ITR/AP: Multiscale Models for Microstructure Simulation and Process - - PowerPoint PPT Presentation

ITR/AP: Multiscale Models for Microstructure Simulation and Process Design

Principal Invest igat ors: Bob Haber (Theor. & Applied Mechs.), Jonat han Dant zig (Mech. & Ind. Engng.), Duane Johnson (Mat l. S cience & Engng.). Universit y of Illinois at Urbana– Champaign Principal Invest igat ors: Principal Invest igat ors: Bob Haber ( Bob Haber (Theor Theor . & Applied . & Applied Mechs Mechs.), .), Jonat han Jonat han Dant zig Dant zig ( (Mech

• Mech. &

. & Ind Ind. . Engng Engng.), .), Duane Johnson ( Duane Johnson (Mat l Mat l . S cience & . S cience & Engng Engng.). .). Universit y of Illinois at Universit y of Illinois at Urbana Urbana– –Champaign Champaign

Faculty Investigators

Cont inuum science

• Jonathan Dantzig (Mech. & Ind. Engrg.)
• Eliot Fried (Theor. & Appl. Mechs.)
• Robert Haber (Theor. & Appl. Mechs.)
• Daniel Tortorelli (Mech. & Ind. Engrg.)

Mat erials (at omist ic) science

• Duane Johnson (Matl. S
• ci. & Engnrg.)

Cont inuum science Cont inuum science

• Jonathan

Jonathan Dantzig Dantzig ( (Mech

• Mech. &

. & Ind Ind. . Engrg Engrg.) .)

• Eliot Fried (

Eliot Fried (Theor

• Theor. &

. & Appl Appl. . Mechs Mechs.) .)

• Robert Haber (

Robert Haber (Theor

• Theor. &

. & Appl Appl. . Mechs Mechs.) .)

• Daniel

Daniel Tortorelli Tortorelli ( (Mech

• Mech. &

. & Ind Ind. . Engrg Engrg.) .)

Mat erials ( Mat erials (at omist ic at omist ic) science ) science

• Duane Johnson (

Duane Johnson (Matl Matl. . S ci S

• ci. &

. & Engnrg Engnrg.) .)

Faculty Investigators

Informat ion science

• Jeff Erickson (Computer S

ci.)

• Michael Garland (Computer S

ci.)

• S

anj ay Kale (Computer S ci.)

• Herbert Edelsbrunner (Computer S

ci., Duke)

Mat hemat ics

• Robert Jerrard (Mathematics) - pde’ s
• John S

ullivan (Mathematics) - geometry

• Martin Bendsøe (Mathematics, Danish Tech. U.)
• topology opt.

Informat ion science Informat ion science

• Jeff Erickson (Computer

Jeff Erickson (Computer S ci S ci.) .)

• Michael Garland (Computer

Michael Garland (Computer S ci S ci.) .)

• S

anj ay Kale (Computer S anj ay Kale (Computer S ci S ci.) .)

• Herbert

Herbert Edelsbrunner Edelsbrunner (Computer (Computer S ci S ci., ., Duke Duke) )

Mat hemat ics Mat hemat ics

• Robert

Robert Jerrard Jerrard (Mathematics) (Mathematics) -

• pde’ s

pde’ s

• John S

ullivan (Mathematics) John S ullivan (Mathematics) -

• geometry

geometry

• Martin

Martin Bendsøe Bendsøe (Mathematics, (Mathematics, Danish Tech. U. Danish Tech. U. ) )

• topology opt.

topology opt.

A joint effort betw een tw o centers

Mat erials Comput at ion Cent er

• Atomistic models
• Prediction of bulk properties

Cent er for Process S imulat ion & Design

• Manufacturing processes
• Continuum models
• S

imulation and optimization of microstructure properties in manufacturing processes

• S

uccessful experience with interdisciplinary collaborations

Mat erials Comput at ion Cent er Mat erials Comput at ion Cent er

• Atomistic

Atomistic models models

• Prediction of bulk properties

Prediction of bulk properties

Cent er for Process S imulat ion & Design Cent er for Process S imulat ion & Design

• Manufacturing processes

Manufacturing processes

• Continuum models

Continuum models

• S

imulation and optimization of microstructure S imulation and optimization of microstructure properties in manufacturing processes properties in manufacturing processes

• S

uccessful experience with interdisciplinary S uccessful experience with interdisciplinary collaborations collaborations

CPSD Funding History

Alcoa (1996 - 2000)

• $20k/ yr seed grant

NS F GOALIE grant wit h Alcoa (1997-2001)

• $120k / year NS

F; $20k / year Alcoa

NS F-DARPA OPAAL grant (1998-2001)

• Math directorates
• ~$800,000 / year over 3 years

NS F ITR grant (2001-2006)

• Division of Materials Research,
• Computer and Informat ion S

cience Engineering

• ~$800,000 / year over 5 years

Alcoa (1996 Alcoa (1996 -

• 2000)

2000)

• $20k/

$20k/ yr yr seed grant seed grant

NS F GOALIE grant wit h Alcoa (1997 NS F GOALIE grant wit h Alcoa (1997-

• 2001)

2001)

• $120k / year NS

F; $20k / year Alcoa $120k / year NS F; $20k / year Alcoa

NS F NS F-

• DARPA OPAAL grant (1998

DARPA OPAAL grant (1998-

• 2001)

2001)

• Math directorates

Math directorates

• ~$800,000 / year over 3 years

~$800,000 / year over 3 years

NS F ITR grant (2001 NS F ITR grant (2001-

• 2006)

2006)

• Division of Materials Research,

Division of Materials Research,

• Computer and Informat ion S

cience Engineering Computer and Informat ion S cience Engineering

• ~$800,000 / year over 5 years

~$800,000 / year over 5 years

CPSD/MCC Mission I: Manufacturing Science

Improve product qualit y t hrough cont rol of microst ruct ure S imulat ion t ools t o predict microst ruct ure evolut ion during processing

• Basic science (atomic to micro scale studies)
• Applied science (micro - macro scale process simulations)

Opt imizat ion t ools f or process design

• Use multi-scale process simulations
• S

ensitivity analysis, optimization of process parameters

– Tool shapes, process rat es, alloy chemist ry, quench, ...

Improve product qualit y t hrough cont rol of Improve product qualit y t hrough cont rol of microst ruct ure microst ruct ure S imulat ion t ools t o predict microst ruct ure S imulat ion t ools t o predict microst ruct ure evolut ion during processing evolut ion during processing

• Basic science (atomic to micro scale studies)

Basic science (atomic to micro scale studies)

• Applied science (micro

Applied science (micro -

• macro scale process simulations)

macro scale process simulations)

Opt imizat ion t ools f or process design Opt imizat ion t ools f or process design

• Use multi

Use multi-

• scale process simulations

scale process simulations

• S

ensitivity analysis, optimization of process parameters S ensitivity analysis, optimization of process parameters

– – Tool shapes, process rat es, alloy chemist ry, quench, ... Tool shapes, process rat es, alloy chemist ry, quench, ...

CPSD/MCC Mission II: Computational Methods

Develop new comput at ional t echniques t o support manufact uring science mission Common requirement s and responses

• Multi-scale physics + optimization = large scale problems

– Parallel comput at ion, adapt ive analysis, mult igrid

• Difficult geometry:

– complex shapes, moving boundaries, variable connect ivit y, – Meshing, phase-field, ALE, spacet ime met hods, “ skin”

• Embedded physical models

– Direct : discont inuous Galerkin, quant um-cont inuum – Linked hierarchical models: homogenizat ion, et c.

Develop new comput at ional t echniques t o support Develop new comput at ional t echniques t o support manufact uring science mission manufact uring science mission Common requirement s and responses Common requirement s and responses

• Multi

Multi-

• scale physics + optimization = large scale problems

scale physics + optimization = large scale problems

– – Parallel comput at ion, adapt ive analysis, Parallel comput at ion, adapt ive analysis, mult igrid mult igrid

• Difficult geometry:

Difficult geometry:

– – complex shapes, moving boundaries, variable connect ivit y, complex shapes, moving boundaries, variable connect ivit y, – – Meshing, phase Meshing, phase-

• field, ALE,

field, ALE, spacet ime spacet ime met hods, “ skin” met hods, “ skin”

• Embedded physical models

Embedded physical models

– – Direct : discont inuous Direct : discont inuous Galerkin Galerkin, quant um , quant um-

• cont inuum

cont inuum – – Linked hierarchical models: homogenizat ion, et c. Linked hierarchical models: homogenizat ion, et c.

Dendritic Solidification

• Jonathan Dantzig, faculty lead
• Controls grain size and morphology in casting
• Jonathan

Jonathan Dantzig Dantzig, faculty lead , faculty lead

• Controls grain size and morphology in casting

Controls grain size and morphology in casting

Scaling with undercooling, grain size Anisotropy due to convective flow

Dantzig Dantzig (M&IE), (M&IE), Goldenfeld Goldenfeld (Physics), Kale (CS ) (Physics), Kale (CS )

Modeling Dendritic Grow th

Microst ruct ure evolut ion wit h f low

• Length scales: nm – mm
• Phase-field method for microstructure
• Parallel, adaptive, Navier-S

tokes solver

Microst ruct ure evolut ion wit h f low Microst ruct ure evolut ion wit h f low

• Length scales:

Length scales: nm nm – – mm mm

• Phase

Phase-

• field method for microstructure

field method for microstructure

• Parallel, adaptive,

Parallel, adaptive, Navier Navier-

• S

tokes solver S tokes solver

Dantzig Dantzig (M&IE), (M&IE), Goldenfeld Goldenfeld (Physics), Kale (CS ) (Physics), Kale (CS )

QuickTime™ and a GIF decompressor are needed to see this picture.

Modeling Dendritic Grow th

Binary Alloy S

• lidificat ion
• Important industrial applications
• Directional solidification (2D and 3D)
• S

pacing selection of interest

• Flow interactions with

complex structures

Binary Alloy S

• lidificat ion

Binary Alloy S

• lidificat ion
• Important industrial applications

Important industrial applications

• Directional solidification (2D and 3D)

Directional solidification (2D and 3D)

• S

pacing selection of interest S pacing selection of interest

• Flow interactions with

Flow interactions with complex structures complex structures

Dantzig Dantzig (M&IE), (M&IE), Goldenfeld Goldenfeld (Physics), Kale (CS ) (Physics), Kale (CS )

QuickTime™ and a GIF decompressor are needed to see this picture.

Parallelization Infrastructure:

Laxmikant Kale, facult y lead 2-prong approach t o user-friendly parallelizat ion

• Parallel Obj ects
• Component Frameworks

S everal successful, diverse applicat ions Laxmikant Laxmikant Kale, facult y lead Kale, facult y lead 2 2-

• prong approach t o user

prong approach t o user -

• friendly

friendly parallelizat ion parallelizat ion

• Parallel Obj ects

Parallel Obj ects

• Component Frameworks

Component Frameworks

S everal successful, diverse applicat ions S everal successful, diverse applicat ions

L.V. Kale, O. L.V. Kale, O. Lawlor Lawlor, G. , G. Kakulapathi Kakulapathi, A. , A. S ingla S ingla, J. Booth , J. Booth

Charm Component Framew orks

Automatic Load balancing

• Auto. Checkpointing

Flexible use of clusters Out-of-core execution

Object based decomposition

Reusable Specialized Parallel Strucutres

Component Frameworks

FEM / Unstructured Grid

• Collision Detection
• NetFEM visualizer

Multi-block TaskGraph: supporting space-time meshes

Charm++

L.V. Kale, O. L.V. Kale, O. Lawlor Lawlor, G. , G. Kakulapathi Kakulapathi, A. , A. S ingla S ingla, J. Booth , J. Booth

Object-based Parallelization

User View System implementation

User is only concerned with interaction between objects

L.V. Kale, O. L.V. Kale, O. Lawlor Lawlor, G. , G. Kakulapathi Kakulapathi, A. , A. S ingla S ingla, J. Booth , J. Booth

FEM Framew ork

Charm++

(Dynamic Load Balancing, Communication)

FEM Framework

(Update of Nodal properties, Reductions over nodes or partitions)

FEM Application

(Initialize, Registration of Nodal Attributes, Loops Over Elements, Finalize)

METIS I/O Partitioner Combiner Collaborators: Jon Dantzig and Coworkers, R. Haber and coworkers, Dan Torterelli and coworkers

Not just FEM:

• Any Unstructured-Grid

app Also being extended for:

• DG method (Haber)
• Implicit solvers

(Torterelli)

• Finite Volume (CSAR)

L.V. Kale, O. L.V. Kale, O. Lawlor Lawlor, G. , G. Kakulapathi Kakulapathi, A. , A. S ingla S ingla, J. Booth , J. Booth

Dendritic Grow th

S t udies evolut ion of solidificat ion microst ruct ures using a phase-field model comput ed on an adapt ive finit e element grid Adapt ive refinement and coarsening of grid involves re-part it ioning S t udies evolut ion of S t udies evolut ion of solidificat ion solidificat ion microst ruct ures using a microst ruct ures using a phase phase-

• field model

field model comput ed on an adapt ive comput ed on an adapt ive finit e element grid finit e element grid Adapt ive refinement and Adapt ive refinement and coarsening of grid coarsening of grid involves re involves re-

• part it ioning

part it ioning

L.V. Kale, O. L.V. Kale, O. Lawlor Lawlor, G. , G. Kakulapathi Kakulapathi, A. , A. S ingla S ingla, J. Booth , J. Booth

Load balancer in action

5 10 15 20 25 30 35 40 45 50 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 Iteration Number Number of Iterations Per second

Automatic Load Balancing in FEM

• 1. Adaptive

Refinement

• 3. Chunks

Migrated

• 2. Load

Balancer Invoked

Restoring Throughput

L.V. Kale, O. L.V. Kale, O. Lawlor Lawlor, G. , G. Kakulapathi Kakulapathi, A. , A. S ingla S ingla, J. Booth , J. Booth

Spacetime discontinuous Galerkin finite element methods

• Bob Haber, faculty lead
• New finite element methods for hyperbolic pde’ s

– S pacet ime formulat ions – Eliminat es nearly all of t he vexing problems of shocks, CFD, et c. ~ w/ o special procedures – Exact balance/ conservat ion at t he element level – O(N) complexit y.

• Applications

– Dynamic fract ure – Cont inuum bulk models of microst ruct ure evolut ion – At omist ic-cont inuum coupling st rat egies

• Bob Haber, faculty lead

Bob Haber, faculty lead

• New finite element methods for hyperbolic

New finite element methods for hyperbolic pde’ s pde’ s

– – S pacet ime S pacet ime formulat ions formulat ions – Eliminat es nearly all of t he vexing problems of shocks, CFD, et c. ~ w/ o special procedures – Exact balance/ conservat ion at t he element level – O(N) complexit y.

• Applications

Applications

– – Dynamic fract ure Dynamic fract ure – – Cont inuum bulk models of microst ruct ure evolut ion Cont inuum bulk models of microst ruct ure evolut ion – – At omist ic At omist ic-

• cont inuum coupling st rat egies

cont inuum coupling st rat egies

Haber, Yin, Haber, Yin, Palaniappan Palaniappan(T&AM); (T&AM); Jerrard Jerrard, S ullivan, , S ullivan, Ko Ko, , Jegdic Jegdic, , Petrocovici Petrocovici(Math); (Math); Erickson, Garland, Zhou, Booth, Kale (CS ) Erickson, Garland, Zhou, Booth, Kale (CS )

Spacetime discontinuous Galerkin finite element methods

• Eliminates oscillations without stabilization
• Eliminates oscillations without stabilization

Eliminates oscillations without stabilization

Position, x Displacement, u

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.2 0.4 0.6 0.8 1 Exact 80 elements 40 elements

Position, x Displacement, u

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.2 0.4 0.6 0.8 1 Exact 1/h=20 1/h=40 1/h=80

Well Well -

• known commercial code

known commercial code S pacet ime S pacet ime DG DG

Spacetime discontinuous Galerkin finite element methods

S pace-t ime DG met hods Nonlinear conservat ion laws Elast odynamics S pace S pace-

• t ime DG met hods

t ime DG met hods Nonlinear conservat ion laws Nonlinear conservat ion laws Elast odynamics Elast odynamics

t x P

∂Q

∫

= 0 ∀ Q ⊂ D dM + b = 0 in Q

DG Method for First-Order Hyperbolic Problems

• S

uccess with linear, first-order problems

– Element -wise conservat ion – S calable element -by-element solut ions – Localized model for subscale physics

• Quench precipitate evolution

– React ion rat e kinet ics

• S

uccess with linear, first S uccess with linear, first -

• order problems
• rder problems

– – Element Element -

• wise conservat ion

wise conservat ion – – S calable element S calable element -

• by

by-

• element solut ions

element solut ions – – Localized model for Localized model for subscale subscale physics physics

• Quench precipitate evolution

Quench precipitate evolution

– – React ion rat e kinet ics React ion rat e kinet ics

Al-Sc syst em I nf low t emp. = 850 K Velocit y = 200 mm/ s

0.2 m 3.0 m 0.0035 m 0.04 m symmetry plane x, flow direction y 1.8 m 0.005 m water-spray quench zone 0.04 m

• Max. size = 1000;

5.12 x 107 unknowns 12.5 hrs. on a PC 5 mins on 64-processors of an SGI Power Challenge

N1 N j−1 N j N j

growth emission

N1 N j−1

Tent-pitcher algorithm

• Cone constraint for space-time grid
• Yields local problem on each element
• Cone constraint for space

Cone constraint for space-

• time grid

time grid

• Yields local problem on each element

Yields local problem on each element

1 1 1 1 2 2 2 3 3 3 3

Q ˆ x x d+1

space time

Tent-pitcher algorithm

• Cone constraint enables O(N) solution
• Progress constraint eliminates locking
• Cone constraint enables O(N) solution

Cone constraint enables O(N) solution

• Progress constraint eliminates locking

Progress constraint eliminates locking

Erickson, Erickson, Guoy Guoy, , S heffer S heffer, , Ungor Ungor (CS ); S ullivan(Math); Haber (T&AM) (CS ); S ullivan(Math); Haber (T&AM)

Simplifying Volumetric Data

115,000 115,000 t et rahedra t et rahedra 10,000 10,000 t et rahedra t et rahedra 2,000 2,000 t et rahedra t et rahedra Michael Garland & Yuan Michael Garland & Yuan Zhou Zhou (Computer S cience) (Computer S cience)

Processing t ime: Processing t ime: ~15 seconds ~15 seconds (1 GHz Pent ium 3) (1 GHz Pent ium 3) Applicat ions: Applicat ions:

• data compression

data compression

• multiscale

multiscale material modeling material modeling

• progressive network transmission

progressive network transmission

• guaranteed interactive display time

guaranteed interactive display time

2D crack-tip w ave scattering

• 2D x time meshing with “ tent-pitcher”
• 2D x time DG implementation
• 2D x time meshing with “ tent

2D x time meshing with “ tent -

• pitcher”

pitcher”

• 2D x time DG implementation

2D x time DG implementation

p t

Haber (T&AM); Haber (T&AM); Jerrard Jerrard, S ullivan(Math); Erickson, Garland, Kale (CS ) , S ullivan(Math); Erickson, Garland, Kale (CS )

QuickTime™ and a Planar RGB decompressor are needed to see this picture.

Monte Carlo Simulations of Elastomers With Variable Functionality

• Eliot Fried & Russell Todres, TAM; David Hardy, CS
• Effects of functionality (# cross-links/ polymer chain)
• n the behavior of an elastomeric network
• simple ball + spring model, randomly remove springs
• Eliot Fried & Russell

Eliot Fried & Russell Todres Todres, TAM; David Hardy, CS , TAM; David Hardy, CS

• Effects of functionality (# cross

Effects of functionality (# cross-

• links/ polymer chain)

links/ polymer chain)

• n the behavior of an
• n the behavior of an elastomeric

elastomeric network network

• simple ball + spring model, randomly remove springs

simple ball + spring model, randomly remove springs

O(N) algorithms for coupling atomistic and continuum models

• Duane Johnson, (MatS

E); Bob Haber, (TAM); Brent Kraczek (Phys.)

• Coupling strategy applicable to wide range of atomistic methods
• Maintains O(N) nature of atomistic and continuum models
• S

patial partition leads to energetically consistent interpolation between atomistic and continuum response models.

• Duane Johnson, (

Duane Johnson, (MatS E MatS E); Bob Haber, (TAM); Brent ); Bob Haber, (TAM); Brent Kraczek Kraczek (Phys.) (Phys.)

• Coupling strategy applicable to wide range of

Coupling strategy applicable to wide range of atomistic atomistic methods methods

• Maintains O(N) nature of

Maintains O(N) nature of atomistic atomistic and continuum models and continuum models

• S

patial partition leads to energetically consistent interpolatio S patial partition leads to energetically consistent interpolation n between between atomistic atomistic and continuum response models. and continuum response models.

Energy Partition Transducer elements (constrained atoms) Atomistic

Standard FE (constrained atoms) Standard FE (no atoms) i α

←atomistic continuum→

Surprising similarity to projection method in topology optimization

• Bob Haber, (TAM); Julian Norato, Dan Tortorelli (M&IE)
• Problem involves optimal shape design of structures allowing

for changes in topology

• Fictitious domain approach requires smooth phase transitions
• Bob Haber, (TAM); Julian

Bob Haber, (TAM); Julian Norato Norato, Dan , Dan Tortorelli Tortorelli (M&IE) (M&IE)

• Problem involves optimal shape design of structures allowing

Problem involves optimal shape design of structures allowing for changes in topology for changes in topology

• Fictitious domain approach requires smooth phase transitions

Fictitious domain approach requires smooth phase transitions

d

ball

Volume fraction computation

px py a b L L

sym sym