Finite element method for structural dynamic

and stability analyses

Module-5

Time integration of equation of motion

Lecture-15 Second order implicit methods

Prof C S Manohar
Department of Civil Engineering
IISc, Bangalore 560 012 India

1
Recall

• Forward Euler scheme: Explicit. Conditionally stable for

damped system. Unstable for undamped system. First
order accuracy.

• Backward Euler scheme: Implicit. Unconditionally stable.

First order accuracy.

• Central difference scheme: Explicit. Conditionally stable.

Second order accuracy. Not self-starting.

2
Critical step size
2 T
tcr  =  Forward difference scheme
 
2 T
tcr  = Central difference scheme
 
When applied to MDOF systems, t must be selected such
that these conditions are met for all particpating modes.
 For direct integrations
2n 2
tcr  min = if n  n  Forward difference scheme
1 n  N n N
2 2
tcr  min = Central difference scheme
1 n  N n N
 For modal superposition method, each mode can be
integrated, in principle, with different step size with higher modes
requiring finer step sizes. 3
Bounds on  N

Let N be the highest natural frequency of an assembled FE

mesh with N e elements.
Let K e & M e be the element matirces for e  1, 2, , Ne .
Consider the eigenvalue problems K   
e e e 2
 M e e ; e  1, 2, , Ne .
Let Ne be the largest frequency of element e.
It can be shown that  N  max  Ne .
1 e  N e

2
Thus, for central difference method, t   tcr
max  N
e
1 e  N e

4
  Ne
 Ne Ne
N   K e  N  Nt K eN  U e
t

 K
t
 e 1 
We have N2  tN N   e 1
 e 1

N M N t 
Ne
 Ne Ne

N   M e  N  N M eN  I e
t

 e 1  e 1 e 1

For the e-th element, we have by property of Rayleigh's quotient

Nt K eN U e
   N   U e   N  I e
e 2 e 2

N M eN I e
t


Ne Ne Ne

U    I
e 2
e N Ie e
 max    max  Ne    N2  max  Ne 
2 2 2
 
2
N
e 1
Ne
 e 1
Ne
e
N
e 1
Ne
1 e  N e 1 e  N e 1 e  N e
I
e 1
e I
e 1
e I
e 1
e

Remarks
Sizes of K e & M e are taken to be that of K and M
In solving element level eigenvalue problem we have removed BCS
and interactions with other elements, 5
 AE  1 1 ml  2 1 L 
Axially vibrating rod K    M  
6 1 2 
t
l  1 1  3 E
AE  1 1 ml 1 0 
Axially vibrating rod K    M t  L
l  1 1  2 0 
1 E
Euler-Bernoulli Beam

 12 6l 12 6l   156 22l 54 13l 

 2
6l 2l 2   4l 2 3l 2 
EI  6l 4l ml  22l 13l
K 3 M
l  12 6l 12 6l  420  54 13l 156 22l 
   
 6l 2l 2
6l 4l 2   13l 3l 2 22l 4l 2 
A 2 
t  L
48I E

6
Average acceleration, linear acceleration, and Newmark β methods

Average acceleration method

1 U n 1 
U    U n 1  U n 
2
 U   
U    U n  U n 1  U n 
2
t Un 
 U n 1  U n  U n 1  U n 
2
2
U    U n   U n  U n 1  U n 

4
tn tn 1
t 2
 U n 1  U n  tU n  U n 1  U n 
4

7
 Average acceleration, linear acceleration and Newmark's methods

Linear acceleration method

 
U    U n  U n 1  U n  U n 1
t

U    U n  U n 
2
U n 1  U n  U   
2t
t Un 
U n 1  U n  U n t  U n 1  U n 
2
2 3
U    U n  U n  U n U n 1  U n 
2 6t 
t 2 t 2 tn tn 1
U n 1  U n  U n t  U n  U n 1  U n 
2 6

8
The Newmark method

Consider U  t 
U  t  t   U  t 
 U  t  t  for some    0,1
t
 U  t  t   U  t   tU  t  t 
Suppose we now approximate U  t  t  as a weighted sum
U  t  t   1   U n   U n 1
 U  t  t   U  t   t 1   U  t    U  t  t  
 U n 1  U n  t 1   U n   U n 1 

9
Now consider the Taylor's expansion
t 2
U  t  t   U  t   tU  t   U  t  2t  for some    0, 0.5 
2
Suppose we now approximate U  t  2t  as a weighted sum
U  t  2t   1  2 U n  2U n 1
 1  
 U n 1  U n  tU n  t    U n  U n 1 
2

 2  
Note:  & are treated as being independent

10
Average acceleration method Newmark's method
t
U n 1  U n  U n 1  U n  1
 = & 
1
2 4 2
t 2  Average acceleration method
U n 1  U n  tU n  U n 1  U n 
4 1 1
Linear acceleration method  = & 
6 2
t  Linear acceleration method
U n 1  U n  U n 1  U n 
2
t 2 Newmark's method
U n 1  U n  U n t  U n 1  2U n 
6 Implicit
Newmark method Self-starting
U n 1  U n  t 1    U n   U n 1  Single step
 1  
U n 1  U n  U n t  t     U n  U n 1 
2

 2  

11
Newmark's method

U n 1  U n  t 1   U n   U n 1  1
 1  
U n 1  U n  U n t  t     U n  U n 1 
2
 2
 2  
MU n 1  CU n 1  KU n 1  Fn 1  3
1 1 1 
 2   U n1  2 U n1  U n  U n t      U n
   4
t  2 
Substitute (4) in (1) to get
   1  
2  n 1
U n 1  U n  t 1   U n  U  U n  U n t     U n   5
 t  2  

12
    1  
2  n 1 
U n 1  U n  t 1   U n  U  U  U t     U n   5
 t n n
 2  
Note: (4) and (5) contain the unknown U n 1 on the RHS.
Substitute of (4) and (5) into (3) to get
 1  
 t 2 M  C  K  U n 1  Fn 1  M  a0U n  a2U n  a3U n 
t 
C  a1U n  a4U n  a5U n 
1  1  1  
a0  ; a  ; a  ; a    1  4
; a   1;
t 2 1
t 2
t 3
 2  
t   
a5    2  ; a6  t 1    ; a7  t
2  

13
Newmark's method - I

(1) n  0; Read U 0 & U 0 ;U 0  M 1  F0  CU 0  KU 0 

1
 1  
(2) U n 1   M C  K
 t t 
2

F n 1 
 M  a0U n  a2U n  a3U n   C  a1U n  a4U n  a5U n 
1 1 
2  n 1
U  U n  U n t      U n
1
(3) U n 1 
t  2 
(4) U n 1  U n  t 1   U n   U n 1 
(5) n  n  1; if nt  t f go to (2); otherwise, end.

14
Newmark's method - II

(1) n  0; Read U 0 & U 0 ;U 0  M 1  F0  CU 0  KU 0 

1
 1  
(2) A=  M  C  K  ; B  AM ; D  AC
 t t 
2

(3) U n 1  AFn 1  B  a0U n  a2U n  a3U n   D  a1U n  a4U n  a5U n 

1 1 
2  n 1 
1
(4) U n 1  U  U  U t     U n
t n n
 2 
(5) U n 1  U n  t 1   U n   U n 1 
(6) n  n  1; if nt  t f go to (3); otherwise, end.

15
Stability of Newmark's method
Consider x  2 x   2 x  0
 xn 1  2 xn 1   2 xn 1  0
xn 1  xn  t 1    xn   xn 1 
 xn 1  xn  t 1     2 xn   2 xn     2 xn 1   2 xn 1  

 1  2t  xn 1   2 t  xn 1  1  2t 1     xn  t 1     2 xn

 1  
xn 1  xn  xn t  t     xn   xn 1 
2

 2  
 1  
 xn 1  xn  xn t  t      2 xn   xn     2 xn 1   xn 1  
2 2 2

 2  

16
  2t 2  xn 1  1   2 t 2  xn 1
 21   21 
  t  2t      xn  1   t      xn
2

 2   2 
1  2t  2 t   xn 1 
 2 
 2t 1   t   xn 1 
2 2

 1  2t 1    t 1     2 
   xn 
 t  2t 2  1    1   2 t 2  1      x 
      n 
2  2 
 AX n 1  BX n
 Characteristic equation B   A  0
17
Jury's criteria

P  z   a0 z 2  a1 z  a2
a0  1  2t   2 t 2
a1  2  2t 1  2    t
2 2

a2  1  2t   1  t 2 2  0.5     
P 1   2 t 2
   
       4  2   4t  2  1
2 2 2
1 P 1 4 t

18
Necessary conditions

P 1  0  P 1   t  0  OK 
2 2

 1 P  1  0 is satisfied if 2  1  0 &  4  2   0

2

1 
   & 
2 2
Sufficient conditions: Number of conditions=2-1=1
 1
a2  a0  2t   t      0.
2 2

 2
1
This is satisfied if   .
2 19
Remarks
The method is unconditionally stable if  >0.5 &   0.5
These requirements are independent of system natural
frequency and damping.

Otherwise the following conditions need to be satisfied

4   2 t 2  4  2   4t  2  1  0
2
2t   t   0.5   0  t 
2 2

  0.5
These requirements depend upon system natural
frequency and damping. 20
Average acceleration method  =0.25 &   0.5 
is unconditionally stable
 1 1
Linear acceleration method   = &    is
 6 2
conditionally stable
For  >0.5& >0.5 the method displays high frequency
dissipation characteristics. That is, the spectral radius drops
below value of 1 for larger  (keeping t fixed).
For  =0.5, the high frequency dissipation characteristics
is not witnessed.
If  =0, the above conditions read as
4   2 t 2  4  2   0 &  2 t 2   0.5   0
   0.5 &   0.5
For   0.5, the accuracy of the method drops to O  t  21
 2.8

2.6
  0.5
2.4 alpha=0.25
alpha=0.24
2.2
alpha=0.2
spectral radius

2 alpha=0.3
alpha=0.4
1.8

1.6

1.4

1.2

0.8 -3 -2 -1 0 1 2
10 10 10 10 10 10
wdt
22
 1

0.9

0.8
spectral radius

delta=0.5
0.7 delta=0.6
delta=0.7
delta=0.8
0.6 delta=0.88

0.5
  0.45
0.4 -3 -2 -1 0 1 2
10 10 10 10 10 10
wdt
23
Error estimates and convergence of forward Euler's method

Reference: S D Conte and C de Boor, Elementary numerical

analysis, 2005, Tata McGraw-Hill, New Delhi

Consider
x  t   a  x  t  , t  ; x  t0   x0 ; x  t  is 2 N 1.
Discrete time approximation: let us take tn  nt  t0 
According to Euler's method
yn 1  yn
 a  yn , tn  , n  0,1, 2,3, 1
t
Let x  tn   true solution of x  t   a  x  t  , t  ; x  t0   x0 at t  t n .
Discretization error en  x  tn   yn
Since we expect to specify y0 exactly, e0  0. 24
 t 2
x  tn 1   x  tn  t   x  tn   tx  tn   x  n  with tn   n  tn 1  2 
2
t 2
x  n   Local discretization error
2
=error committed in the single step from tn to tn 1 assuming
that x  t  & x  t  are known exactly at t  tn .
Assumption: neglect round off errors due to finite precision calculations.
yn 1  yn
 a  yn , tn  , n  0,1, 2,3, 1
t
From 1 and 2 we get
t 2
x  tn 1   yn 1   x  tn   yn   t  x  tn   a  yn , tn   x  n 
2
t 2
 
 en 1  en  t a  x  tn  , tn   a  yn , tn  
2
x  n 

25
 t 2

en 1  en  t a  x  tn  , tn   a  yn , tn   2
x  n 

Consider a  x  tn  , tn   a  yn  x  tn   yn , tn 
 a  yn , tn    x  tn   yn  a  xn , tn  for some xn between x  tn  & yn
 a  x  tn  , tn   a  yn , tn   a  xn , tn  en
t 2
 en 1  en  ta  xn , tn  en  x  n  with e0  0 (3)
2
Let a  xn , tn   L & x  n   Y with L, Y  0.
t 2
 3  en1  en  tL en  Y
2
t 2
 en 1  en 1  tL   Y
2
26
 t 2
en 1  en 1  tL   Y  4
2
Consider the difference equation
t 2
 n 1   n 1  tL   Y with  0  0  5 
2
At n  0,  0  0 & e0 =0   n  en for n =0.
Let  n  en be true for some n>0.
t 2
en 1   n 1  tL   Y
2
 en 1   n 1
Thus to obtain a bound on the error, we need to solve
the difference equation  5  .
27
 t 2
 n 1   n 1  tL   Y with  0  0
2
Consider  n 1   n 1  tL  . Let  n  c n  c n 1  c n 1  tL     1  tL 
Try  n  c 1  tL   B
n

n 1
 n

c 1  tL   B  c 1  tL   B 1  tL  
t 2
2
t
Y B Y
2L
t
  n  c 1  tL  
n
Y
2L
t t

 0  0  c  Y   n  Y 1  tL   1
2L 2L
n

x2
Recall: exp  x   1  x  exp    exp  x   1  xx
2
 1  tL  exp  tL   1  tL   exp  ntL 
n

t t
 n  Y exp  ntL   1  Y exp  tn L   1
2L 2L
t
 en  Y exp  tn L   1 28
2L
Remarks

1 The global error in Euler's method is O  t  , that is, the error  0,
as t  0, like ct for some constant c if t  tn is held fixed. It should be
noted that the local error is O  t 2  .

t
 2  en  Y exp  tn L   1 is an upper bound; it may not be realistic.
2L
Its value lies in providing an idea about the nature of the error.

29
Order of accuracy in Newmark's method
Exact solutuion: U  t  ,U  t  ,U  t  ,
Newmark's approximation:U n ,U n ,U n
U n 1  U n  t 1   U n   U n 1 
Consider U  t  t   U  t   tU   ; t    t  t
 U n 1  U n  t 1   U n   U n  tU    ; t    t  t (1)
t 2
We have U  t  t   U  t   tU  t   U   ; t    t  t (2)
2
(2) - (1) 
1 
U  t  t   U n 1  U  t   U n  t U  t   U n   2
t U    
 2
  

1 
en 1  en  ten  t 2U      
2 
30
Similarly
 1  
U n 1  U n  tU n  t     U n  U n 1 
2

 2  
 1  
U n 1  U n  tU n  t     U n   U n  tU   ; t    t  t
2

 2  

t 2 t 3
U  t  t   U  t   tU  t   U t   U    ; t    t  t
2 6
t 2
U  t  t   U n 1  U  t   U n   t U  t   U n   U t   U n 
2
31 
 t     U   ; t    t  t
6 

31
 1 
en 1  en  ten  t 2     U  
2 
t 2 31 
en 1  en  ten  en  t     U   ; t    t  t
2 6 

   0  velocity is O  t 3   local 
1
2 
 Linear acceleration method
   0  displacement is O  t 4   local  
1
6 
Summary
1
If  > , the method displays high frequency dissipation
2
characteristics (desirable) but the global error would be of
O  t  (not desirable).

32
 HHT-α method and generalizations
Motivation
To develop an unconditionally stable, second order
accurate method with acceptable high frequency
dissipation characteristics.

Idea: Introduce a few more algorithmic parameters

H M Hilber, T J R Hughes, R L Taylor, 1977, Improved numerical

dissipation for time integration algorithms in structural dynamics,
Earthquake Engineering and Structural Dynamics, 5, 283-293.

J Chung and G M Hulbert, 1993, A time integration algorithm for

structural dynamics with improved numerical dissipation, Journal of
Applied Mechanics, 60, 371-375. 33
HHT  method and generalizations
The system of equillibrium equations is written as
1   m  MU n1   m MU n  1   f  CU n 1   f CU n  1   f  KU n 1   f KU n
= 1   f  Fn 1   f Fn
The velocity and displacement vectors are represented as in the Newmark method as
U n 1  U n  t 1    U n   U n 1 
U n 1  U n  U n t  t 2  0.5   U n  U n 1 

1   f  K 1    C 1    M  U 

f m n 1
 
 0 1 t  U n 1  
 t 2  U 
 1 0   n 1 

  f K  f C  m M  U n  1   f  Fn 1   f Fn 
     
 0 1 t 1     U n    0 
 1 t t 2  0.5     U n   0 
 
 AX n 1  BX n  f n 34
Implementation of the generalized HHT  method
1 Read algorithmic paramters t,  m ,  f ,  , &
 2  Read initial conditions: U 0 , U 0 & forcing vector; Evaluate U 0  M 1  CU 0  KU 0  F0 
1   f  K 1    C 1    M    f K  f C  m M 

f m
 
 3 Form A   0 1 t &B   0 1 t 1    
 1 0 t 2   1 t t 2  0.5    
 
 4  Evaluate A  A-1B
 5 Set n  0
1   f  Fn 1   f Fn 
 
 6  X n1  AX n  A  0 
 0 
 
 7  n  n  1; if nt  t f , exit; otherwise go to step 6.

35
Stability analysis of the generalized HHT  method
Consider x  2 x   2 x  0

1   m  xn1   m xn  1   f  2 xn1   f 2 xn  1   f   2 xn1   f  2 xn  0

xn 1  xn  t 1    xn   xn 1 
xn 1  xn  xn t  t 2  0.5    xn   xn 1 

1   f   2 2 1   f  1   m    xn1    f  2 2 f   m   xn 
     
 0 1 t   xn 1    0 1 t 1      xn 
 1 0 t 2   xn 1   1 t t 2  0.5      xn 
 
 AX n 1  BX n
Characteristic equation Az  B  0
Do the roots lie within the unit circle in the complex plane?

36
f  z   a0 z 3  a1 z 2  a2 z  a3
a0  1   m  2 1   f  t   1   f   2 t 2
a1  2  3 m  2t 1  2   f  3 f  
a2  1  3 m  2t  1    2 f  3 f     2 t 2  0.5      2 f   3 f 
1 
a3   m  2 f t   1   f  2 t 2      
2 
Necessary conditions
P 1   2 t 2  0
 1 P  1  4  2 m  1  4t  2 f  1  2  1  2 2 t 2  2     2 f  1  0
3

This condition is satisfied if  m  0.5,  f  0.5,   0.5 &   0.5

Sufficient conditions
a3  a0 & b0  b2
Parameter range for unconditional stable scheme
 m   f  0.5;   0.5   m   f ;   0.25  0.5  f   m 
37
 Newmark's algorithm
u  2u   2u  0; u  0   0; u  0   1
  2 rad/s; t =0.1s; =0.5; =0.3
0.2 0.2
Exact Exact
0.15 Newmark 0.15 Newmark

0.1 0.1

0.05 0.05
u(t)

u(t)
0 0

-0.05 -0.05

-0.1 -0.1

-0.15 -0.15

-0.2 -0.2
0 2 4 6 8 10 0 2 4 6 8 10
time t time t

 0   0.05

38
 Newmark's algorithm
u  2u   2u  0; u  0   0; u  0   1
  2 rad/s; t =0.01s; =0.5; =0.3
0.2 0.2
Exact Exact
0.15 Newmark 0.15 Newmark

0.1 0.1

0.05 0.05
u(t)

u(t)
0 0

-0.05 -0.05

-0.1 -0.1

-0.15 -0.15

-0.2 -0.2
0 2 4 6 8 10 0 2 4 6 8 10
time t time t

 0   0.05

39
 HHT- algorithm
u  2u   2u  0; u  0   0; u  0   1
  2 rad/s; t =0.1s; =0.5; =0.25; m  0.25;  f  0.25

0.2 0.2
Exact Exact
0.15 HHT 0.15 HHT

0.1 0.1

0.05 0.05

u(t)
u(t)

0 0

-0.05 -0.05

-0.1 -0.1

-0.15 -0.15

-0.2 -0.2
0 2 4 6 8 10 0 2 4 6 8 10
time t time t

 0   0.05

40
 HHT- algorithm
u  2u   2u  0; u  0   0; u  0   1
  2 rad/s; t =0.01s; =0.5; =0.25; m  0.25; f  0.25
0.2 0.2
Exact Exact
0.15 HHT 0.15 HHT

0.1 0.1

0.05 0.05
u(t)

u(t)
0 0

-0.05 -0.05

-0.1 -0.1

-0.15 -0.15

-0.2 -0.2
0 2 4 6 8 10 0 2 4 6 8 10
time t time t

 0   0.05

41
Example of recent developments
Y K Ping, 2008, A new family of generalized- time integration algorithms
without overshoot for structural dynamics, Earthquake Engineering and
Structural Dynamics, 37, 1389-1409

The system of equillibrium equations is written as

1   m  MU n1   m MU n  1   f  CU n1   f CU n  1   k  KU n1   k KU n
= 1   k  Fn 1   k Fn
The velocity and displacement vectors are represented as
U n 1  U n  t  U n   U n 1 
U n 1  U n  U n t  t 2 U n  U n 1 

 m ,  f ,  k ,  ,  ,  , &  : seven algorithmic parameters
42
Desirable features of numerical integration schemes

At least second order accuracy

Unconditional stability when applied to LTI systems
Controllable algorithmic damping in higher modes
Investigate spectral radius as frequency  
Spectral radius  1 as driving frequency  0
 No overshoot
Excessive oscillations during the first few steps
Self-starting
 No more than one set of implicit equations to be solved at each step

43