1 non-linear model reduction strategies for generating data driven stochastic input models nicholas...
TRANSCRIPT
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
1
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
NON-LINEAR MODEL REDUCTION STRATEGIES FOR GENERATING DATA DRIVEN STOCHASTIC INPUT MODELS
Nicholas ZabarasMaterials Process Design and Control Laboratory
Sibley School of Mechanical and Aerospace Engineering188 Rhodes HallCornell University
Ithaca, NY 14853-3801
Email: [email protected]: http://mpdc.mae.cornell.edu/
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
2
- Thermal and fluid transport in heterogeneous media are ubiquitous
- Range from large scale systems (geothermal systems) to the small scale
- Most critical devices/applications utilize heterogeneous/polycrystalline/functionally graded materials
ANALYSIS OF HETEROGENEOUS MEDIA
- Properties depend on the distribution of material/microstructure
- But only possess limited information about the microstructure/property distribution (e.g. 2D images)
Incorporate limited information into stochastic analysis:- worst case scenarios - variations on physical properties
Hydrodynamic transport through heterogeneous permeable media
Thermal transport through polycrystalline and functionally graded materials
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
3
FRAMEWORK FOR ANALYSIS OF HETEROGENEOUS MEDIA
Extract properties P1, P2, .. Pn, that the structure satisfies.
These properties are usually statistical: Volume fraction, 2 Point correlation, auto correlation
Reconstruct realizations of the structure satisfying the correlations.
Construct a reduced stochastic model of property variations from the
data. This model must be able to approximate the class of structures.
Solve the heterogeneous property problem in the reduced stochastic space for computing property variations.
1. Property extraction 2. Microstructure/property reconstruction
3. Reduced model4. Stochastic analysis
B. Ganapathysubramanian and N. Zabaras, "Modelling diffusion in random heterogeneous media: Data-driven models, stochastic collocation and the variational multi-scale method", Journal of Computational Physics, in press
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
4
INPUT STOCHASTIC MODELS: LINEAR APPROACH
Methodology for creating linear models from data
- Given some limited information (either in terms of statistical correlation functions or sample microstructure/property variations)
- Utilize some reconstruction methodology to create a finite set of realizations of the property/microstructure.
- Utilize this data set to construct a model of this variability
Use Proper Orthogonal Decomposition (POD), Principal Component Analysis (PCA) to construct a reduced order model of the data.
= a1 a2 +..+ an+
Convert variability of property/microstructure to variability of coefficients.
Not all combinations allowed. Developed subspace reducing methodology1 to find the space of allowable coefficients that reconstruct plausible microstructures
B. Ganapathysubramanian and N. Zabaras, "Modelling diffusion in random heterogeneous media: Data-driven models, stochastic collocation and the variational multi-scale method", Journal of Computational Physics, in press
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
5
LINEAR APPROACH TO MODEL GENERATION
Successfully applied to investigate effect of heterogeneous media in diffusion phenomena in two-phase microstructures.
PCA based methods are easy to implement and are well understood.
But PCA based methods are linear projection methods
Only guaranteed to discover the true structure of data lying on a linear subspace of the high dimensional input space
# of samples#
of e
igen
vec
tors
PCA works very well when the input space is linear
What about when the input space is curved/non-linear?
PCA based techniques tend to overestimate the dimensionality of the model
Further related issues:- How to generalize it to other properties/structures?
Can PCA be applied to other classes of microstructures, say, polycrystals?
- How does convergence change as the amount of information increases? Computationally?
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
6
TOWARDS A NON-LINEAR APPROACH TO MODEL GENERATION
- Find structure of data lying on a possibly non-linear subspace of the high-dimensional input data
- PCA finds a low-dimensional embedding of the data points that best preserves their variance as measured in the high-dimensional input space. Variance is measured based on Euclidian distance. This results in a linear subspace approximation of the data
3D data
PCAPt A
Pt B
- But what if the data lie on nonlinear curve in high-dimensional space?
- Have to unfold/unravel the curve
- Need non-linear approaches
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
7
NONLINEAR REDUCTION: THE KEY IDEA
Set of images. Each image = 64x64 = 4096 pixels
Each image is a point in 4096 dimensional space.
But each and every image is related (they are pictures of the same object). Same object but different poses.
That is, all these images lie on a unique curve (manifold) in 4096.
Can we get a parametric representation of this curve?
Problem: Can the parameters that define this manifold be extracted, ONLY given these images (points in 4096)
Solution: Each image can be uniquely represented as a point in 2D space (UD, LR).
Strategy: based on the ‘manifold learning’ problem
Different images of the same object: changes in up-down (UD) and left-right (LR) poses
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
8
NONLINEAR REDUCTION: THE KEY IDEA (Contd.)
Based on principles from psychology and cognitive sciences.
How does the brain interpret/store/recall high-resolution images at different poses, locations?
A low-dimensional parameterization of this very high-dimensional space
Effectively constructs a surrogate space of the actual space.
Performs all operations on this surrogate low-dimensional parametric space and maps back to the input space
These ideas are being increasingly used in problems in vision enhancement, speech and motor control, data compression Different images of the same
object: changes in up-down (UD) and left-right (LR) poses
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
9
NONLINEAR REDUCTION: EXTENSION TO INPUT MODELS
Different microstructure realizations satisfying some experimental correlations
Given some experimental correlation that the microstructure/property variation satisfies.
Construct several plausible ‘images’ of the microstructure/property.
Each of these ‘images’ consists of, say, n pixels.
Each image is a point in n-dimensional space.
But each and every ‘image’ is related.
That is, all these images lie on a unique curve (manifold) in n.
Can a low-dimensional parameterization of this curve be computed?
Strategy: based on a variant of the ‘manifold learning’ problem.
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
10
A FORMAL DEFINITION OF THE PROBLEM
State the problem as a parameterization problem (also called the manifold learning problem)
Given a set of N unordered points belonging to a manifold embedded in a high-dimensional space n, find a low-dimensional region d that
parameterizes , where d << n
Classical methods in manifold learning have been methods like the Principle Component Analysis (PCA) and multidimensional scaling (MDS).
These methods have been shown to extract optimal mappings when the manifold is embedded linearly or almost linearly in the input space.
In most cases of interest, the manifold is nonlinearly embedded in the input space, making the classical methods of dimension reduction highly approximate.
Two approaches developed that can extract non-linear structures while maintaining the computational advantage offered by PCA1,2.
1. J. B. Tenenbaum, V. De Silva, J. C. Langford, A global geometric framework for nonlinear dimension reduction Science 290 (2000), 2319-2323.
2. S Roweis, L. Saul., Nonlinear Dimensionality Reduction by Locally Linear Embedding, Science 290 (2000) 2323--2326.
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
11
AN INTUITIVE PICTURE OF THE STRATEGY
- Attempt to reduce dimensionality while preserving the geometry at all scales.
- Ensure that nearby points on the manifold map to nearby points in the low-dimensional space and faraway points map to faraway points in the low-dimensional space.
3D dataPCA
0 20 40 60 80 100-10
0
10
20
30
40
50
60
-10
-5
0
5
10
15
-10
-5
0
5
10
15
20
40
Linear approach Non-linear approach: unraveling the curve
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
12
KEY CONCEPT
Pt A Pt B
Euclidian dist
Geodesic dist
1) Geometry can be preserved if the distances between the points are preserved – Isometric mapping.
2) The geometry of the manifold is reflected in the geodesic distance between points
3) First step towards reduced representation is to construct the geodesic distances between all the sample points
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
13
MATHEMATICAL DETAILS: AN OVERVIEW
Start from N samples lying on a curve that is embedded in a high- dimensional space.
What are the properties of this curve ?
Next step is to extract some knowledge of the geometry of this curve. Utilize the concept of geodesic distance to encode this knowledge of the geometry.
How is this geodesic distance defined and computed?
Generate an isometric transformation (parameterization) from this curve to a low-dimensional region. Proof of existence of this mapping
How is this information used in numerically constructing the required mapping?
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
14
PROPERTIES OF THE CURVE Given a set of N sample points lying on { }ixMust have a notion of distance between these sample points. Define an appropriate function
: [0, ) D:M M [that determines the difference between any two points Ensure that it satisfies the properties of positive-definiteness, symmetry and the triangle inequality.
Lemma 1: (, ) is a metric space
Lemma 2a: (, ) is a bounded metric space
Lemma 2b: (, ) is dense
Lemma 2c: (, ) is completeTheorem: (, ) is compact1.
1. B. Ganapathysubramanian and N. Zabaras, "A non-linear dimension reduction methodology for generating data-driven stochastic input models", Journal of Computational Physics, submitted.
2. V. de Silva, J. B. Tenenbaum, Unsupervised Learning of Curved Manifolds, In Proceedings of the MSRI workshop on nonlinear estimation and classification. Springer Verlag, 2002.
A compact manifold embedded in a high-dimensional space can be isometrically mapped to a region in a low-dimensional space2.
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
15
THE ISOMETRIC MAPPING: COMPUTING THE GEODESIC
A compact manifold embedded in a high-dimensional space can be isometrically mapped to a region in a low-dimensional space.
Isometry encoded into the geodesic distances
Have no notion of the geometry of the manifold to start with. Hence cannot construct true geodesic distances!
( , ) inf{length( )}i j
Mm
D
Approximate the geodesic distance using the concept of graph distance G(i,j) : the distance of points far away is computed as a sequence of small hops.
This approximation, G, asymptotically matches the actual geodesic distance . In the limit of large number of samples1,2. (Theorem 4.5 in Ref. 1)
1 2(1 ) ( , ) ( , ) (1 ) ( , )i j i j i j GM Mm m mD D D
1. B. Ganapathysubramanian and N. Zabaras, " A non-linear dimension reduction methodology for generating data-driven stochastic input models ", submitted to Journal of Computational Physics.
2. M.Bernstein, V. deSilva, J.C.Langford, J.B.Tenenbaum, Graph approximations to geodesics on embedded manifolds, Dec 2000
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
16
THE ISOMETRIC MAPPING: COMPUTING THE GEODESIC
Proof of theorem is based on ideas from geometry, differential algebra and simple probability arguments
Using this approximation – can compute the pair wise geodesic
distance matrix, M, between all the N sample points
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
17
FROM THE MATRIX, M, TO A LOW-DIMENSIONAL MAP
Have N objects, the preceding developments provided a means of computing the pair wise distances between them.
Convert this set of pair wise distances into points in a low-dimensional space
This is a straightforward problem in multivariate statistical analysis1.
Can easily solve the problem using the concept of Multi Dimensional Scaling2 (MDS)
1. J. T. Kent, J. M. Bibby, K. V. Mardia, Multivariate Analysis (Probability and Mathematical Statistics), Elsevier (2006).2. T. F. Cox, M. A. A. Cox, Multidimensional scaling, 1994, Chapman and Hall
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
18
MULTI DIMENSIONAL SCALING
Given the N x N matrix of the geodesic distances, M, with elements dij.
Compute the symmetric N x N matrix, A, with elements aij = -1/2 dij2.
Suppose that the N low-dimensional points that we are interested in finding (the parameters of the N objects in high-dimensional space) are { }iy
Denote the N x N matrix of the scalar product of these N low-
dimensional points as B,
1
.d
Tij ik jk i j
k
b y y y y
Can show that A, and B, are related as B HAHWhere H is the centering matrix, 1/ij ijh N
TB YY
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
19
MULTI DIMENSIONAL SCALING (contd.)
B is a positive definite matrix and can be written in terms of its eigen values and eigen vectors as:
B
Where Λ is the diagonal matrix of the eigenvalues and Γ is the
corresponding matrix of eigenvectors. B will have a decaying eigen-spectrum. From and we get: TB YY B
1/ 2d d Y
The above equation is an estimate of Y in terms of the largest d eigenvalues of the eigenvalue decomposition of the squared geodesic distance matrix
From N points in a high-dimensional space to N points in a low-dimensional space
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
20
THE NONLINEAR MODEL REDUCTION ALGORITHM
Given a set of N sample points lying on { }ix
The complete algorithm consists of three simple stepsStep 1:
- Construct the neighborhood graph: Determine which points are neighbors, Construct the weighted graph with edges given weights corresponding to the distances, , between the pointsStep 2:
- Estimate the shortest intrinsic path- the geodesic lengths: Compute
the shortest path length, M, between the points in the weighted graph, using, say, Floyd’s algorithm
Step 3:
- Construct the d-dimensional embedding: Perform classical Multi
Dimensional Scaling on the geodesic distance matrix, M, to extract a set of N points in a low-dimensional space
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
21
CHOOSING THE OPTIMAL DIMENSIONALITY, d
The MDS procedure in the dimension reduction methodology constructs the eigen-value decomposition of a matrix. The low-dimensional representation correspond to the largest d eigenvalues of this matrix.
How is the value of d chosen?
Use ideas from recent development on estimating the intrinsic dimensionality of manifolds (Costa and Hero1)
Elegant ideas from differential geometry and graph theory.
Connect dimensionality to the rate of convergence of a graph theoretic quantity1. J.A.Costa, A.O.Hero, Geodesic Entropic Graphs for Dimension and Entropy Estimation in Manifold Learning, IEEE
Trans. on Signal Processing, 52 (2004) 2210--2221.
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
22
CHOOSING THE OPTIMAL DIMENSIONALITY, d
Based on very elegant ideas linking graph theory with differential geometry.
Based on the theorem of Beardwood-Halton-Hammersley1,2.
Based on concepts in geometric probability, the theorem relates the structure/geometry of embedded manifolds to the graph structure of a finite set of points on these embedded manifolds
1. J.A.Costa, A.O.Hero, Manifold Learning with Geodesic Minimal Spanning Trees, arXiv:cs/0307038v12. J. Beardwood, J. H. Halton, and J. M. Hammersley, “The shortest path through many points,” Proc. Cambridge
Philosophical Society, 55 (1959) 299–327.3. B. Ganapathysubramanian and N. Zabaras, " A non-linear dimension reduction methodology for generating data-driven
stochastic input models ", submitted to Journal of Computational Physics.
The theorem states that the length functional of the minimal spanning graph is related to the entropy of density of distribution of a finite set of points in a embedded manifold (with some specific properties)
Costa and Hero1 extended/applied this theorem to general embedded manifolds
Ganapathysubramanian and Zabaras3 applied it to the model reduction problem
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
23
CHOOSING THE OPTIMAL DIMENSIONALITY, d (Contd.)
The length functional of the minimal spanning tree of the geodesic matrix,
M, is related to the intrinsic dimensionality of the low-dimensional representation of the manifold 1,2.
1. B. Ganapathysubramanian and N. Zabaras, "A non-linear dimension reduction methodology for generating data-driven stochastic input models", submitted to Journal of Computational Physics.
2. J.A.Costa, A.O.Hero, Geodesic Entropic Graphs for Dimension and Entropy Estimation in Manifold Learning, IEEE Trans. on Signal Processing, 52 (2004) 2210--2221.
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
24
CHOOSING THE OPTIMAL DIMENSIONALITY, d (Contd.)
log( ) log( )L a N 1da
d
with
From the above theorem, can easily extract the dimensionality.
Taking the logarithm of the main result of the algorithm,
Where L is the length functional of the minimal spanning tree of the geodesic matrix, N is the number of samples and d is the optimal dimensionality.
Use readily available algorithms (Kruskal’s algorithm or Prim’s algorithm) to compute L for different sample sizes
Perform a least squares fit for the value of a
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
25
THE NONLINEAR MODEL REDUCTION FRAMEWORK
Given N unordered samples
N points in a low dimensional space
n. d
The procedure results in N points in a low-dimensional space. The geodesic distance + MDS step (Isomap algorithm1) results in a low-dimensional convex, connected space2, d.
1. J. B. Tenenbaum, V. De Silva, J. C. Langford, A global geometric framework for nonlinear dimension reduction ,Science 290 (2000), 2319-2323.
2. B. Ganapathysubramanian and N. Zabaras, "A non-linear dimension reduction methodology for generating data-driven stochastic input models", submitted to Journal of Computational Physics.
Using the N samples, the reduced space is given as
serves as the surrogate space for .
Access variability in by sampling over .
BUT have only come up with → map …. Need → map too
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
26
THE REDUCED ORDER STOCHASTIC MODEL
Only have N pairs to construct → map. Various possibilities based on specific problem at hand. But have to be conscious about computational effort and efficiency.
Illustrate 3 such possibilities below. Error bounds can be computed1.
n d
n d
n
d
1. Nearest neighbor map 2. Local linear interpolation
3. Local linear interpolation with projection1. B. Ganapathysubramanian and N. Zabaras, "A non-linear dimension reduction methodology for generating data-driven
stochastic input models", submitted to Journal of Computational Physics.
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
27
THE LOW DIMENSIONAL STOCHASTIC MODEL
Given N unordered samples
Compute pairwise geodesic distance
Perform MDS on this distance matrix
N points in a low dimensional
space
Algorithm consists of two parts.
1) Compute the low-dimensional representation of a set of N unordered sample points belonging to a high-dimensional space
For using this model in a stochastic collocation framework, must sample points in →
2) For an arbitrary point ξ € must find the corresponding point x €. Compute the mapping from →
n. d
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
28
NUMERICAL EXAMPLE
Problem strategy:
Extract pertinent statistical information form the experimental image
Reconstruct dataset of plausible 3D microstructures
Construct a low-dimensional parametrization of this space of microstructures
Solve the SPDE for temperature evolution using this input model in a stochastic collocation framework
Given an experimental image of a two-phase metal-metal composite (Silver-Tungsten composite).
Find the variability in temperature arising due to the uncertainty in the knowledge of the exact 3D material distribution of the specified microstructure.
T= -0.5 T= 0.5
1. S. Umekawa, R. Kotfila, O. D. Sherby, Elastic properties of a tungsten-silver composite above and below the melting point of silver J. Mech. Phys. Solids 13 (1965) 229-230
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
29
r (um)
g(r)
0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
r (um)
g(r)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
Experimental image Experimental statistics
GRF statistics
Realizations of 3D microstructure
TWO PHASE MATERIAL
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
30
NON LINEAR DIMENSION REDUCTION
The developments detailed before are applied to find a low-dimensional representation of these 1000 microstructure samples.
The optimal representation of these points was a 9-dimensional region
Able to theoretically show that these points in 9D space form a convex region in 9.
This convex region now represents the low- dimensional stochastic input space
Use sparse grid collocation strategies to sample this space.
10
15
20-15 -10 -5 0 5 10 15
-10
-5
0
5
10
15
Log(Samples)
Log
(len
gth
ofM
ST
)
100 300 500
10000
20000
30000
40000
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
31
# of collocation points
Err
or
100 101 102 103 10410-6
10-5
10-4
10-3
10-2
10-1
100
101
Computational domain of each deterministic problem: 65x65x65 pixels
COMPUTATIONAL DETAILS
The construction of the stochastic solution: through sparse grid collocation level 5 interpolation scheme used
Number of deterministic problems solved: 26017
Computational platform: 50 nodes on local Linux cluster (x2 3.2 GHz)
Total time: 210 minutes
Total number of DOF: 653x26017 ~ 7x109
32Cornell University- Zabaras, FA9550-07-1-0139
Coupling data driven model generation with a Multiscale stochastic modeling framework
Seamlessly couple stochastic analysis with multiscale analysis.
Multiscale framework (large deformation/thermal evolution) + Adaptive stochastic collocation framework
Provides roadmap to efficiently link any validated multiscale framework
Coupled with a data-driven input model strategy to analyze realistic stochastic multiscale problems.
r (um)
g(r)
0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
T= -0.5 T= 0.5
Limited data
Statistics extraction + model generation
Stochastic multiscale framework
Mean statistics
Higher-order statistics
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
33
OTHER APPLICATIONS OF THIS FRAMEWORK
From unordered samples in high-dimensional space to a convex space representing the parameterization
n d
- This methodology has significant applications to problems where working in high-dimensional spaces is computationally intractable. - Can pose the problem in a low-dimensional space
3.05 3.06 3.07 3.08 3.09 3.1 3.11 3.123.03
3.04
3.05
3.06
3.07
3.08
3.09
3.1
3.11
3.12
Taylor factor along RD
Tay
lor
fact
or a
long
TD
b
2. Rolling
3. Rolling followed by drawing
1. Drawing
R
C
ad
fe
Property-process space
-1
-0.5
0
0.5
-1-0.5
00.5
1-1.8
-1.7
-1.6
-1.5
-1.4
-1.3
Process-structure space
Process pathsA100
A1000
A80
Property-structure spacevisualizing property
evolution, process-property maps, searching and contouring, representing input uncertainty, data mining …
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic design/optimization framework
Most critical components in many applications/devices are usually fabricated from polycrystalline/ functionally graded/ heterogeneous materials
The properties at the device-scale (thermal and electrical conductivity, elastic moduli and failure mechanisms) depend on the microstructure and material distribution at the meso-scale
Have to design the operating conditions/parameters taking into account this limited information about the microstructure.
Robust design in the presence of topological uncertainty
Given limited topological information about the microstructure, design the optimal (stochastic) heat flux to be applied on one end of the device such that a required (stochastic) temperature is maintained at the other end.Design criterion:
- Maintain a specified thermal profile on the right wall
- This thermal profile is given in terms of a pdf, usually specified in terms of moments Additional uncertainties:
- Do not know the exact microstructure that the device is made up of.
- Only know certain statistical correlations that the microstructure satisfies.
- Consider the microstructure to be a random field .Design variables:
- Must design the pdf of the optimal heat flux
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic design/optimization framework : Some details
Convert the problem into an optimization problem
Denote the stochastic space in which the flux, q(x) exists as q. Convert the problem of constructing the stochastic function q € q into designing the values of the stochastic function at a finite number of points in the stochastic space (collocation based design)
Solution exists to the problem in the sense of Tikhonov.
2 2 2 2 21 2
1({ }) ( ) ( ) ... ( )
2r
k ki kJ q T T T dx
1({ }) [ ( ) ( )i i
r
q i qD J q T D
2 2 22 ( ) ( ) ... ( ) ( )]
i i
k k kq k qT D T D dx
Use gradient based minimization methods. Need to compute gradient. Notion of the directional derivative
1 1 1 1
( )q qk k
i i
n nn nm j m j
q q k q k qm j m j
D w w w w D
Need to compute the directional derivative of the temperature
1
1 1
q k
i
n nm jq k q
m j
w w D
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic design/optimization framework (contd.)
Leads to the definition of the continuum stochastic sensitivity equations.
Linear in ‘design variable perturbation’ part of the dependant variables
( , : ) ( , : ) ( , : , ) . .i i i i ix q q x q x q q h o t
Novel, highly-efficient way to compute stochastic sensitivities based on parallel sparse grid collocation schemes.
Major advance in non-intrusive stochastic optimization – any stochastic optimization problem can now be posed as a deterministic optimization problem in a large-dimensional space. Classical gradient and gradient-free optimization methods are directly applicable.
The stochastic direct problem is converted into a set of n-decoupled deterministic equation (the collocation based approach to solve SPDEs)
The stochastic temperature sensitivity equations are simply directional derivatives of the n deterministic that form the above stochastic direct problem. The stochastic sensitivity is then represented in terms of these decoupled stochastic equations as:
1 1
( , , : ) ( , , : )q kn n
m j m jm m i q k q k i
m j
x q L L x q
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Stochastic design/optimization framework
Mean temperature (right wall) Standard deviation of temperature
Physical domain: 20 μm x 20μm x 20μm
Computational domain: 65x65x65
Microstructure stochastic space: 1177 collocation points
The design stochastic heat flux represented as a Bezier surface: number of parameters: 25
Designed mean heat flux (left wall)
Bounds on the flux variation (left wall)
Run optimization problem on 36 nodes of our Linux cluster
Each iteration ~ 3 hours
Solve 10593 deterministic problems in each iteration
Each deterministic problem has 65x65x65 = 275625 DOF.
Total number of design variables = 225
Zabaras et al. JCP, 2007b
38Cornell University- Zabaras, FA9550-07-1-0139
1. Desired product shape
2. Desired microstructure-sensitive properties
Z
X
Y
3.052.922.792.662.532.412.282.15
Impact: 3D Multi-scale Design of Deformation Processes
- A two-length scale continuum sensitivity framework has been developed that allows the efficient and accurate computation of the effects of perturbations of macroscopic design variables (e.g. dies and preforms) on microscopic variables (slip system resistances, crystallographic texture, etc.)
- Using this sensitivity framework, the first 3D multiscale deformation process design simulator has been developed for the control of texture-dependent material properties and applied to complex processes.
Micro-scale representation (Orientation distribution function)
Bo
X
Bn+1
B’n+1
Macro-sensitivity problem driven by perturbation to macro-design variable ()
Micro-sensitivity problem driven by sensitivity of deformation gradient
39Cornell University- Zabaras, FA9550-07-1-0139
Multi-scale design: Design for uniform yield strength during 3D forging process
(a) Iteration 1
Yield Strength (MPa)
75
85
95
105
115
125
135
(b) Iteration 2
(c) Iteration 3 (d) Iteration 4: Optimal solution
ODF evolution at the center of the workpiece
Strong z-axis <110> fibers (compression texture)
Distribution of (z- axis tension) yield strength obtained at various design iterations
Initial guess: Large variation in yield strength, incomplete fill
Uniform strength, Desired shape obtained
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Impact: Application to designing critical components with extremal properties
Forging velocity
Die/workpiece friction
Die shape
Initial preform shape
Uncertainty in constitutive
relationUncertainty in Texture
Uncertainty in grain sizes
- Significant developments towards the design of processing paths and materials for manufacturing critical components that have extremal properties.
- Dimension reduction strategy has significant applications to problems where working in high dimensional spaces is computationally intractable: visualizing property evolution, process-property maps, searching and contouring.
- Stochastic multiscale framework provides an approach to incorporate in the multiscale design framework shown earlier operational variabilities across multiple scales. Provide rigorous failure criteria and the associated probabilities.
CCOORRNNEELLLL U N I V E R S I T Y
CCOORRNNEELLLL U N I V E R S I T Y
Future Plans1. Address a number of critical issues on non-linear model reduction techniques:
basis representation, projection, solution of PDEs, SPDEs, etc.
2. Utilize model reduction strategy to compute realistic input stochastic models of polycrystalline materials based on limited information.
3. Develop a multiscale framework for modeling the effect of microscale (topological) uncertainty on the performance of macro-components. Investigate uncertainty propagation and interaction of uncertainties across multiple scales.
4. Utilize model reduction strategy as a technique for reducing problems in high- dimensional spaces (visualizing property evolution, process-property maps, searching and contouring) into more tractable low-dimensional surrogate spaces.
5. Address many open issues on stochastic optimization using sparse grid collocation. Many technologically important applications.
6. Multi-element collocation wavelets as an alternative to Lagrange polynomials (extension to Smolyak quadrature selection rule).