Sei sulla pagina 1di 50

Time Integration:

Fundamentals

Thomas J.R. Hughes


Outline

1) Semi-discrete Methods

Heat equation

Structural dynamics

Nonlinear systems

1) Space-time Methods
Semi-discrete Heat Equation
Generalized Trapezoidal Method
Implementation: Combine (1)-(3):

For n 1,2,3, ... , d0 given,

(M +tK)dn1
(M (1 )tK)dn tFn

=0 Forward Euler
1
= Trapezoidal Rule
2
=1 Backward Euler
Remarks
If 0 and M is diagonal ("lumped"), then no equations
need to be solved to advance the solution: "explicit"

dn1 dn tM 1{Fn Kdn}

If equations need to be solved to advance the solution:


"implicit." For 0, this is the case.

Explicit algorithms economical per step but stability


limits size of time step.

Implicit algorithms more expensive per step, but


more stable, so larger time steps may generally be
taken.
Commutative Diagram
Stability
Numerical Stability
Stability | dn1 | | dn | | A| 1 -1 A 1

1 (1 )t
1 1
1t
Therefore, the stability condition is satisfied if either of the following is
true:

1. 1 / 2, for all t 0, unconditional stability


2.
Significance of Stability Concept
dn And0

n
100 1000
A
0.99 0.37 4.3210-5

1.01 2.70 2.09104

0.90 2.6610-5 1.7510-46

1.1 1.39104 2.471041


The largest eigenvalue of M 1 K determines allowable t
for a system in the conditionally stable case.
t 2 / (1 2 ) , is the largest eigenvalue
Consistency and Convergence
Rearrange algorithm:

dn1 Adn Ln 0
tFn
Ln
(1t )

Arrange the exact solution in a similar fashion:

d(tn1 ) Ad(tn ) Ln t n ,
where n local truncation error

If | | ctk with k 0, the algorithm is consistent.


k is the order of accuracy or rate of convergence.

1
If (Trap.) k 2
2
1
If k1
2

Theorem: Stability + Consistency Convergence

e(tn ) dn d(tn ) 0 as t 0 .

In fact,
e(tn ) O(tk ) .
Semi-Discrete Equations of Motion

Newmark Algorithm

Nathan Newmark (Courtesy of


the University of Illinois archives)
Various forms are useful:
Predictor/multicorrector Newmark algorithm
(H. et al, 1979, H. 2000) Start

Predictor
i 0
d
i
n 1 d n 1
v
i
n 1 v n 1
a
i
n 1 an 1

M* =M t C t2 K:
standard Newmark F ni 1 F n 1 M a in 1 C v in 1 K d in 1
M a F ni 1

M* =M, and M diagonal: Corrector


a i1
a in 1 a i = i+1

explicit for all ,


n1

v in 11 v in 1 t a
d in 11 d in 1 t 2 a

Test no
i 0
F n1 F n1

?
yes
Stop
Examples

1. Average Acceleration Method


1 1
,
4 2
Implicit

Unconditionally Stable

Second-order accurate ( 1/ 2)
Average acceleration method is equivalent to
the trapezoidal rule applied to first-order form
of the equation of motion:
4. Central Difference Method
1
0 ,
2
M,C diagonal explicit

Conditionally stable

Second-order accurate

Man Cvn Kdn Fn


vn (dn1 dn1 ) / (2t)
an (dn1 2dn dn1 ) / t2
Stability for Newmark
1 Always!
2

Unconditional stability: 2 ( 1 )
4
Conditional stability: (2 < )

t crit
If 0 or 1 , crit ( 1 )1/ 2
2 2
Also sufficient for 0, 1
2
O(h1 ), where h mesh-length, t O(h), not so small
(This is the reason explicit, conditionally stable algorithms
are effective in structural dynamics.)
Survey of Structural Dynamics Algorithms

Implicit, unconditionally stable, (usually) second-order accurate,


linear multi-step methods.

Two-step displacement-difference equation algorithms:

Newmark

+ Simple

- High-frequency dissipation requires ( 1 / 2)


implying first-order accuracy
Three-step displacement-difference equation algorithms

Houbolt
+ Very strong high-frequency dissipation
- Asymptotic annihilation
- Poor second-order accuracy

Collocation/Wilson
+ Fair combination of low-frequency accuracy and high-frequency
dissipation
- Overshoot pathology

Hilber-Hughes-Taylor (HHT -method)


+ Good combination of low-frequency accuracy and high-
frequency dissipation

Chung-Hulbert (Generalized -method)


+ Similar to HHT -method
+ Asymptotic annihilation
Generalized -method
S ta rt

P r e d ic to r
i 0
v in 1 v n
( 1)
a in 1 an

( t)2
d in 1 d n tv n
2
(1 2 ) a n 2 a in 1

i = i+ 1

d in dn ( d in 1 dn) C o rre c to r
f i1
a in 1
f
a n 1 a
v i
n vn f (v
i
n 1 vn)
f
v in 11 v in 1 t a
a i
a i
m (a
i
an )
n m n 1 n 1
d in 11 d in 1 t 2 a
R i
n 1 R (d i
n f
,v i
n f
,a i
n m
)

Test no
i 0
R n 1 R n 1

dR i ?
a R i
n 1
d an 1
yes
S to p
Hilber-Hughes-Taylor (HHT -method)

David Hilbert (not Hilber!) Hans-Martin Hilber


Nonlinear Systems
Outline

1. Semi-discrete equations of nonlinear mechanics

2. A simple class of nonlinear problems

3. Newmark algorithms

4. Predictor-corrector algorithms

5. Implicit-explicit finite element algorithms


(mesh partitions)
Semi-Discrete Equations of Nonlinear Mechanics
Ma F F ext F int ()

M Generalized mass matrix


a Generalized acceleration vector
F Generalized force vector

i () is to be thought of as arising from a finite element


spatial discretization of a solid, fluid, structure,
or combined multiphysics system.

i Explicit characterization of M, a, F depends on


the particular system under consideration.
A Simple Class of Nonlinear Problems
F int N(d, v)

- Includes nonlinear elasticity and some nonlinear rate-


type viscoelastic materials

KT N / d Tangent stiffness
CT N / v Tangent damping

Assume:

M is constant
M, KT , CT are symmetric
M, KT are positive-definite
CT is positive-semidefinite
Step 1: Implicit Algorithm
Newmark Algorithms ( 0)

Man1 N(dn1 , vn1 ) Fnext


1
Implementation by Newton-Raphson: Displacement Form

1 Man1 N dn1 , vn1


(2) F Fnext (i ) (i ) (i )

yes, i i 1, go to (2).
(residual or "out-of-balance" force) More iterations?
no, continue

1
K* M CT KT dn1 dn1
(i1)
, vn1 vn1
( i1)
, an1 an1
(i1)
t2 t
("effective stiffness,"
reform and factorize only if required) n n 1 , go to (1).
Step 2: Explicit Algorithm

Predictor-Corrector Algorithm
M diagonal

Same as Newmark, except

predictors
Implementation: same, except

1
1. K* M
t
2

2. One pass through iterative phase


dn1 dn(1)1
vn1 vn(1)1
an1 an(1)1
Step 3: Synthesis
Implicit-Explicit FE Algorithms

Elements are divided into two groups: implicit group and explicit group

Notation:

M I , N I assembled mass and internal force


for implicit group

M E , N E assembled mass and internal force


for explicit group; M E is diagonal
Implementation: same, except
Note
Convergence and Accuracy

1. If N E (d, v) N E (d) , then 1/ 2 O(t2 ) .

2. If v is an argument of N E , then 1/ 2 O(t) .


Explicit Predictor/Multicorrector Algorithms

Man( i11) N dn( i)1 , vn( i)1 Fnext


1

1/ 2, plus one additional iteration gives O(t2 )


(with no adverse effect on stability).

Implementation: same, except

F Fn1
ext
Man1
(i)
N I dn1
(i)
(i)
,vn1 N E dn1
(i)
(i)
,vn1
Note: arguments of N E may be "frozen"
at a previous iterate to save calculations.
Implementation by Newton-Raphson: Displacement Form

(2) F Fn1
ext
Man1
(i)

N dn1
(i) (i)
,vn1
(residual or "out-of-balance" force) yes, i i 1, go to (2).
More iterations?
Only changes compared no, continue
with implicit algorithms
1 dn1 dn1
(i1)
, vn1 vn1
( i1)
, an1 an1
(i1)

K* 2 M CTI KTI
t t
("effective stiffness," n n 1 , go to (1).
reform and factorize only if required)
Space-time Formulations

Example: Initial-Value Problem of Elastodynamics


Space-time Formulations
Discontinuous Galerkin Method in Time
Remarks:
1.

2. Continuity of the solution across time slabs is weakly enforced.

3. A complete mathematical convergence theory exists.

4. The issue of time integrators is eliminated by the choice of


space-time interpolation.

5. Unconditional stability in all cases.

6. A system of linear algebraic equations on each time slab.

7. See Hughes-Hulbert: Vol. 36, pp. 339-363 (1988)


Computer Methods in Applied Mechanics and Engineering.
Features of space-time discontinuous
Galerkin finite element methods

Inter-element discontinuous
basis functions
Weak enforcement of
balance/conservation conditions
in space-time (e.g., Rankine-
Hugoniot conditions in for
conservation laws)
Enables exact conservation per
element and O(N) complexity for
hyperbolic problems
Features of space-time discontinuous
Galerkin finite element methods

Inter-element discontinuous
basis functions
Weak enforcement of
balance/conservation conditions
in space-time (e.g., Rankine-
Hugoniot conditions for
conservation laws)
Enables exact conservation per
element and O(N) complexity for
hyperbolic problems
Causal space-time mesh and O(N)
advancing-front solution strategy

xt

3 3 3 3

1
2
1
2
1
2
1 x E d
Tent Pitcher: Causal space-time meshing
Given a space mesh, Tent Pitcher
constructs a space-time mesh such that
every facet on sequence of advancing
fronts is space-like (patch height
bounded by causality constraint)

Similar to CFL condition, except entirely


local and not related to stability
causality
(required for O(N) solution)
constraint

tentpitching sequence
Patchbypatch meshing and solution

Patches (tents) of tetrahedra; solve immediately for O(N)


method with rich parallel structure
Maintain space mesh as advancing, space-like front with
non-uniform time coordinates

1 2

3 4
Space-time Discontinuous Galerkin Methods
for the Dynamics of Solids
Robert B. Haber
University of Illinois at UrbanaChampaign

Structured Integration Workshop


California Institute of Technology
7-8 May 2009
Center for Process NSF: ITR/AP DMR 01- Materials
Simulation & 21695 ITR/AP DMR 03- Computation
Design 25939 Center
Crack-tip Wave Scattering
Crack-tip Wave Scattering
Crack-tip Wave Scattering
Crack-tip Wave Scattering

Potrebbero piacerti anche