A new algorithm for non-linear dynamic structural analysis

B.W. Golley
School of Civil Engineering, UNSW@ADFA, Australian Defence Force Academy, Canberra, ACT
2600, Australia.

A.D. Mitchell
Royal New Zealand Engineers, Palmerston, New Zealand.

J. Petrolito
Department of Physical Sciences and Engineering, LaTrobe University, Bendigo, VIC 3550,
Australia.

ABSTRACT: The analysis of structures under dynamic loading is generally carried out
using time stepping methods, of which the most popular is Newmark's method. In most
non-linear methods, dynamic equilibrium is satisfied at the ends of the time step, although
in some methods it is satisfied at the midpoint. In this paper, linear structures are first
considered, and properties of the exact amplification matrix are noted. An approximate
procedure is then developed using cubic interpolation, with coefficients being chosen
using weighted residuals with unspecified weight functions. Properties of the weight
functions are then determined such that the amplification matrix is of a similar form to the
exact matrix, and are unconditionally stable, permits controlled algorithmic damping and
is of the maximum available accuracy. The procedure is shown to be equivalent to a
weighted sum procedure based on two pairs of sampling points, thus satisfying dynamic
equilibrium in an average sense. The weighted sum procedure is ideal for solving non-
linear problems.

1 INTRODUCTION
In structural dynamics, the equations of motion for a spatially discretised multi-degree of freedom
(MDOF) linear system are of the form
[ M ]{ a( t )} + [C ]{v (t )} + [ K ]{ x(t )} − {F (t )} = {0} (1)
where [M], [C] and [K] are the mass, damping and stiffness matrices, {a(t)}, {v(t)} and {x(t)} are ac-
celeration, velocity and displacement vectors, {F(t)} is a prescribed force vector and t is time. If there
are N degrees of freedom, [M], [C] and [K] are (N x N) symmetrical matrices, and {a(t)},{v(t)}, {x(t)}
and {F(t)} are (N x 1) vectors. If there are material or geometric nonlinearities, the governing equa-
tion is

[ M ]{ a( t )} + [C ]{v (t )} + {FS ( x (t ))} − {F (t )} = {0}

where {FS (x(t))} are internal forces which are functions of displacement. Equation 1 or 2 is to be
solved subject to the initial conditions that at time t=0, the displacement and velocity vectors {x(0)}
and {v(0)} are prescribed.

Paper Number 20
Exact solutions may be obtained for the linear case, which involve the determination of 2N generally
complex eigenvalues and corresponding eigenvectors (see for example Meirovitch (1980)). With non-
linear cases, numerical time stepping procedures are adopted. Many time stepping algorithms are
available with a number of these being discussed by Hughes and Belytshko (1983) and Dokainish and
Subbaraj (1989a and 1989b), and Fung has recently developed a range of time stepping procedures
based on Padé approximations of the exact solution of single degree of freedom (SDOF) systems
(Fung 1999a, 1999b). An extensive mathematical treatment is given in the text by Wood (1990). In
single-step two-stage algorithms, the displacement and velocity are specified at time t=0 and are esti-
mated during the period 0<t<=∆t where ∆t is the time step. Values of displacement and velocity are
calculated at t=∆t, and become the initial values for the next time step. The most popular time step-
ping method is Newmark's method (Newmark 1959). At the recent Twelfth World Congress on
Earthquake Engineering, Newmark's method was the procedure used for the numerical solution of
Equation 1 in fifty seven papers. Wood (1990) has carried out extensive studies on Newmark's
method and the θ method (not to be confused with the Wilson θ method (Bathe and Wilson (1973)),
and as a mathematician lists four reasons for the popularity of Newmark's method, the last being that
"It has been in use since 1959''.

In this paper, the exact amplification matrix used with single -step two-stage algorithms is presented in
a form thought to be new, which is shown to have a number of properties which may be used to reduce
computation time. The single degree of freedom solution is then derived in a similar form, and some
interesting properties of the amplification matrix are observed. An unconditionally stable approximate
method is then developed using cubic polynomials as interpolation functions. A weighted residual
procedure using two unspecified weight functions is followed. "Areas" and "moments of area'' of the
weight functions are selected such that the algorithm has the following features:
• The approximate amplification matrix is of a form similar to the exact amplification matrix.
• Algorithmic damping may be incorporated by specifying a single parameter.
• The algorithm has maximum accuracy.
• Numerical operations are minimised.

Although the weight functions are not specified, equivalent weight functions consisting of two Dirac
delta functions are also determined. Use of these latter functions is equivalent to a weighted sum
method, which is then applied to solve non-linear problems.

2 EXACT AMPLIFICATION MATRIX

The displacement and velocity at time t may be expressed in terms of values at time t=0 by the
equation

 {x( t )}  [ A11 (t )] [ A12 ( t )]   {x (0)}  {F}1 

[ A (t )] [ A12 (t )]
[ A(t )] =  11  (4)
[ A21( t )] [ A22 ( t )]

is the amplification matrix. With linear problems, [A(t)] may be determined exactly (Meirovitch
(1980)). The solution involves the introduction of N subsidiary equations, and requires the
determination of 2N eigenvalues and eigenvectors. The eigenvectors have two important
orthogonality properties, which are used to uncouple the equations. Initial conditions are generally
incorporated by using one of the othogonality conditions. If both are considered however, it is
possible to show that the amplification matrix may be expressed in the form

2
 [ A11e (t )] [ A12e (t )]  [ R2e (t )][ C ] + [ R3e ( t )][ M ] 1[ R2e (t )][ M ]
[ A (t )] =  e  =  
e
t (5)
[ R3 (t )][ M ] 
e
 21
[ A ( t )] [ A22 (t )]   − t [ R2 ( t )][ K ]
e e

where the superscript e indicates the solution is exact. The matrices [R2(t)] and [R3 (t)] are symmetri-
cal. In the SDOF case, the amplification matrix is
 e 1 e 
 R2 (t ) c + R3e (t ) m R2 (t ) m
[ A (t )] =
e
t (6)
 − tRe ( t ) k
 2 R3e (t ) m 

where m, c and k are mass, damping coefficient and stiffness of the SDOF system. In the SDOF case
the matrices [R2 (t)] and [R3 (t)] become scalars R2 (t) and R3 (t) which may be determined analytically as

2 4km − c 2
R2e (t ) = sin( t )e − ct / 2 m (7)
4 km − c 2 2m
and

1 4 km − c 2 c 4km − c 2
R3e (t ) = (cos( t) − sin( t ))e −ct / 2m (8)
m 2m 4km − c 2 2m
Important properties of any approximate method are obtained from the eigenvalues of the amplific a-
tion matrix. It should be noted that the eigenvalues of the amplification matrix given by Equation 6
are
1 1
( R2 (t )c + 2 R3 (t )m ) ± R2 (t ) 4 km − c 2
2 2 (9)

for any functions R2 (t) and R3 (t). Thus any approximate amplification matrix which can be written in
the form of Equation 6 has complex eigenvalues if c2 -4km<0, ie the system is underdamped, and has
real eigenvalues if c2 -4km>=0. The resulting solution using such an amplification matrix therefore
automatically produces an oscillating displacement for underdamped systems, and a non-oscillating
displacement for overdamped systems. The single-step two-stage form of Newmark's method cannot
be written in the form of Equation 6 with general parameters β and γ.

3 APPROXIMATE AMPLIFICATION MATRIX

In this section, an approximate algorithm is briefly described. The displacement of a SDOF system,
x(t), is approximated by a cubic function. Two coefficients are chosen to satisfy the initial conditions,
and the remaining two coefficients are chosen to satisfy two weighted integral equations with
unspecified weight functions (for a description of this method with first order problems see
Zienkiewicz (1977)). The approximate amplification matrix is then determined, with lengthy
algebraic terms of the form
Aika = Aika ( m, c, k , ∆t , u10 , u11 , u12 , u13 , u 20 , u 21, u22 , u23 ) (10)

where ∆t is the time step, and the coefficients u ij are "areas" and "moments of area" of the two weight
functions. The eight coefficients are now selected in the following series of steps.
3.1 Establish desirable form of the amplification matrix

Solve the four equations implied by

3
  a 1 a 
R (∆t )c + R3a ( ∆t ) m R2 ( ∆t ) m 
[ A (∆t )] =  2
a
∆t (11)
 − ∆tRa ( ∆t ) k
 2 R3a ( ∆t ) m 

for two coefficients and the two approximate functions R2 a (∆t) and R3 a (∆t).
3.2 Increase order of accuracy of the algorithm

Eliminate terms O(∆t3 ) in the expansions of the differences R2 a (∆t) − R2 e(∆t) and R3 a(∆t)− R3 e(∆t).
3.3 Incorporate algorithmic damping

This is achieved by introducing a damping parameter ρb which is the spectral radius as ∆t tends to
infinity. Wood (1990) suggests this should be in the range 0.5-0.7 to algorithmically damp high
frequencies. If ρ b =1 there is no algorithmic damping, and if ρ b =0 there is algorithmic annihilation. It
is also convenient to introduce a related parameter
1 − ρb
β= (12)
1+ ρb

where β ranges from 0 (no algorithmic damping and maximum accuracy) to 1 (algorithmic
annihilation and minimum accuracy)
3.4 Increase numerical efficiency

A coefficient matrix requires decomposition during the numerical solution process. Efficiency can be
increased by selecting two terms uij such that the coefficient matrix is symmetrical.
3.5 Simplify expressions and scale for convenience

Expressions are simplified with no change in accuracy by scaling.

After these steps, which require extensive algebra using a symbolic computer language, we obtain the
following simple coefficients: u10 =12, u11 =6, u12 =4, u13 =3-β, u20 =-2β, u21 =u22 =u23 =-(1+β).
3.7 Stability

It may be shown that with the coefficients given in Para 3.6 the algorithm is unconditionally stable for
0<=β<=1

4 COMPARISON BETWEEN EXACT AND APPROXIMATE SOLUTIONS

Exact and approximate solutions may be compared by considering the functions R2 (∆t) and R3 (∆t). To
enable graphical comparison, the non-dimensional functions
1
G 2 = G2 ( β ,ζ , ω∆t ) = 4km − c 2 R2 ( ∆t ) (13)
2

G3 = G3 ( β ,ζ , ω∆t ) = mR3 (∆t ) (14)

are defined, where ω2 =k/m and ζ2 =c2 /4km. These functio ns are plotted for ζ=0 and ζ=0.1 in Figures 1
to 4. It will be observed from these graphs that the algorithm produces accurate results for time steps
up to a quarter of the period of vibration with no algorithmic damping and up to a sixth of the period
of vibration with algorithmic annihilation.

4
 Figure 1. The function G2 , ζ=0.0 Figure 2. The function G2 , ζ=0.1 .

Figure 3. The function G3 , ζ=0.0 Figure 4. The function G3 , ζ=0.1 .

5 WEIGHTED SUM EQUIVALENT

5.1 General case

With linear problems, the "areas" and "moments of area" of the weight functions are required, but not
the weight functions. However, to apply the method to non-linear problems it is necessary to prescribe
weight functions with the same values uij . It is a simple procedure to determine cubic functions with
the same values u ij , but of more use are combinations of Dirac functions at prescribed sampling times.
The weight function
Wi (t ) = wi1δ (t − τ i1∆t ) + wi2δ (t − τ i2 ∆t ), i = 1,2 (15)

5
where δ is the Dirac delta function, therefore consists of concentrated weights wi1 and wi2 at the
sampling times τi1 ∆t and τi1 ∆t. We find
w11 3β w 3β w w
= 6(1 − ), 12 = 6(1 + ), 21 = 1 − β , 22 = −1 − β (16)
∆t 3 + 9β 2 ∆t 3 + 9β 2 ∆t ∆t

1 1 1 1
τ 11 = (1 − β ) − 3 + 9β , τ 12 = (1 − β ) + 3 + 9β , τ 21 = 0, τ 22 = 1
2 2
(17)
2 6 2 6

These weight functions enable the method to be viewed as a weighted sum procedure. Using the
weights and locations given by Equations 16 and 17, and noting τ21 =0 and τ22 =1 for all β, the
weighted integral equations reduce to
w11r (τ 11∆t) + w12r (τ 21∆t ) = 0 (18)

and
w21r (0) + w22r (1) = 0 (19)

where r(τ∆t) is the residual at time τ∆t.

With no algorithmic damping, ie β=0, the weights and sampling times become
w11 w w w
= 6, 12 = 6, 21 = 1, 22 = −1 (20)
∆t ∆t ∆t ∆t

and

1 3 1 3
τ 11 = − , τ 12 = − , τ 21 = 0, τ 22 = 1 (21)
2 6 2 6

The sampling times are therefore the beginning and end of the time step and two Gauss points. Thus
in the linear case with no algorithmic damping, the method equates to zero the sum of the residuals at
the Gauss points, and also equates the residuals at the beginning and end of the time step. If the
applied force is approximated by a polynomial up to cubic order, the residual is a cubic function, and
consideration of cubic functions with these characteristics reveals the residual is zero twice within the
time step. The equivalent weighted sum procedure is readily applied to non-linear problems.

6 APPLICATIONS

The method has been successfully applied to the solution of the SDOF Duffing equation, and to
Bathe's pendulum (Bathe and Wilson, 1973). The latter problem has been considered in detail by
Crisfield and Shi (1994) , as most numerical solutions lead to a phenomenon called locking. Using the
proposed method, locking is avoided, although a large number of time steps are necessary for high
accuracy due to the contribution of high frequencies in the solution as discussed by Golley and
Petrolito (2002). The algorithm has also been used successfully by Mitchell (2001) in a detailed
investigation of lead core hysteretic bearings used in seismic isolation of buildings.

7 CONCLUSIONS

A new unconditionally stable algorithm for the linear and non-linear dynamic analysis of structures
has been presented. The algorithm is developed using two unspecified weighted functions, with

6
coefficients chosen to satisfy a number of criteria, including optimising accuracy and numerical
efficie ncy. Comparison of the approximate algorithm with exact solutions for the linear case shows
that large time steps are feasible. An equivalent weighted sum method is derived from the optimum
solution, which may be used to solve MDOF problems with material and geometric nonlinearities.

8 REFERENCES

Bathe, K.-J. & Wilson, E.L. 1973. Stability and accuracy analysis of direct integration methods, Earthquake
Engineering and Structural Dynamics, Vol 1, 283-291.
Crisfield, M.A. & Shi, J. 1994. A co-rotational element/time -integration strategy for non-linear dynamics,
International Journal for Numererical Methods in Engineering, Vol 37, 1897-1913.
Dokainish, M.A & Subbaraj, K. 1989a. A survey of direct time -integration methods in computational structural
dynamics---I. Explicit methods, Computers and Structures, Vol 32, 1371-1386
Dokainish, M.A & Subbaraj, K. 1989b. A survey of direct time -integration methods in computational structural
dynamics---II. Implicit methods, Computers and Structures, Vol 32, 1387-1401
Fung, T.C. 1999a. Weighting parameters for unconditionally stable higher-order accurate time step integration
algorithms. Part 1-First-order equations, International Journal for Numererical Methods in Engineering,
Vol 39, 941-970.
Fung, T.C. 1999b. Weighting parameters for unconditionally stable higher-order accurate time step integration
algorithms. Part 2-Second-order equations, International Journal for Numererical Methods in Engineering,
Vol 39, 971-1006.
Golley, B.W. & Petrolito, J. 2002. An investigation of Bathe's pendulum using new algorithms for non-linear
analysis, 2nd International Conference on Structural Dynamics and Stability, Singapore.
Hughes, T.J.R & Belytschko, T. 1983. A precis of developments in computational methods for transient
analysis, Journal of Applied Mechanics, ASME, Vol 50 1033--1041.
Meirovitch, L. 1980. Computational methods in structural dynamics, Rockville, Sijthoff & Noordhoff.
Mitchell, A.D. 2001. Seismic isolation, Thesis submitted to the School of Civil Engineering, UNSW@ADFA in
partial fulfilment of the requirements of the BE degree.
Newmark, N.M. 1959. A method of computation for structural dynamics, Journal of the Engineering
Mechanics Division, ASCE, Vol 85, No EM3, 67-94.
Wood, W.L. 1990. Practical time-stepping schemes, Oxford, Clarendon Press.
Hughes, T.J.R & Belytschko, T. 1983. A precis of developments in computational methods for transient
analysis, Journal of Applied Mechanics, ASME, Vol 50 1033-1041.
Zienkiewicz, O.C. 1977. The finite element method. London, McGraw-Hill.