '" 1 1

The lemma provides an orthogonal basis for the pressures in

which each well-disposed pressure is obtained by taking the
dilatation of a corresponding dilatational mode in similar
fashion to the manner in which pressures are calculated in a
penalty method [lOJ. Other choices of basis for the pressures
are of course possible.
COROLLARY 3.2: Let the Pi and qi be as in lemma 3.2.1,
then
a) The infinite eigenvectors can be taken to be
 705

x. -- [__C2._]
1 Pi

b) Each pure pressure mode in (a) is generated by a

cyclic vector

in the sense that

I"V
Kx.1 = [M 4'
~
O]y.1
f'V
(25)
c) A Q-orthogonal basis of the ill-disposed pressure
modes is

[-~~-] 1

d) Each mean isochoric mode is of the form

[-;~-l 1

where si is some pressure vector.

COROLLARY 3.3: The pure pressure modes satisfy the follow-
ing static or steady-flow problems

a) !\ [-:~-l [-~~i-'-l (26)

b) JS [-~~-l = 0 (27)
1

ThlS implies, in particular, if m> 0 JS is singular

with rank deficiency m.

Interpretation of THEOREM 3 and COROLLARIES

Each of the k-m independent constraints removes a dilata-
tional mode M- l Kl p.1 from the eigenvector space by "tying it
~ ~

up" in a nonlinear divisor. The fact that for any pure pres-
sure mode [ 0:. p T]T

[0 ! pTJ[!'J e oj [-;-] = 0 (28)

reflects the fact that such a mode is transmitted without

kinetic energy. The infinite frequency of the k-m well-
706

disposed pressure modes represents their infinite acoustic

velocity. If O'i = 0, the acoustic velocity is indeterminate.
If the uniformity condition (7) fails, then there is one or
more ~i ~ O. This implies that, in the limit, the eigen-
problem A is tending to a more poorly disposed limiting pro-
blem which has proportionately fewer instantaneously propa-
gating pressure modes, and more pressure modes of indeter-
minate frequency. These are extraneous artifacts of the
f. e. m. model.
With each instantaneously propagating pressure mode, be-
sides the dilatational mode it removes from the eigenvector
-T : 0 JT . This can be
space, there is the closely related [ PiT ~l
considered as a system of distributed loads which generates
the corresponding pure pressure mode as a solution to a
static or steady-flow problem, according to (26). Then (27)
is the negative of the familiar heuristic interpretation of
the LBB condition as the assurance that no pressure mode can
be transmitted by zero load - except perhaps the ambient con-
stant pressure in a Dirichlet problem. Equation (27) says
that the ill-disposed pressures are transmitted with zero
load. Again the failure of the uniformity condition (7) has
the obvious interpretation for the limiting problem.

The Eigenstructure of Problem B

Eigenproblem B is non-defective because ~ is positive
definite. In fact, making use of the Cholesky decomposition
of ~, it is easy to see that there is a complete set of
eigenvectors associated with real, finite eigenvalues. In
more detail
THEOREM 4: There are n+k eigenvectors of problem B.
They are classified as follows
a) Al = ),2 = ••• = Am = 0 with eigenvectors

[-~~-l
1

b) "\Am+1 = A m+ 2 = • •• = An- k +m = 1 with eigenvectors

[-~~-l
 707

such that K1T u.l = 0 (mean isochoric).

A:
~

c) k-m dilatational modes with l

> 0,

-1 - T
Q
~
Kl u.l
~

1 + vll + 4cri
with A = A~ = 2

d) k-m dilatational modes with Ai < 0,

ui

-1 - T
Q
~
Kl u.l
~

1 - vll + 4cri
wi th /,. = 1\ i = 2

COROLLARY 4.1: The eigenstructure of problems A and B

can be put in 1-1 correspondence as follows:
a) The m ill-disposed eigenvectors of A and the eigen-
vectors with A = 0 in eigenproblem B can be taken
to be the same.
b) The Bh parts of the mea~isochoric modes of problem
B can be chosen as the u parts of the mean isochoric
modes of problem A. ~
c) The 2 x 2 Jordan blocks of problem A can be put in
correspondence with the invariant subspaces of di-
mension 2 in problem B generated by the eigenvectors
of A~, I\i
pair. Each such invariant subspace of
problem B contains a pair of vectors

[9=1-~Y::-] and [-~:-l

Note that the (c) part of the corollary does not establish
the identity of M- l Kl p.l in problem A with the u.l here.
~ ~

But these two vectors pay a closely corresponding algebraic

role in each eigenproblem: 1) Each dilatational moie is
"tied up" in an invariant subspace with a corresponding
pressure; 2) each pressure is generated by the dilatation
of the corresponding u i ' in the sense of lemma 3.2.1; 3)
the well-disposed pressures are HO-orthogonal to the ill-
708

disposed pressures. Note that in the convergence eigen-

problem B, the orthogonality of the ill-disposed to the well-
disposed ones arises naturally, rather than by the arbitrary
basis choice of problem A.

Interpretation of THEOREM 4 and COROLLARY

For each <Jl = <J2 = ••• = <Jm = 0, there is a Ai = o. If
Of ~ 0, this shows up as a negative eigenvalue, Ai, which
tends to zero as the mesh is refined. The deterioration in
the disposition of eigenproblem A and the corresponding limit-
ing indeterminacy of the frequency of a pressure mode shows
up here in the failure of one or more negative eigenvalues of
problem B to be bounded away from zero. Since the signature
of eigenproblem B is the same as the signature of ~, these
negative eigenvalues correspond to negative eigenvalues of ~.

Penalty Methods
The work of this author and T. J. R. Hughes [lOJ has shown
that the penalty/reduced-selective integration technique pro-
duces a static or steady-flow problem
(~2 + z ~l)u = F (29)
~2 is the same matrix as in (9), and z is a large penalty
parameter. When ~ is constructed based on the integration
points of Sh used to evaluate ~l as described in [lOJ, then
relation (21) holds, and (29) produces the same solutions as
a slightly perturbed version of (9) [4,10J. More recently
Hughes and Malkus [12J have established the correspondence
between penalty methods and Lagrange multiplier methods,
which allows a much broader choice of pressure trial spaces
~, yet yields relation (21) relating penalty methods to
Lagrange multiplier methods. So in what follows assume that
~l in the penalty method is given by (21) for Rl and g de-
fined for an associated Lagrange multiplier method which also
has the same ~2 and F.
The penalty method (29) has the following eigenproblem
which determines the norm of the inverse of the discrete
operator:
 709

(30)
This rearranges to
,lS1 u = cr,lS2 u (31)
where
A -1
z (32)

Therefore the LBB eigenproblem C is the appropriate discrete

stability condition for the penalty method. In more detail:
THEOREM 5: The eigenproblem in (30) determines the
uniform invertibility of the operator in (29). If
0- 1 < ()" 2 < ••• < c),k are the non-zero eigenvalues of
m+ - m+ - -
the LBB eigenproblem, the eigenvalues of (30) are
T
a) Al = ).,2 = ••• = An-k+m = 1 with u i such that ,lS1 u i = 0
(mean isochoric).
b) k-m eigenvalues with ).1 = 1 + z O""i and dilatational
eigenvectors.
Because pressures in a penalty method are computed from the
dilatational part of the solution to (29), the mill-disposed
pressure modes are" invisible" to a penalty method.
COROLLARY 5.1: All pressures computed from a solution
to (29) are in the span of the vectors [0: p,TJT which
span Colm jh. 1

Furthermore it is clear that the matrix ,lS2 + z,lSl is directly

invertible for any m.

Interpretation of THEOREM 5 and the COROLLARY

Penalty methods avoid the non-invertibility of,lS. They
avoid the possible contamination of the computed pressures
with ill-disposed modes. Therefore the direct violation of
the stated LBB condition by having m > 0 is not necessarily
fatal to such a method. B. Mercier [llJ claims success
with his family of crossed triangular elements in a variety
of problems. A simple calculation for linear crossed tri-
angles shows that m is equal to the number of circumscribed
quadrilaterals, with ill-disposed modes having a variety of
oscillating patterns. Mercier uses a penalty method for his
calculations. He provides a proof for one member of the
-T
family, which shows that in Ker ,lS1 ' a di vergenceless field
710

Bh can be approximated with optimally consistent accuracy.

With other arrangements of linear triangles the dimension
-T
of Ker ~l would not be large enough.
The work presented in this paper reminds us of two more
results which need to be established for such a penalty
method. First, the remaining k-m well-disposed pressure
modes must be capable of approximating arbitrary pressures
in HO to appropriate accuracy. Second, the resulting method
must still be stable, in the sense that the LEE condition
must be satisfied on {Ker &1\1, as would be reflected in
the uniform boundedness away from zero of CJm+l ~ CJm+2 ~ ...
~ ~ with decreasing h. If this condition is satisfied,
note that THEOREM 5 implies that the dilatational modes can
be given a frequency Ai = 1 + z "i which is arbitrarily
large, independent of h. If any of the ~+l' .•. 'Ok tend to
zero, then the cardinal sin of penalty methods is committed:
For a fixed z, the separation between the two components of
the operator spectrum is a function of h.

References
1. Eabu~ka, I.; Oden, J. T.; Lee, J. K.: Mixed-hybrid
finite element approximations of second order elliptic
boundary-value problems. Compo Meth. !2E!.. Mech. ~.
11 (1977) 175-206.
2. Erezzi, F.: On the existence, uniqueness and approxi-
mation of saddle point problems arising from Lagrangian
multipliers. R.A.I.R.O. 8 (1974) 129-151.
3. Eabu~ka, I.; Aziz, A. K.: Mathematical Foundations of
the Finite Element Method. New York: Academic Press 1972.
4. Eercovier, M.: Perturbation of mixed variational pro-
blems. Application to mixed finite element methods.
R.A.I.R.O. 12 (1978) 211-236.
5. Hlava~ek, I.; Ne6as, J.: On inequalities of the Korn
type. Arch. Rat. Mech. Anal. 36 (1970) 305-334.
6. Malkus, D. S.: Preprint, available from author.
7. Gantmacher, F. R.: Matrix Theory II. New York, Chelsea
1971.
8. Moler, C.; Stewart, G. W.: An algorithm for generalized
matrix eigenvalue problems. S.I.A.M. ~ Numer. Anal.
10 (1973) 241-256.
 711

9. Stewart, G. W.: On the sensitivity of the eigenvalue

problem Ax = ABx. S.I.A.M. J. Numer. Anal. 9 (1972)
669-686. - ---
10. Malkus, D. S.; Hughes, T. J. R.: Mixed finite elements-
reduced and selective integration techniques: a unifi-
cation of concepts. Compo Meth. ~ Mech. Engng. 15
(1978) 63-81.
11. Mercier, B.: A conforming finite element method for
two-dimensional incompressible elasticity. ~ ~ ~
~ ~ 14 (1979) 942-945.

12. Hughes, T. J. R. and Malkus, D. S.: Manuscript in

prepara ti on.
Part VI:
Computational Algorithms
Data Management in Finite Element Analysis

P.l. PAHL
Techmsche Umversitiit Berlin, Germany

Summary
The global data for finite element analyses are conven~ently
stored in matrices and table trees. In order to optim~ze the
use of the available storage capacity, the matrix and table
tree data structures are ~mplemented on the basis of a common
storage structure. The paper describes the properties and
functions of storage, matrix and table tree managements for
fin~te element programs.

1. PROPERTIES OF FINITE ELEMENT DATA

1. 1 Introduct~on

During the past years, several technical application programs

us~ng different storage and data structures have been devel-
oped at the Institut fUr Allgemeine Bau~ngenieurmethoden [1-8].
In connect~on w~th a proposed program to develop software for
nonlinear finite element methods in a cooperative effort at
several independent institutions, we have evaluated our ex-
periences with these programs. In addition, we have reviewed
publications of comparable programs [9-17]. Th~s paper con-
tains an extract of those aspects of global data management
which we consider important for a jo~nt research effort in
nonlinear finite element analysis.

1.2 Global Data

The global data of finite element analyses can be categor~sed

as follows:

Descriptive data ~n tabular form, ego nodal coordinates,

element topology and loading conditions.
716

computed element properties, ego element stlffness matrices

and element load vectors.

System equations ln matrix form, ego stlffness matrices, in-

cremental and total dlsplacement and load vectors.

Computed element results in tabular form, ego incremental

and total strains and stresses.

Data created by postprocessing, ego for graphics, in tabular

form.

ThlS list shows that tabular and matrix data structures are of
equal importance ln finite element programs.

1.3 Working Set of the Global Data

The worklng set of the global data consists of that part of

the global data which resides in primary memory at a particu-
lar stage of the analysls. In general, the required working
set is only a subset of the avallable working set. For an In-
core solution, all required working sets are subsets of the
avallable working set. If storage for an in-core Solutlon is
not available, the algorithms should keep the working set as
constant as possible to mlnimize the access to secondary sto-
rage.

The following table shows that the required worklng set chan-
ges signlficantly during the analysls. In particular, the re-
latlve volume of tables and matrlces in the working set chan-
ges. The use of the available prlmary memory for out-of-core
solutions is therefore optimized if the table and matrix data
structures are implemented on the basis of a common array sto-
rage structure. Space released by matrices can then be used
for tables, and vice versa. This concept is the main feature
of the data management proposed in this paper.
 717

Worklng Sets of Global Data for Flnlte Element Analysls

Phase of the Analysls Tables Matrlces

Descrlptlon of the structure x

Checklng of the lnput values x
Computatlon of the element propertles x x
Assembly of the system equatlons x x
Solution of the system equatlons x
Incrementation of the varlables x x
Postprocesslng x

2. STORAGE MANAGEMENT

2.1 Storage Structure

The flnlte element program and its data are stored in a pro-
gram area on prlmary memory and in data sets on secondary
memory, which are asslgned through the operatlng system.

In prlmary memory, the global data are separated from the

procedures elther by definlng a COMMON array, or by passlng
a large FORTRAN array through the argument llStS of the sub-
routlnes. This array is divided lnto an administratlve zone
and a data pool. The admlnlstrative zone contains admlnlstra-
tive tables of the storage management, the matrlx management
and the table management. If the operating system is vlrtual,
real memory lS subdlvided into pages. These are automatlcally
managed by the operatlng system.

On secondary memory, the global data are stored in one or

more data sets. The data sets are subdlvided into records,
which are accessed wlth a data management system of the opera-
ting system. The storage management of the finlte element pro-
gram dlvldes the data sets into flies. Each file is a consecu-
tive set of records contalning a specific type of data, ego a
global matrlx or table trees. The flle lS ldentified by a
unique number.
718

All f1les are subdivided into arrays of equal size, as shown

in F1g. 1. The array Slze should be a multiple of the record
size. The arrays of a file are numbered consecutively. Thus
every array of the finite element program is identified by a
unique duplet (file number, array number). Arrays are assigned
by the storage management upon request by the matrix and table
managements, as described in sections 3 and 4 of this paper.

The data pool is subdivided into blocks, whose size equals

that of the arrays. For virtual systems, the block size should
be a multiple of the page Slze. The blocks of the pool are num-
bered consecutively. During execution, blocks are assigned by
the storage management upon request of the matrix and table
managements. These blocks contain the working set of the arrays.
Two special blocks, the scratch pad and the auxiliary array, are
reserved for the matrix management.

2.2 Administrative Data of the Storage Management

The administrative data of the storage management are stored
1n the file table and the pool block table shown in F1g. 2.
The f1le table 1S 1ndexed w1th the f1le number. Each llne des-
cribes the location, Slze and occupation of one file. The pool
block table is indexed with the block number. Each line des-
cribes the status, location and present contents of one block
1n the pool.

The status parameter of the blocks is used to control the re-

sidence of selected arrays. It differentiates between empty,
free and f1xed blocks. An empty block is assigned without sto-
ring its present contents. A free block can only be reassigned
after its contents has been saved in a file. A fixed block can-
not be reassigned. The matrix management, for instance, controls
the pool residence of hypermatrices by storing them in fixed
blocks.

2.3 Ass1gnment Algor1thms for Arrays

The arrays of a file are assigned sequentially. Individual ar-
rays cannot be deleted or reassigned. The file can, however, be
deleted and reused as a whole.
 719

The assignment of blocks ~n the pool ~s controlled with pr~ori­

ties. In a first loop, the attempt ~s made to assign an empty

block. The search starts w~th the block follow~ng the last
assigned block and ends after all blocks have been considered.
If the end of the pool (except the reserved blocks) is reached,
the search recommences at the front of the pool.

If the pool does not contain an empty block, the attempt is

made in a second loop to assign a block which contains a free
array. The sequence of the search is identical with that of
the f~rst loop. If a free block ~s encountered, its present
contents is wr~tten into its file before the block ~s reass~g­

ned. The transfer to the f~le ~s suppressed ~f the array was

previously stored in the file, and has not been changed by wri-
ting.

If the pool contains neither empty nor free arrays, the exe-
cution ~s term~nated due to lack of storage space.

2.4 Funct~ons of the Storage Management

The storage management is called to execute the follow~ng func-

t~ons:

DEFINE POOL
An empty pool block table ~s set up.

STORE POOL
All arrays ~n the blocks of the pool are stored ~n the~r files.

DEFINE FILE
The adm~nistrat~ve data for a specif~ed f~le are entered in
the f~le table.

RELEASE FILE
All arrays of the specified f~le are released in the pool. The
file ~s set to empty.

DELETE FILE
The specified f~le is released. Its definition ~n the file
table ~s nullified.

STATUS FILE
The parameters describing the file ~n the file table are re-
turned.
720

ASSIGN BLOCK
A block is assigned in the pool by the algorithm described in
section 2.3. Its number is returned.

FIX BLOCK
The contents of the specifled block is fixed ln the pool: the
block lS not available for reassignment.

FREE BLOCK
The specified block is avallable for reassignment.

RELEASE BLOCK
The status of the speclfled block lS set to the value empty.

STATUS BLOCK
The parameters describlng the block in the pool table are re-
turned.

ASSIGN ARRAY
A new array lS assigned in the specified flle by the algorlthm
of section 2.3. Its number is returned.

GET ARRAY
The specifled array lS transferred from the flle to a specl-
fled block ln the pool.

PUT ARRAY
The array stored in the speclfled block lS transferred to the
flle reglstered in the pool block table.

STATUS ARRAY
If the array resides in the pool, the number of the block ln
which it is stored is returned, otherwise the value zero.

3. MATRIX MANAGEMENT

3.1 Hypermatrix Data Structure

Nonlinear finlte element analysls by incremental techniques

involves programmlng and executlon of volumlnous matrix opera-
tions. For out-of-core solvers, the matrix operatlons are
assoclated wlth conslderable data flow. Sparslty and symmetry
properties of matrices must be consldered in the algorlthms
ln order to mlnimize the volume of data and the number of
operations. Frequently, thlS lS achieved through the well-
known hypermatrix data structure [13].
 721

In the hypermatrix data structure, each matrix is divided lnto

rectangular or square submatrlces. Usually, all submatrices of
a given matrix are of the same Slze. The coefficients of a sub-
matrix are stored consecutively (columnwlse). Profiles are
stored to descrlbe the populated parts of the submatrlx. The
hypermatrix contains the submatrlx numbers. The prlnciple is
lilustrated In Flg. 3.

It is useful to introduce the hypermatrlx scheme as a common

data structure for all maJor matrlces of the finite element
program. The data flow and the matrlx algebra can then be pro-
grammed through a standardlsed set of functlons (eg. subrou-
tlne calls) WhlCh are readlly lmplemented In different enVlron-
ments by replaclng mlnor subroutlnes of the matrlx management
with well-defined functlons and parameter 11StS. The program
therefore becomes hlghly portable and adaptable to dlfferent
computer capacltles.

3.2 Storage of Matrlces In Arrays

Each matrlx of the flnlte element program lS identifled by a

unlque matrlx number and lS assoclated wlth a speciflc flle.
As shown In Flg. 4, thlS flle contalns only data for the asso-
clated matrlx. The file can readlly be reused lf the matrlx lS
deleted, ego durlng lteratlon.

The submatrlces of a matrlx are ldentifled by consecutlve num-

bers, WhlCh are automatlcally asslgned when the submatrlces
are deflned. These submatrlx numbers are stored In the hyper-
matrlx In the locatlon correspondlng to the hyperindlces of
the submatrlx (see Fig. 3).

Slnce different matrlces of a set of equatlons may be dlvided

into submatrlces of slgnlficantly different size (eg. S * S
for matrix K and S * L for matrlx U In Flg. 3), whereas all
arrays are of equal size, a varlable number of submatrlces lS
stored per array. A speciflc submatrlx lS located by ltS array
number and the displacement d l shown In Flg. 4.
722

In order to mark empty rows and columns and to avoid unnec-

essary operations, the matrix management sets up profiles of
hypermatrices and submatrices (Fig. 3), which are stored to-
gether with the coefficients (Fig. 4).

3.3 Admin~strative Data of the Matrix Management

All matrices of a fin~te element system are recorded in the

matrix table shown in Fig. 5. Each line of the table describes
the hypermatr~x, submatrices and administrative tables of one
matrix. During execution, the matrix table resides in the ad-
ministrative zone of primary memory.

For each defined matrix, the management builds up a hyperma-

tr~x containing the submatr~x numbers, and a submatrix table
containing the submatrix locat~ons. During execution, a block
table is built up to record the numbers of the blocks in which
the arrays of the matrix are stored ~n the pool.

The hypermatrix, the submatrix table and the block table are
usually stored in fixed blocks of the pool. Between runs, the
hypermatr~x and the submatrix table are stored ~n the f~le

associated with the matr~x.

3.4 Access Algorithm for Matrices

The submatr~x, which is to be accessed, is identified by the

matrix number and its hyper~ndices. These are used to find
the submatrix number in the hypermatrix.

If the submatrix number is zero, the submatrix is not defined.

A new array is then defined (if requ~red), the next submatr~x

number is assigned, the displacement in the array is determ~ned

and the administrative data are entered in the tables.

If the submatrix is defined, its array number is read in the

submatr~x table. Th~s ~s used to read the block number in the
block table of the matrix. If the block number is zero, a block
is assigned and the array loaded from the file. For positive
block numbers, the file and array numbers are checked against
 723

the entrles in the block table of the storage management. If

the block has been reassigned since the last access to the sub-
routlne, they will not cOlncide. In thls case, a new block must
be assigned and the array reloaded from the flle.

Finally, the pool lndices of the first coefficient and of the

profiles of the submatrix are returned. Hypermatrlces are
accessed In a simllar manner.

3.5 Functions of the Matrix Management

3.5.1 Conventions

All functlons of the matrlx management are subJect to a com-

mon set of conventlons:

Matrices, for which the executlon of an operatlon lS re-

quested, must prevlonsly have been deflned.

All matrlx functions (with the exception of SET, ADD and

GET GROUP) refer elther to complete submatrices or to a
complete matrix.

If a submatrlx requlred to perform a requested functlon

is not available in the pool, It is automatlcally deflned
(if applicable) or loaded from the flle.

The hypermatrlces, proflles and admlnlstratlve tables are

updated automatically.

3.5.2 General Functlons

In additlon to the matrix algebra performed on submatrlces,

the management supports a number of general functlons for ma-
trlces:

DEFINE MATRIX
The propertles of the matrlx are entered in the matrlx table.

RELEASE MATRIX
All arrays of the matrlx are deleted. The flle is set to empty.

DELETE MATRIX
The matrlx lS released. Its deflnltlon In the matrlx table lS
nullifled.
724

STORE MATRIX
The arrays of the matrix are stored 1n the associated f1le. All
blocks assigned to the matr1x are released.

PRINT MATRIX
The hypermatr1x and adm1n1strat1ve tables of the matr1x are
pr1nted.

STATUS MATRIX
The contents of the assoc1ated row in the matrix table is re-
turned.

INDEX MATRIX
The pool indices of the first element of the hypermatr1x and
its profiles are returned.

DEFINE SUBMATRIX
The spec1f1ed submatr1x 1S def1ned and recorded.

SET SUBMATRIX
All coeff1c1ents of the submatr1x are set to a g1ven value.
The profiles are set to zero.

PRINT SUBMATRIX
The coeffic1ents and the prof1les of the spec1fied submatr1x
are printed.

INDEX SUBMATRIX
The pool 1nd1ces of the f1rst element of the submatr1x and of
its prof1les are returned.

SET GROUP
The coeff1c1ents of a specified group are set to g1ven values.

ADD GROUP
The coeff1cients of a specified group are mod1fied by adding
given values.
GET GROUP
The values of the coefficents of a specified group are re-
turned.

The group funct10ns are requ1red primar1ly to assemble matr1-

ces and to read results selectively (eg. d1splacement compo-
nents at a node). The group is specified as shown in Fig. 6.
It must lie totally within a submatrix.
 725

3.5.3 Matrix Algebra

The algebra~c functions of the matr~x management are always

performed on complete submatr~ces. S~nce the solution algo-
r~thms are frequenthy programmed for symmetric matr~ces, all
funct~ons can be requested both for the stored submatr~ces A
or R and for their transposes. The d~mens~ons of the submat-
rices are automatically read in the data structure. The user
must make sure that the requested operat~ons are perm~tted

by the rules of matrix algebra.

The proposed set of funct~ons ~s shown ~n Fig. 7. It is as-

sumed that all values are real and of the same accuracy. The
programmer can store intermediate results in the scratch pad.
In addition, he can use one of the operands to store the re-
sults. The matrix management automat~cally recogn~ses this
condition and uses the auxil~ary array to store temporary re-
sults ~f necessary.

4. TABLE MANAGEMENT

4.1 Table Tree Structure

Tables are rectangular data structures. A spec~fic value is

identified by a triplet: table number, line number and column
number. It will be assumed that all values ~n a table are of
the same type and length. In add~tion, it will be assumed that
the number of columns is f~xed at definition, whereas the num-
ber of lines is dynam~cally adapted to the space requirements.

Typical global values of a finite element program are ~dent~­

f~ed by two or more ind~ces, for example:

Nodal coordinates node, component

Nodal loads node, loading condit~on, component
Element loads element, load~ng condit~on, face, component

Some of the ~nd~ces assume consecut~ve values, ego the compo-

nent numbers of a vector. Other ind~ces may be sparsely popu-
lated, ego external node and element numbers.

Due to the variable number of indices and the sparcity of some

of the indices, the descript~ve data and the computational re-
726

suIts of a finite element analysis are not conveniently stored

1n 1nd1v1dual tables. Instead, related tables are cha1ned w1th
pointers to form table tree structures. This 1S 111ustrated in
F1g. 8 for nodal data.

The root of a table tree is always a number table. It 1S used

to compress a sparse external 1ndex (eg. node numbers) into a
consecutive set of internal 1ndex values. All other tables of
the tree are data tables conta1n1ng values and p01nters (eg.
nodal data such as coord1nates, loads and d1splacements).

4.2 Storage of Trees 1n Arrays

For iterative nonlinear analyses, the data are so voluminous

that secondary storage 1S frequently requ1red. The tables are
therefore divided into subtables containing a mUlt1ple of com-
plete Ilnes of the table. Each subtable is stored 1n an array
of 1tS own. In the pool, the table management competes w1th
the matr1x management for storage space 1n the ava1lable blocks.

Each table of a finite element program 1S 1dent1f1ed by a

un1que table number and 1S assoc1ated w1th a spec1f1c f1le. The
table is stored 1n arrays of th1S one f1le only. The file may
also contain other tables. The arrays ass1gned to a table in a
f1le cannot be reused, unless the file as a whole is deleted.

4.3 Administrat1ve Data of the Table Management

The relationship between the arrays of a table tree are de-

scribed with pointers, which are stored in the subtables and
conta1n two values:

pOinter (1) number of the array conta1ning the dependent

subtable
pointer (2) number of a Ilne w1th1n th1S subtable

In add1t1on to the p01nters, the table management uses a table

to store administrative data describing the trees. Each line
of this table conta1ns the following parameters for a spec1f1c
table:
 727

Administrat~ve Table for the Table Management

Column Contents of one l~ne of the table

1 f~le number of the table

2 type of the table
3 number of columns ~n the table
4 number of l~nes per subtable
5 number of the last ass~gned array
6 number of the last l~ne assigned ~n the subtable
7 number of the preceed~ng table ~n the table tree
8 column number of pointer ~n the preceed~ng table
9-20 6 pa~rs of descriptors: array number,block number

Number tables consist of at most 6 subtables. Both the array

number and the block number (~f resident ~n the pool) of all
subtables are stored ~n the adm~n~strative table.

Data tables cons~st of an arb~trary number of subtables. At

most 6 subtables are stored in the pool at a given t~me. The
numbers of these arrays, as well as the numbers of the blocks
~n wh~ch they are res~dent, are stored in the admin~strat~ve

table.

4.4 Access Algorithm for Table Trees

The table management accesses one l~ne of a spec~fied table

at a t~me. Th~s l~ne ~s ~dent~fied by the table number and
one or more ind~ces. The locat~on of the l~ne in the pool is
determined by a backward search ~n the adm~n~strative table,
followed by a forward search ~n the subtables. Th~s ~s ~llus­

trated in Fig. 9.

The backward search starts w~th the specified table. The num-
ber of the preceeding table ~s read ~n column 7 of the adm~n­

istrative table. The search ~s cont~nued w~th this table un-

til a number table ~s encountered. For all tables, the file
number as well as the type and the column number of the poin-
ter in the preceeding array (~f applicable) are recorded, as
shown ~n the f~gure.
728

The first specified index is used to determine the appropr1ate

number subtable and the local line 1ndex K in this subtable.
The line contains a single forward pointer (array number and
line number). The f1le and array numbers of the next subtable
1n the forward search are therefore known. This subtable con-
tains the next pointer in the chain. Depend1ng on the type of
the array, the column number of this pointer is e1ther read
during the forward search or computed from a spec1fied 1ndex.
The search ends with the table specified in the access request.

The arrays containing the cha1ned subtables are made available

through the storage management. If an empty pointer is encoun-
tered during an access for reading in the table, the search 1S
terminated unsuccessfully. If the access is for writ1ng, a new
line is assigned in the next table. The array number and line
number are recorded in the pointer. The search is continued
unt11 the specif1ed values can be entered in the specif1ed
line.

4.5 Functions of the table management

All funct10ns of the table management are requested through a

common subroutine call:

CALL TABLE] (funct1on, code, 1ndices, data)

j 1dent1f1er for data type and length
function 1/2/3/4 = read/write/status/define
code complet1on code of the operat1on
1ndices table number and 1ndex values for access
data FORTRAN-vector, equal in type and dimension
with one line of the specif1ed table

The table management consists of def1nition and access func-

tions:

DEFINE TABLE
The propert1es of the table are entered in the adm1nistrative
table.
 729

PUT LINE
The data are wrltten from the FORTRAN-vector lnto the specl-
fled llne. If necessary, space for pOlnters and for the data
lS automatlcally asslgned ln the table tree.

GET LINE
The contents of the speclfled llne is read ln the table and
stored ln the FORTRAN-vector. If the llne lS deflned, the
code is set to 1, otherwlse to 2.

STATUS LINE
The code is set to the status of the speclfied llne (defined
or undeflned). The functlon lS used to check the eXlstence of
speciflc data, ego a node wlth a partlcular number.

5. TUNING OF FINITE ELEMENT PROGRAMS

Flnlte element programs are lnstalled on computers of dlffer-

ent capacity and conflguration. They are used to solve prob-
lems of dlfferent type and Slze. ThlS leads to significant
varlatlons ln the total requlred memory, in the available prl-
mary memory and in the data management by the operatlng system.
If the Solutlon lS to be efflclent, the flnlte element program
must be tuned to these condltions.

The tuning parameters describing the lndlvldual installation

and applicatlon all are parameters of the storage management.
The total avallable memory and ltS assignment to different stor-
age requlrements wlthin the program are controlled by the fol-
lowlng parameters:

number and length of data sets

number and length of files ln a data set

The avallable prlmary memory as well as the data flow between

primary and secondary memory are controlled by additional pa-
rameters:

length of the data pool

length of an array
730

The storage management is read11y implemented in such a way

that the user can adjust all of these parameters to the re-
qU1rements. The management parameters for the data structures,
ego submatrix and subtable d1mensions, can be computed inter-
nally from the array size and need not be spec1fied by the
user. The separation of the data structure from the storage
structure thus simplifies the tun1ng of a finite element pro-
gram.

Acknowledgement

The concepts described in this paper have been influenced

strongly by cooperation and discuss10ns with my colleague at
the Institut fur Allgeme1ne Bau1ngenieurmethoden, Prof. Rudolf
Damrath.
 731

1
Data Set

File 4

Data Set 2 Q)
C)
co
11 0
+-'
en
~
co
"'C
c
Data Set N 0
u
E
IIIIII
Q)
Q)
+-' en
en
>
en
I

i
File 6
C)
c
.;::;
e
Q)
0-
o
>
.0
Virtual Memory
o
+-'

0- Q)
en C)
c co
co 0
.= Admlnlstr Pool +-'
en
S A
~
co
E
~

Real Memory

j
FIGURE 1 . Storage areas of a finite element system

Adress In secondary storage file number, array number

Adress In prl mary storage block number

Special arrays In the pool scratch pad S

auxIliary array A
732

File table Pool block table

2 3 4 5 2 3 4 5

J' file number K block number

Column Contents of file table

1 number of the data set containing the file

2 number of the first record of the file

3 number of arrays In the file

4 number of records per array

5 number of the last assigned array

Column Contents of pool block table

1 status of the block . empty / free / fixed

2 Index of the first word of the block In the pool

3 file number of the stored array

4 array number of the stored array

5 access to the block: reading / writing

FIGURE 2 Administrative tables of the array management

 733

K U p

I 1 3 5 7 9 11 13 15 3 1 3 I
1 1
I 1
3
'" 3
'"
5 5

7 7
*
'" 9 9 '"
11 11

'" 13 13 '"

]15 15 I
I
t--- s s s -- -f- L - - L - f-

1 2
Structure of matrix K 3 4
5
1 2 31 4 5
I I I I I
1 1 1 1

3 3 3 3 3
,
5 5 5 5 5
3 5 3 5 3 5 3 5 3 5

-.!.
2-
2+ 35
~ lg1 ~
1 3
1 1 I 1 ' 4
1 1 1 2 3 · 4

3 3 3 3~ 5 ~ ~
5 5 5
3 3 3 Base

Structure of matrix U Profiles of a submatrix

FIGURE 3 Hypermatrlx structure for symmetric equations

734

T : array containing submatnx table and block table

H array containing hypermatnx with profiles

o array containing submatnces with profiles

array

submatrix n1

t submatrix n2 submatrix n3

Arrangement of submatnces In an array S

coefficients profiles

Arrangement of coefficients and profiles of a submatnx

FIGURE 4 Storage of a matnx In arrays

 735

Matrix table Submatrlx table Block table

2 17 2

M~

J matrix number K . submatrlx number M . array number

Column Contents of matrix table

1 status of matrix J . undefined / defined

2 number of the file associated with matrix J
3 array number of submatrlx table
4 array number of hypermatrlx
5,6 dimensions of hypermatnx
7, .,10 displacements of the hypermatrlx profiles
11 displacement of the block table
12,13 dimensions of the submatrlces
14, .,17 displacements of the submatnx profiles

Column Contents of submatnx table

1 array number of submatrlx K

2 displacement of submatrlx K

Column Contents of block table

1 number of block containing array M

FIG U R E 5· Administrative tables of the matrix management

736

K = 13 J JHYP

l
J =2_t--t--HI--II--IH-t-t-t-t--+-f' i3'ROb"P·...-+-+-+-+-+-.\--I~-I
1-:"", -. 3

5
~-+~+-~~~+-~-+~~~~~~-+'~,-
,,'~~
:}' 6
b'' "
1M " .~ SUBMATR I X ¥, ,?S' 6 2
., :LI' .,
i.= k ,d. 10

t--+-+-+-~~~~-r-r-r-+-+-+-+-+-f-4-4-4-4~~~~ 11

K 1 3 5 7 11 13 15 19 21 23
1
KHYP 2 3

CoeffIcIent mode: specIficatIon of a group of coeffIcIents

Number of the matrix M =9

Indices of apex element (J, K) = (2, 13)

Dimensions of group (JDIM, KDIM) = (2, 3)

Submatrlx mode: specIfIcatIon of all coefficIents of a submatrlx

Number of the matrix M = 9

Hypermdlces of submatrlx . (JHYP, KHYP) = (2, 3)

FIGURE 6. SpecIfIcatIon modes for matrix coeff IcIents

 737

Set: X=A X=D

Scale: X = c * A X = c * D

Add. X = A1 ± A2 X=A±D

Multiply' X = A1 * A2 X=D*A

X=A*D

Solve X = R- 1 * A X = D- 1 * A

X = A * R- 1 X = A * D- 1

Trlangularlze: RT*R=Q determine R

L * R = Q determine L, R

ST*D*S=Q determine D,S

A scratch pad or submatrlx

X scratch pad or submatrlx, can be Identical with A

D diagonal matrix, stored as a vector

L left triangular matrix, coefficients 1.0 on diagonal

R right triangular matrix

S right triangular matrix, coefficients 1.0 on diagonal

Q symmetric submatrlx, only diagonal and upper triangle stored

FIGURE 7 Functions for matrix algebra

CALL MATOP (function, 0P1, 0P2' oP3)

 r--- .....
~
P P, P2 P3 Boundary conditions System variable numbers Incident element numbers
r----
r- -
r- -
r--- -
r- l - r-
r--- l - I I
L
r- I - I Node property table
- I- ;--- L r---- L
- I-
- l- X, X2 X3 P1 P2 ... ... PlO P, P2 .. .. . PlO
- I-

N
r---
.- -
r-
r- -----
--
- r---
- r--- - -
- r-
- r--- Loadin g cond ition table Load in condition table
- -
Coordinate table
I- U1 U2 U3
- -- F1 F2 F3
I- -
I- - ~
I- -
I- - FIGURE 8 :
I- - Table tree for nodal data (simplified)
r- -
r- -
'-- -
Node number table
 739

Request CALL TABLE (1, NCODE, NPOOL, 65, 815)

Result The line of table 65 assigned to node 815 IS supplied In the pool

Administrative data for table trees

1 2 3 4 5 6 7 8 9 10 11 12 20

13
- 4 2 4 .256 27 1 42 28

-
-
27 4 1 2 512 - - 17 72 38 75

65 9 4 3 192 13 3 18 33

Chaining of arrays

2 2 3 4 2 3

K,- 42 2 LJ- 18 5 r---

Table 27 13 65
Type' number I * 4 R * 8
File 4 4 9
Array 38 42 18

Block 75 28 33
Line, 303 2 5
Column. 3

FIGURE 9: Example of a table tree chain for nodal data

740

L~terature

Pahl, P.J.: Datenverwaltung des Programmsysteros DISKRET.

Interner Bericht, Institut fur AIIgeme~ne Bauingen~eur­

methoden, TU Berlin, Mai 1978.

2 Pahl, P.J.; Damrath, R.: Finite Elemente. Vorlesungs-

skript, Institut fUr Allgemeine Bauingenieurmethoden,
TU Berlin, 1979.

3 Pahl, P.J.; Beilschmidt, L.: Informationssystem Technik,

Programmierhandbuch. CAD-Bericht 81, Kernforschungszentrum
Karlsruhe, Juli 1978.

4 Beilschmidt, L.: Data and F~le Management in the Informa-

t~on System Technology. Proceedings, Conception assistee
par Ordinateur, Toulouse, 1974, CERT, p. 32-49.

5 Pahl, P.J.; Damrath, R.; Beilschm~dt, L.: Prozeduren zur

Verwaltung Nummerierter Elemente. Forschungsgruppe Infor-
mat~onssystem Techn~k, TU Berlin, Ber~cht 77-07, Aug. 1977.

6 Beilschmidt, L.: Datengruppenhandbuch, IST-Bausteingruppe

Hochbau. Forschungsgruppe Informat~onssystem Techn~k, TU
Berlin, Bericht 77-11, Dezember 1977.

7 Beilschmidt, L.; Lebrecht, F.; Sturm, U.: Datengruppen-

handbuch, IST-Bauste~ngruppe He~zung, LUftung, Kl~ma,

Kalte (HLKK, grafische Darstellung). Forschungsgruppe

Informationssystem Technik, TU Berl~n, Bericht 79-02,
Juni 1979.

8 Damrath, R.; Pahl, P.J.: Data and Storage Structures for

a System of Finite Element Programs. U.S.-Germany Sym-
posium on Formulations and Computat~onal Algorithms in
Finite Element Analysis, MIT Press, 1977, p. 140-162.

9 Roos, D.: ICES System Design. MIT-Press, Cambr~dge, 1967.

10 Logcher, R.D.; Connor, J.J.; Nelson, M.F.: ICES STRUDL II,

MIT, Reports 68-91, 70-35, 70-77; Internal Logic Manual;
1968-1971.

11 Alcock, Shear~ng and Partners: GENESYS Reference Manual,

The GENESYS Centre, Loughborough, 1971.
 741

12 Enderle, G.; Schlechtendahl, E.G.: The Design of the In-

tegrated CAD-System Regent. Proceedings, Concept~on

assistee par Ordinateur, 1974, CERT, p. 50-79.

13 Schrem, E.: Computer Implementation of the Fin~te Element

Procedure. Numerical and Computer Methods in Structural
Mechanics, Academic Press, New York, 1973, 79-121.

14 Schrem, E.: Die Konzip~erung e~nes allgemeinen Rechenpro-

gramms fUr d~e Anwendung der Methode der Fin~ten Elemente.
Kolloquium Finite Elemente in der Statik, 302-320.
Wilhelm Ernst & Sohn, Berlin, 1973.

15 Bronlund, 0.; Kiesbauer, A.; Wilhelmy, v.: TOPAS- e~n

neues FEM-System, Proceedings, Finite Element Congress,

Baden-Baden, 1975, 23-54. IKOSS, Stuttgart.

16 Werner, H.; Axhausen, K.; Katz, C.: Programmaufbau und

Datenstrukturen ~n EntwurfsunterstUtzenden Programmketten.
Finite Elemente in der Baupraxis, Verlag Wilhelm Ernst und
Sohn, Berlin, 1978, 328-337.
Performance of Finite Element Algorithms on an Array
Processor-Minicomputer Based System

H. A. KAMEL, J. M. TAN
Umversity of Anzona, Tucson, USA

1. In~roductlon

Finite element analysis is a marriage of classical mechanics,

numerical methods and computer technology. As new computer
hardware is developed, solution algorithms are affected. In
order to be able to take advantage of such advances one has
to understand more accurately the behavior of new devices
and the interaction between them.
As the technology progresses, issues become more complex. A
simple count of the number of floating point operations re-
quired to perform certaln algorlthms may be sufficient for
the purpose of comparing alternative core resident algorithms,
but is not effective in evaluating the performance of a com-
plex algorithm on a sophistlcated system. The overhead im-
posed by increasingly elaborate operating systems, and the
intensive flow of data from one component of a computer sys-
tem to another, and the contentions and conflicts between
such devices and operations make an in-depth study mandatory.
Examples of such situations are microcomputer networks [lJ,
array processor/minicomputer combinations [2J, as well as
mainframe distributed processors [3J. The primary aim of
this paper is to attempt an examination of the benefits ac-
cured by adding an array processor as a part of a computing
system performing finite element analysis and develop a
general methodology for the evaluation of algorithm perform-
ance on distributed processors.
In order to evaluate the effectiveness of say, the addition
of an array processor to an existing system, one may proceed
 743

by simply converting programs which have run before on the

system to a configuration incorporatlng such a device. This
is indeed one of the goals of this work, and some preliminary
results of such an effort are included. However, in order
to interpret the results appropriately, and obtain the answers
to more important fundamental questions, it is essential that
we approach the problem on a more abstract level. The fund-
amental questions raised here are:

1. How to design an algorithm to run efficiently on a given

system configuration.

2. How much advantage is there in installing an array pro-

cessor?

3. How to configure a computer system for optimum perform-

ance, given a particular algorithm.

An approach based on a simplified hardware model, coupled

with some average performance measurements is proposed. It
is based on simulation, recognizes hardware contention fact-
ors, allows for parallel computation and has look-back and
ahead features. This program may be used to predict the per-
formance of certain classes of algorithms on specific hard-
ware configurations and design optimum systems for specific
algorithms. This simulation program may be used as a "proto-
type" for an executive for implementing finite element algo-
rithms on such a system.

It would be appropriate here to mention some related efforts.

For example, the use of "supercomputer" [4J has been the sub-
ject of many other investigations. Noor and Larnbiotte [5J
have examined the effect of vectorization on finite element
analysis of dynamics problems. The use of simulators to pre-
dict algorithm performance has been reported by Forslund and
Nielson [6J, who 5ucceeded in balancing I/O operations with
central processor computation, producing a CPU bound algorithm
using 2 disk channels and triple buffering. Other examples
of simulation may be found in papers by Klein [7J, Chlamtac
and Franta [8J, and Jones et. al. [9J.
744

2. Performance Predictions in a Complex System

A simple model is proposed, in which the main computer ap-

pears as a single entity encompassing CPU, floating point pro-
cessor and memory. Other components of the model are the disk
unit(s), including both disk controller(s) and drive(s), the
array processor(s), and data busses. Communication between
the various components is included in the model as well as
the characteristics of each unit. The model allows for de-
vice contention. If more than one device communicate over
the same bus, a conflict arises which blocks a new data trans-
fer from taking place until an ongoing transfer is ready. The
model also allows for parallel computation, which is an im-
portant feature of a distributed processor. It is planned
that the assumptions underlying this model be tested on a real
system in order to validate them.

Once the hardware model has been created, an algorithm is

chosen for consideration. The algorithm is then broken down
into primitives on a somewhat high level. Examples of these
would be matrix multiplications or data block transfers from
one device to another. Test programs are then written to mea-
sure resources required by each primitive. These experiments
have to be selected carefully, and conducted in a controlled
environment. Once this is accomplished, one is ready to pre-
dict algorithm performance on the given system, estimate
changes in performance due to modifications in the hardware
configuration, and evaluate alternative algorithms provided
they use the same primitives. It is of course possible to
define new primitives, as needed, and include them in the
model in order to simulate different types of algorithms.

The proposed method of performance prediction in a multi-

processor system is conducted by simulation. Two programs
are needed. The "coder" translates the algorithm, written
in a higher order language recognizing standard primitives
into a set of instructions, indicating interdependence be-
tween operations. The inclusion of interdependence is neces-
sary to include parallel computation. The second program,
called the "simulator", mimics the execution of these in-
 745

structions on a given hardware configuration. The simulator

must include a definition of hardware characteristics. It
has a look-ahead feature, enabling it to utilize system par-
allelism. In addition, it initiates "execution" of an in-
struction only if the prerequisites are satisfied and the ap-
propriate device is free. The simulator also serves as a
prototype for an "executive" to be developed later for the
actual execution of the algorithm.

The first algorithm chosen to test the effectiveness of an

array processor-minicomputer combination is that of hyper-
matrix decomposition. This particular operation was chosen
for many reasons. For one thing, It is a representative com-
ponent of a typical complex computation, whether static analy-
sis, vibrational mode extraction, transient response computa-
tion or nonlinear behavior. Secondly, it has distinct prim-
itives, such as matrix multiplication and block data transfers
from and to the backing storage. The overall algorithm is,
therefore, well suited to simulation as described here, and
the submatrix operations are ideal candidates for array pro-
cessing. The algorithm performance may be influenced by sev-
eral variable control parameters, including the size of the
partitions, and the amount of buffer space allocated within
the minicomputer and array processor, in order to minimize
input-output operations.

3. Definition of Primitives

The algorithm should be broken down into a number of well-

defined primitives, amenable to coding and measurements as
independent tasks. Such primitives would include computa-
tional as well as data transfer operations. Examples are
given in the following sections.

3.1 Disk File Commands

The code generator should allow the declaration of storage

areas (files) on various devices, including the disk. These
files may be either one- or two-dimensional hypermatrices,
where all the submatrices are of identical maximum size. The
submatrices may be addressed by their position within the
746

file. It is assumed that the file will incorporate headers

which define the sizes of the blocks under consideration.
Such hyperarrays may be created by the program at the time
of program execution, or attached if they already exist. The
following commands may be defined at this point:

OEFINE,device,filename,filetype,hypersize,recordsize

defines a new file on a system component. The "device" is

a predefined mnemonic, such as OK (disk), MC (main computer)
or AP (array processor). "filetype" may be 10 or 20. "hyper-
size" may be 10 for 10, or 20,30 for 20. "recordsize" may be
30 (10) or 40,40 (20). The command:

ATTACH,device,filename,filetype,hypersize,recordsize

Attaches an already resident file on a device. It should be

noted here that no real distinction is made between the vari-
ous devices. Once a OEFINE or ATTACH command is given, both
the device and filename will be declared active until the
initialization of the file is complete. Any subsequent com-
mands which address this device or file will have to wait the
completion of the initialization process. Here, it may be
necessary to engage another device, namely the main computer,
for overhead purposes.

3.2 Transfer Commands

A transfer command results in the copying of a record of a

particular file on a device to another record on another file
and device. For example:

TRANSFER,devicel,filel,recordl,device2,file2,record2

has the effect of copying the contents of recordl, suitably

addressed, of filel on devicel to record2 of file2 on device2.
While the transfer is being conducted, both devices and re-
cords, as well as the appropriate data busses, are declared
active, and subsequent transfers to and from them are sup-
pressed. Both devices may be the same, and both files may
be the same.
 747

3.3 Examples of Computational Primitives

The following instruction has the effect of mUltiplying a sub-

matrix, entered as a record on a certain file by a record of
another file, and placing the result on a record in a th~rd

file. It is assumed that all three files reside on the same

device. During the operation it will be assumed that the de-
vice and the three records are busy. Depending on the device
characteristics, input and output to other files and records
on the same device may be allowed. The command format may
read:

MULT,filel,recordl,file2,record2,file3,record3

Other examples of computational instructions are the matrix

subtaction command, SUBT, the matrix inversion command, MINV,
and the matrix transposition command, TPOS. The following
samples are self-explanatory:

SUBT,arrayl,recordl,array2,record2,array3,record3
MINV,arrayl,recordl,array2,record2
TPOS,arrayl,recordl,array2,record2

3.4 Logical Control Statements

In order to handle hypermatrices of arbitrary size, it must

be possible to construct loops using arbitrary variables.
Standard notations, as used in typical structured languages,
such as the WHILE DO construct, are adopted here.

3.5 Measurement of Primitive Parameters

Estimates of resource requirements for the chosen primitives

must be obtained before a simulation can be attempted. It was
natural to model and measure first the matrix multiplication
operation, and compare measured values with predictions.

The results of the comparison between estimates based on know-

ledge of the hardware characteristics and algorithm demands are
quite surprising, showing that the operating system overhead
is substantial (see Table 1). As a matter of fact, the amount
of time spent doing multiplications and additions is an order
of magnitude less than the total CPU time required. The rest
~s consumed by loop counting, pushing variables on the stack,
748

and other overhead activities. From measurements, it is ob-

served that the computation needs more than eight times the
predicted times to run if the computer is lightly loaded, and
over nine times the simulation times when the computer is rroder-
ately loaded. One observes the large overhead in copying a
single doubly-subscripted real variable, whihc is four to five
times the time required to multiply two such real variables to-
genter. It becomes, therefore, a difficult task to model sys-
time performance without a thorough knowledge of the internal
organization of the operating system, which is tedious and un-
reliable, and was abandoned in favor of direct measurements.

Several lessons may be learned from this example. For instance,

the actual time measured is an order of magnitude higher than
predicted, even though the estimates included an estimate of
"direct" overhead. The experiment also shows that data trans-
fers, even from memory to CPU, are a major factor, and should
be even more significant if an array processor is used. Also,
it is apparent that a recoding of such heavily used portions of
the machine language is highly desirable, an already well-known
fact.

4. Algorithm Simulation, The Coder and the Simulator

As llEIltioned before, an algori tlun will be chosen, and a coder and simulator
will be developed to predict its performance on specific hardware con-
figuration. The following gives this device:

4.1 Example Algorithm

The program, shown in Figure 1, represents the code necessary

to triangulate a hypermatrix, using the Gaussian elimination
scheme. It uses the language and primitives defined under 3.

4.2. The Coder

The "coder" accepts a program written in a higher-order langu-

age, such as the one shown in Figure 1, and produces code such
as that shown in Figure 2. It does not need detailed informa-
tion about the system which will be used, except for the decla-
ration of the device mnemonics, and the assignment of each in-
struction type to a particular device.
 749

Submatrl.x Sl.mulatlon Measured Ratio Loop Loop Floatlng Copy

Slze Tlme Tlme Tl.me Plus Operation Tlme
Copy Time
(sec.) (sec. ) (sec. ) (sec. ) (mlrco sec.)

5 x 5 .00151 .0l3 8.61 .002 .012 4 36.4

.014 9.27

10 x 10 .01151 .097 8.43 .011 .083 7 34.3

.107 9.30

15 x 15 .03827 .320 8.36 .035 .272 7.11 34.0

.353 9.22

20 x 20 .09002 .749 8.32 .083 .634 7.19 33.6

.824 9.15

25 x 25 .17503 1.449 8.29 .161 1. 226 7.14 33.4

1.595 9.11

30 x 30 .30153 2.510 8.32 .273 2.104 7.52 33.4

2.744 9.10

35 x 35 .47779 3.996 8.36 .439 3.326 7.81 33.2

4.393 9.19

40 x 40 .71204 5.941 8.34 .650 4.944 7.79 33.1

6.469 9.09

45 x 45 1. 01254 8.405 8.30 .926 7.0l3 7.64 33.0

9.191 9.08

50 x 50 1. 38755 11.553 8.33 1.256 9.608 7.78 33.1

13.05 9.41

Table 1. Simulation and Measurenents of Matrix Multiplication.

A few points may be made regarding the generated code. Each

instruction is given a sequence number, with which it is
identified later. The prerequisites. for each instruction are
determined by examining all operands involved in the instruc-
tion. The last time an operand was referenced by a previous
instruction indicates that that instruction is a prerequisite
for the instruction under consideration. A "look-back" is,
therefore, used in the coder.

4.3 The Simulator

The simulator accepts an instruction stream such as that shown

in Figure 2, and simulates its execution on a particular sys-
tem configuration. The program must be supplied with a
7~

ATTACH,DK,STIF,2D,M,M,S,S
DEFINE,MC,WS,1D,5,S
DEFINE,AP,BUF,1D,5,8
D=1
WHILE D<M DO
TRAN,DK,STIF,D,D,MC,WS,1
TRAN,MC,WS,1,AP,BUF,1
MINV,BUF,1,BUF,2
I=D+1
WHILE I <=M DO
TRAN,DK,STIF,D,I,MC,WS,2
TPOS,WS,2,WS,3
TRAN,MC,WS,3,AP,BUF,3
J = I
WHILE J <= M DO
TRAN,DK,STIF,I,J,MC,WS,4
TRAN,MC,WS,4,AP,BUF,4
TRAN,DK,STIF,D,J,MC,WS,5
TRAN,MC,WS,5,AP,BUF,5
MULT,BUF,1,BUF,5,BUF,3
SUBT,BUF,4,BUF,3,BUF,4
TRAN,AP,BUF,4,MC,WS,4
TRAN,MC,WS,4,DK)STIF,I)J
J = J+1
ENDWHILE J
I = 1+1
ENDWHILE I
D = D+1
ENDWHILE D
Figure 1. Program to Triangulate a Hypermatrix

description of the system devices, and their interconnection,

as well as the assignment of primitives to specific devices,
and estimated execution times for each instruction of the ap-
propriate device.

The model represented here is obviously greatly simplified.

For one thing, the logical layout of a modern computer is
quite complex. It is assumed here that the system is dedicated
to one user, which may not be the case. The current model is
deterministic, whereas many events, such as disk access times,
are random events. One of the assumptions made is that each
instruction is executed on a single device, whereas for ex-
ample, the CPU will always incur a certain amount of overhead
in controlling and setting up of operations executed on other
components.

On the other hand,we must state that the approach here may be
generalized to include many of the effects described above.
 751

DIJ ## PRE OPCODE DEV DEVI FILE REC DEV2 FILE REC DEV3 FILE REC
1-- 1 o TRAN D-M DK STIF 1,1 MC WS 1
1-- 2 1 TRAN M-A MC WS 1 AP BUF 1
1-- 3 2 INV AP AP BUF 1 AP BUF 2
12- 4 o TRAN D-M DK STIF 1,2 MC WS 2
12- 5 4 TPOS MC MC WS 2 MC WS 3
12- 6 5 TRAN M-A MC WS 3 AP BUF 3
12- 7 6, 3, 2 MULT AP AP BUF 3 AP BUF 2 AP BUF 1
122 8 o TRAN D-M DK STIF 2,2 MC WS 4
122 9 8 TRAN M-A MC WS 4 AP BUF 4
122 10 0 TRAN D-M DK STIF 1,2 MC WS 5
122 11 10 TRAN M-A MC WS 5 AP BUF 5
122 12 2,11, 6 MULT AP AP BUF 1 AP BUF 5 AP BUF 3
122 13 9, 6,12 SUBT AP AP BUF 4 AP BUF 3 AP BUF 4
122 14 13, 8 TRAN M-A AP BUF 4 MC WS 4
122 15 14, 8 TRAN D-M MC WS 4 DK STIF 2,2
123 16 15 TRAN D-M DK STIF 2,3 MC WS 4
123 17 16,14 TRAN M-A MC WS 4 AP BUF 4
123 18 11 TRAN D-M DK STIF 1,3 MC WS 5
123 19 18,12 TRAN M-A MC WS 5 AP BUF 5
123 20 12,19,13 MULT AP AP BUF 1 AP BUF 5 AP BUF 3
123 21 17,20 SUBT AP AP BUF 4 AP BUF 3 AP BUF 4
123 22 21,17 TRAN M-A AP BUF 4 MC WS 4
123 23 22,16 TRAN D-M MC WS 4 DK STIF 2,3
133 24 18, 5 TRAN D-M DK STIF 1,3 HC WS 2
133 25 24, 6 TPOS MC MC WS 2 MC WS 3
133 26 25,20 TRAN M-A MC WS 3 AP BUF 3
133 27 26, 7,20 MULT AP AP BUF 3 AP BUF 2 AP BUF 1
133 28 23 TRAN D-M DK STIF 3,3 MC WS 4
133 29 28,22 TRAN M-A MC WS 4 AP BUF 4
133 30 24,19 TRAN D-M DK STIF 1,3 MC WS 5

KEY INFORMATION:
DIJ Current va~ues of loop control parameters
## Instruction sequence number
PRE Prerequisites
OPCODE Operational code TRAN=transfer, INV=invert
MULT=multiply, SUBT=subtract, TPOS=trans-
pose
DEV Device on which instruction is executed,
MC=main computer, DK=d1sk, AP=array
processor, D-M=disk/computer bus, M-A=
computer/AP bus.
DEVl Device used with first operand
FILE File where first operand resides
REC Record address for first operand
(DEV2,FILE,REC), (DEV3,FILE,REC)
Device, file and record for second and
third operands
Figure 2. List of F1rst 30 Instructions Generated by the
Code Generator for Simulation Purposes
752

It is not appropriate at this time, however, to do this, since

our main concern is to develop a methodology and obtain answers
to some fundamental questions.
During execution or simulation, the executive or simulator,
examines the instruction stream constantly for an instruction
which has all the necessary prerequisites for execution. For
an instruction to be activated, all previous instructions ad-
dressing one or more of its operands must have been executed,
and the device on which it is to be executed must be free.
It is obvious, therefore, that the simulator, and later on the
executive, has a look-ahead feature, which may span a large
number of instructions, as well as the ability to execute in-
structions in parallel and out of the original program sequence.
It is also obvious that such an executive may not necessarily
be as effective on a virtual memory system.
4.4 Choice of Primitive Execution Parameters
Before a simulation may be conducted, it is necessary to estim-
ate, or measure, the performance of specific system components
in executing the given primitives. The times shown in Table 2
are for previously defined primitives, assuming SOxSO sub-
matrices, in double precision. Some of these values are based
on extrapolated measurements, and some on educated estimates.
These are used here for the purpose of illustration

INSTRUCTION DEVICE TIME, IN SECS.

TRAN D-M .0785
TRAN H-A .0785
TPOS MC .1
MULT AP .188
INV AP .184
SUBT AP .0025

Table 2. Instruction Times for Typical Primitives

 753

4.5 Simulation Results

Two simulations are attempted here. The first deals with a

conventional system, in which three devices are present: a
main computer, an array processor (AP) and a disk (DK) -- Con-
figuration 1. Communication between the disk and the main
computer (D-M), and between the main computer and the array
processor (M-A) both take place via the same bus. The results
of the simulation are presented in Figure 3. As expected, the
communications bus is the bottleneck. It is fully utilized,
whereas the array processor is reasonably busy. The main com-
puter is lightly loaded. However, it can not be used for other
activities (in a multiprocessor environment, for example) since
input and output to the disk is blocked. In such a configura-
tion, multiprocessing is impractical, although the array pro-
cessor 1S well ut1lized, the main computer is wasted to a
great degree.

The second system simulated is one where the array processor

and disk may communicate directly with the main computer with-
out conflict -- Configurat10n 2. In other words, there are
two separate communication busses. The busses are assumed to
have the same speed. Other system characteristics remain un-
changed. The results are presented in Figure 4. In this
situation the array processor appears to be fully utilized,
whereas the ~ain computer and the two busses, particularly the
memory-array processor bus, appear to be reasonably free, thus
allowing for other activities. This clearly demonstrates the
importance of a proper design of communication paths in such
a system.

Other system configurations may also be attempted here. For

example, the array processor may be interfaced with the main
computer via a faster bus. It is also possible that a disk
be attached directly to the array processor, or that a disk
may be accessed by both the main computer and array processor
independently. Additional primitives may be defined in order
to investigate other types of algorithms.
754

AP @ I 0· I@
..
1
OK- MC - AP CD jl® 0 ® I@ I® ® I® I® I@ 1\
Me @ ·1 ~
.2 .3 .4 5 .6 .7 (sec r

Figure 3. Execution Histogram, Configuration 1

® 10
~ ..
AP

MC - AP
10 ® I® I@ I
I@
I""
®7
MC ® ® 1\
MC - OK (2) 10 ® ® I T@ I@· \
.1 .2 .3 .6 .7 (sec)

Figure 4. Execution Historgram, Configuration 2

5. Implementation of a Cholesky Matrix Decomposition on a

Minicomputer-Array Processor Combination

Finall~ the paper presents some measurements undertaken at

the laboratory showing the behavior of a hypermatrix decom-
position routine, using the Cholesky method, as executed on
an ordinary minicomputer, and on the minicomputer with an
attached array processor.

The presented measurements were obtained using a 16 bit DGC

Eclipse minicomputer, with a CSPI MAP-200 array processor,
attached to memory via a data channel, as a user defined de-
vice. Transmission of data and instruction packages to con-
 755

trol the MAP-200 was done via direct memory access (DMA).

The Eclispe computer operates under the Advanced Operating

System (AOS), which is a multi-user timesharing system. This
arrangement forced the implementation of the MAP driver as
part of the program, instead of imbedding it in the operating
system, as is usually possible with a real time operating sys-
tem. The driver uses a multitasking algorithm in order to
perform control functions and communication with the user pro-
gram. It was necessary to overlay the user program to provide
enough space for the code.

The hypermatrix Cholesky decomposition algorithm uses fairly

small submatrices (18x18). The computational primitives are
matrix multiplication, inversion and transposition. Only the
matrix mUltiplication routine was implemented on the array
processor. Instruction overlap (parallel processing) is not
fully possible at this time.

5.1 Measurement Results

Two problems were chosen for measurement purposes. They re-

present shell structures of different sizes. The problem
parameters and measured results are shown in Table 3.

Looking over the results, it becomes apparent that the array

processor is not well utilized in this particular case. The
CPU t1me 1S practically unchanged for the small problem, but
is reduced by 22 percent for the large problem. The number
of blocks transferred from and to disk are approximately the
same, which is to be expected. Observing the minicomputer
performance, via a special system monitoring program, showed
that the CPU was only 40% to 60% active during the computa-
tion. One fact which is so far unexplainable is, the elapsed
time increased substantially using the array processor over
times observed for the minicomputer alone. From the figures
measured, it is most probable that the communication with
the array processor must be time consuming, pointing out to
the need for further optimization of the MAP driver.
756

. Large Problem Small Problem

Without With Without With
A.P. A.P. A.P. A.P.
Elapsed Time
(secs) 6,666 25,735 67 97

CPU Time 7,194 5,607 122 122

(secs)
No. of Blocks 138,005 138,033 1102 1074
Transferred
Problem Parameters
Number of 3,840 200
unknowns

lNumber of 688 40
nodes
lNumber of 575 64
~lements

Half-bandwidth
Maximum 318 43
r.m.s 210 32

Table 3. Performance Measurements for Cho1esky Decomposition

on Mlnicomputer with and without an Array Processor
6. Discussion and Conclusions
The paper consists of two parts and presents a report on work
in progress. A simulation methodology has been proposed,
which should be helpful in developing finite element algorithms
for distributed processors. An important factor here is the
proper modeling of the system, and the accurate measurement
of primitive execution times.
The second part of the paper describes another effort, so
far not directly connected to the simulation approach. It
presents results comparing performance of a decomposition
algorithm on a minicomputer with and without an array pro-
cessor. No effort has been as yet devoted to optimizing
either the a1gorlthm or the array processor driver. These
results are only to be regarded as a starting point for fur-
ther work. If anything, it does point out that the problem
 757

is not straightforward, and must be addressed with care.

Particular emphasis is placed on understanding hardware
characteristics.

7. Acknowledgements

The authors gratefully acknowledge the support of the Office

of Naval Research under contract no. N00014-75-C-0837, and
CSP Inc., without which this research would not have been
possible. A debt of gratitude is also owed to Ms. Juanita
R. Alvarez who typed the manuscript.

8. References

1. Jordan, H.F.; Sawyer, P.L.: A Multiprocessor System for

Finite Element Structural Analysis. Computers and
Structures, Vol. 10, Nos. 1 & 2, (1979) 21-32.

2. Grosch, C.E.: Poisson Solver on a Large Array Computer.

Proc. of the 1978 LASL Workshop on Vector and Parallel
Processors. Los Alamos, New Mexico: 1978.
3. Satyanarayanan, M.: Commercial Multiprocessing Systems.
Computer, Vol. 13, No.5, (1980).

4. Kascic, M.J. Jr.: Vector Processing on the Cyber 200.

Infotech State of the Art Report "Supercomputers", Info-
tech Ltd., Maidenhead, U.K., 1979.

5. Noor, A.K.; Larnbiotte, J.J. Jr.: Finite Element Dynamic

Analysis on CDC STAR-100 Computer. Computers and Structures,
Vol. 10, Nos. 1 & 2, (1979) 7-20.

6. Forslund, D.W.; Nielson, C.W.: Vectorized PIC Simulation

Codes on the CRAY-l. Proc. of the 1978 LASL Workshop on
Vector and Parallel Processors. Los Alamos, New Mexico:
1978.

7. Klein, D.: MMPS -- A Reconfigurable Multi-microprocessor

Simulator. Proc. of the National Computer Conference: 1979.

8. Chlamtac, I.; Franta, W.R.: Aids to the Development of

Network Simulators. Proc. National Computer Conference:
1979.

9. Jones, A.K.; Chansler, R.J. Jr.; Durham, I.; Feiler, P.H.;

Scelza, D.A.; Schwans, K.; Vegdahl, S.R.: Programming
Issues Raised by a Multiprocessor. Proc. IEEE Vol. 66,
No.2, 1978.
High-Speed Processors and Implication for Algorithms and
Methods

G.F.CAREY
UmVerSlty of Texas, Austin, USA

Summary

In this paper we consider the interrelation of numerical

methods and algorithms with new parallel and vector
processor capabilities. A brief historical note on the
development of computer processors establishes the perspec-
tive of this study and leads into an examination of two
proposed new processors: (1) an array processor and (2) a
vectorized processor with some parallelism. Next we examine
typical linear, nonlinear and transient finite element compu-
tations, noting where parallelism and vectorization are
possible. Substructuring and splitting are taken as special
examples, and we summarize some important features concerning
d~rect and iterative methods for linear systems. Finally,
the extens~on to nonlinear systems and the implication for
a specific iteration scheme are cons~dered.

Introduct~on

Development of new computer processors and of methods and

algorithms go hand in hand. C. Babbage invented the first
"modern" calculating machine in the mid-1800's. H~s "analy-
tical engine" was capable of tabulating the values of a
function and printing the results. Only a part of the
machine was constructed due to excessive costs to the British
Admiralty, and it was consigned to the Kensington museum.
However, the underlying principles were not pursued extens~ve­

ly until the first half of this century. Advanced scientific

computing has its beginnings in the 1940's. The first
studies of difficult numerical problems are those of Hartree
at Manchester. His work represented the first concerted
attempt to solve extremely complex physical problems by
computational means.
Prior to Hartree's work, the main motivation for
developing more advanced automated calculators had been the
 759

desire for efficlent sustained computatlon - to do several

weeks' work in a day. To some degree this is stlll valld
today as we attempt to solve ever-larger problems. At the
same time we should address the important questlon: do
proposed new advances In computer processors open up new
fields of numerical and mathematical analysls? Thls question
remains partlcularly relevant today as we consider the possible
impact of parallel and vector processors on sClentific
computing.
Von Neumann captured the essence of the sltuatlon
prophetically in 1945 In his introductory memorandum on the
Advanced Instltute Machine, today termed the "von Neumann
Machine." Many of his observatlons are equally relevant now
and we quote segments of the text below:

A machine as descrlbed will certainly revolutionize

the purely mathematlcal approach to the theory of
nonllnear differentlal equations. It wlll also
(practically for the first time) permit extenslon of
. compressible, hydro and aerodynamlcs, as well
as the more complicated problem of shock waves . . .
It is likely to make the theorles of elasticlty and
plasticity much more accessible than they have been
up to now. It wlll probably help a great deal in
three-dimenslonal electrodynamical problems . . .
Apart from these things, the machine, if intelligently
used, will completely revolutionize our computing
techniques, or to formulate it more broadly, the field
of approximation mathematics. Indeed this devlce lS
likely to be, at the very least ten thousand times
faster (actually probably more) than the present
human computer and multlplying machlne methods.
However, our present computing methods are developed
for these, or for stlll slower (purely human) proce-
dures. The projected machlne wlll change the possi-
bllities, the dlfflcultles, the emphases, and the
whole lnternal economy of computlng so radically, and
shlft all procedural optlons and equillbria so com-
pletely, that the old methods wlll be much less effi-
cient than the new ones whlch have to be developed.
These new methods will have to be based on entirely
new criterla of what is mathematlcally simple or
compllcated, elegant or clumsy.

It is, unequlvocally, thls development of the hlgh-

speed digital computer since the late 1940's that has made
pract~cal numerlcal methods such as flnlte element methods
760

for boundary- and initial-value problems. From this view-

point we wish to begin to address the implications of evolv1ng
computer processors for finite element methods and related
numerical algorithms. In particular we shall examine the
impact of array and vector processor designs on element
calculations and assembly and solution algorithms.

Processor Development and Limitatious

An electrical signal (electron) can traverse approximately
15 cm 1n 1 nanosecond (1 ns = 10- 9 sec). The maximum "speed"
of the central processor is determined by the time required
for a gate to change state and by the time required to pass
from one gate to another. The worst possible case (path)
determines the minimum cycle time interval for the master
clock of the computer in question. Most high-speed processors
currently in operation have a cycle time of 30 to 50 ns. In
the case of the CRAY 1 the maximum signal path has been
reduced to a few meters and the cycle time is of the order of
12 ns. In terms of hardware design for increased speed, one
goal is a processor with cycle time of 1 ns. This would g1ve
the increase in speed requ1red for studying such problems as
plasma fus1on, reactor failure, weather prediction, NaV1er
Stokes problems, nonlinear three-dimensional so11d mechanics
problems, fluid-structure interactions, and so on. *
Some exploratory studies on exotic computer hardware
with cycle times of a few ns are being conducted using super-
conductors and Josephson junction circuits [1]. As the
junctions must be cooled within a few degrees of absolute
zero the dev1ce would operate immersed 1n liquid helium.

* In such applications as weather prediction, war games, and

reactor safety, we are really interested in very high speed
processing. Most standard engineering computations can be
more appropriately treated using small computers with calcu-
lat10ns that may take days or even weeks on larger problems.
While we can anticipate a grow1ng demand for less expensive
dedicated small computers, there are applications of the form
cited above when waiting several weeks, days or even hours
for a response may not be appropriate.
 761

Practical operat1ng computers of th1S type necess1tate a new

technology and are clearly still in the early exploratory
stage.
More 1mmed1ate developments are imminent in the area
of array and vector processors. An example is the CRAY 1
ment10ned above. Conceptually, the design principle is now to
design the computer architecture to exploit parallel1sm of
computat1ons and/or vector manipulations. Both of these 1deas
are very logical when one cons1ders the nature of approximate
numer1cal Solut1on of boundary and initial-value problems.
There is a cons1derable proportion of parallelism and vector
operations in approximate solution using techniques such as
f1n1te element analysis. The objective 1S to exploit these
features 1n designing a new generation of high-speed process-
OrS. The design may be highly vector1zed or array-oriented
or perhaps be intermediate between the extreme forms. This
observat1on has led recently to the idea of developing speci-
f1C "PDE Machines," "Finite Element Machines" and "Navier
Stokes Machines."
There are immediate ramif1cations for algorithms: in
an array processor 1n which extensive parallelism is to be
exploited, certa1n algorithms w1ll be preferred that may not
function well on a h1ghly vectorized machine that has less
parallelism. This has particular bearing on the development
of llnear and nonl1near systems solvers, a p01nt we shall
develop fUrther 1n the second part of this art1cle. At a
somewhat subtler level 1t may have a d1rect impact on the
cho1ce of method 1tself - e.g. finite d1fference versus f1nite
element, global weighted residual methods, and so on.

Two Proposed Configurations

The general objective is to obtain numerical solutions to
fully three-dimensional unsteady nonlinear problems, using
fin1te d1fference, finite element or other methods. Reliable
and accurate results might be expected in a time frame of 15
minutes on a new scient1fic computer with a sustained rate of
10 9 operations per second. The simulation may, in fact, be
one step in an integrated preliminary design procedure so
7~

that several such computations m~ght be requ~red as part of

the overall optimization strategy. In [2] it ~s est~mated

that most advanced high speed computers in use are approxi-

mately a factor of 50 slower than needed for solut~on of com-
plex three-dimensional non1~near problems.
There are several other considerations that are pert~­

nent. As processing time for large problems is sign~ficant ~n

these calculations, slow-speed access must be avo~ded so that

the demand on memory ~s increased accordingly. This d~ctates

requirements for parallelism of hardware and the abi1~ty to

contain the entire problem on-line without having to utilize
low-speed mass storage peripherals. It also ~mp1ies that the
process should conta~n a fast access memory of s~gnificant

size (8 million words of working storage and 256 m~ll~on words

of secondary storage have been estimated).
Let us briefly examine two processors that have been
proposed by Burroughs and CDC, respectively, to meet the above
requirements. A more detailed d~scuss~on of the designs and
performance est~mates are given in [2 J.
The Burroughs device is a h~ghly parallel process~ng

array consisting of a system of 512 computat~ona1 processors,

each w~th its own local data program memor~es. These are
coordinated by a s~ng1e control unit and connected by a trans-
portat~on network to 512 modules of extended memory. An
innovat~ve feature of the proposed design concerns the syn-
chronization of processor arrays: between symchron~zation

po~nts the ~nd~vidua1 process~ng elements are to operate

asynchronously, allowing them a degree of freedom in schedul-
ing instructional sequences. A strategy for eliminat~on of
non-productive time resulting from reordering or transpos~t~on

of data ~s to be achieved by means of a transpos~tion network.

The CDC device approaches the problem from the stand-
point that it w~ll be more rel~able to build a super processor
with a minimum number of parallel units ~mplemented with the
fastest ex~st~ng or evolving technology. The processor con-
sists of vector units, map unit, scalar unit and a swap unit
which can operate concurrently. Eight vector units are
ut~l~zed in this des~gn. The map unit performs memory access
 763

operations for restructuring data and the vector units at the

same time are computing on data within the buffers of the
vector units.

Finite Element Computations

Let us examine the processing logic for representative finite
element analyses of linear, nonlinear and time-dependent
problems. We identify, at least qualitatively, those aspects
of the computations where parallelism and vectorization can
be exploited with relative ease.

(1) Linear Problems

The main processing steps are listed as follows: (1) Input
element and nodal data (Parallel); (2) Construct matrix con-
tributions for each element (Parallel); (3) Assemble element
contributions (Parallel and Vector); (4) Include boundary
data; (5) Linear system solution (Vector or Parallel).
In the input stream we have some opportunities to use
parallelism as each element has identified with it specific
properties, node numbers, and so on, but this is of lesser
significance. The element matrix and vector calculations
are almost entirely local. That is, they require only lnfor-
mation pertainlng to the element ln question, and are not
serially dependent on information from other elements. In an
array processor, such as Processor 1 in the previous section,
each small processor ln the array can be asslgned a subset of
elements and compute element contributions in parallel. For
example, in a three-dimensional problem with 0(50 3 ) elements,
the 512 array processor could treat 0(500) elements in
parallel so that each processor would compute 0(250) element
contributions. There would be some minor modificatlons to
the data structure to avoid, for instance, duplicate access
of data at a common node of adjacent elements. This could be
achieved by includlng pertinent nodal data, such as nodal
loads in the element data structure. However, the probability
of simultaneous access is obviously small and delay would be
minimal even with present data structures. Since the assem-
bly process simply requires addition of element contributions
764

into the global matrix and vector locations, each lndividual

processing unit can add the present element contribution
directly. Thus the parallel logic can again be exploited,
but there will be some slight delay when common global array
locations are accessed. Vector operations cannot really be
utilized efficiently in assembly, since element variables are
stored in a compact form and only related to global variables
through a boolean transformation identifying local and global
node numbers. Essential and natural boundary data would
probably be handled separately and in a sequential manner,
although there is some parallelism here.
The final step - the system solver - lS most crltical
as most of the processor effort is in this calculation. Most
linear finite element programs employ sparse elimination algor-
ithms and use scalar arithmetic. In large three-dimensional
problems and in algorithms where adaptive-mesh refinement,
multigrids, and acceleration procedures are used, iterative
methods are competitive. When one also introduces the ques-
tion of array processors and vectorization, the issue becomes
even more uncertain. In the latter parts of this article, we
conslder In greater detall some of these issues. At this
point, we shall simply note that: for vector computations
with direct solvers, long vectors are desirable and special
storage and operatlon technlques are needed; some iterative
methods are eminently suited to array processors - only local
information at a node or grid point is utillzed and not
altered during an iterative sweep.

(2) Nonlinear Problems

The general organization of the program for stationary non-
linear problems does not differ substantially from that of
the linear problem. The data structure is essentially un-
changed and the major distinction now is the nonlinear solu-
tion algorithm, which must be imbedded in the computation.
If a standard Newton-Raphson scheme with dlrect solution of
the linear Jacobian system is used, then within each iteration
we compute element contributions to the Jacobian, assemble,
and solve as before. The parallel and vector processlng
 765

opportunities may be followed in each Newton iteration. In-

cremental load approaches, including iteration, can be con-
sidered in like manner. Other solution schemes such as SOR-
Newton may also be used. We examine some of their properties
later.

(3) Transient Problems

Depending now on the choice of explicit, semi-impliclt or
implicit time integration, quite different solutlon algorithms
are obtained. If the system is lumped and explicit integra-
tion used, the solution value at a node for the next time
level will depend on the nodal values at the adjacent nodes
and current time level. Because of the inherent sparslty of
the system, this implies that there is a high degree of para-
llelism. On the other hand, implicit methods require system
solution at each time level so that the level of parallelism
is reduced. However, if certaln iterative solution algorithms
are introduced to solve the resulting linear or nonlinear
systems, the proportlon of parallelism increases.

Special Techniques
There are a few special techniques used in flnlte element and
finite difference methods that merit particular consideratlon
in the context of parallel and vector processors. Here we
llmit the treatment to substructuring and splittlng strategles.

(1) Substructuring
Substructuring was introduced In the 1960's to solve large
scale structural analysls problems, such as aircraft "wing-
body" interaction, wlthout heavy dependence on slow out-of-
core solution algorlthms. The basic idea is to consider the
entire structure S as made up of a set of substructures
S., i
l
= 1, • • . ,n. For example, the wing and body may be
treated as separate structures in the "wing-body" problem.
Treating each substructure as a separate domain, the element
contributions can be calculated and assembled to the sub-
structure stiffness K
_l. .
Using statlc condensation (pre-
ellmlnatlon) the substructure stiffness can be reduced In
766

slze, retaining only those degrees-of-freedom of lnterest in

the interior and on the interface with the neighboring sub-
structures. Reduced substructures systems are then comblned
and the final merged system solved.
We see that, to a large extent, each substructure can
be handled independently in parallel through much of the cal-
culation. Only a few (less than ten) substructures would
typlcally be encountered in practice. Thus, a design such as
the CDC processor with 8 sophisticated vector unlt could be
used to great advantage. Many of the assembly and reduction
calculations are vector oriented (being pre-eliminatlon steps)
and could be coded to use the vector capabilities within each
of the units. The general principles to be followed are:
(l) prescribe the data for each substructure; (2) assign each
substructure to a vector unit; (3) element assembly and sub-
structure reduction are vectorized; (4) final reduced total
system solved using vector computations.
Remark: Substructuring in actuality corresponds to a
specific form of total system partitioning wlth pre-ellmina-
tion. Hence, we can equally view the above procedure as a
speclal system solver lndependent of any physical identifica-
tion with a structure and substructures. However, the phYS1-
cal structural problem does provide an easy means for explain-
ing the ideas.

(2) Splitting
Splitting techniques have been more extensively utllized in
finite difference computations, but they can also be applied
in conjunction with finite element methods and other approxi-
mation techniques. Splltting can be defined in a general
manner to reduce the number of dimensions and/or to separate
individual contributions ln the governing equation. [3 - 5].
In the former instance, the splitting can be carried out on
the differential equation by replacing the differential
operator by a sequence of one dlmensional dlfferential opera-
tors. For example, let the governing dlfferential equation
of motion have the general form
 767

au + aF + aG 0
at ax ay

We can Spl1t this equation by cons1dering the operators Lx

and Ly and advanc1ng the solution 1n time first through the
x-der1vative terms alone and then through ~he y-derivative
terms. Th1S yields the sequence

o o

If the numerical solution of these evolution equations is

performed using a predictor-corrector strategy, then in the
y-split equat10n

lit
u 1' J' - -x-
uy (G ..
1J - G1J-
.. 1)

determines a predicted value of the s61ut1on vector. In the

corrector we have

(c) 1 { + u~I?) lit (G (p)

u ..
1J 2 u ij 1J - lIy ij+l

In a similar manner, the splitt1ng 1n the x-direction can be

treated by a predictor-corrected integration. The scheme is
explic1t and comb1n1ng operations we see that the Solut1on at
time level (n+l) is

(n+l)
u ..
1J

If the problem is such that stability will necessitate too

small a t1mestep, then we may use a split implicit method.
Then, at each time step, the splitting produces a sequence
of tridiagonal one-d1mensional systems for solution. Since
the split tridiagonal systems can be treated individually, we
can use processors for each of the reduced problems in para-
llel. Within each processor the solver can be vectorized as
l.ndicated later.
768

In ADI methods the algebraic system is factored to a

product of algebraic operators in each of the coordinate direc-
tions. Then the algebraic problem is treated as a sequence
of lower-dimensional problems through the factorization.
Similar strategies to those described above can be applied
within each ADI sweep: let Au = b be the linear system
to be solved and now split the matrix A to the sum of two
matrices

A A
~x
+ A
~y

where ~x is the discrete operator corresponding to the x-

derivatives and similarly A~y
corresponds to Ly u. For
example, in torsion and potential flow we encounter the
Laplacian ~ = L = Lx + Ly where Lx u u xx' Ly u = u yy
Thus, the discrete problem for a uniform mesh on a rectangle
can be factored accordingly. We solve the one-dimensional
systems for

(aA + I)u k +l/ 2 k k

u a(A~y~u - b)
~x

and

(aA + I)u k +l u k +l / 2 _ a(A u k +l / 2 - b)

~y ~ ~ ~x~

where a is a relaxation parameter.

The methods can be generalized to less regular grids
of tensor-product elements [ 6]. These systems are typically
tridiagonal for many applications. In the second case, the
equations need reordering to achieve the tridiagonal form.
This introduces an additional constraint on the processor
design if high processing speeds are to be sustained. In
the next section we summarize some points related to linear
system solution.

Linear System Solution

(1) D1rect (Elimination) Methods

 769

Since much of the processing time may be related to repeated

system solutlon, It is essentlal that the vector capablllties
be explolted fully. Ideally, to mlnlmize "startup" time for
operations wlth vectors, we would like to use few vectors of
large Slze. However, thls should be obtained wlthout exceSSlve
increase in operatlons or data manipulation.
Let us examlne a sltuatlon where spllttlng has reduced
the problem to a sequence of many tridiagonal systems. If
we elect to vectorize the trldiagonal elimination scheme, the
most obvlous scheme is to take the system matrix stored as
N consecutlve vectors by columns. Each vector contains at
most 2w + 1 = 3 non-zero entries where w is the half-
bandwidth excluding the diagonal. In a simple matrlx * vec-
tor product we have N vector operations of length 2w + 1 .
If, lnstead, the system matrlx is stored by diagonals, we
obtain only 3 vectors and each vector is essentlally full.
The matrix * vector multiply requires 2w -1 = 3 vector
operations of length N-l .
Crout decomposition of a symmetric global stiffness
matrlx can be obtained in n steps where r is the number
of nodes In the structure. Let d be the number of degrees
of freedom at a node so that each nodal entry is a d x d
block and let 6 be the block half bandwidth. Each decompo-
sitlon step lnvolves one vector divide, S vector repllca-
tlons, 6+1 vector multipllcatlons and S vector subtrac-
tions (fewer In the flnal 6 steps) . The average length of
the vectors in the multipllcatlon and subtractlon operations
1
lS 26d.
Finally, the solutlon u is obtained by substltutlon
sweeps uSlng the decomposed system LT D L u bEach
step of the forward sweep consists of a vector multiplication
1
and a vector subtractlon of average length 2Sd. The back
substitutlon follows wlth vector multiplicatlons followed by
vector subtractions. A complete algorithm and further details
are given In [ 7 1 for a study of the vector capabilities of
the STAR 100.
770

(2) Iterative Solution

Large sparse systems were initially treated using iterative
methods, but since large core storage became available, direct
methods have been extensively used. This is partlcularly the
case in two-dimensional finite element programs for linear
structural mechanics. For very large two-dimensional prob-
lems and particularly for three-dimensional problems, itera-
tive methods again warrant consideration [ 8 l.
Rates of convergence depend upon the spectral radius
of the iteration matrix concerned and may be very slow in
some problems. We can precondition these systems using
approximate factorization (again with vectorization) to
accelerate convergence. If reasonable bounds on the eigen-
values can be obtained inexpensively, then the relaxation
factors of the iterative schemes can be optimized to produce
dramatic improvements in rates of convergence. Semi-iterative
or Chebyshev acceleration methods need acceleration para-
meters based on estimates, as does the ADI method.
The Jacobi method is perhaps the simplest linear
iteration and provides an apt example of the use of vector
and parallel concepts in connection with iterative methods.
The Jacobi method is defined by

n+l -1 n
u (I - D A)u + k

-1
where k = D b The components of the right-hand side can
be computed independently in parallel or as blocks in parallel.
The calculation involves formation of the rows of the itera-
tion matrix on the right, followed by a matrix x vector
product and a vector addition. We can use the vector product
form discussed previously for direct solvers. Thus the in-
efficient scalar Jacobi scheme can be easily implemented as
a vectorized parallel scheme in an array processor. However,
the Jacobi iteration is inherently so slow that it would re-
quire unusual circumstances to make it really practlcal. SOR,
which is superior for scalar processors, lS sequential by
nature and cannot be so readily adapted to parallel processing.
Conjugate gradient and accelerated conjugate gradient
 771

techniques appear well-suited to parallel processors since

the right-hand sldes in the iteration system can be computed
slmu1taneous1y and some vectorization can be employed. Some
Block methods, as noted for Block Jacobi and for substruc-
turing earlier, are eVldent1y well-suited to parallel process-
ors.
In [9] several iterative methods are compared for solu-
tion of the Dirichlet problem for Laplace's equation on a unit
square using the ASC single pipe machine designed by Texas
Instruments. On a uniform mesh with h = 1/80 the total time
for Jacobl lteration is half the time required using scalar
calculations. All points are treated in parallel so the time
per Jacobi iteration is less than that of any other method.
The importance of parallelism of the data structure is evident
when one compares the total tlme of Jacobi and Gauss-Seidel
iterations on the ASC. Let the standard grld numbering scheme
be used on the unit square - sequentially left-to-right in rows
from bottom to top.
The Jacobl iteration on the plpeline computer takes 85%
of the time required by scalar Gauss-Seidel. Although the
Gauss-Seidel iteration requires only approximately half as many
iterations as Jacobi, the computer time in this test is four
times as large for Gauss-Seidel as for Jacobl. This lneffi-
ciency is readily identified as due to the data organizatlon.
As the most recently computed values are used in the Gauss-
Seidel iteration, the arithmetic unit must wait until a grid
point value is fully calculated before lt can proceed wlth the
next point. This difficulty can be circumvented by restruc-
turing the grld data. For example, we can process the grid
diagonally along the lines x + y constant (i.e. on i + j
constant for mesh lines x.
1
ih, u.
J
= jh). This strategy is
advocated also for the Gauss-Seidel-Newton nonlinear iterations
in the next section, as the basic arguments apply directly to
nonlinear point-iterative methods.

Non11near Solution
As indicated ear11er, straightforward Newton-Raphson itera-
tion leads to a sparse linear Jacobian system to be solved at
772

each ~terat~ve step. Both direct and iterative methods can be

employed to solve the Jacobian system, and the problem reduces
to the l~near case discussed above. Thus, we have Newton-
elimination and Newton-iteration techniques. For example,
if SOR is used to solve the Jacobian system, the scheme is
termed Newton-SOR. Alternatively, one can develop a scalar
iteration directly for the nonlinear system and solve each
nonlinear equation by Newton's method to produce a Jacobi-
Newton or SOR-Newton method. A particular iterative method
that has proven effective in nonlinear membrane calculat~ons

has been classified as belonging to a class of SOR-Newton mk

step iterations [10,11 j.
Let us write the nonlinear system as

F(u) 0

and componentwise as

... , u
~-
l'u.,u.+ l , . • • . , u )
~ ~ n
o

for ~ = 1, • . .
, n.
Mot~vated by linear iterative schemes, let u~k) be the
J
estimate of component u. at iterate k • Now consider the
J
i'th equation and solve

. .. , (k)
u i _l ' u i '
u (k)
~+l'

which ~s a nonlinear scalar equat~on for u We call th~s

~
the outer (or primary) iteration. The standard Newton ~tera-

tion for finding the root u.

~
= a of the scalar equation
F(u) 0 is a possible inner (or secondary) iteration. We
write

u(k+l,m) = u(k+l,m-l) _ F. (u~k+l,m-l»/

~ ~ ~ ~

{aF (u(k+l,m-l»/au(k+l,m-l)}
~ ~ ~
 773

(k+l 0) (k)
where we have chosen as startlng lterate ul ' = ul
the current estimate of ul a. Note that ln thlS formula
the dependence on u, j I 1 has been suppressed for nota-
J
tional convenience.
In this Jacobl-Newton scheme, each equatlon of the
outer (Jacobi) iteratlon can be assigned to a processor and
inner Newton iteratlons computed independently. Thus the
scheme is very suitable for parallel array processing. If we
accelerate the scheme and use most re ently calculated values,
convergence rates improve but parallellsm is lost. In [ 11 )
we dlSCUSS the convergence and other properties of thl~ f?mily
of iterative methods.

The llmltatlons on the data structure noted above for

linear systems must be observed for Gauss-Seldel-Newton and
SOR-Newton schemes if array or pipeline processors are to be
used to advantage. These nonllnear point-iterative methods are
employed in (11) for the catenary problem ln one dimension and
the minimal surface (membrane) problem ln two dimensions. In
the catenary problem beginning wlth an lnitial vector u = 0
~O
at points ln the interior, we obtain 2-figure accuracy in 11
outer lterations and 6 flgures in 33 outer iteratlons, using
an inner l-step Newton at each point, so the rate is linear.
As shown in (11) the asymptotic rate is the same for the
SOR-Newton mk step method (mk > 1) as for the SOR-Newton-l
step method. Thus a very effective form of the nonlinear
pOlnt-iteration scheme is the I-step Newton method on an array
processor ln the Jacobi form or with parallel data structure
in the Gauss-Seldel or SOR forms. An investigation of the lm-
plementation and performance of these nonllnear pOlnt-lterative
schemes on advanced computer archltecture lS ln progress.
Work estlmates glve an indlcation of the relative poten-
tial of an array processor. Let FE be the number of func-
tion evaluations and C be the number of multiply/divide oper-
ations ln the scalar Newton step. The work per inner iteration
is 2FE + 2C. For k outer iterations, n equatlons and ~

lnner lteratlons the work estlmate W lS

774

k
W: L n(mk (2FE + 2C))
k=l

In the 1-step method,

W : 2Kn(FE + C)

The work estlmate for an array processor having p processors

lS

k
W mk ~(2FE + C)
k=l P

and for 1

W _ 2k ~(FE + C)
p

The iteration work estimates alone, however, are insufflclent

to ascertaln the effectiveness of advanced scientific computer
processors. Reorganizatlon of data and the suitability of a
chosen algorithm together with careful and appropriate pro-
gramming are less tangible factors that must be included in
the evaluation.

Conc1udlng Remarks
The difficulties concernlng design of advanced computers for
scientlflc computing go signiflcant1y beyond the design of
processors, data bases and support systems. The numerlca1
methodology as well as algorithms are still evolving and to
achieve the required processlng speed there must be a strong
lnterdependence between these and the actual processor design.
Parallel array processors and vector processors with some
parallelism provide a means to achleve the desired improve-
ment In processor speed. Thls lmp1les that the deslgn must
 775

be more profoundly t1ed to the class of problems of 1nterest

and to associated methodology and algorithms. On the other
hand, the poss1bil1ty of computing w1th vector calculations
and parallel1sm leads one to re-evaluate algor1thms from
th1S standpo1nt and this 1S the focus of the present study.
We exam1ne br1efly the structure of f1n1te element programs,
1nclud1ng system Solut1on, by d1rect and 1terat1ve methods,
and features such as substructur1ng and Spl1tting.

Acknowledgment
This study has been supported 1n part by NASA grant NSG-2338
and by ONR contract N00014-78-C-0550. I wish to express my
apprec1at1on to Dav1d Young and L1nda Hayes for their helpful
d1SCUSS1ons.

References
1 Mat1soo, J.: The Superconduct1ng Computer, SC1ent1f1c
Amer1can, 242, 5, 50-65, 1980.

2 Petersen, V. L. et al.: Future Computer Requ1rements for

Computat1onal Aerodynam1cs, Workshop Proceed1ngs, NASA
Conference Publ1cat1on 2032, Feb. 1978.

3 Yanenko, N. N.: The Method of Fractional Steps, (trans M.

Holt), Spr1nger, Berl1n, 1971.

4 MacCormack, R. W.: An Efficient Numer1cal Method for

Solving the Time-Dependent, Compress1ble, Nav1er-Stokes
Equat10ns at High Reynolds Number, NASA TN X-73, 1976.

5 Carey, G. F. and R. Krishnan: NaV1er Stokes Problems:

Methods, Solut1on and Computatlon, TICOM Report 79-11,
Texas Institute for Computatlonal Mechanlcs, Unlversity
of Texas at Austin, Austln, Texas, Sept. 1979.

6 Hayes, L. J.: Galerkln Alternating-D1rection Methods for

Non-Rectangular Regions USlng Patch Approxlmations (to
appear) .

7 Noor, A. K. and J. J. Lamblotte, Jr.: Finlte Element

Dynamic Analysis on CDC Star-100 Computer, Computers and
Structures, 10, 7-19, 1979.

8 Young, D.: Iteratlve Solutlon of Large Llnear Systems,

Academlc Press, New York, 1971.

9 Hayes, L. J.: Tlmlng Analysls of Standard Iterativ~

Methods on a pipeline Computer, CNA Report 136, CNA, Univ.
of Texas, April 1978.
776

10 Perrone, N. and R. Kao: A General Non11near Relaxation

Technique for Solving Nonlinear Problems in Mechanics,
J. Appl. Mech., 38, 371-376, 1971.

11 Carey, G. F. and R. Krishnan: On a Nonlinear Iterative

Method in Applied Mechanics, Part I and II, CMAME, (to
appear) •

Additional References
B1rkhoff, G. and A. George: Elimination by Nested Dissection,
Complexity of Sequential and Parallel Numerical Algor1thms,
J. F. Traub, ed., Academic Press, New York, 1973, pp. 221-269.

3uzbee, B. L.:A Fast Poisson Solver Amenable to Parallel

Computation, IEEE Trans. Comput. C-22 (1973), pp. 793-796.

Calahan, D. A.: Complexity of Vectorized Solution of 2-

Dimens10nal Finite Element Grids, Report 91, SEL, Univ. of
Mich1gan, 1975.

Calahan, D. A., W. G. Ames and E. J. Sesek: A Collection of

Equation-Solving Codes for the CRAY 1, SEL Report No. 133,
University of Michigan, Ann Arbor, Michigan, 1979.

Csanky, L.: Fast Parallel Matrix Inversion Algorithsm, SIAM

J. Comput., 5 (1976), pp. 618-623.

George, A., W. G. Poole, Jr. and R. G. Voigt: Analysis of

Dissection Algor1thms for Vector Computers, ICASE, Hampton,
Virg1nia, 1976.

Graham, W. R.: The Parallel and Pipeline Computers, Datama-

tion, 68-71, April, 1970.

Heller, D.: A Survey of Parallel Algorithms in Numerical

Linear Algebra, SIAM Review, 20, 4, 740-777, 1978.

Heller, D., D. K. Stevenson and J. F. Traub: Accelerated

Iterative Methods for the Solut10n of Tridiagonal Linear Sys-
tems on Parallel Computers, J. Assoc. Comput. Mach., 23 (1976),
pp. 636-654.

Lambiotte, J. J.,Jr. and R. G. Voigt: The Solut10n of Tri-

diagonal Linear Systems on the CDC STAR-IOO Computer, ACM
Trans. Math. Software, 1 (1975), pp. 308-329.

Lazarus, R. B.: Computer Arch1tecture and Very Large Problems,

AFIPS Conference Proceedings, 40, 45-49, 1972.

Madsen, N. K., G. H. Rodrigue and J. I. Karush: Matrix Multi-

plicat10n by Diagonals on a vector/Parallel Processor, Infor-
mation Processing Lett., 5 (1976), pp. 41-45.

Miranker, W. L.: A Survey of Parallelism in Numerical Analy-

sis, SIAM Review, 13, 524-545, 1971.
 717

Morice, Ph.: Calcul Parallele et D~composition dans la R~sol­

ution d'Equations aux Derivees Partielles de Type Elliptique,
IRIA, Rocquencourt, France, 1972.

Ortega, J. M. and W. C. Rheinboldt: Iterative Solution of

Nonlinear Equations in Several Variables, Academ1c Press, New
York, 1970.

Pease, M. c.: Matrix Inversion Using Parallel Processing, J.

Assoc. Comput. Mach., 14 (1967), pp. 757-764.

Poole, W. G., Jr. and R. G. Voigt: Numerical Algorithms for

Parallel and Vector Computers: An Annotated Bibliography, CR,
15 (1974), pp. 379-388.

Sameh, A. H.: On Jacobi and Jacobi-like Algorithms for a Par-

allel Computer, Dept. of Computer Sci., Univ. of Illinois,
Urbana, 1975.

Sameh, A. H., S. C. Chen and D. J. Kuck: Parallel P01sson and

Biharmonic Solvers, Computin~ 17 (1976} pp. 219-230.

Stone, H. S.: Parallel Tridiagonal Equation Solvers, ACM

Trans. Math. Software, 1 (1975), pp. 289-307.

Traub, J. F.: Iterative Solution of Tridiagonal Systems on

Parallel and Vector Computers, Complexity of Sequential and
Parallel Numerical Algorithms, J. F. Traub, ed., Academic
Press, New York, 1973, pp. 49-82\

Watson, W. J.: The TI ASC, a Highly Modular and Flexible

Super-Computer Arch1tecture, AFIP~ Fall 1972, AFIPS Press,
Montvale, New Jersy, vol. 41, pt. 1, pp. 221-229.