You are on page 1of 14

Tenth U.S.

National Conference on Earthquake Engineering


Frontiers of Earthquake Engineering
July 21-25, 2014
10NCEE Anchorage, Alaska

MATHEMATICAL PROGRAMMING
APPROACH TO STRUCTURAL COLLAPSE
SIMULATION

Z. Lotfian1 and M. V. Sivaselvan2

ABSTRACT

In this article, a novel approach is presented for the simulation of failure processes of structures
in extreme events such as earthquakes. Models of material behavior are represented using a
Generalized Standard Material framework, and the governing equations are formulated using a
Mixed Lagrangian Formalism. This results in the failure simulation computations being cast as
different types of mathematical programs. Efficient strategies are devised to solve these
mathematical programs in the context of large-scale problems of practical interest. Numerical
examples are presented to illustrate the approach.

1
Graduate Student, Dept. of Civil, Structural and Environmental Eng., University at Buffalo, Buffalo, NY 14260
2
Assistant Professor, Dept. of Civil, Structural and Environmental Eng., University at Buffalo, Buffalo, NY 14260

Lotfian Z, Sivaselvan MV. Mathematical programming approach to structural collapse simulation. Proceedings of
the 10th National Conference in Earthquake Engineering, Earthquake Engineering Research Institute, Anchorage,
AK, 2014.
Tenth U.S. National Conference on Earthquake Engineering
Frontiers of Earthquake Engineering
July 21-25, 2014
10NCEE Anchorage, Alaska

Mathematical Programming Approach To Structural Collapse


Simulation

Z. Lotfian1 and M. V. Sivaselvan2

ABSTRACT

In this article, a novel approach is presented for the simulation of failure processes of structures in
extreme events such as earthquakes. Models of material behavior are represented using a
Generalized Standard Material framework, and the governing equations are formulated using a
Mixed Lagrangian Formalism. This results in the failure simulation computations being cast as
different types of mathematical programs. Efficient strategies are devised to solve these
mathematical programs in the context of large-scale problems of practical interest. Numerical
examples are presented to illustrate the approach.

Introduction

Numerical solution of solid and structural mechanics problems with nonlinear material behavior
is usually performed using a displacement-based method. The state of a spatially discretized
model consists of the displacements and velocities (in dynamic problems) at the nodes, and the
stresses and internal variables at material points (commonly quadrature points) and lumped
models of components. The task in each time increment of the numerical solution, is to compute
the new state of the model, given the current state and the increments of external forces and
boundary conditions. In the following, we refer to this task as the incremental state update. In a
displacement-based method, the incremental state update is carried out in a nested fashion. In
each time increment, a displacement increment is computed from a linearized global momentum-
conservation equation. This displacement increment is then used to determine compatible strain
increments at material points. The strain increments are in turn used to compute increments of
stresses and other internal variables at material points using constitutive equations. These
updated stresses result in an imbalance in the global momentum-conservation equation, as well
as in a new linearization of this equation. A Newton-type method is used to iteratively reduce
this imbalance to within a tolerance. The aspects in which different displacement-based methods
vary include the strategies used to update the states at the material points, the manner in which a
linearization of the global momentum conservation equation is constructed, the types of
Newton's method used etc.
A different strategy, one which is the focus of this paper, takes an alternate approach to the

1
Graduate Student, Dept. of Civil, Structural and Environmental Eng., University at Buffalo, Buffalo, NY 14260
2
Assistant Professor, Dept. of Civil, Structural and Environmental Eng., University at Buffalo, Buffalo, NY 14260

Lotfian Z, Sivaselvan MV. Mathematical programming approach to structural collapse simulation. Proceedings of
the 10th National Conference in Earthquake Engineering, Earthquake Engineering Research Institute, Anchorage,
AK, 2014.
incremental state update problem. The incremental state update problem is cast as a
mathematical program. This approach was first proposed by Maier [1]. Mathematical programs
include a large class of problems related to optimization, including convex and non-convex
optimization, second order cone complementarity problems, and mathematical programs with
equilibrium constraints. Mathematical program formulations are possible for three classes of
problems (1) incremental state update (2) limit analysis, and (3) shakedown. The focus in this
paper is on the first class – incremental state update. The reader is refered to [2] for a detailed
review of the literature in this area.
Attractive features of casting the incremental state update problem as a mathematical program
include
 Complex behavior of materials and components such as frictional contact, plasticity, damage
and phase changes can be described using different formalisms of nonsmooth mechanics.
Some of these formalisms are nonsmooth energy and dissipation potentials, variational
inequalities and hemi-variational inequalities. These nonsmooth models naturally lead to
mathematical programs for the incremental state update (see for example, [3]).
 Mathematical program formulations allow asking questions about the well-posedness of the
problem – existence, uniqueness and sensitivity.
 Once the incremental state update has been cast into a particular type of mathematical
program, solvers or algorithms designed specifically for that type of program can be invoked.
These solvers or algorithms typically have effective global search strategies for their specific
type of problem.

Generalized Standard Material Framework

A wide class of material and component behaviors can be represented by suitable stored energy
functions and dissipation potentials. Such a representation is referred to as the Generalized
Standard Material Framework [4,5]. Schematics of some common behaviors encountered in
structural modeling in earthquake engineering with corresponding stored energy functions and
dissipation potentials are shown in Table 1. The generalized standard material framework
provides a natural variational interpretation of the incremental constitutive update of a material
point (Gauss integration point in finite element analysis). This variational interpretation has been
widely recognized in the context of plasticity, a particular instance of a generalized standard
material model. Such a variational formulation is also possible for the incremental update of the
state of an element or the full structural state, and leads to a mathematical programming
approach. In the following, a brief description of the generalized standard material framework is
presented, including an adaptation to represent strength degradation behavior. For more details,
the reader is referred to [2,5,6,7].
Here, we describe the constitutive equations in terms of stress-resultants such as axial forces,
bending moments etc. denoted by the symbol F. Strains refer to generalized strains such as axial
strains, curvatures etc. and are denoted by . Internal hardening and softening variables are
denoted by . In the small deformation theory, the rates of strain and the kinematic internal
variables at a material point can be additively decomposed into elastic and plastic parts as

 = e   p
(1)
 = e   p
We further partition the kinematic internal variables  as h, the hardening variables, s and the
softening variables. Following the generalized standard material formalism, we define a stored
energy function, , and a dissipation potential, , which are functions respectively of the elastic
components, and of the rates of the plastic components of the strain, , and the internal variables,
. The specific form of the stored energy function (e,he,se) considered here is convex in the
arguments e and the hardening variables he, and concave in the softening variables se. The
dissipation function , is convex. Using an extension of Fenchel's duality for saddle functions,
we define c, the complementary stored energy function and c complementary dissipation
function. These are generalizations of the strain energy and the complementary strain energy
functions familiar to us from elementary structural mechanics. Then the constitutive relationships
may be written in dual form as

 e   F c  F ,  h ,  s 
 he    c  F ,  h ,  s  (2)
h

 se     c  F ,  h ,  s  
s

and

 p   F c  F ,  h ,  s 
 hp    c  F ,  h ,  s  (3)
h

 sp    c  F ,  h ,  s 
s

Governing Equations – The Mixed Lagrangian Formalism

When the material or component behavior is represented using equations (2) and (3), the
governing equations of a spatially discretized model can be written as


d
dt
  F  F ,  h ,  s     h  F ,  h ,  s   Bv  0 deformation compatibility
c c


total strain rate
elastic strain rate plastic strain rate

Mv  Cv  BT F = P  momentum conservation (4)


d
dt
 
  c  F ,  h ,  s     c  F ,  h ,  s   0
h h


reversible internal variable rate irreversible internal variable rate 
 internal variable evolution
d
dt
 
   c  F ,  h ,  s      c  F ,  h ,  s   0
s s


reversible internal variable rate irreversible internal variable rate 
Table 1: Example of the construction of one table.

Component Schematic Lagrangian Dissipation


Linear-elastic k
 ( e )  12 k e2  c ( F )  21k F 2 N/A
spring
Nonlinear  ( e )  strain energy -
elastic spring  ( F )  compl. energy
c

Linear viscous c
N/A  ( p )  12 c p2
 c  F   21c F 2

  p    11 c  p
Nonlinear N/A  1
viscous damper or
 1
 c  F    1 1
1

F 

Maxwell k c  c ( F )  21k F 2  c  F   21c F 2


element

Kelvin element k  ( e )  12 k e2  ( p )  12 c p2
c

 c ( F )  21k F 2  c  F   21c max  F  Fy ,0 


Fy 2
Viscoplastic
element k

Elastic-plastic  c ( F )  21k F 2 c F  = F


C
element Fy k

C  F | F  Fy 

 c  F ,    21k F 2  21k  2  c  F ,   F  
Kinematic k2
1 2 C
Hardening
C   F ,   | F    Fy 
k1

Tension only  ( e )  C  e  C   |    max  N/A


with gap  max  F F 0
 c ( F )   max
  otherwise

where F is the vector whose components are internal forces or stresses (or stress resultants) at
material points, h is the vector of conjugates of the internal variables in which the stored energy
function is convex (hardening variables), s is the vector of conjugates of the internal variables in
which the stored energy function is concave (softening variables), v is the vector of velocities at
the free degrees of freedom (DOF), B is the strain-displacement matrix, M and C are respectively
positive semi-definite mass and damping matrices, and P is the external load vector. The first of
equations (4) expresses deformation compatibility in the elements, the second represents
momentum conservation, and the third and fourth, the evolution of the constitutive internal
variables. Equation (4) can be obtained as the Euler-Lagrange equations of a generalized
Hamilton's principle [6] through the Mixed Lagrnagian Formalism. When specialized to the
piecewise linear equations models [2], equations (4) become
AF   F ELASTIC  F ,  h ,  s   Bv  0
Mv  Cv  BT F = P
Gh h  
h ELASTIC  F , h , s   0 (5)

Gs s 
d
dt
 RES  s     s ELASTIC  F , h , s   0
These equations are next discretized in time.

Time Discretization

Equations (5) are discretized using the midpoint rule. The choice of the midpoint rule is
motivated by the fact that it arises out of using a variational integrator starting from Hamilton’s
principle [6]. For linear elasticity, the resulting discretization reduces to Newmark's method with
constant average acceleration.
 F  Fn  1  vn 1  vn 
A  n 1     F ELASTIC ( Fn 1 ,  h ,n 1 ,  s ,n 1 )   F ELASTIC ( Fn ,  h,n ,  s ,n )   B  0
 h  2  2 
 v v  v v   F  Fn   Pn 1  Pn 
M  n 1 n   C  n 1 n   BT  n 1 = 
 h   2   2   2  (6)
   h ,n  1
Gh  h ,n 1
 h  2
 
   h ELASTIC ( Fn 1 ,  h ,n 1 ,  s ,n 1 )   h ELASTIC ( Fn ,  h ,n ,  s ,n )  0

   s ,n    RES ( s ,n 1 )   RES ( s ,n ) 
Gs  s ,n 1  
 h   h 
1
2 s
 
 ELASTIC ( Fn 1 ,  h ,n 1 ,  s ,n 1 )   ELASTIC ( Fn ,  h, n ,  s , n )  0
s

where subscripts denote discrete time, and h is the time increment. The following substitutions
can be made for the differentials in equation (6).
 RES ( s ) = TRES ;  F ELASTIC ( F ,  h ,  s ) = TF 
(7)
 ELASTIC ( F ,  h ,  s ) = Th  ;  ELASTIC ( F ,  h ,  s ) = Ts 
h s

where  and  are Lagrange multipliers corresponding respectively to the residual strength and
plastic yield constraints, and satisfying the following complementarity conditions.
  0,   RES s  bRES  0,
 T ( RES s  bRES ) = 0
(8)
  0,   F F   h h   s s  bELASTIC  0,
 T ( F F   h h   s s  bELASTIC ) = 0
Substituting equation (8) into the first, third and fourth of equations (6), the following equations
for the stress resultants and stress-like internal variables are obtained in “predictor-corrector”
format. The predictor equations are
h 
Fn = Fn  K  Bvn  TF n 
2 
 h,n =  h ,n  K h  h n
T
(9)
 s ,n =  s ,n  K s (TRES n  Ts n )
where K = A-1 is the block diagonal matrix of element stiffnesses, Kh = Gh-1 and Ks = Gs-1. The
corrector equations are
h 
Fn 1 = Fn  K  Bvn 1  TF n 1 
2 
 h ,n 1 =  h ,n  K h  h n 1
T
(10)
 s ,n 1 =  s ,n  K s ( TRES n 1  Ts n 1 )

Fn ,  h,n ,  s ,n are predictors for internal forces, hardening and softening internal variables.

Mathematical Programming Problems

At this stage, two approaches present themselves:


1. In the absence of softening (degrading behavior), the predictor-corrector equations together
with the complementarity conditions constitute the Karush-Kuhn-Tucker (KKT) conditions
of a convex minimization problem. The forces and velocities at time n+1 may be obtained by
computing a solution to this convex minimization problem [7].
2. When softening is present, the predictor-corrector equations and the complementarity
conditions are not the KKT conditions of any optimization problem. Rather, they constitute a
Mixed Complementarity Problem (MCP) [2].
These possibilities are shown schematically in Figure 1.

Numerical Examples

We consider is a model of a 15-story structure with 4 bays in each plan direction. It is roughly
30m  30m in plan and 55m tall. The model consists of 400 nodes with 2250 free DOF and 1215
elements. The frequencies of the first three modes of vibration of the linearized model are 1.64
Hz, 1.65 Hz and 1.91 Hz. No inherent damping is modeled. The first two examples are from [7]
and the second from [2].

Collapse analysis with ground acceleration input

In this simulation, gravity loads are applied on the structure, and the structure is subjected to a
five times scaled El Centro ground motion. Geometric nonlinearity is considered, and the
structure collapses as a result of this excitation. The deflected shape of the collapsing structure is
shown in Figure 2(c). The simulation is carried out with three different time increments, and the
convergence of displacements in the horizontal and vertical directions is shown in Figures 2(a)
and (b), respectively. While the large time increment results in significant error as seen in this
plot, Figure 2(d) shows that even with a hundred-fold increase in time increment, the number of
iterations remains small.

Collapse analysis with element removal

In this simulation, the structure is loaded with gravity loads corresponding to self weight,
followed by instantaneous removal of two outer rows of columns simulating an event such as an
explosion. The structure collapses as a result as shown in Figure 3(c). The convergence of
displacements with decreasing time increments is shown in Figures 3(a) and (b). Once again it is
seen in Figure 3(d) that the iteration count remains even small with a hundred-fold increase in
time increment.

MLF representation
Variational integrator (or other)
time discretization

System of nonlinear equations


and generalized inequalities

Some restrictions Other restrictions


Other restrictions

Mixed Nonlinear
Complementarity Problem
Second Order
Cone Program

(MNCP) (SOCP)
No deterioration Quadratic (non-convex) energy,
(no softening, Piecewise planar yield surfaces

∙∙
convex energy), (and possibly non-associated flow)
Associated Flow


Convex
Minimization
Quadratic energy,
Piecewise planar
yield surfaces

Mixed Linear
Convex Quadratic
Complementarity
Program (QP) Problem (MLCP)

Figure 1: Mathematical programming problems arising from MLF


(b)
(a)

(c) (d)

Figure 2. Earthquake collapse analysis (a) X displacements (b) Z displacements (c) Deformed
shape at collapse (d) Iteration count
(a) (b)

(d)

(c)

Figure 3. Element removal simulation (a) Z displacement of outer row above removed columns
(b) Y displacement of outer row above removed columns (c) Deformed shape (d) Iteration count
Model with strength degradation (MCP approach)

This example is from [2]. Three different hardening-softening cases are considered, and the
analysis is performed using the Complementary Pivot Algorithm (CPA). In each case, the model
is subjected to a ground acceleration input corresponding to a measured ground acceleration
curve from the 1940 El Centro earthquake.
Analysis case 1 – Isotropic softening only: In this first analysis case, the plastic hinges are
modeled with no hardening, but with isotropic softening with modulus 10% of elastic, and with a
residual strength of 50% of yield. There are 2250 free degrees of freedom, and 11250 yield and
residual constraints, so that the size of the complementarity problem is 13500. The dynamic
analysis is carried out with three different time increments - 0.02s, 0.002s and 0.001s, and with
two different covering vectors to solve the incremental state update complementarity problem -
Lemke and Residual. Figure 4(a) shows displacements of three nodes picked from the 5th , 10th
and 15th stories. The curves corresponding to time increments of 0.002s and 0.001s are
indistinguishable, indicating convergence of these curves. Figure 5(a) shows a typical plastic
hinge moment-rotation curve. The CPU times are summarized in Table 3. It is seen that residual
start leads to smaller CPU times than Lemke start, due to fewer pivot steps.
Analysis case 2 – Ideal plastic: In this analysis case, the plastic hinges are modeled as elastic-
ideal plastic with no hardening or softening. This results in 9300 yield constraints so that the size
of the complementarity problem is 11550. Once again, the six combinations resulting from the
three time increments and two covering vectors are considered. Figure 4(b) shows the
displacements of the nodes. The numerical behavior is similar to that in Analysis case 1. Figure
5(b) shows a typical plastic hinge moment-rotation curve for both residual and Lemke starts. The
CPU times are summarized in Table 3, where it is again seen that residual start leads to smaller
CPU times than Lemke start.
Analysis case 3 – Kinematic hardening and softening: In this final analysis case, the plastic
hinges are modeled with kinematic hardening of 2% of elastic and isotropic softening of 10% of
elastic, with a residual strength of 50% of yield, and peak-to-yield ratio of 1.025. The number of
yield and residual constraints is 25200, so that the size of the complementarity problem is 27450.
As shown in Table 3, the complementary pivot algorithm with residual start sometimes cycles,
and fails to find a solution. The analysis is terminated at the first time increment at which this
occurs. With Lemke start however, a solution is always found. Displacements are shown in
Figure 4(c), and typical plastic hinge behavior in Figure 5(c). CPU times are shown in Table 3.

Table 2. CPU times for the optimization examples (using a system with a 1.8 GHz Intel Pentium
M processor and 1 GB RAM)

Increment 0.2 s 0.02 s 0.002 s


Earthquake collapse simulation 17.60 s 121.29 s 1038.39 s
Removal of bottom rows of columns 25.21 s 129.83 s 1163.65 s

Concluding Remarks

Using the Generalized Standard Material framework to represent material behavior, and the
Mixed Lagrangian Formalism (MLF) to described the governing dynamics, the simulation of
failure processes in structures is cast in a novel way as different mathematical programming
problems. Efficient algorithms have been devised to solve these mathematical programming
problems in the context of large-scale problems of practical interest. Numerical examples
havebeen presented to illustrate this new approach. The MCP approach has also been applied in

(a) (b)

(c)
Figure 4. Displacement for MCP analysis (a) case 1 (b) case 2 (c) case 3
the analysis of tensegrity structures [8]. The mathematical programming approach serves as a
useful educational tool [9].

(a) (b)

(c)
Figure 5. Moment-rotation curves for a typical plastic hinge in MCP analysis (a) Analysis case 1
(b) Analysis case 2 (c) Analysis case 3

Table 3. CPU times for MCP examples (using system with computer with an 2.5 GHz Intel
Core2 Duo processor and 4 GB RAM)

Time increment Covering vector Number of CPU time


increments
Analysis case 1
0.02 s Lemke 500 130 s
0.02 s Residual 500 88 s
Analysis case 2
0.02 s Lemke 500 130 s
0.02 s Residual 500 85 s
Analysis case 3
0.02 s Lemke 500 269 s
0.02 s Residual N/A Cycling at increment 149
Acknowledgements

The author gratefully acknowledges financial support from the National Science Foundation
through grant number CMMI 0847053 and from Kajima Corporation, Japan.

References

[1] Maier, G. (1970). A Matrix Structural Theory of Piecewise Linear Elastoplasticity with
Interacting Yield Planes, Meccanica, 5(1):54-66.
[2] Sivaselvan, M. V. (2011). Complementarity framework for nonlinear dynamic analysis of
skeletal structures with softening plastic hinges, International Journal for Numerical
Methods in Engineering, 86(2), 182-223.
[3] Acary, V. and Brogliato, B. (2008). Numerical methods for nonsmooth dynamical systems :
applications in mechanics and electronics. Springer, Berlin.
[4] Halphen, B. and Nguyen, Q. S. (1975). Sur les matériaux standard généralisés. Journal de
Mécanique, 14(1):39-63.
[5] Sivaselvan, M. V. (2013). “Hysteretic models with stiffness and strength degradation in a
mathematical programming format”, International Journal of Non-linear Mechanics, 51,
10-27, doi:10.1016/j.ijnonlinmec.2012.12.004.
[6] Sivaselvan, M. V. and Reinhorn, A. M. (2006) Lagrangian approach to structural collapse
simulation. Journal of Engineering Mechanics-ASCE, 132(8):795-805.
[7] Sivaselvan, M. V., Lavan, O., Dargush, G. F., Kurino, H., Hyodo, Y., Fukuda, R., Sato, K.,
Apostolakis, G. and Reinhorn, A. M. (2009) Numerical collapse simulation of large-scale
structural systems using an optimization-based algorithm. Earthquake Engineering &
Structural Dynamics, 38(5):655—677.
[8] Olivetto, N. and Sivaselvan, M. V. (2011) “Dynamic analysis of tensegrity structures using
a complementarity framework”, Computers and Structures, 89(24-25), 2471-2483.
[9] Sivaselvan, M. V. (2011). “Failure analysis of engineering structures in undergraduate
courses using optimization”, Computer Applications in Engineering Education, Electronic
early view, doi:10.1002/cae.20555.

You might also like