You are on page 1of 32

# UCSD

## Nonlinear Finite Element

Simulation:
To do [it] or not to do [it],
that is not the question.

Petr Krysl

http://Hogwarts.ucsd.edu/~pkrysl

P. Krysl
UCSD Outline

## Models, approximation, errors.

Shell modeling w/subdivision
surfaces.
Optimal models.
Conclusions.

P. Krysl
UCSD Modeling
Physical problem

Mathematical Equations,
idealization assumptions

Discretization
Numerical
(finite elements),
approximation algorithms (solvers)

Accurate solution
of mathematical model?

Agreement w/
physics of problem?
P. Krysl
UCSD Predictive?
Experiment

Simulation

## Is this simulation capability predictive?

Not in this case: non-physical material
properties had been used to make it
look right (computer graphics ;).
P. Krysl
UCSD Is the world linear?
Linearization is a (very useful)
approximation of the way real world
processes work.
Nonlinear simulations are just a more
sophisticated way of modeling.
If used to predict how things work, the
prediction will be more accurate.
Accuracy should be thought of in
terms of uncertainties.

P. Krysl
UCSD Are computers creative?
Linear matrix equation: Ax=b
Solution x is not fundamentally new.
Understanding.
Computers transform.
What error is involved in the
transformation?
Choose the right algorithm.
Quantify the error.

P. Krysl
UCSD Modeling
Physical problem
Conceptual
errors
Mathematical Equations,
idealization assumptions
Discretization
errors
Discretization
Numerical
(finite elements),
approximation algorithms (solvers)
Transformation
errors
Accurate solution
of mathematical model?

Agreement w/
physics of problem?
P. Krysl
UCSD Simulation Errors
Errors:
Modeling
Conceptual (omitted physics, range, )
Approximation
Discretization (space, time); and
Transformation (finite-precision
arithmetics!)
Error control:
Quantify (error estimation), and

P. Krysl
UCSD Example: Modeling errors

## Static force: G=gM ~ 0.8Fy

w
Support Time

Wire Computed
Measured
Mass M Force
w In Wire
Fy
table

Time
1. Range of validity (elasticity); and
2. Neglected effect (damping).
P. Krysl
UCSD Quality Assurance

Validation Verification

## Knowledge of the Field,

Math & Computer Science
Engineering Science

## It is impossible to validate a computer code,

only a single simulations may be validated.

P. Krysl
UCSD Selected Topics in Modeling
Thin Shells
Discretization
Model reduction (optimal models)

P. Krysl
UCSD Shell Analysis

## Severe inelastic deformation, contact,

fragmentation, varied spatial and time scales.

P. Krysl
UCSD Subdivision Surfaces
Sequence of control grids which all lead to the
same limit surface.
Courtesy Leif Kobbelt

and so on
Properties:
Compactly supported basis;
C1;
Nested approximations (multiresolution).

P. Krysl
UCSD Subdivision Shell Elements
Subdivision well suited for
Kirchhoff-Love approximation
of thin shells. Cirak, Ortiz, Schrder 1999
1.2

## Normalized max. displacement

Pinched Hemisphere 1
Benchmark 0.8

0.6

0.4

0.2

0
0 500 1000 1500 2000
Number of degrees of freedom

## Present ASM DKT

P. Krysl
UCSD Nonlinear Problems
Cirak, Grinspun, Krysl, Ortiz, Schrder 2000

## Plastic buckling Crumpling

P. Krysl
1. Remeshing
2. Local refinement
FEM: Split finite elements, but ensure
compatibility
Constraints
Lagrangian multipliers or penalty
methods
Irregular splitting of neighboring
elements

P. Krysl
UCSD Conceptual Hierarchy
Infinite sequence of globally
refined spaces
(0) (1)
Mesh M is globally refined to form M
and so on
(0)
M
(1)
M
( 2)
M
( 3)
M
Strict nesting of M (n )

P. Krysl
UCSD 1D Example
One dimensional grid
Subset B * is a particular refined basis

(0)
M
(1)
M
( 2)
M
( 3)
M

## Nodes associated with active basis functions

P. Krysl
UCSD Refinement Implementation

## To refine: replace the nodal

patch to be refined by
other, finer patches;
select appropriately
degrees of freedom.

## 1D linear Subdivision surface

P. Krysl
UCSD A Little Bit of Motivation
Big, ~10,000 degree-of-freedom model, right?

## Wrong, 18 degree-of-freedom model, visually

indistinguishable from the big one.
P. Krysl
UCSD Cost of Simulations
Asymptotic behaviour e CN p
Example: linear statics

log(Error)
Assembly of K: O(N),
solution of Kx=f: O(N2). log(N)
Total cost = aN2 + bN + c
Fixed budget (given grid)
Select modes as linear
combinations of

log(Error)
hat functions optimal

log(#modes)

P. Krysl
UCSD Optimal Representation
Example: Mechanical system w/ 2 dofs

## Optimal coordinates in configuration

space: best-fit linear subspace
P. Krysl

P. Krysl
UCSD RC Bridge under seismic loads
El Centro accelerogram
Isotropic damage model

P. Krysl
UCSD RC Bridge: continued

P. Krysl
UCSD Grid Convergence Studies
Nonlinear analyses: need answers with some indication of accuracy.
Approach: establish convergence by a series of analyses on refined
grids.

## Full FE model: between 20 and 4,320

hexahedral elements.
pulse, J-2 plasticity with combined
hardening.
h=1, 1/2, 1/3, 1/4, 1/5, 1/6

## Contours of equivalent plastic strain

P. Krysl
UCSD
Reduced Model: Lower Cost
Approximate optimal representation on a finer grid
by re-interpolation of modes on a coarse grid.

h=1/5 to h=1/6
Run time

Improved solution on a
finer grid costs only a
fraction of the cost of
solution on the coarser
grid.
P. Krysl
UCSD Monte Carlo Simulation
Notched plate with uncertain material Contour of plastic
properties under low-cycle fatigue load deformation at the end of
Displacement of the
Copper plate 24x20x2 mm.
Finite element model of 1/8 of
the plate: 415 hexahedra. J-2
elastoplastic material.
Elastic and hardening
modulus, yield stress are
uncertain with standard
deviation 0.025.
Evolution of the
equivalent plastic
strain at the notch

Statistics of the
response due to
uncertainty?

P. Krysl
UCSD
Monte Carlo Simulation: cont

## Monte Carlo: Vary the

material properties for each Plastic strain at the notch:
repeated run. Problem: may mean value
have to run thousands of
simulations. Hence: use an
optimal (reduced) model.

## Reduced dynamic model:

3 Ritz modes yield accuracy
of plastic strain better than Plastic strain at the notch:
2%. standard deviation

## Full FE model: 3,250 CPU sec

Reduced FE model: 170 CPU sec

P. Krysl
UCSD
Engineering Design
Customer Technical
Preferences Specifications

Design Synthesis
Refinement

## Design Fabrication Prototype Operation Function/

Description Hardware Behavior

Modeling and
Manufacturing Simulation

Product

P. Krysl
UCSD Design Space Exploration

Point-by-point

Design Performance
Variable Variable Space
Space
Set-based

## Inexpensive tangent approximation

with reduced models.
P. Krysl
UCSD Conclusions
Robust, efficient, and scalable
simulations in constant (and
increasing) demand.
Much more attention needs to be
paid to verification & validation.
More attention should be given to
Integration with design and
manufacturing; and
Synergy with experiments.

P. Krysl