You are on page 1of 41

CHAPTER THREE

Numerical Solutions of
Dynamics of Structural
Problems
3.1 Governing Equations of Dynamics of
Structural Systems
To solve structural or continuum mechanics transient dynamics problems
numerically, the governing hyperbolic partial differential equations are first
discretized in space. Either finite element or finite difference methods can be
used to semi-discretize the governing equations. The semi-discretization will
reduce the problem to a system of second order ordinary differential equations
in time, which in turn must be integrated to complete the solution.
The resulting semi-discrete equations of motion can be viewed as a set
of static equilibrium equations of a structure at time t, including the effect of
acceleration-dependent inertia forces and velocity-dependent damping forces
in addition to externally applied discretized forces and internal nodal forces of
the continuum or structure, which may be written as:
F I (t )  F D (t )  F Int . (t )  F Ext . (t ) (3.1)

where

F I (t ) , F D ( t ), F Int. ( t ), and F Ext. ( t ) are inertial, damping, internal, and


external force vectors, respectively, all of which may be time dependent.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 14

The internal nodal forces F Int . ( t ) include forces due to initial stresses in

the system, which may depend upon displacements and velocities at the nodes
or their histories.
For linear systems, the static equilibrium equations (3.1), after the
application of finite element spatial-discretization, will take the form:
  CU
MU   KU  R (3.2)
t t t t

Where, M is the discrete mass matrix, C is the viscous damping matrix, K is


the linear stiffness matrix, R t is the vector of external discrete forces which

include body forces, surface tractions, and concentrated loads acting on the
 , and U
system at time t. U t , U  are the nodal displacements , velocities, and
t t

accelerations vectors at time t respectively, where the super-script denotes


differentiation with respect to time.
In general, M , C , and K are constant symmetric and banded matrices. M
and K are positive definite, C is positive semi-definite, and Rt is a given
continuos function of time.
Mathematically, Eq. (3.2) represents a coupled system of linear second
order differential equations. The solution of the initial-value problem
described by (3.2) consists of finding the time history response, U  U( t ), t  0 ,

satisfying both Eq. (3.2) and the initial conditions:


U(0 )  U o
(3.3)
 (0)  U
U 
o

 are the given initial displacement and velocity vectors. If


Where, U o and U o

 ,is not specified, it can be calculated by application


the initial accelerations, U o

of the equilibrium equation (3.2) at time to which gives:


  M 1 R  CU
U   KU  (3.4)
o o o o

In the absence of the effects of inertia and damping forces on the system, Eq.
(3.2) will reduce to the finite element equilibrium equation of quasi-static
problems i.e.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 15

KU t  R t (3.5)

Here, the displacement response Ut, is a linear function of the applied forces
vector Rt, noting that these forces may vary with time, and accordingly the
displacement response will vary with time also, i.e. Eq. (3.5) represents a
statement of static equilibrium for any specific station in time.
The choice for static or dynamic analyses, i.e. for the inclusion or
negligence of velocity and acceleration dependent forces in the analyses is
apparently decided by engineering judgement of the rates at which the loads
are applied. However, if the loads are applied rapidly, inertial forces need to
be considered and a truly dynamic problem must be solved.

3.2 Mass Matrix, M


The mass matrix may be formulated either as a consistent, nondiagonal mass
matrix, or as a lumped diagonal mass matrix. Previous work indicated that
consistent mass matrices tend to yield more accurate frequencies for flexural
members, such as beams and shells [33]. But, consistent mass matrices usually
over-estimate the frequencies, whereas, lumped masses under-estimate the
frequencies [33]. The nature of the frequency error for those two types of mass
matrices differs. This should be born in mind when choosing a specific method
for the integration of Eq. (3.2). Since it is desirable that the errors introduced
by the time-integration operator compensate for the error due to the mass
matrix rather than amplify it. Many studies recommend the use of lumped
mass matrices with the explicit central difference operator and associate
consistent mass matrices with implicit methods [28]. This is tested in this
investigation. The use of consistent mass matrices does not always lead to
improved accuracy and involves always, additional computational work [34].
Key and Beisinger [35] presented a method for deriving a diagonal mass
matrix from the standard consistent mass matrix, for elements with linear or
Chapter Three Numerical Solutions of Dynamics of Structural Problems 16

cubic displacement functions. Hinton et. al. [36] recommended a procedure of


mass lumping for parabolic isoparametric elements to compute diagonal terms
so as to preserve the total mass of the element. Numerical examples of the
application of this mass lumping scheme are given by Shantaram [37].

3.3 Damping matrix, C


In practice and for general finite element assemblage it is difficult, if not
impossible, to determine the element damping parameters in particular,
because the damping properties are frequency dependent. For this reason, the
matrix is in general not assembled from element damping matrices. The most
conventional method is to use Rayleigh damping, which is related to the mass
matrix M, and the stiffness matrix K of the complete element assemblage. The
damping matrix would then take the form:
C  aM  bK (3.6)
Where a and b are the given Rayleigh constants.

3.4 Step-By-Step Numerical Direct Integration


Technique
The step-by-step numerical direct integration technique appears as a most
powerful technique to solve equations of motion. Its importance stems from its
generality and ability to deal with nonlinear conditions, and general types of
loading.
The term “direct” means that prior to the numerical integration, no
transformation of the equations of motion into a different form is carried out,
such as that needed in mode superposition for example.
In essence, direct integration is based on two principles. Firstly, instead
of trying to satisfy Eq. (3.2) at any time t, the equation is only satisfied at
Chapter Three Numerical Solutions of Dynamics of Structural Problems 17

discrete time intervals, t apart. This is akin to solutions of, basically, “static”
equilibrium equations, which include the effect of inertia and damping forces
at discrete time stations. Secondly, the direct integration method is based on
the assumption of a given variation of displacements, velocities, and
accelerations within each time interval, t. It is the form of the assumption of
the variations of displacements, velocities, and accelerations within each time
interval that determines the accuracy, stability, and cost of the solution
procedure.
Direct integration procedures can be further subdivided into explicit
and implicit methods each with distinct advantages and disadvantages. Each
approach employs difference equivalence to develop recurrence relations,
which may be used in a step-by-step computation of the response. The critical
parameter in the use of each of these techniques is generally the largest value
of the time step which can be adopted to provide sufficiently accurate results,
as this is directly related to the total cost of satisfactory analysis.

3.5 Explicit Methods


The first approach is based on the so called explicit time integration techniques
employing, especially, finite difference methods. This approach is,
particularly, well studied for dynamics problems of short duration, such as
structures subjected to blast or high velocity impact.
In the explicit methods, the solution at time t+t is obtained
incrementally by considering the equilibrium conditions at time t. Such
integration schemes, generally, do not require factorization of the (effective)
stiffness matrix in the step-by-step solution. Hence, the computational cost per
time step is generally much less for explicit methods and less storage is
required than that for implicit methods. Computer operations for explicit
Chapter Three Numerical Solutions of Dynamics of Structural Problems 18

methods are relatively few in number and independent of the finite element
band width, since no factorization of the effective stiffness matrix is required.
However, explicit time marching schemes are only conditionally stable
and generally require small time steps to be employed to ensure numerical
stability. Here, the step size restriction is often more severe than accuracy
considerations requirements. This restriction limits the effectiveness of the
approach when used in analyses of dynamic problems of moderate duration,
e.g., earthquake response problems. The conditionally stable algorithms
require the time step size employed to be inversely proportional to the highest
frequency of the discrete system.

3.6 Implicit Methods


The second approach utilizes implicit time-integration schemes. Its primary
strengths are exactly in the areas where the first approach, i.e. explicit, has its
weakness, namely, in treating inertial problems such as fast impact and blast
problems.
In the implicit methods, the calculations of displacements at the current
time step involve the velocities and accelerations at the current time step itself,
t+t. Hence, the determination of the displacements at t+t usually involve the
solution of the structure stiffness matrix at every time step.
Many implicit methods are unconditionally stable for linear systems and
the maximum length of the time step that can be employed is governed by
accuracy of the solution and not by the stability of the integration process.
Although the implicit methods require considerable effort per time step as
compared to the explicit methods, the time step may be much larger since it is
restricted in size only by accuracy requirements. The time step in most explicit
methods, on the other hand and as presented previously, is restricted only by
numerical stability requirements. This may result in a time step much smaller
Chapter Three Numerical Solutions of Dynamics of Structural Problems 19

than that needed for the required accuracy, thus increasing the cost of
computation.

3.7 The Two Families of Direct Integration


Methods
The direct integration methods can be divided from the point of view of the
derivation of the base recurrence formulae into two main families, namely the
finite difference family and the finite element family.

3.7.1 Finite Difference Family of Methods for the Solution


of the Equations of Motion

Systems of linear or nonlinear second order ordinary differential equations,


which represent the equations of motion, can effectively be dealt with by a
general finite difference approach.
Many finite difference methods have been developed over the years to
solve equations of motion. Among these methods, standard finite difference
methods like the central and backward difference methods as well as other
variations with their own assumptions fall under the general finite difference
concept.
In the present work, four of the most well known and frequently used
methods are considered. These methods are:
1- The explicit central difference method.
2- The Houbolt backward difference method.
3- The generalized Newmark method.
4- The Wilson- method.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 20

3.7.1.1 The Explicit Central Difference Method

The second order explicit central difference method is one of the most widely
used methods among explicit numerical integration techniques. The central
difference method is claimed to have the highest accuracy and maximum
stability limit for any explicit method of order two [26]. However, it has the
disadvantage of requiring small time steps to ensure stability.
The central difference method is based on expressions which
approximate each of velocities and accelerations in terms of displacements as:

  1  U
U t t  t  U t  t  (3.7)
2 t
and
1
 
U t U  2 U t  U t  t  (3.8)
t 2 t  t
Where, t-t, t, and t+t are three successive time stations.
The expressions (3.7) and (3.8) are of a second order of error, i.e. the
error in U is quartered when t is halved.
For linear problems, the displacements at time t+t is obtained by
considering the discrete equations of motion (3.2) at time t, that is:
  CU
MU   KU  R
t t t t

 and U
Substituting the relations for U  in (3.7) and (3.8) into Eq. (3.2) and
t t

arranging terms, the following recurrence formula is obtained:


 1 1   2   1 1 
 M C U t  t  R t   K  M U t   M C U t  t (3.9)
 t  t    t 
2 2 2
2 t   2 t 

Which can be solved for U t  t .

The solution of U t  t using Eq. (3.9) is based on the equilibrium

conditions at time t. For this reason, the integration procedure is explicit and
in such schemes there is no need for factorization of the effective stiffness
matrix in the solution process.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 21

Secondly, the calculation of U t  t requires each of U t and U t  t to be

known. Therefore, a special starting procedure is needed to initiate the


solution process.
 are naturally known, while U
In dynamic problems, both U o and U 
o o

can be calculated using the equilibrium equation (3.2) at time to. That leaves
U  t to be determined. That can be achieved by matching equations (3.7) and

(3.8) which yields:

  t 2 
U  t  U o  tU o Uo (3.10)
2
The effectiveness of this procedure depends on the use of a diagonal mass
matrix, which can be obtained by a lumping process. There is a need also to
neglect the general viscosity-dependent damping forces, or the use of a
diagonal damping matrix. Otherwise, the solution requires a factorization of
the effective coefficient matrix. This is one of the shortcomings of the central
difference method apart from time step size restriction due to stability
considerations.
For structural systems without physical damping Eq. (3.9) reduces to

1  2  1
MU t  t  R t   K  M U t  MU t  t (3.11)
t  2
 t  
2
t 2
which can be further simplified to

U t  t  t 2 M 1 R t  KU t   2 U t  U t  t (3.12)

Therefore, if the mass matrix is diagonal, the undamped structural dynamic


equations can be solved without factorization, i.e., no solution of simultaneous
equations is needed.
Even if a nondiagonal mass matrix is used, i.e. when a consistent
formulation of the semi-discrete equations is used, the mass matrix needs to be
inverted only once. Hence, explicit procedures tend to be computationally
more efficient per time step than implicit methods.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 22

Damping may be included and the explicit form can still be preserved,
provided M is diagonal. If a backward rather than central difference
 ; i.e. Eq. (3.7) is replaced by
approximation is used to approximate velocity U t

  1  U
U t t  t  U t  (3.13)
t
The latter equation modifies Eq. (3.9) into

 
U t  t  M  1 t 2 R t  KU t   tCU t  U t  t   2 U t  U t  t (3.14)

This explicit formulation suggested by Warbuton [38] require no matrix


factorization if M is diagonal, even if damping is included.
The critical time step, which must not be exceeded, has been determined
by many researchers [21,25,39] to be:
Tn
t  t crit.  (3.15)

where Tn is the smallest period of the system modes.

3.7.1.2 Houbolt Method

Houbolt integration scheme is similar to the previously discussed central


difference method in that standard finite difference expressions are used to
approximate the velocity and acceleration components in terms of
displacement components. Houbolt operator was obtained by fitting a cubic
polynomial through values of the current displacement, to be found, and the
three previous values. The first and second derivatives of this polynomial are
evaluated at current time t+t using backward difference expressions, as
1

U t  t  11U t  t  18U t  9U t  t  2 U t  2 t  (3.16)
6 t
1

U t  t  2 U t  t  5U t  4U t  t  U t  2 t  (3.17)
t 2
Considering the equilibrium equation (3.2) at t+t i.e.

MU  (3.18)
t  t  CU t  t  KU t  t  R t  t
Chapter Three Numerical Solutions of Dynamics of Structural Problems 23

Substituting both Eqs. (3.16) and (3.17) into Eq. (3.18) and arranging all known
vectors on the right hand side yields

 2 11   5 3   4 3 
 M C KUtt  Rtt   M CUt  M CUtt
 t
2
6t   t
2
t   t
2
2t 
(3.19)
 1 1 
 M CUt2t
 t
2
3t 

This recurrence formula determines the solution of U t  t by knowing each of

U t , U t  t , and U t 2 t which band three steps of t increments. These required

values might be calculated using some other means, i.e. a special starting
procedure must be employed. The central difference scheme may be used with
a time step taken as a fraction of t adopted in Houbolt method.
In Eq. (3.19), the appearance of the stiffness matrix K as a factor of the
unknown displacement U t  t , is due to the formulation of the equilibrium

equations at the current time t+t. That makes Houbolt operator an implicit
operator. Hence, it requires a solution of a system of simultaneous equations.
A noteworthy point regarding Houbolt operator is that the method
reduces the problem directly to a static problem when the mass and damping
effects are neglected. The analysis of Houbolt method proved its unconditional
stability. But, the most significant drawback associated with Houbolt
integrator is the algorithmic damping, which is inherent in the numerical
procedure and is introduced into the response when large time steps are used.
In linear problems, this damping can cause the transient response to decay so
severely that a static response is obtained.

3.7.1.3 Newmark Generalized Acceleration Method

Newmark [11] presented a family of step-by-step direct integration methods,


which is the most widely used method for time integration of the semi-discrete
equations of motion. The solution at the end of a time step is expressed in
Chapter Three Numerical Solutions of Dynamics of Structural Problems 24

terms of a Taylor series with the remainder given by a quadrature formula.


Two free parameters of integration,  and  , indicate how much the
acceleration at the end of an interval enters into the relations for velocity and
displacement at the end of that interval.
The two approximations for the expressions of displacements and
velocities are

   t 2  1   U


U t  t  U t  tU    U
 
t  2 t t  t     
  
 
U  
t  t  U t  t 1   U t  U t  t 
       

The parameters  and  determine the stability and accuracy of the algorithm.
The value of  =0.5 ensures that no spurious damping forces arise. Choosing
 =0.5 and  =0.25 is equivalent to a constant acceleration assumption and

gives an unconditionally stable method for linear problems [11]. For  =1/6
and  =1/2, the relations (3.20) and (3.21) correspond to the linear acceleration
method [11].

Eqs. (3.20) and (3.21) can be matched to express U 
t  t and U t  t in

terms of U t  t which are substituted in the equilibrium equations. These two

equations

      

U t  t  U t  t  U t    1  U t  t 1  U t (3.22)
 t    2 

1 1   1  

U t  t  U t  t  U t   U t   1  U t (3.23)
t  2  t  2 

Considering the equilibrium equations at time t+t and using Eqs. (3.22)
and (3.23) gives

 1    1 1  1   
 M C  KU  R  M  U  U  
  1Ut  
 t2  t  tt t t  t2 t
 t
t
2
     
(3.24)
       
C Ut    1U t  t
  1U t
  t     2  
Chapter Three Numerical Solutions of Dynamics of Structural Problems 25

The widely used Newmark method of  =1/2 and  =1/4 (constant


average acceleration method) does not introduce any artificial damping. The
only error associated with this formulation is period elongation increasing
t
with the ratio where T is the time period of a given mode of vibration. This
T
implies that it is impossible to suppress undesirable higher modes with highly
inaccurate time periods resulting from discretization of continuous problems
t
and large value of . This can be reduced by choosing a value of  greater
T
than 1/2 such as
1
  (3. 25)
2
where  is a positive value that can be used to control the artificial
viscosity introduced in the solution. The value of  should then be greater than
1/4 to preserve unconditional stability for linear systems.
The stability condition for the two parameters  and  is
1 1 1
 ;   (  ) 2 (3.26)
2 4 2
The advantages of using  damping with Newmark method over the
Houbolt and Wilson- methods, both of which possess inherent damping, is a
greater degree of control over the degree of damping that is introduced.
It can be seen that if  is taken equal to zero, negative damping results,
which will involve a self-excited vibration arising slowly from the numerical
procedure. Similarly if  is grater than 1/2 i.e. for positive value of , a positive
damping is introduced which will reduce the magnitude of the response even
without the existence of real damping in the problem.
The methods that can be derived from the general Newmark method as
well as some of their characteristics are shown in the Table (3.1).
Chapter Three Numerical Solutions of Dynamics of Structural Problems 26

Table (3.1). Properties of well-known members of the Newmark family


methods.
Method Type   Stability Order of
Condition accuracy
Average Acceleration Implicit 1/4 1/2 Unconditional 2
(trapezoidal rule)
Linear Acceleration Implicit 1/6 1/2 t  0.5641 Tn 2
Fox-Goodwin Implicit 1/12 1/2 t  0.3898 Tn 4
Central Difference Explicit 0 1/2 t  0.3183 Tn 2

3.7.1.4 Wilson- Method

The linear acceleration method as a member of the Newmark family


(=0.5,  = 1/6) was modified by Wilson [12] to generate an implicit algorithm.
This modification assumed a linear variation of acceleration to exist over a
time interval 2t, at the end of which a solution is obtained using standard
linear acceleration method. The solution at the end of the interval t is, then
obtained by the use of the kinematic relations originating from the assumption
of linear acceleration variation. This method is shown to be unconditionally
stable for linear systems, but possess considerable amount of inherent
damping, more than sufficient to suppress the superior oscillation of
discretized system. A further extension of this method was made by
Farhoomand [13] by using t rather than 2t and then obtaining the optimal
value of . It was shown that  must be greater than 1.377 for unconditional
stability. Reduction of  from the previous value of 2 had the effect of reducing
the amount of damping introduced into the system.
The Wilson- method equations are based on the following
assumptions, for any time  ,as 0    t . The linear variation of acceleration
at this interval yields
 

U 
t  U t 
t
U t t  U t  (3.27)
Chapter Three Numerical Solutions of Dynamics of Structural Problems 27

Integrating (3. 27) results in


2
   U
t  t  U t 

U    U
U   (3.28)
t t t
2 t
and

2  3 
 
U t  U t   U t
2
Ut 
6t
U tt  U t  (3.29)

using (3.28) and (3.29) at time  = t yields


t 

U 
t  t  U t 
2
U tt  U t  (3.30)

  t 
2
U t  t  U t  tU t
2
U t t  U t  (3.31)


which can be solved for U 
t  t and U t  t in terms of U t  t as

6 6 

U t  t  U t  t  U t   
U t  2U t (3.32)
t  2 t

3

U t  t  U t  t  U t   2 U t  t U

t (3.33)
t 2
To obtain the solution for displacements, velocities, and accelerations at time
t+t, the equilibrium equations of motion are considered at time t+t

MU 
  t  t  CU t  t  KU t  t  R t  t     
Substituting Eqs. (3.32) and (3.33) into Eq. (3.34) and rearranging gives
 6 3   6 6   
 M C  KUtt  Rtt  M U
2 t
Ut  2U t
 t  t
2
t  t  (3.35)
3   t U
 C Ut  2U t
 
t
 t 2 
Solving the above simultaneous system for U t  t and using (3.32) and (3.33)
6

U t  t  U t  t  U t   26 U t   1  3 U

t (3.36)
  t 
3 2
 t  
t 

U 
t  t  U t 
2
  
U t  t  U t (3.37)

  t  U
2
U t  t  U t  tU t
6
  
t  t  2 U t  (3.38)
Chapter Three Numerical Solutions of Dynamics of Structural Problems 28

3.7.2 The Finite Element Family of Methods


A trial function finite element discretization according to the general weighted
residuals method can be applied to the time domain. As the time dimension is
of an infinite extent, it can be represented by a finite domain of time and
repeated subsequently by domains with new initial conditions. The process
will thus lead to step-by-step or recurrence calculations which are recognized
to be similar to many time stepping procedures which were generally derived
by the finite difference method. These will, however, turn out to be much
more general.
For the system of equations of motion, various two- and three-step
schemes can be developed using “finite elements in time”. This will give rise
to new formulations with special merits in addition to several well known
“finite difference” expressions which can be obtained.
For prismatic problems (i.e. the problems that do not change their space
boundary with time), the solution of finite elements in time will give the same
results obtained in a full space-time finite element solution [40]. Even in this
case, i.e., for prismatic problems, the choice of achieving spatial finite element
discretization and then the time finite element solution, is preferable.
It has been proved that many of the standard recurrence formulae that
are alternatively derived can also be arrived at by using finite elements in
time. Indeed, the process of applying finite elements in the time domain
appears to encompass all alternative methods of time discretization and in
addition presents many new interesting possibilities [40].
In the present work, two-time marching schemes for the solution of
second order differential equations of motion are presented. Both were
derived from the general weighted residuals method, these are:
1- The Direct Weighted Residuals Algorithms.
2- The Unified Single-Step Algorithms SSpj.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 29

3.7.2.1 The Direct Weighted Residuals Algorithms

The discretization of displacements in the time domain using finite element


convention can be written as
 
U( t )  U( t )   N l ( t )U l (3.39)
l 1

Where U l is the nodal set of displacements U at time node l. The base function

Nl (t) is an approximate function in time, which approximates the field

variable variation, U(t), within the elements. The base functions Nl(t), are
assumed to be the same for each component of a vector U, and therefore Nl (t)
is a scalar.
The lowest order of polynomials necessary for the expression of base
function Nl (t) is a second order polynomial, because of the appearance of a

second order derivative as the highest derivative in the equations of motion.


That leads to the lowest type of element to be a quadrilateral element.
Two particular members of this general method will be considered in
present study, those are:
1- The Weighted Residuals Three-point Recurrence Scheme.

2- The Weighted Residuals Four-point Recurrence Scheme.

3.7.2.1.1 The Weighted Residuals Three-point Formula

This formula is derived by the assumption that the time domain of interest is

broken into a finite number of one-dimensional isoparametric elements of

length (2t). Figure (3.1) shows a typical element, n, with three nodes
(n-1, n, and n+1). The corresponding nodal displacements are ( U n 1 , U n , and

U n 1 ). The figure also shows nodal base functions variations, as well as the

most frequently used weights functions.


Chapter Three Numerical Solutions of Dynamics of Structural Problems 30

The interpolation functions and their first and second derivatives can be

written in terms of a local variable , as

 1
  
  1     2   1
N n 1  ; N n 1   ; N n 1 
2 t t 2
N n  1  2 ; N   2  ; N    2
n n
t t 2 (3.40)
 1
  
 1     2   1
N n1  ; N n1   ; N n1 
2 t t 2
  t  t n   t ;  t  t n  1  t n  t n  t n  1

The displacement distribution within the element (n) according to Eq. (3.39)

will take the form

 n n n
U(t )  N n 1 U n 1  N n U n  N n  1 U n  1 (3.41)

A typical weighted residuals equation can be written assuming that a

full domain of interest corresponds to the one element of size (2t) and
assuming that the values of vectors Un-1 and Un are prescribed. Thus, leaving
only Un+1 to be determined. A single weighted residuals equation needs to be

considered.

tn1    
2
 d U dU 

  dt 2
W M  C
dt
 K U  R dt  0 (3.42)
t n 1
 

Matching Eqs. (3.40) and (3.41) and substituting into Eq. (3.42) yields
Chapter Three Numerical Solutions of Dynamics of Structural Problems 31

  1 2 1  
M  U
  2 n1  U  U   
 t t2 n t2 n1  
 
 1
1    
1  
 
  2 2 2 
WC

1  
t
Un1  Un 
t t
Un1 


d 0
 (3.43)
   
 1 1 
 
K Un1 1 Un 
2
Un1 R
  2 2  
 
Various weighting functions can be inserted in this general equation, which

can be simplified for any weight function as

 1 Z   2 12Z 1  
 M C KUn1  M C 2 K
t2 
 t
Z  t2 
t

2
 Z Z
 Un
 Z   
 1 1Z 1  
t2 M C  Z ZK
  Un1 R
 t 2  
(3.44)

in which

1 1
 1
Z 

1
W    d
 2  Wd
1
1 1
Z 
 W1  d  Wd
1 1
(3.45)
1 1
R
1
 WRd  Wd1

Arranging Eq. (3.44) and putting all known terms in the right hand side yields
Chapter Three Numerical Solutions of Dynamics of Structural Problems 32

 1 Z   2 1 
   
t2 M  t CZKUn1 RMt2 Un  t2 Un1 
 Z   
 2 1 1 Z 
C z Un  Un1 
 t t  (3.46)
 1  1 
K2Z  Z  Un Z Z  Un1 
 2  2 

This final recurrence equation calculates the current displacements U t  t from

U t and U t t . For this reason it is called a three-point formula.


1 1
1 
 z   W  d  Wd
1  2  1

11 1
z   W1  d  Wd
2 1 1

z  z Method

-1/2 0 Forward

1/2 0 Central explicit

3/2 1 Backward

1/2 1/6 Linear acceleration

1/2 1/10 Galerkin, 2nd point

3/2 4/5 Galerkin, 3rd point

1/2 1/12 Fox-Goodwin

1/2 1/4 Average acceleration

Fig. (3.1). Shape and weight functions for three-point recurrence


formulae [29].
Chapter Three Numerical Solutions of Dynamics of Structural Problems 33

3.7.2.1.2 The Weighted Residuals Four-Point Formula

The weighted residuals approach in the previous section can easily be


extended to derive a higher order formula by using a larger number of time
stations within the element (higher order element).
Figure (3.2) shows a typical element of length 3t with four-nodal
points (n-2, n-1, n, and n+1). The figure also shows two of the weighting
functions.
The interpolation functions, as well as their first, and second derivatives
are expressed in terms of local coordinates as follows:
  1  2  3   32  12  3 
Nn2   ; Nn2  ; Nn2    2
6 6
  2  3  2
3  10  6 
Nn1  ; Nn1  ; Nn1  3  5
2 2
   1  3   32  8  3 
Nn  ; Nn  ; Nn    4 (3.47)
2 2
  1  2  32  6  2 
Nn1  ; Nn1  ; Nn1    1
6 6
  t t

Substituting Eq. (3.42) in the governing equation of motion, and proceeding as


was done in deriving Eq. (3.46) from Eq. (3.42). The cubic formula with three
parameters  z ,  z , and z is obtained:

  1 
  1   
 z 2 M 2 3 C z  z  z KU  RM 3z 4U  53z U  z 2U 
    
 t t  6 2 3 
n1  t2 n t2 n1 t2 n2 
   
 
 3z 3 3 
 4z  5z  z 3 

 z 1
C 2 2U 
n
2 U 
n1  2z   n2 
U
 t t 2 6 
 
 
  3    5    11  
K z  z 2z Un  z 3z  z Un1  z  z z 1Un2 
 2 2   2 2  6 6  

(3.48)
where
Chapter Three Numerical Solutions of Dynamics of Structural Problems 34

3 3

 
3
 Z  W  d  Wd 
0 0
3 3

 
2
 Z  W  d  Wd 
0 0 (3.49)
3 3
 Z  
0
W d   0
Wd 

3 3
R  
0
WRd  
0
Wd 

3 3

  Wd
3
 z  W  d
30 30

  Wd
2
 z  W  d
03 30


 z  W d
0
 Wd
0
z z z

30 9 3

702 36 13
35 5 5

Fig. (3.2). Shape and weight functions for four-point recurrence


formulae [29].

3.7.2.2 The Unified Single-Step Algorithm (SSpj)

A general algorithm, SSpj, for single-step time marching scheme for use in the
solution of second order dynamic equations or first order diffusion equations
or similar problems was presented by Wood [31] and by Zienkiewicz et.
al.[32]. The subscript p denotes the desired order of approximation, while j
denotes the order of the equation to be solved. The algorithm is based on the
general weighted residuals theory, and has many additional desirable features over
Chapter Three Numerical Solutions of Dynamics of Structural Problems 35

the approach of the previous section (the direct weighted residuals


formulation), which can be listed below:
1- The generality of the algorithm extends to cover the first order diffusion
equation, the second order dynamic equation and higher order equations
In all these problems any desired order of approximation can be chosen.
2- The algorithm can easily be programmed in its universal form for all
orders of approximations without considerable additional effort.
3- The algorithm has the inherent advantage, that it is a simple matter to
alter the time step as the requirements of the solution dictate.
4- The algorithm generates a series of general algorithms of order p which
embrace the standard Method in first order problems and Newmark
family of methods in second order problems.
The derivation of the general formula can be achieved by considering a
single time interval t, and approximating the displacement function U by a
polynomial of degree p in time (taken as zero at point n), i.e. 0  t  t ,
Fig. (3.3).

Fig. (3.3). A second order time step approximation [32].

The displacement function will take the form

  t2 (p) t
p
U(t )  U n  U n t  U n  ...........   n (3.50)
2 p!
Chapter Three Numerical Solutions of Dynamics of Structural Problems 36

or more compactly as
p1 q
tq (p) t
p
U(t ) 
q 0
Un 
q!
 n
p!
(3.51)

q
where U n is the q-order derivative of the vector U at node n with respect to
p 1
 , ......., U n are known values at the
time. The algorithm assumes that U n , U n

start of the interval and  np  is the unknown vector from which,
p 1

U n1 , U n  1 , ............, U n  1 will be determined as

p 1
t q t p t p

q 
U n1  Un   n (p)  U n1   n (p)
q! p! p!
q 0
p 1
t q1 t p1  n  1   ( p ) t 
 p 1

U
q
 (p)
U n1  n  n U
q1
q  1! p  1! n
p  1!
. (3.52)
.
etc.
 
 n  1 , ......... etc. can be viewed simply as extrapolated
The values of U n  1 , U

(predictor) values and  n (p ) as an average of the pth derivative in the interval


[Fig. (3.3)].
To determine the unknown parameter  n (p ) a weight function is
multiplied by the equation of equilibrium of motion i.e.
t

 W MU (t)  CU (t)  KU(t)  R(t)dt  0


0
(3.53)

t t

defining

0
Wt dtq
 0
Wdt  qt q (3.54)

q=1 to p; o = 1; 0   q  1

Matching Eqs. (3.54) and (3.51) gives


Chapter Three Numerical Solutions of Dynamics of Structural Problems 37

t t p1 q
t q t p
 WUdt  Wdt  
q 0
Un
q!
 q   n p 
p!
p
0 0
t t p1 q
t q1 p  t 
p1


0
 dt
WU
0 Wdt  
q 1
Un  
q  1! q1 n p  1! p1

(3.55)
t t p1 q
t  q2
t  p 2

  dt
WU  Wdt  U
q2
n
q  2 !
 q  2   n p  
p  2 ! p2
0 0
t t

 WRdt  Wdt  R
0 0

t

Substituting Eq. (3.55) into Eq. (3.53) and dividing by Wdt gives

0

 p1 q t q  2 t p2 



M U n
 q  2 q  2 !
 q  2   n p 
p  2 !
 p 2  

 p1 q t q 1 t p1 

C U n
 q 1 q  1!
 q 1   n p 
p  1!
 p1  

(3.56)

 p1 q t q 
p  t 
p


K U n
 q 0 q!
q   n 
p!
p   R  0


From the above equation,  n p  can be found as

 
1
p 

tp2tp1 tp  ~
 ~
 ~
n  p2M p1C pK RMUn1 CUn1 KUn1 (3.57)
p2! p1! p! 

~ ~ ~

where U n  1 , U n  1 , and U n 1 can be considered as a mean predictor values of
 
U n 1 , U n  1 , and U n  1 in the interval which can be defined by
Chapter Three Numerical Solutions of Dynamics of Structural Problems 38

p1 q
~ t q
U n 1  
q 0
Un
q!
q

p1 q
~ t q1
U n 1  
q 1
Un 
q  1! q1
(3.58)

p1 q
~ t q2

U n 1  
q2
Un 
q  2 ! q2

The final step is to substitute the value of  (np) obtained in Eq.(3.57) into

p1

Eq.(3.52) and obtain the required vectors ( U n  1 , U 
n  1 , U n  1 ,....., U n  1 ). As

shown above, the pth order algorithm requires values of U o and all initial
values of the time derivatives up to order p-1 to be defined for the solution to
proceed.
It should be noted that having chosen the polynomial of order p, the
error associated with the displacement U in specific time interval, will be of
(p+1) order and the algorithm is at least p-order accurate in a given time step,
provided exact starting values are used. The lowest order of approximation for
dynamic problems is p=2, to be able to approximate the second derivative in
the governing equation.
Two particular cases of the general algorithm SSpj dealing with
dynamic equations will be considered in the present study, SS 22 and SS32.

3.7.2.2.1 Quadratic Algorithm SS22

For this algorithm, and from the previous general derivation, two parameters
 1 and  2 need to be specified together with the starting values of U o and
 that need to be given. The sequence of the solution is represented by three
U o

stages:
(i) Eq.(3.58) gives
~  t
U n 1  U n  U n 1 (3.59)
~ 
UU n (3.60)
Chapter Three Numerical Solutions of Dynamics of Structural Problems 39

(ii) Eq.(3.57) gives

f  C U~  K U~ 
1
2    t  2 
(3.61)
n   M   t 1 C   2 K 
 2 

(iii) Eq.(3.52) gives

   2  t 2
U n  1  U n  tU n n (3.62)
2

  2 
U n  1  U n   n t (3.63)

This particular algorithm has identical stability properties as the Newmark


algorithm, if the constants  and  of Newmark algorithm are identified as
follows:

  1
(3.64)
  0.5 2
Thus, the requirment for unconditional stability is
 1  0.5
(3.65)
2  1
Although SS22 is similar to Newmark, it has considerable computational
as well as physical advantages:

1- Only the natural initial conditions (displacement and velocity) are


needed to start the computations.

2- The values of  (np) represent the average acceleration U


 in the interval
n  
t , which is more realistic than point values that are generally computed.

The numerical answers, although not identical with Newmark, are very
comparable.
3- The governing equilibrium equation is satisfied in the mean throughout
the interval.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 40

3.7.2.2.2 Cubic Algorithm SS32

Again, the general derivation gives a particular algorithm SS 32 which is


essentially identical to the 4-points weighted residuals operator presented
before, but a more convenient single-step form is obtained.
Three parameters  1 ,  2 and  3 are obviously necessary and starting

 and U
values of U o , U  are needed. As in SS22 the sequence of solution
o o

equations follows the same three stages:


(i) Eq.(3.58) gives

~   t 2
U n  1  U n  U n t 1  U n 2 (3.66)
2
~  
UU n  U n t 1 (3.67)

~
 
U n 1  U n 1 (3.68)

(ii) Eq.(3.57) gives

f  MU~ 
1
3   t 2 t 3  ~ ~
n   t 1 M  C  3 K  n 1  CU  KU (3.69)
 2 6 

(iii) Eq.(3.52) gives

 t     3  t 
2 3
 U
U n  1  U n  tU (3.70)
n n n
2 6

 t    3  t 
2

U   U
U (3.71)
n 1 n n n
2
  3 
U n  1  U n   n t (3.72)

The algorithm SS32 embraces a series of well-known algorithms among them

(a) Houbolt algorithm (b) The Wilson-  procedure (c) Hilber-Hughes-Taylor


damped method.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 41

3.8 Computer Programs


Two computer programs were utilized throughout the present work, and both
programs were coded in FORTRAN.
The first program was developed to solve the discrete equations of
motion, using each of the eight integration operators investigated in the
present work (see the Appendix). The program was coded to study and
compare the performance of the two families of time marching schemes
(finite difference and finite element), and to exhibit their intrinsic
characteristics as stand alone schemes without the interaction with a finite
element program. This objective was met by addressing a system of vibrating
mass-spring components, where the equations of motion could be solved
directly, i.e. without the use of finite element discretization.
The second program was utilized to investigate the performance of the
investigated time stepping methods within a general finite element program.
The integration schemes were coded and implemented in a published finite
element program. The finite element program used was MIXDYN originally
presented by Owen and Hinton [41].
A brief summary of the main features of the program follows:
1- The program is designed to solve plane stress/strain and axisymmetric
dynamic problems.
2- Both material and geometrical nonlinearities are dealt with in the
program.
3- The program is written in modular form, which allows freedom of
choice as to the time integration operator.
4- The time integration operator originally adopted by the program is the
predictor corrector Newmark operator.
5- Three types of two dimensional isoparametric elements are presented in
the program. The 4-node linear, 8-node serendipity, and the 9-node
Lagrangian elements.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 42

6- The solution and storage scheme adopted in the program is the active-

column heights technique.

7- The program allows the choice of consistent or lumped mass matrices.

8- The adopted Gaussian quadrature numerical integration schemes are 3-

point formula for the mass matrices and 2-point formula for the stiffness

matrices.

The program was modified to adopt each of the eight time integration

schemes, which are investigated in the present work (see the Appendix).

3.9 Example Problem


A simple mass-spring system subjected to a piecewise linear forcing term was

selected for the direct numerical experiments [Fig. (3.4)]. As explained earlier,

the objective of this example is to investigate the characteristics of the eight

time integration schemes presented in this work prior to their implementation

in a finite element program. This example is taken from [16, 42].

m1=m2=m3=1.0
k1=1.0
k2=100.0
k3=10000.0
k4=2.0
k5=200.0

Fig. (3.4). Identification of the example problem [16, 42].


Chapter Three Numerical Solutions of Dynamics of Structural Problems 43

The physical parameters shown in Fig. (3.4) are of nondimensional form,


while the dimension of time is in units of seconds.

The differential equations of motion which govern this system, put in


matrix form, appears as follows:

0 U   k  k  k
m1 0 1  1 2 5  k2  k5 U1  0 
 0 m  
 2 0 U   
2   k2 k2  k3  k 4  k3 U2   0  (3.73)
 0 0 m3 U    k5  k3 k3  k5 U3  F
 3 

where

2 t 0t5

F t   20  2 t 5  t  10 (3.74)
0 t  10

This particular problem was chosen to ensure the development of a
wide spectrum of natural frequencies [16, 42]. These natural frequencies were
calculated and are listed together with their associated natural periods in
Table (3.2).
Table (3.2). Natural frequencies for the example problem.
Number f(Hz) Period (sec.)
1 0.159 6.2893
2 3.379 0.2959
3 22.593 0.0442

The frequency of the forcing function was constructed to be on the order of


lowest natural frequency so that the integration time step size may be chosen
as large as possible.

The results of a series of numerical experiments with the different


methods are presented as the displacement- time history of the third mass, m3.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 44

3.10 Discussion
3.10.1 Finite Difference Family
3.10.1.1 Central Difference Method

Figure (3.5) shows the instability phenomenon associated with the explicit
central difference operator as a function of time-step length. The critical
time-step length shown in the figure is 0.014088 sec which satisfies the
criterion (3.15) exactly. The criterion relates the critical time step with the
highest natural frequency of the system
1
t  t crit . 
fn

where fn is the highest natural frequency of the system, here (fn = 22.593 Hz).
The figure shows how a slight increase in time step above the critical limit
leads to the deterioration of the solution and produces unbounded response.

3.10.1.2 Houbolt Method

Figure (3.6) shows the behavior of the Houbolt operator solution with the
variation of time step length. The figure shows that increasing the time step
size produces various errors, i.e. the attenuation of the response as well as
phase shift.
However, the figure shows good results for time step length up to
(t=0.1 sec.) which is approximately 7 times the critical time step of the explicit
central difference method. In other words, 300 time steps were needed to
arrive at the response shown in the figure by the use of Houbolt method, while
the central difference solution, Fig. (3.5), required 2130 time steps.

3.10.1.3 Wilson- Method

Figure (3.7) shows the effect of the variation of the parameter  on Wilson-
method for a constant time step (0.3 sec.). The Wilson- operator needs a value
Chapter Three Numerical Solutions of Dynamics of Structural Problems 45

of 1.377 or greater for the method to be unconditionally stable. The figure


shows how the instability delayed as the value of  approaches the critical
limit. As depicted in Fig. (3.7) a value of =1.34 was needed to produce the
instability phenomenon. On the other hand, values of  greater than 1.377,
=2.0 for example, causes period elongation of the response.
In Fig. (3.8) with  made equal to 1.377 (the smallest value to achieve
unconditional stability), the variation of time step length causes both artificial
damping and period elongation, however, these two errors were less severe in
the Wilson- method as compared to the Houbolt method.

3.10.1. 4 Newmark Method

Figures (3.9) and (3.10) show the effect of varying  and  parameters on
Newmark operator. The two figures tested the criterion (3.26) for
unconditional stability:
  0.5 ;   0.25  0.52
Figure (3.9) shows the effect of variation of  keeping =0.25. The values of 
greater or smaller than 0.5 do not satisfy the criterion above and thus yield
unstable response as shown in the Figure.
Figure (3.10) where is varied while  is kept equal to 0.5, shows that
for =0.246 (which do not satisfy the criterion above), the solution oscillates
and is unbounded, while for >0.25, the solution exhibits period elongation
and gives arithmetic negative damping (self excitation), however, this
behavior need a value more greater than 0.25 (1.5 in the figure) to be noticed.
On the other hand, keeping =0.5 and =0.25 (equivalent to constant
average acceleration) produce very good results for a relatively large time step
size (0.3 sec.) shown in Fig. (3.11), which is 21 times the critical time step of the
explicit central difference method.
The behavior of Newmark constant average acceleration operator as a
function of time step length is presented in Fig. (3.11). The response exhibits
no artificial damping even when t takes relatively large values. The only
error associated with large t is period elongation. The accuracy of this
Chapter Three Numerical Solutions of Dynamics of Structural Problems 46

5.0 5.0

4.0 Exact 4.0 Exact


t=0.1 sec.
Displacement of mass 3

Displacement of mass 3
t=0.014088 sec.  t=0.3 sec.
3.0 t=0.014089 sec. 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.5). Explicit central difference Fig. (3.8). Wilson- method using
method using two time-steps different time-steps increments (=1.377)
5.0 5.0

4.0 Exact 4.0 Exact


t=0.1 sec.  = 0.50
Displacement of mass 3
 = 0.40
Displacement of mass 3

 t=0.3 sec.
3.0 t=0.5 sec. 3.0  = 0.52
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.6). Houbolt method using Fig. (3.9). Newmark method with
different time-step increments different ’s keeping =0.25 (t=0.3 sec.)
5.0 5.0

4.0 Exact 4.0 Exact


 =1.40  = 0.250
Displacement of mass 3
Displacement of mass 3

 =1.34  = 1.500
3.0  =2.00 3.0  = 0.246

2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.7). Effect of -variation on Fig. (3.10). Newmark method with
Wilson- method different ’s keeping =0. 5 (t=0.3 sec.)

Chapter Three Numerical Solutions of Dynamics of Structural Problems 47

unconditionally stable implicit operator is the best when compared with


Houbolt and Wilson- implicit unconditionally stable methods.

3.10.1.5 Other Variations of Newmark Methods

The other three well-known operators covered by Newmark method each are
conditionally stable. These are the explicit central difference, the Fox-
Goodwin, and the linear acceleration operators.
Figure (3.12) shows the explicit central difference operator as a
particular case of Newmark method (=0.5, =0.0). The behavior is exactly
identical to the explicit central difference method shown in Fig. (3.5).
The second operator is the Fox-Goodwin operator (=0.5, =1/12) which
is shown in Fig. (3.13). The critical time step value is that presented in
Table (3.1), i.e., tcrit. =0.1725 sec. which is 1.225 times that of the central
difference method.
The linear acceleration operator (=0.5, =1/6), on the other hand has
tcrit.=0.0244 sec. and also satisfies the value in Table (3.1), and is 1.772 times
that of the central difference method.

3.10.2 Finite Element Family


3.10.2.1 The Three-Point Formula

Figures (3.15), (3.16), (3.17), and (3.18) shows how this operator embraces the
Newmark operator by introducing the constant average acceleration, explicit
central difference, Fox-Goodwin, and linear acceleration operators
respectively.
Two additional operators are presented in Figs. (3.19) and (3.20), the
backward point collocation and the Galerkin method. Each of these operators,
as depicted by the figures, exhibits both artificial damping and response shift
errors. It is noticed that the imposed damping of these operators is very
severe. However, the backward point collocation operator is better than the
Galerkin with respect to both of these two errors.
Chapter Three Numerical Solutions of Dynamics of Structural Problems 48

5.0 5.0

4.0 Exact 4.0 Exact


t=0.1 sec. t=0.024400sec.
Displacement of mass 3

Displacement of mass 3
 t=0.3 sec. t=0.024415sec.
3.0 t=0.5 sec. 3.0

2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.11). Constant average acceleration Fig. (3.14). Linear acceleration version of
version of Newmark’s method (=0.5, Newmark’s method (
=0.25)
5.0 5.0

4.0 Exact 4.0 Exact


t= 0.014088 sec. t=0.1 sec.
Displacement of mass 3

Displacement of mass 3
t= 0.014089 sec.  t=0.3 sec.
3.0 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.12). Explicit central difference Fig. (3.15). Constant average acceleration
 version of Newmark’s method (=0.5, =0) version of 3-point formula (z=0.5, z=0.25)
 
5.0 5.0

4.0 Exact 4.0 Exact


t=0.017250sec. t=0.014088sec.
Displacement of mass 3

Displacement of mass 3

t=0.017257sec. t=0.014089sec.
3.0 3.0

2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.13). Fox-Goodwin version of Fig. (3.16). Explicit central difference
Newmark’s method (=0.5, =1/12) version of 3-point formula (z=0.5, z=0)

Chapter Three Numerical Solutions of Dynamics of Structural Problems 49

3.10.2.2 The Four-Point Formula

Figures (3.21), (3.22), and (3.23) show how the finite element four-point

formula covers each of Houbolt, Wilson-, and Fox-Goodwin operators,


respectively.
The formula also offers an additional operator that is the Galerkin
operator. Figure (3.24) shows the characteristics of this operator as a

conditionally stable operator with tcrit.=4.188 times that of the central


difference method. This critical time step value is the highest among the three

conditionally stable operators presented earlier. The other operator in Fig.


(3.25) gives bounded but oscillatory response. This oscillation takes place

when the time step is made greater than 0.059 sec., which is the same as the
Galerkin operator critical time step.

3.10.2.3 The Single-Step Method (SS22)


The quadratic single-step formula SS22 encompasses the generalized Newmark
method. Thus it can cover the constant average acceleration, the explicit
central difference, the Fox-Goodwin, and the linear acceleration operators.
This is illustrated in Figs. (3.26), (3.27), (3.28), and (3.29), respectively.

3.10.2.4 The Single-Step Method (SS32)

The cubic single-step formula covers the Houbolt, the Wilson-, and the linear
acceleration operators as special cases, Figs. (3.30), (3.31), and (3.32),
respectively.
Two additional unconditionally stable members presented in Figs. (3.33)
and (3.34). Figure (3.33) produce unconditionally stable operator with very
encouraging characteristics. The accuracy of this operator with the variation of
Chapter Three Numerical Solutions of Dynamics of Structural Problems 50

5.0 5.0

4.0 Exact 4.0 Exact


t=0.017256sec. t=0.1 sec.
Displacement of mass 3

Displacement of mass 3
t=0.017257sec. t=0.3 sec.
3.0 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.17). Fox-Goodwin version of Fig. (3.20). Galerkin version of 3-point
3-point formula (z=0.5, z=1/12) formula (z=1.5, z=0.8)

5.0 5.0

4.0 Exact 4.0 Exact


t=0.024400sec. t=0.1 sec.
Displacement of mass 3

Displacement of mass 3
t=0.024415sec. t=0.3 sec.
3.0 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.18). Linear acceleration version of Fig. (3.21). Houbolt version of 4-point
3-point formula (z=0.5, z=1/6) formula (z=27, z=3, z=9) 
5.0 5.0

4.0 Exact 4.0 Exact


t=0.1 sec. t=0.1 sec.
Displacement of mass 3

Displacement of mass 3

t=0.3 sec. t=0.3 sec.


3.0 t=0.5 sec. 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.19). Backward point collocation Fig. (3.22). Wilson- version of 4-point
version of 3-point formula (z=1.5, z=1.0) formula (z=16.224, z=2.4, z=6.093)

Chapter Three Numerical Solutions of Dynamics of Structural Problems 51

5.0 5.0

4.0 Exact 4.0 Exact


t=0.017256sec. t=0.1 sec.
Displacement of mass 3

Displacement of mass 3
t=0.017257sec.  t=0.3 sec.
3.0 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.23). Fox-Goodwin version of Fig. (3.26). Constant average acceleration
4-point formula (z=9, z=2, z=4.1666) version of SS22 method (1=0.5, 2=0.5)
5.0 5.0

4.0 Exact 4.0 Exact


t=0.059 sec. t= 0.014088 sec.

Displacement of mass 3
Displacement of mass 3

t=0.065 sec. t= 0.014089 sec.


3.0 3.0

2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.24). Galerkin version of 4-point Fig. (3.27). Explicit central difference
formula (z=20.05714, z=2.6, z=7.2) version of SS22 (1=0.5, 2=0)
5.0 5.0

4.0 Exact 4.0 Exact


t=0.059 sec. t=0.017256sec.
Displacement of mass 3

Displacement of mass 3

t=0.065 sec. t=0.017257sec.


3.0 3.0

2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.25). Weighted residual 4-point Fig. (3.28). Fox-Goodwin version of SS22
formula (z=15.25, z=2.5, z=3.0) method (1=0.5, 2=1/6)
Chapter Three Numerical Solutions of Dynamics of Structural Problems 52

5.0 5.0

4.0 Exact 4.0 Exact


t=0.024400sec. t=0.017256sec.
Displacement of mass 3

Displacement of mass 3
t=0.024415sec. t=0.017257sec.
3.0 3.0

2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.29). Linear acceleration version of Fig. (3.32). SS32 with (1=0.6, 2=0.4333,
SS22 method (1=0.5, 2=1/3) 3=0.3)
5.0 5.0

4.0 Exact 4.0 Exact


t=0.1 sec. t=0.1 sec.
Displacement of mass 3

Displacement of mass 3
t=0.3 sec. t=0.3 sec.
3.0 t=0.5 sec. 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.30). Houbolt version of SS32 Fig. (3.33). SS32 with (1=0.78, 2=1.05,
(1=2.0, 2=11/3, 3=6.0) 3=1.4)
5.0 5.0

4.0 Exact 4.0 Exact


t=0.1 sec. t=0.1 sec.
Displacement of mass 3

Displacement of mass 3

t=0.3 sec. t=0.3 sec.


3.0 t=0.5 sec. 3.0 t=0.5 sec.
2.0 2.0

1.0 1.0

0.0 0.0

-1.0 -1.0

-2.0 -2.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (sec.) Time (sec.)
Fig. (3.31). Wilson- version of SS32 Fig. (3.34). SS32 with (1=0.6, 2=0.7666,
(1=1.4, 2=1.96, 3=2.744) 3=0.9)
Chapter Three Numerical Solutions of Dynamics of Structural Problems 53

time step is greater than the average acceleration method. The figure shows
period elongation, but only slight damping appears.
The other operator in Fig. (3.34) exhibits good characteristics for large
time steps, the operator overestimates the response by producing negative
damping. The operator is also accompanied by very small phase shift error.

3.11 Conclusions
It is clearly apparent from the numerical experiments thus conducted that the
finite element method in both of its versions: the direct weighted residuals and
the unified single-step SSpj methods contain as special cases all finite
difference methods presented in this work. The finite element approach also
offers many additional operators with special merits.

You might also like