Needs and Trends of Crash Simulations in the Next 10 Years

PDB

Needs and trends of crash simulations in the next 10 years

Paul Du Bois

June 2014

PDB

Overview

• The big question

• Building larger models

• Building more models

• Modeling the process chain

• Conclusions

predicting failure is far more difficult than predicting ductile deformation

J. Jergeus, 2012

PDB

The big questions

• An increase of a factor of 25 in CPU availability could be achieved in

somewhat less then 10 years by Moore’s law

• How far will the capability to run 100M elements rather then 4M elements

get us with respect to predictive potential of crash simulations ?

• How do we best use that capability in the context of automotive

development ?

• How to define industrial reliability ?

A simulation result is reliable if it

drives the design in a direction that

improves the subsequent test result

This must be achieved in 99% of all cases

PDB Copyright © 2013 Altair Engineering, Inc. Proprietary and Confidential. All rights reserved.

6577

3865

1785

1036 909 495 347 355 542

325 267 349

0

1000

2000

3000

4000

5000

6000

7000

128 256 512 1024

Engine Performance (DP) on Curie super computer

E5-2680@2,7GHz

Pure MPI 8 OpenMP 16 OpenMP

RADIOSS : PRACE project 15 Melts model

• Pure MPI scales very well up to 1024 domains

• Test Hybrid with 8 (optimal data locality on Sandy Bridge) and 16 OpenMP (max per node)

• Excellent scalability up to 4096 cores ( 512 * 8 and 256 MPI x 16 OpenMP)

• Maximum performance achieved using 8192 cores (512 x 16)

• First time ever run with 16384 cores ! Need bigger model (less than 1000 solids per core in this case )

#domains

Elapsed (s)

16384 cores 4096 cores 8192 cores

63.24267

6577

PDB

Mercedes-Benz Cars, Dr. Markus Feucht (EP/SPB) / Chrysler TC 14-05-2012 5

Crashsimulation History

W220 (1998)

250.000 E. W168 (1996)

130.000 E.

1990

W124 (1988)

25.000 E.

W211 (2000)

600.000E.

1995 2000 2005

Model size Crash/ Mio. El.

1

W210 (1994)

75.000 E.

W221 (2003)

>1.1 Mio. E.

W251 (2006)

1.8 Mio. E.

(EH-Modell)

W212 (2010)

2.8 Mio. E.

PDB

Mercedes-Benz Cars,

Crash simulation 2014 Example: Frontcrash

(Euro NCAP, 64 km/h)

Show model: 6 Mio elements

Element size: 3…5mm

Turn around time: 10-12 hours

(MPP-Cluster 192 CPU,

FEM-Code LSDYNA)

PDB

7

Crashsimulation History analysis & Prognosis

1

No mesh convergence close to convergence failure

Prony series fit

PDB

State of the art in vehicle component modelling 4PB test

PDB

State of the art in vehicle component modelling

4pb 5mm

4pb 2.5mm

PDB

observations

• Convergence in terms of displacement and force does not necessarily inly

convergence in terms of stress and strain

• Failure models without regularisation cannot work on non-homogeneous

meshes as failure will be biased towards the smaller elements

PDB

11

Crashsimulation History analysis & Prognosis

1

Trendline predicts 70M elements by 2024

PDB

What are our options ?

• Bigger models

• Shells

• solids

• More models

• More load cases

• Stochastic analysis

• Process Chain

• Manufacturing simulations

• Mapping

• Unification

PDB

Bigger models

PDB

Thin shell, thick shell, 3D shell or solid ?

• Assumptions of thin shell theory are fulfilled for curvature radii up to 5t

• Thick shell theory should be invoked for radii<5t

• Although fibers are still straight, thin shell theory will over/under predict

strains in the outer layers due to changes in lamina length

• Reasons to go to solid elements are : T-joints, local necking, through-the-

thickness shear failure ( all induce a 3D state of stress )

12t

Thin shell, inner=outer=middle fiber

PDB

Model size for a car body :

• Estimate of converged meshes for a car body for a small and large car :

• Mesh refinement beyond this point will add limited value with thin shell elements

• Estimate of ‘converged’ meshes using solid elements : multiply by 1000,

corresponding to 10 elements through the thickness

• Studies in ballistics suggest convergence may take 100-1000 elements through the

thickness

• Recently seen the first model > 1 billion elements

Car body 20m**2 2mm sheet 40m**2 1mm sheet

# shell elements 4M 40M

PDB

Speed-up techniques :

• More and more speed-up techniques are developed

• Adaptivity

• Subcycling / multi-scale

• advanced mass scaling

• …

• Usually these techniques work very well for displacement driven problems (

problems where we know the final deformed shape ) but must be very carefully

assessed for bifurcation problems as they tend to favor certain deformation modes

• In brief : bigger crash models will mean more cpu

PDB

More models

PDB

about 30 loadcases are currently investigated

regulators are very inventive


PDB

Crash simulation results with mapped data (thickness, plastic strain)

Crack

Local very high damage (90%)

but still no crack initiation

No damage mapping With mapped pre-damage

PDB

20

Validation of forming simulation Analysis of tension tests from B-pillar

B – Säule

Innen

IN 1 IN 2 IN 3 IN 4

0

50

100

150

200

250

300

350

400

450

500

0,0 5,0 10,0 15,0 20,0 25,0 30,0 35,0 40,0 45,0 50,0

Dehnung e [%]

Sp

an

nu

ng

s [

MP

a]

•Real component shows

softer behavior than

crash-material card (ca.

10%)

•Reduction of local failure

strain is captured well by

pre-damage

Experiment

Simulation with

strain mapping

Simulation

without strain

mapping

Allowed tolerance band (Rm)

PDB

21

Crash simulation results with mapped data (thickness, plastic strain, damage)

• Crack initiation is very sensitive to small changes in

tensile strength (10%)

• Same failure strain in both simulations

(no change GISSMO card)

Standard material card material strength -10%

PDB


Material cards in crash simulation Work practice with tolerances

• Virtual experimental curves must be

created as base for the definition of

„min/max“ cards

• Daily development work in CAE Passive

Safety: Use of „average material cards“ or

individually „min/max“ cards for worst case

scenarios

• Problem: What is worst case?

•=> Robustness investigations on a

stochastic base are neccesary!

„Min“

„Max“

„Mean“

Experiment

Tension test curves

PDB

Modeling the process chain

The failure of a high strength steel part is often preprogrammed in the manufacturing

process

PDB

24

Advanced High Strength Steels Reduced ductility

22MnB5

CP800

TRIP800

ZE340 Aural

TWIP

PDB


Local distribution of mechanical properties Variations in yield stress and failure strain

0

200

400

600

800

1000

1200

1400

1600

1800

0,00 0,05 0,10 0,15 0,20

etechnisch

ste

chnis

ch

[MPa]

MN1-Rz-S1

MN1-Rz-S2

MN1-Rz-S3

MN1-Rz-S4

MN1-Rz-S5

MN1-Rz-S6

MN1-Rz-S7

MN1-Rz-S8

MN1-Rz-S9

MN1-Rz-S10

MN1-Rz-S11

MN1-Rz-S12

Technische Sigma-Epsilon-Kurve

0

100

200

300

400

500

600

700

800

900

0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70

e

ste

ch

nis

ch

[M

Pa]

QH2-1-Fz-S1

QH2-1-Fz-S2

QH2-1-Fz-S3

QH2-1-Fz-S4

QH1-Fz-S1L

QH1-Fz-S1Q

QH1-Fz-S1D

Technische Sigma-Epsilon-Kurve

Press hardened steel 22MnB5 Micro alloyed steel ZStE340

1 1

PDB


Influence of the manufacturing process on material properties

PDB

27

Process chain B-pillar

Plast. strain Thickness damage

Forming simulation

Anisotropic plasticity (Barlat)

*MAT_ADD_EROSION

(GISSMO)

Crash simulation:

J2 plasticity (Mises)

*MAT_ADD_EROSION

(GISSMO)

Mapping

PDB

Does the crash group need it‘s own manufacturing team ?

• Metalforming, hotforming and casting simulations are far advanced , however

do not always provide exactly the type of output needed for the crash model

(e.g. initial damage , microporosity…)

• Special-purpose manufacturing simulations may be needed

PDB

Anisotropic plasticity : thin shell vs 3D shell

Thin shell midplane 3D shell midplane

Increase of damage in the midplane due to 3D state of stress

PDB

Anisotropic plasticity : thin shell vs 3D shell

Thin shell maximum 3D shell maximum

decrease of maximum damage due to 3D state of stress

PDB

Isotropic vs anisotropic plasticity

Isotropic midplane anisotropic midplane

Computed damage values differ by almost 50%

PDB

Isotropic vs anisotropic plasticity

Isotropic maximum anisotropic maximum

Computed damage values differ by 25%

PDB

Material law and/or failure law ?

• Too much emphasis has been put on failure laws and too little on material

laws

• Adding a failure law to J2 plasticity will not always do the job

• Higher order anisotropic plasticity with distortional hardening may be needed

• In this case mapping the material directions from the forming results is critical

• Mapping vectors is hard : a sensitivity study will be needed to learn about the

required accuracy

• A common material law for crash and forming would be ideal : currently no

material law that has the required accuracy and robustness under bifurcation

problems (=crash)

• Need to finetune a generalized metals plasticity law with the needed

efficiency, accuracy, robustness and user-friendliness

PDB

Material Data Set

• An data set was provided for material

DBL4919.10, extruded aluminium

• This data set included global tensile

test measurements for three different

angles

• 0°, 45° and 90°

/Presentation/MAT135OPT

PDB

Material Anisotropy

• At first glance, the selected

material does not look

anisotropic based on the yield

stress

• Failure strain varies, however it

can be attributed to

measurement scatter

• R00 was measured using the

Aramis system to be 0.49

indicating strong anisotropic

flow 0

50

100

150

200

250

300

0.00 0.05 0.10 0.15 0.20 0.25 0.30

e

s [

MPa]

VP3-Fz-S1L

VP3-Fz-S2L

VP3-Fz-S3L

VP3-Fz-S1Q

VP3-Fz-S2Q

VP3-Fz-S3Q

VP3-Fz-S1D

VP3-Fz-S2D

VP3-Fz-S3D

0°

90°

45°

PDB

Yield curves

Extrusion Direction

PDB

Material data for anisotropic plasticity model :

• Reference material shows R values as:

• R00 = 0.48, R45 = 0.29, R90 = 1.76

• “Bumper Beam Longitudinal System Subjected to Offset Impact Loading” Kokkula

(PhD Thesis)

• AA-6060 T1 Aluminum

• Optimized R values for AW-6060 T66 are:

• R00 = 0.49, R45 = 0.27, R90 = 1.69

PDB

Crashworthiness Application

• This model was tested to improve the response/failure prediction of an

extruded tube profile

• Original model was Material 24 in LSDYNA

• Initial simulations provide excellent force vs. deflection results however the

simulation lacks the necessary plastic strain to create element failure

PDB

Profile Bending Simulation

Three point bending test

PDB

Summary and conclusions

Start your own metalforming department

PDB

Summary and conclusions

• Assuming we achieve a 25 fold increase, we will be as short of CPU in 2024

as we are today (The need for cpu goes up with the square of the availability (

T. Belytschko ))

• The need of predictability with respect to failure will force a unification of

methods between manufacturing and crash

• Unlike manufacturing, crash is a bifurcation problem and therefore may have

stochastic aspects, this is particularly true where failure is concerned

• Failure related research currently likely puts too much emphasis on failure

and damage models and too little on material laws

PDB

Thank you very much