Tobias Erhart, 30.

Juni 2005

DYNAmore GmbH
Industriestr. 2, D-70565 Stuttgart
Tel. 07 11 - 45 96 00 - 18 Fax 07 11 - 45 96 00 - 29
e-mail: [email protected]
Internet: www.dynamore.de
 Motivation: Why Implicit?

prestressed, quasi statically loaded structures

long duration analysis > 200 ms

different time scales in process

e.g. static loading followed by transient loading
or transient loading followed by static loading

applications
e.g. metalforming, roof crush, door sag, dummy seating, ...

LS-DYNA provides explicit and implicit solution schemes

one data structure
one input / output

2
 Explicit vs. Implicit (Dynamics)
ρ u,tt = ∇ ⋅ + f
explicit implicit
Ma n = f next − fnint M ∆a n +1 + K ∆u n +1 = f next+1 − f nint − Ma n

solution: directly solution: iteratively

decoupled: efficient, fast linearization necessary
many small time steps few large time/load steps
conditionally stable (Courant) unconditionally stable
equilibrium? equilibrium! convergence?

short time dynamics: structural dynamics:

high frequency response, low frequency response,
wave propagation vibration, oscillation

impact, crash, ... earthquake, machines, ...

3
 Types of Implicit Analyses

Linear Analysis
static or dynamic
single, multi-step

Eigenvalue Analysis
frequencies and mode shapes
linear buckling loads and modes
modal analysis: extraction and superposition

Nonlinear Analysis
Newton, Quasi-Newton, Arclength solution
static or dynamic
default LS-DYNA: static and nonlinear!

4
 Simplified Implicit Flowchart: Terms

0. Initialisierung u 0 , u 0 , u 0
1. Zeitschleife (t n +1 = t n + ∆t )
Bestimme ‘Lasten’ (Dirichletwerte, Lasten, ...) für t n +1
Initialisiere Iterationszähler i = 0
2. Setze Prädiktor-Größen u0n +1 = u n ; u0n +1 = u n ; u 0n +1 = u n
3. Berechne iterationsunabhängige RHS-Anteile nonlinear problem
4. Iterationsschleife
5. Berechne und assembliere die effektive RHS R̂
6. Berechne und assembliere die effektive Steifigkeitsmatrix K̂
7. Löse das Gesamtgleichungssystem ∆u = K ˆ −1R ˆ linear problem
8. Aktualisiere die Verschiebungen, Geschwindigkeiten
und Beschleunigungen
9. Konvergenzcheck: if (Residuum>TOL) goto 4. else goto 10.
10. n = n + 1 goto 1.

5
 Activating Implicit Analysis

Use *CONTROL_IMPLICIT_GENERAL to activate implicit

specify time step size
all other *CONTROL_IMPLICIT keywords are optional
default is nonlinear, static analysis

Use a double precision executable for implicit analysis

better convergence for nonlinear
mandatory for linear, eigenvalue accuracy
ls-dyna i=input.k memory=200m
Stiffness Matrix requires lots of memory 200,000,000 words:
huge speed penalty for out-of-core jobs 800 Mbytes in single precision
1600 Mbytes in double precision

Most keywords apply to explicit and implicit

*NODE, *ELEMENT, *SECTION, *MAT, ...
easy to run a model with either method, but: carefully inspect input deck

6
 Activating Implicit Analysis

Three types of analyses can be performed

fully explicit (default)
fully implicit
switching: explicit - implicit, implicit - explicit (prescribed or automatic)

All keywords for implicit

*CONTROL_IMPLICIT_GENERAL *CONTROL_IMPLICIT_SOLVER
*CONTROL_IMPLICIT_SOLUTION *CONTROL_IMPLICIT_AUTO
*CONTROL_IMPLICIT_STABILIZATION *CONTROL_IMPLICIT_DYNAMICS
*CONTROL_IMPLICIT_MODES *CONTROL_IMPLICIT_EIGENVALUE
*CONTROL_IMPLICIT_BUCKLE

Proper selection of LS-DYNA features

not all features are available in implicit mode
warning & error messages, feature substitution

7
 Implicit Keywords

*CONTROL_IMPLICIT_GENERAL (required for implicit)

activates implicit mode, explicit-implicit switching
defines implicit time step size (standard LS-DYNA termination time is used too)

*CONTROL_IMPLICIT_SOLVER (optional)
parameters for linear equation solver, which inverts stiffness matrix: [K]{x}={f}

*CONTROL_IMPLICIT_SOLUTION (optional)
parameters for nonlinear equation solver (Newton-based methods)
controls iterative equilibrium search, convergence
"linear" analysis selected here (a special case where no iterations are
performed)

*CONTROL_IMPLICIT_AUTO (optional)
activates automatic time step control
default is fixed time step size, error termination if any steps fail to converge

8
 Implicit Keywords

*CONTROL_IMPLICIT_DYNAMICS (optional)
M ∆a n +1 + K ∆u n +1 = f next+1 − f nint − Ma n
include inertia terms
problem “time” must now be real, physical time
can improve convergence, especially when rigid body modes are present

*CONTROL_IMPLICIT_EIGENVALUE (optional)
signals LS-DYNA to perform eigenvalue analysis, then stop
number of eigenvalues/vectors, optional frequency shift
great for debugging/model checking

*CONTROL_IMPLICIT_STABILIZATION (optional)
Allows multi-step springback

User‘s manual contains helpful notes on each input parameter,

Appendix M gives survey of LS-DYNA/implicit features

9
 Linear Static Analysis Ku = R
Activate the implicit method
*CONTROL_IMPLICIT_GENERAL: imflag = 1

Select a stepsize and termination time

for static analysis, choice of time is arbitrary
*CONTROL_IMPLICIT_GENERAL: dt0 = 1.0
nstep = term/dt0 = 1
*CONTROL_TERMINATION: term = 1.0

Select linear solution method (no equilibrium iterations)

*CONTROL_IMPLICIT_SOLUTION: nsolvr = 1

Select a linear element type

shell # 18, 20, 21
brick # 18

Use a double precision LS-DYNA executable

10
 Element Formulations for Linear Analyses

Linear and nonlinear element formulations are different

linear: integrate stress over undeformed geometry
infinitesimal deformation eliminates some locking problems
enhanced strain fields accurately represent linear elasticity

Brick Elements
f = BT dΩ
type 18: linear solid
Ω0

Shell Elements
type 18: linear thin shell (Kirchoff)
type 20: linear thick shell (Mindlin)
type 21: linear enhanced shell (CQUAD4)

11
 Boundary Constraints

Boundary conditions and rigid body modes

static implicit simulation requires boundary constraints
rigid body modes must be eliminated
(otherwise stiffness matrix is singular / not invertible)
apply translational constraints to three nodes

a - reference node, dx=dy=dz=0

eliminates all translational modes
b - node along X-axis, dy=dz=0
eliminates rotations about y- and z-axis Z c
c - node along Y-axis, dz=0 Y
eliminates rotation about x-axis a
X

12
 Example: Linear Static Analysis
2000 Chrysler PT Cruiser
Body Assembly

courtesy of Daimler-Chrysler
Auburn Hills, Michigan, USA

13
 Example: Linear Static Analysis
Static Linear Torsion Analysis model details
NSOLVR=1 240,000 nodes
shell type 21 3,300 type 100 spotwelds
1,450,000 equations
double precision memory=740m (~6 Gbyte)
SMP parallel performance
1 cpu: 753 sec
4 cpu: 485 sec

Attention: larger models demand 64-bit O/S to exceed 2 Gbyte memory limit

14
 Linear Equation Solver
ˆ ∆u = R
During each nonlinear iteration, the linear system K ˆ is solved.
Direct Methods LS-DYNA default
gaussian elimination
inexpensive backsolve (quasi-Newton)
The sparse direct solver
costly: CPU and memory F
robust and reliable

Iterative Methods good performance for

iteration: improve approximate solution solid, massive structures
potentially low operation count
convergence difficult for some problems
promising future developments

Keyword *CONTROL_IMPLICIT_SOLVER

15
 Direct Solvers

A sparse, direct linear equation solver is used by default (LSOLVR=4)

serial or SMP parallel execution
automatic out-of-core mode if insufficient memory available for incore
double precision version also available (LSOLVR=5)
• improved convergence for a few models
• 2x memory penalty
• better to use a double precision version of LS-DYNA
BCSLIB-EXT solver from Boeing also available (LSOLVR=6)
• double precision only
• best for very large models (excellent out-of-core performance)

All sparse direct solvers execute in three phases

symbolic factorization
numeric factorization
forward elimination / back substitution

16
 Iterative Solvers

LS-DYNA offers six iterative linear equation solvers

LSOLVR = 10: “best” iterative solver (currently activates #16)
LSOLVR = 11: Conjugate Gradient method
LSOLVR = 12: CG with Jacobi preconditioning
LSOLVR = 13: CG with Incomplete Choleski preconditioning
LSOLVR = 14: Lanczos method
LSOLVR = 15: Lanczos with Jacobi preconditioning
LSOLVR = 16: Lanczos with Incomplete Choleski preconditioning

All iterative solvers use the sparse matrix storage scheme

eliminates all zero entries inside bandwidth
minimizes total storage requirement
Boeing Harwell format for portability

17
 Dynamic Implicit Analysis M ∆a n +1 + K ∆u n +1 = f next+1 − f nint − Ma n

Newmark Method relates displacement, velocity, acceleration

u n +1 = u n + v n ∆t + [( 12 − β )a n + β an +1 ] ∆t 2

v n +1 = v n + [(1 − γ ) a n + γ a n +1 ] ∆t

β = 0, γ = 1 / 2 : explicit central difference method

β = 1 / 4 , γ = 1 / 2 : implicit undamped trapezoidal rule

γ > 1/ 2 : numerical damping LS-DYNA default

convergence may be possible with large DT

small DT may be needed to resolve high frequency response
ˆ =K+ 1
stabilizing effect to nonlinear equilibrium iterations K M
β ∆t 2
rigid body modes OK! (mass terms eliminate singularity)

18
 Activating Dynamic Implicit Analysis

Activating Dynamic Analysis

Implicit Dynamic Analysis may be linear or nonlinear

inertia terms are simply added to stiffness matrix and residual vector
very efficient when only one stiffness matrix factorization is performed
• earthquake response analysis: long periods of nearly linear behavior
• same stiffness matrix used for many nonlinear steps
if time step size changes, a new stiffness will automatically be formed

ˆ =K+ 1
K M
β ∆t 2

19
 Example: Sheet Metal Gravity Loading

20
 Example: Car gravity loading

21
 Eigenvalue Analysis

Compute Natural Frequencies and Mode Shapes

linear analysis
infinitesimal deformation (K − λ M ) =0
(magnified for display)

Accuracy Requires Special Considerations

linear elements (type 18, 20)
double precision executable

Applications
frequency analysis
model integrity check
extract modes for modal analysis

22
 Activating Eigenvalue Analysis
Required Input Parameters
non-zero termination time
*CONTROL_TERMINATION term=1.0

implicit analysis
*CONTROL_IMPLICIT_GENERAL isolvr=1, dt0=1.0

number of eigenvalues
*CONTROL_IMPLICIT_EIGENVALUE neigv=30

Eigenvalue analysis with an existing implicit input deck

just add one keyword, one input parameter:
*CONTROL_IMPLICIT_EIGENVALUE
$ neigv center
30 0.0

LS-DYNA computes 30 lowest modes, terminates

23
 Eigenvalue Input / Output
Input Options
number of eigenvalues/modes
center frequency, frequency range
eigenvalue extraction method: Lanczos eigensolver (default)

New output databases

d3eigv: - binary plot database similar to d3plot
- each state shows one mode shape
- “State times” give circular frequencies f
eigout: - ASCII text file
- summary of frequencies found
λ ω= λ f = ω / 2π T = 1/ f

MODE EIGENVALUE RADIANS CYCLES PERIOD

1 1.619382E+02 1.272549E+01 2.025325E+00 4.937478E-01
2 6.292547E+03 7.932558E+01 1.262506E+01 7.920756E-02
3 1.922690E+04 1.386611E+02 2.206860E+01 4.531325E-02

24
 Mode Extraction *CONTROL_IMPLICIT_MODES

Modal Analysis
approximate structural deformation using a set of modes
modal amplitudes become the unknowns
greatly reduced problem size
superposition principle assumes linearity

Flexible Rigid Bodies

large rigid body motion + superimposed modal deformation
apply to a subset of parts, treat others as fully nonlinear

Analysis Procedure
1. compute modes for subset of parts (d3eigv, d3mode in v970)
2. rigidize, merge these parts
3. define *PART_MODES for master part
4. perform explicit transient dynamic analysis

25
 Nonlinear Implicit Analysis
σ
Material Nonlinearity
plasticity, damage, failure ε
rate dependence
slope of stress-strain curve gives stiffness,
should usually be monotonic ε

Geometric Nonlinearity
large displacement, large rotation

Contact Nonlinearity
normal force is sharply discontinuous
frictional effects elastic-perfectly-plastic

26
 Nonlinear Implicit Analysis
Implicit governing equations contain two problems to solve

M ∆a n +1 + K ∆u n +1 = f next
+1 − f int
n − Ma n

Nonlinear Equilibrium Problem: *CONTROL_IMPLICIT_SOLUTION

find displacements u which satisfy equilibrium fext=fint
both K, fext and fint can be nonlinear functions of u
iterative search employed using Newton-based method
interactive switch "<ctrl-c> nlprint" toggles diagnostic output

Linear Algebra Problem: *CONTROL_IMPLICIT_SOLVER

solve system of linear algebraic equations
must solve during every nonlinear iteration
great CPU and memory cost
interactive switch "<ctrl-c> lprint" toggles diagnostic output

27
 Nonlinear Equation Solver - Newton Method

ˆ (u ( i ) ) ∆u = R
K ˆ ( i ) = f ext − f int (u ( i ) )
Force Norm

K̂(u )
int
1
R̂ 2
f
ext
f n+1
1
R̂

K̂(un )
“Full Newton”
.
(LS-DYNA offers also Modified Newton
ext
fn and Quasi-Newton=“BFGS”)

Displacement Norm
1 2 3
un u(n+1) u(n+1) u(n+1) u(n+1)

Equilibrium is reached when iterations converge: ∆ ˆ (i ) → 0

u →0, R

28
 Input Parameters for Nonlinear Solver
NSOLVR: nonlinear solution method
=1: linear approximation (no equilibrium iterations)
=2: BFGS quasi-Newton method (DEFAULT)

ILIMIT: equilibrium iteration limit before re-evaluating K̂

=1: new K ˆ each iteration (“Full Newton" method)
=11: use cheap BFGS update for 11 iterations, reform if not yet converged

MAXREF: maximum reformation count before abandoning step

.

if AUTO is active, dt will be reduced and step will be re-tried,

so MAXREF can be smaller (~5)
if AUTO not active, error termination occurs when MAXREF is reached
so MAXREF should be larger (~15, default)

DCTOL, ECTOL: convergence tolerances

use NLPRINT=1 or "<ctrl-c> nlprint" to monitor progress of iterations

29
 Example: Vickers Hardness Test

F [kN]

t [s]
10 14 24

implicit
1 cpu: 4 h
Viert
e lsyst memory=140m (1.1 GB)
em
explicit
1 cpu: 270000 h (2.4E10 cycles)

30
 Automatic Time Step Control

Automatic time step control adjusts stepsize during the simulation

very persistent, reliable
*CONTROL_IMPLICIT_AUTO
After successful steps
compare iteration count to target value ITEOPT
increase/decrease size of next step if difference exceeds window
ITEWIN

After failed steps

.

decrease step size

back up, repeat failed step with new DT

Exponential algorithm for adjusting step size

increase stepsize by 1/5 decade until DTMAX is reached
decrease stepsize by 1/3 decade until DTMIN is reached
error termination if convergence fails when DT=DTMIN

31
 Implicit Capabilities: Element Types

Brick Elements: 1, 2, 3, 4, 10, 15, 16, 18

Beam (and 2D Shell) Elements: 1, 2, 3, 4, 5, 6, 7, 8, 9
Shell (and 2D Solid) Elements: 2, 4, 6, 10, 12, 13, 15, 16, 17, 18, 20, 21

recommended elements for implicit, e.g.

s
.

t
r

Hughes-Liu: fully integrated: fully integrated S/R:

type 1 (default) type 16 (not default!) type 2 (not default!)

Alternate shell element formulations are substituted if requested

elements are not available for implicit

32
 Implicit Capabilities: Material Models

stiffness matrix terms require extra evaluation of δσ

σ/δεε
some material models only available for selected element formulations

3D Solid Elements
• 1-7, 11, 12, 13, 14-17, 18, 19, 20, 21-23, 24, 26, 27, 30, 31, 33, 35, 36, 38,
41-50, 51-53, 57, 59, 60-62, 63, 64, 65, 70, 72, 73, 75-80, 83-85, 87-89,
91, 92, 96, 98, 100, 102, 103, 104, 105, 106, 107, 109-112, 115, 124, 126-
129, 141-145, 161, 177, 178, 192, 193

Shell Elements
.

• 1-4, 6, 9, 18, 20, 21-23, 24, 27, 32, 36, 37, 41-50, 54, 55, 60, 76, 77, 91,
92, 98, 99, 103, 104, 106, 107, 116-118, 123

Beam Elements
• 1, 3, 4, 6, 9, 18, 20, 24, 41-50, 100

2D Solid Elements
• 1-7, 9, 12, 13, 18, 20, 24, 26, 41-50, 57, 63, 103, 104, 106, 107

33
 Implicit Capabilities: Contact Interfaces
Several contact interfaces are available for implicit analysis
*CONTACT_SURFACE_TO_SURFACE
*CONTACT_NODES_TO_SURFACE
*CONTACT_ONE_WAY_SURFACE_TO_SURFACE
*CONTACT_FORMING_… (three variations)
*CONTACT_AUTOMATIC_… (three variations)
*CONTACT_TIED, …TIED_OFFSET
.

*CONTACT_AUTOMATIC_SINGLE_SURFACE
All implicit contact interfaces use the penalty method, except TIED
Shooting node logic is automatically disabled for implicit
SNLOG=1 on optional contact interface card "B"
Oriented normal vectors are recommended
Automatic contact types can fail due to large implicit stepsize

34
 Nonlinear Convergence Problems
Convergence trouble is the most common problem
Error messages displayed by LS-DYNA, e.g.
iteration limit reached
displacement and energy tolerances were not satisfied, abandon step
divergence
out-of-balance force R is growing, reform K and continue iterations
negative eigenvalues
error from linear equation solver while computing K-1
.

Procedures for solving convergence problems

determine reason for termination (examine error messages)
activate print flags to get more information
view deformed geometry during iteration process using "d3iter" database
carefully inspect input deck
see user’s manual: Appendix M

35
 Current and Future Developments
MPP Implicit is nearing completion with all capabilities implemented
Implicit simulation (statics and dynamics)
Springback
Vibration and Buckling analysis Version 971
Constraint and Attachment modes

Transition from dynamic to static

e.g. gravity loading, roof crush
.

Implementation of still missing features, e.g.

e.g. consistent tangent stiffness for more materials
elements, e.g. type 13 tetrahedron for bulk forming
contact types
seatbelts
airbags: fabric materials, inflator models, soft=2 - contact

36