Model order reduction
POD and Homogenisation
Stéphane P.A. Bordas - Pierre Kerfriden
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
Ways to reduce the models
• Homogenisa+on (FE^2, etc.) - Hierarchical
• Concurrent and hybrid (bridging domain, ARLEQUIN, etc.)
• Enrichment (PUFEM, XFEM, GFEM)
• Model reduc+on (algebraic)
2
Reduction methods based on homogenisation
3
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
4Coupling of macroscopic and microscopic levels The volume averaging theorem is postulated for:
1) Strain tensor:
2) Virtual work (Hill-Mandel condition):
3) Stress tensor:
Definition of an RVE
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Hierarchical multi-scale approaches (FE^2)
5
In softening regime:
• Lack of scale separation
• Macroscale mesh dependence The macroscopic constitutive law is not
required
Non-linear material behaviour can be simulated Microscale behaviour of material is monitored at each load step
Advantages and abilities: Drawbacks:
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
6bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
7bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
8Details in Phil. Magazine, 2015, Akbari, Kerfriden, Bordas
Reduction methods based on algebraic reduction
16
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Illustration of the method of separated representation
17
0 20
40 60 80
100
0 200 400 600 800 1000
−1
−0.5 0 0.5 1
0 20
40 60 80
100
0 200 400 600 800 1000
−1
−0.5 0 0.5 1
X X
C1 = sin(0.01x)
C2 = (x 500)3
↵1 = e 0.02t
↵2 = cos(p t)
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Illustration of the method of separated representation
18
C1 = sin(0.01x)
C2 = (x 500)3
↵1 = e 0.02t
↵2 = cos(p t)
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Illustration of the method of separated representation
19 +
=
C1 = sin(0.01x)
C2 = (x 500)3
↵1 = e 0.02t
↵2 = cos(p t)
C1↵1 +C2↵2
Very rich approximations!
(Cxi, Cyi)i2J1,ncK = argminX
xi
X
yj
(u(xi, yj) u(x¯ i, yj))2
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Data compression: get the nose with the POD!
20
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500
50 100 150 200 250 300 350 400 450
nc = 1
¯
u(xi, yi) =
nc
X
i=1
Cix(xi)Ciy(yi)
(Cxi, Cyi)i2J1,ncK = argminX
xi
X
yj
(u(xi, yj) u(x¯ i, yj))2
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Data compression: get the nose with the POD!
21
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500
50 100 150 200 250 300 350 400 450
nc = 1 nc = 2
¯
u(xi, yi) =
nc
X
i=1
Cix(xi)Ciy(yi)
Got the nose (rectangle, approximation of order 2 is enough)
(Cxi, Cyi)i2J1,ncK = argminX
xi
X
yj
(u(xi, yj) u(x¯ i, yj))2
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Data compression: get the nose with the POD!
22
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500 50
100 150 200 250 300 350 400 450
50 100 150 200 250 300 350 400 450 500
50 100 150 200 250 300 350 400 450
nc = 1 nc = 2
nc = 5 nc = 10
nc = 20
¯
u(xi, yi) =
nc
X
i=1
Cix(xi)Ciy(yi)
Converges slowly locally (idem fracture) Got the nose (rectangle, approximation of order 2 is enough)
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
a posteriori model order reduction. Idea: search for the solution as a linear combination of a set of pre-calculated representative solutions
23
−5 0 5 10 15 20 25
−3
−2
−1 0 1 2 3 4 5 6
UD
FD1 FD2
S1
nS loads
−5 0 5 10 15 20 25
−2 0 2 4 6 8 10
−5 0 5 10 15 20 25
−2 0 2 4 6 8 10
−5 0 5 10 15 20 25
−2 0 2 4 6 8 10
Reduced basis: family of representative solutions
C1
C2
C3
C = U1 U2 ... UnC
(1) Solve FINE for n_S parameters (EXPENSIVE!)
S = S1 S2 ... SnS
(3) Trunca+on
Solution
Coefficients Family of
representative solutions
Approximation of the solution in a space of small dimension (nc)
F
Int(U) + F
Ext= 0
Initial set of equations
(4) Galerkin orthogonality
S = U⌃VT =
nS
X
k=1
⌃k UkVkT
(⌃k)k2J1 nSK
where in decreasing order
nS solutions, sorted by relevance
(2) Singular value decomposi+on
•
P. Kerfriden, P. Gosselet, S. Adhikari, and S. Bordas.Bridging proper orthogonal decomposi2on methods and augmented Newton-Krylov
algorithms: an adap2ve model order reduc2on for highly nonlinear mechanical problems. Computer Methods in Applied Mechanics and Engineering, 200(5-8):850-866, 2011.
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Limitations: case of highly non-linear fracture mechanics pbs
−5 0 5 10 15 20 25
−2 0 2 4 6 8 10
−5 0 5 10 15 20 25
−2 0 2 4 6 8 10
−5 0 5 10 15 20 25
−2 0 2 4 6 8 10
Reduced Ritz basis
C1
C2
C3
0 5 10 15 20 25 30 35 40 45 50
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Error
Time step
Maximum damage /20
−5 0 5 10 15 20 25
−3
−2
−1 0 1 2 3 4 5 6
FD
UD
24 This solution is not in the
snapshot !
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Application to a parametric fracture problem
25
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Application to a parametric fracture problem
26
‣ The POD solu+on is not able to reproduce the
solu+on in the cracked area
‣ Due to lack of correla+on introduced by crack growth
‣ Leads to a local projec+on
error
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Parametric / stochastic multiscale fracture mechanics
27
➡
Reduced order modelling?➡
Direct numerical simulation: efficient preconditioner?➡
Adaptive coupling?First realisation Second realisation
Highly correlated solution fields
Localisation of fracture, uncorrelated
THE RETURN OF THE MONKEY!
28
50 100 150 200 250 300 350 400 450 500
50 100 150 200 250 300 350 400 450
What can we do to address this lack of separation of scales/reducibility?
29
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut How we got to this point...
30
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3672853/
http://orbilu.uni.lu/bitstream/10993/12454/2/presentationUSNCCM.pdf
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Data compression: fracture
31 POD order 1
POD order 3
“Exact” solution
Snapshot POD (snapshot space is spanned by the ensemble of solutions at all time steps)
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Partitioned POD/DDM
32
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Reduced DDM-POD
33
‣ Decompose the structure into subdomains
‣ Perform a reduc+on in the highly correlated region
‣ Couple the reduced to the non-
reduced region by a primal Schur
complement
Order of the POD transforms
Cross-validation error estimate
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Choice of the reduced subdomains: local error estimation by “leave
one out cross-validation” (LOOCV)
•
Reduced subspaces are independent and we assume a snapshot is a priori available‣
(1) Dimension of the local space for each subdomain?‣
(2) Is a given subdomain is reducible?•
(1) and (2) will be treated by cross-valida+on (e.g. W. J. Krzanowski. Cross- valida+on in principal component analysis. Biometrics, 43(3):575-584, 1987.)‣
Training set: snapshot‣
Valida2on set: set of addi+onal finescale solu+ons‣
Independent training/valida+on avoids overfijng‣
Cross valida+on emulates independence. Error calculated using the local reduced basis obtained by a snapshot POD transform of all the available snapshot solu+ons except the one corresponding to the value of the summa+on variable.•
NOTE: If the snapshot is not assumed a priori then‣ Assess whether the snapshot contains sufficient informa+on, and generate addi+onal, suitable, data if required
‣ Most analysis (mostly by sta+s+cians) assume the snapshot is known a priori.
Recent review: Hervé Abdi and Lynne J. Williams. Principal component analysis.
Wiley Interdisciplinary Reviews: Computa+onal Sta+s+cs, 2(4):433{459, 2010.
34
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
35Order of the POD transforms
Cross-validation error estimate
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Performance: load angle 40 | 27 - 121 nodes
•
Rela+ve error36
40 o 27 o
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Performance: load angle 40 | 27 - 256 nodes
37
40 o 27 o
•
Rela+ve errorbordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Performance: load angle 40 | 27 - 441 nodes
•
Rela+ve error38
40 o 27 o
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut Performance: load angle 40 | 27 - 961 nodes
•
Rela+ve error39
40 o 27 o
•
Rela+ve errorM A M
Institute of Mechanics
& Advanced Materials
I
40
Applica+ons to surgical simula+on
with INRIA, France; Karol Miller, UWA.
RealTcut
1Interactive multiscale
cutting simulations
Surgical simulation (real time/interactivity)
Learning PlanningSim Assistance
Precision
RealTcut
bordasS@cardiff.ac.uk, stephane.bordas@alum.northwestern.edu
RealTcut
iMAM
The ERC RealTcut project
‣ Reduce the problem size while controlling error in solving very large multiscale mechanics problems
complex microstructure
Courtecuisse et al. PBMB 2011 Discretise
41
ERC Star+ng Grant - RealTcut - 2012-2016
‣ reduce the problem size but preserve relevant mechanical information, control the error
complex
microstructure
discretizationReal-2me/interac2vity for non-linear problems involving topological changes
appren+ssage planifica+on
Simassistance
précision requise
complexité des calculs
RealT cut
42
Interac2ve simula2on of cuEng in soF 2ssue
Approach
Concrete objec2ve: compute the response of organs during surgical procedures (including cuts) in real +me (50-500 solu+ons per second)
43
Two schools of thought
‣ constant +me
➡ accuracy oqen controlled visually only
‣ model reduc+on or “learning”
➡ scarce development for biomedical problems
➡ no results available for cujng
Proposed approach: maximize accuracy for given computa+onal +me. Error control
A
4.30 A
10
72.6 ⇥
U Q
0 0.02 0.04 0.06 0.08 0.1
0 20 40 60 80 100 120 140
U Q
15%
3.4 U
Q
x
cW
c1%
25 U
[Courtecuisse et al., MICCAI, 2013]
Collabora+on INRIA
First implicit, interac2ve method for cuEng with contact
Model
reduc2on
Error control
• interac+vity
• space-+me discre+za+on?
• op+mize use of compute resources
Complex geometries from
medical images Topological changes & contact
Verifica2on & Valida2on Four main difficul+es
RoI MRI
RoI
segmenta2on
-
RdI Region of interest (RoI)
in té gr a+on e n t em ps, e rr eur
discre2za2on
adap
2vit é
44
Model reduc2on
Advanced discre2za2on
6 =
enriched zone
calculs offline calculs online: interac+vité
ac+on de l’instrument généra2on solu+ons par+culières
tri
pré-opératoire
“mapping”
spécifique pa+ent
~10^3 snapshots
POD
O(10) fonc2ons
espace réduit de pe+te
dimension
approxima+on POD globale
Local (FE) Local (FE)
~10^6 snapshots
enrichissement “pointe de coupe”
r u
calcul champs asympto+ques
représenta+on locale
45
Results - Dr Hadrien Courtecuisse, PhD INRIA
46
47
48
NUMERICAL SOLUTION GEOMETRICAL MODEL DISCRETISATION
MATERIAL MODELS Phenomenological Elasticity/Plasticity Crack growth law (Paris…)
Fracture energy
Maximum tensile strength Multi-scale
Debonding, Fibre pull-out Fibre breakage, interface fracture, grains, dislocations, MD, quantum…
…
A POSTERIORI ERROR CONTROL
EXPERIMENTS
Validation & parameter identification
Verification
CONVENTIONAL APPROACH
49
NUMERICAL SOLUTION GEOMETRICAL
MODEL
DISCRETISATION
LEARN MATERIAL MODELS
which scales?
what models?
what parameters?
what scale transition?
what data is missing?
A POSTERIORI ERROR
REAL SYSTEM
DIGITAL TWIN OF THE SYSTEM
DATA INFORMATION
Strain
Structural Health
Cracks
Environment Conditions Scales of
interest
Crack growth rate
Worst load combination
Inspection
interval
Mission?
a-priori knowledge
upscaling techniques
automatic model selection
data &
measurements
VISION
Scales++
50
Machine Learning
data Models
General approach No predefined model
fracture patterns in a composite panel
Digital Twins…
51Characterisation Monitoring
Multiscale models are unreliable
Quantitative predictions
?
Fracture/lack of scale separation
Learn better models
Measure Analyse & Learn
from data Improved model
Identify missing
data Validate
52
10010110100011110100101010100010100100101 10101010101010101101111100011010100001110 1010000100000001110101111001010010111101 01110101011110100000011111000010000000111
10100001000000011101011110010100101111010101010101110100101010010010101010000001111010101010000000111010101010010100100101001001010110111111100011 1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100010100
1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100010100
10100001000000011101011110010100101111010101010101110100101010010010101010000001111010101010000000111010101010010100100101001001010110111111100011000101 1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100 1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100101000010000000111010111100101001011110101010101011101001010100100101010100000011110101010100000001110101010100101001001010010010101101111111
53
10010110100011110100101010100010100100101 10101010101010101101111100011010100001110 1010000100000001110101111001010010111101 01110101011110100000011111000010000000111
10100001000000011101011110010100101111010101010101110100101010010010101010000001111010101010000000111010101010010100100101001001010110111111100011 1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100010100
1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100010100
10100001000000011101011110010100101111010101010101110100101010010010101010000001111010101010000000111010101010010100100101001001010110111111100011000101 1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100 1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100101011011111110001100101000010000000111010111100101001011110101010101011101001010100100101010100000011110101010100000001110101010100101001001010010010101101111111
54
10010110100011110100101010 10101010101010101101111100 10100001000000011101011110 01110101011110100000011111
1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100 1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100
1010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100 10100001000000011101011110010100101111010101010101110100101010010010101010000001111010101010000000111010101010010100100101001001010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100100
10100001000000011101011110010100101111010101010101110100101010010010101010000001111010101010000000111010101010010100100101001001010000100000001110101111001010010111101010101010111010010101001001010101000000111101010101000000011101010101001010010010100
• Experience every event that its flying twin experiences
• Will revolu+onise cer+fica+on, fleet management and support (mirrors life of the “as-built” state)
• Will decrease weight
• no reliance on sta+s+cal distribu+on of material proper+es
• no reliance on heuris+c design methods
• less reliance on physical tes+ng (environment?)
• no assumed similitude between tes+ng and opera+onal condi+ons
56
NUMERICAL SOLUTION GEOMETRICAL MODEL DISCRETISATION
MATERIAL MODELS Phenomenological Elasticity/Plasticity Crack growth law (Paris…)
Fracture energy
Maximum tensile strength Multi-scale
Debonding,Fibre pull-out Fibre breakage, interface fracture, grains, dislocations, MD, quantum…
…
A POSTERIORI ERROR CONTROL
EXPERIMENTS
Validation & parameter identification
Verification
CONVENTIONAL APPROACH
57
58
NUMERICAL SOLUTION
IMAGE/MODEL DISCRETISATION
MATERIAL MODELS Phenomenological Neo-Hookean, Ogden, …
Multi-scale cutting, fracture,
???
Patient specific ???
A POSTERIORI ERROR CONTROL
EXPERIMENTS ???
Validation & parameter identification
Verification
Alex Bilger
59
REAL PATIENT
DIGITAL TWIN OF THE PATIENT
DATA INFORMATION
Organ state
Health
Disease
Environment Conditions Scales of
interest
Disease evolution
“Inspection”int
erval
Fitness
Treatment
simulation
Model
Data
Global error
Local errors
Multi-physics Multi-scale Data-enriched
> E g
< E g
Global single scale model selection
h, p refinement
analytical enrichment experimental enrichment atomistics
meso-scopic continuum
a priori
a posteriori real-time
Data-aware
mechanics
Papers on fracture
http://orbilu.uni.lu/handle/10993/26421 http://orbilu.uni.lu/handle/10993/22289 http://orbilu.uni.lu/handle/10993/20721 http://orbilu.uni.lu/handle/10993/24170 http://orbilu.uni.lu/handle/10993/21427 http://orbilu.uni.lu/handle/10993/21295 http://orbilu.uni.lu/handle/10993/16323
http://orbilu.uni.lu/handle/10993/22420 http://orbilu.uni.lu/handle/10993/19535 http://orbilu.uni.lu/handle/10993/21330 http://orbilu.uni.lu/handle/10993/18262
http://orbilu.uni.lu/handle/10993/19509 http://orbilu.uni.lu/handle/10993/19371 http://orbilu.uni.lu/handle/10993/17536 http://orbilu.uni.lu/handle/10993/17647 http://orbilu.uni.lu/handle/10993/14135 http://orbilu.uni.lu/handle/10993/16842
Papers on fracture
https://orbilu.uni.lu/bitstream/10993/22331/2/paper.pdf http://orbilu.uni.lu/handle/10993/25048
http://orbilu.uni.lu/handle/10993/20721 http://orbilu.uni.lu/handle/10993/22420 http://orbilu.uni.lu/handle/10993/19960 http://orbilu.uni.lu/handle/10993/12316 http://orbilu.uni.lu/handle/10993/15109 http://orbilu.uni.lu/handle/10993/14067 http://orbilu.uni.lu/handle/10993/13879 http://orbilu.uni.lu/handle/10993/13876 http://orbilu.uni.lu/handle/10993/12523 http://orbilu.uni.lu/handle/10993/10965 http://orbilu.uni.lu/handle/10993/21442 http://orbilu.uni.lu/handle/10993/12107 http://orbilu.uni.lu/handle/10993/12026 http://orbilu.uni.lu/handle/10993/12026 http://orbilu.uni.lu/handle/10993/12089 http://orbilu.uni.lu/handle/10993/12113 http://orbilu.uni.lu/handle/10993/12116 http://orbilu.uni.lu/handle/10993/21337 http://orbilu.uni.lu/handle/10993/15234 http://orbilu.uni.lu/handle/10993/19960