STOCHASTIC MODELING OF MULTISCALE SYSTEMS NICHOLAS ZABARAS Materials Process Design and Control Laboratory Sibley School of Mechanical & Aerospace Engineering 188 Frank H T Rhodes Hall Cornell University Ithaca, NY 14853 Email: zabaras@cornell.edu URL: http://mpdc.mae.cornell.edu CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory OUTLINE Motivation: coupling multiscaling and uncertainty analysis Mathematical representation of uncertainty Variational multiscale method (VMS) Stochastic support method Stochastic convection-diffusion equations Computing PDFs of microstructures (Maximum Entropy) Sparse grid collocation methods (Smolyak quadrature) Natural convection on rough surfaces Diffusion in stochastic heterogeneous media Future research directions CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory NEED FOR UNCERTAINTY ANALYSIS Uncertainty is everywhere From NIST Porous media From Intel website Silicon wafer From GE-AE website Aircraft engines From DOE Material process Variation in properties, constitutive relations Imprecise knowledge of governing physics, surroundings Simulation based uncertainties (irreducible) CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory WHY UNCERTAINTY AND MULTISCALING ? Uncertainties introduced across various length scales have a non-trivial interaction Current sophistications – resolve macro uncertainties Micro Physical properties, structure follow a statistical description CORNELL U N I V E R S I T Y Meso Macro Use micro averaged models for resolving physical scales Imprecise boundary conditions Initial perturbations Materials Process Design and Control Laboratory UNCERTAINTY ANALYSIS TECHNIQUES Monte-Carlo : Simple to implement, computationally expensive Perturbation, Neumann expansions : Limited to small fluctuations, tedious for higher order statistics Sensitivity analysis, method of moments : Probabilistic information is indirect, small fluctuations Spectral stochastic uncertainty representation Basis in probability and functional analysis Can address second order stochastic processes Can handle large fluctuations, derivations are general CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory RANDOM VARIABLES = FUNCTIONS ? Math: Probability space (W, F, P) Sample space Probability measure Sigma-algebra Random variable F W W W : Random variable W : (W) A stochastic process is a random field with variations across space and time X : ( x, t , W) CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SPECTRAL STOCHASTIC REPRESENTATION A stochastic process = spatially, temporally varying random function X : ( x, t , W) CHOOSE APPROPRIATE BASIS FOR THE PROBABILITY SPACE HYPERGEOMETRIC ASKEY POLYNOMIALS GENERALIZED POLYNOMIAL CHAOS EXPANSION SUPPORT-SPACE REPRESENTATION PIECEWISE POLYNOMIALS (FE TYPE) SPECTRAL DECOMPOSITION KARHUNEN-LOÈVE EXPANSION COLLOCATION, MC (DELTA FUNCTIONS) SMOLYAK QUADRATURE, CUBATURE, LH CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory KARHUNEN-LOEVE EXPANSION X ( x, t , ) X ( x, t ) X i ( x, t )i ( ) i 1 ON random variables Deterministic functions Stochastic Mean process function Deterministic functions ~ eigen-values , eigenvectors of the covariance function Orthonormal random variables ~ type of stochastic process In practice, we truncate (KL) to first N terms X ( x, t , ) fn( x, t , 1 , CORNELL U N I V E R S I T Y ,N ) Materials Process Design and Control Laboratory GENERALIZED POLYNOMIAL CHAOS Generalized polynomial chaos expansion is used to represent the stochastic output in terms of the input X ( x, t , ) fn( x, t , 1 , ,N ) Stochastic input Z ( x, t , ) Z i ( x, t ) i (ξ( )) i 0 Stochastic output Askey polynomials in input Deterministic functions Askey polynomials ~ type of input stochastic process Usually, Hermite, Legendre, Jacobi etc. CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SUPPORT-SPACE REPRESENTATION Any function of the inputs, thus can be represented as a function defined over the support-space A ξ (1 , , N ) : f (ξ) 0 FINITE ELEMENT GRID REFINED IN HIGH-DENSITY REGIONS X Xˆ L2 2 ˆ ( X (ξ ) X (ξ )) f (ξ )dξ A Ch q 1 JOINT PDF OF A TWO RANDOM VARIABLE INPUT CORNELL U N I V E R S I T Y – SMOLYAK QUADRATURE OUTPUT REPRESENTED ALONG SPECIAL COLLOCATION POINTS – IMPORTANCE MONTE CARLO Materials Process Design and Control Laboratory VARIATIONAL MULTISCALE METHOD WITH ALGEBRAIC SUBGRID MODELLING Application : deriving stabilized finite element formulations for advection dominant problems CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory VARIATIONAL MULTISCALE HYPOTHESIS EXACT SOLUTION COARSE SOLUTION INTRINSICALLY COUPLED SUBGRID SOLUTION H COARSE GRID RESOLUTION CANNOT CAPTURE FINE SCALE VARIATIONS THE FUNCTION SPACES FOR THE EXACT SOLUTION ALSO SHOW A SIMILAIR DECOMPOSITION In the presence of uncertainty, the statistics of the solution are also coupled for the coarse and fine scales CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory VARIATIONAL MULTISCALE BASICS DERIVE THE WEAK FORMULATION FOR THE GOVERNING EQUATIONS PROJECT THE WEAK FORMULATION ON COARSE AND FINE SCALES COARSE WEAK FORM SOLUTION FUNCTION SPACES ARE NOW STOCHASTIC FUNCTION SPACES FINE (SUBGRID) WEAK FORM COMPUTATIONAL SUBGRID MODELS REMOVE SUBGRID EFFECTS IN THE COARSE WEAK FORM USING STATIC CONDENSATION MODIFIED MULTISCALE COARSE WEAK FORM INCLUDING SUBGRID EFFECTS CORNELL U N I V E R S I T Y ALGEBRAIC SUBGRID MODELS APPROXIMATE SUBGRID SOLUTION NEED TECHNIQUES TO SOLVE STOCHASTIC PDEs Materials Process Design and Control Laboratory VMS – ILLUSTRATION [NATURAL CONVECTION] Mass conservation Momentum conservation Energy conservation Constitutive laws v vg v0 v v v Ra( ) Pr( ) eg t v 2 t pI 2 Pr( ) (v) 1 (v) [v (v)T ] 2 D VMS DERIVE DERIVE DEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID CORNELL U N I V E R S I T Y g OBTAIN FINAL COARSE ASGS FORMULATION Materials Process Design and Control Laboratory WEAK FORM OF EQUATIONS Energy function space Test Trial E : L2 (W; L2 (T ; H 1 ( D))), g on E0 w : w L2 (W; H 1 ( D)), w 0 on Energy equation – Find the following holds E such that, for all w E0 , (t , w) (v. , w) ( , w) v 0 VMS hypothesis: Exact solution = coarse scale solution + fine scale (subgrid) solution ' E E E ', E0 E0 E0 ' VMS DERIVE DERIVE DEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID CORNELL U N I V E R S I T Y OBTAIN FINAL COARSE ASGS FORMULATION Materials Process Design and Control Laboratory ENERGY EQUATION – SCALE DECOMPOSITION Energy equation – Find E and w E0 such that, for all E and w E0 , the following holds Coarse scale variational formulation ( t t ', w) (v. v. ', w) ( ', w) v 0 Subgrid scale variational formulation ( t t ', w ') (v. v. ', w ') ( ', w ') v 0 These equations can be re-written in the strong form with assumption on regularity as follows t ' v. ' 2 ' ( t v. 2 ) R VMS DERIVE DERIVE DEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID CORNELL U N I V E R S I T Y OBTAIN FINAL COARSE ASGS FORMULATION Materials Process Design and Control Laboratory ELEMENT FOURIER TRANSFORM Element Fourier transform D (e) gˆ (k , ) exp(i D( e ) SPATIAL MESH RANDOM FIELD DEFINED IN WAVENUMBER SPACE k x ) g ( x, )dx h RANDOM FIELD DEFINED OVER THE DOMAIN Addressing spatial derivatives kj kj g k x n j exp(i ) g ( x, )d i gˆ (k , ) i gˆ (k , ) x D( e ) h h h NEGLIGIBLE FOR LARGE WAVENUMBERS SUBGRID VMS DERIVE DERIVE DEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID CORNELL U N I V E R S I T Y APPROXIMATION OF DERIVATIVE OBTAIN FINAL COARSE ASGS FORMULATION Materials Process Design and Control Laboratory ASGS [ALGEBRAIC SUBGRID SCALE] MODEL t ' v. ' 2 ' ( t v. 2 ) R STRONG FORM OF EQUATIONS FOR SUBGRID 1 t f n ( f n 1 f n ), t f n f n 1 (1 ) f n CHOOSE AND APPROPRIATE TIME INTEGRATION ALGORITHM t n' L( n' ) Rn TIME DISCRETIZED SUBGRID EQUATION TAKE ELEMENT FOURIER TRANSFORM 2 1 ' k vk 1 ˆ' i 2 ˆn Rˆn n t h h t 2 1 1 ' 1 v t Rn n , t c1 2 c2 h t h t 2 n' CORNELL U N I V E R S I T Y 1 2 Materials Process Design and Control Laboratory MODIFIED COARSE FORMULATION Assume the solution obeys the following regularity conditions (v ', w) ( ', v w), ( ', w) v ( ', 2 w) v By substituting ASGS model in the coarse scale weak form ( t n , w) (v n , w) ( n , w) v ( q0 , w) ht Nel ( t n v n 2 n , t ( w v w 2 w) e 1 Nel ( n' , t w /( t ) t v w) 2 w w) 0 e 1 w w /( t ) A similar derivation ensues for stochastic Navier-Stokes VMS DERIVE DERIVE DEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID CORNELL U N I V E R S I T Y OBTAIN FINAL COARSE ASGS FORMULATION Materials Process Design and Control Laboratory FLOW PAST A CIRCULAR CYLINDER NO-SLIP TRACTION FREE 6 Y RANDOM UINLET 8 4 2 0 0 5 10 15 20 X NO-SLIP INLET VELOCITY ASSUMED TO BE A UNIFORM RANDOM VARIABLE KARHUNEN-LOEVE EXPANSION YIELD A SINGLE RANDOM VARAIBLE THUS, GENERALIZED POLYNOMIAL CHAOS LEGENDRE POLYNOMIALS (ORDER 3 USED) Investigations: Vortex shedding, wake characteristics CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory FULLY DEVELOPED VORTEX SHEDDING Mean pressure Second LCE coefficient First LCE coefficient Wake region in the mean pressure is diffusive in nature Also, the vortices do not occur at regular intervals [Karniadakis J. Fluids. Engrg] CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory VELOCITIES AND FFT 0.6 0.15 0.4 0.2 0.09 V Amplitude 0.12 0 0.06 -0.2 0.03 0 0.1 -0.4 -0.6 0.2 0.3 Frequency 0.4 0.5 FFT YIELDS A MEAN SHEDDING FREQUENCY OF 0.162 FFT SHOWS A DIFFUSE BEHAVIOR IMPLYING CHANGING SHEDDING FREQUENCIES CORNELL U N I V E R S I T Y Deterministic Mean 5 8 11 X 14 17 20 MEAN VELOCITY AT NEAR WAKE REGION EXHIBITS SUPERIMPOSED FREQUENCIES Materials Process Design and Control Laboratory VARIATIONAL MULTISCALE METHOD WITH EXPLICIT SUBGRID MODELLING FOR MULTISCALE DIFFUSION IN HETEROGENEOUS RANDOM MEDIA CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory MODEL MULTISCALE HEAT EQUATION D u ( K u ) f in D t u ( x, 0) u0 ( x) in D u u g on THE DIFFUSION COEFFICIENT K IS HETEROGENEOUS AND POSSESSES RAPID RANDOM VARIATIONS IN SPACE OTHER APPLICATIONS – DIFFUSION IN COMPOSITES – FUNCTIONALLY GRADED MATERIALS FLOW IN HETEROGENEOUS POROUS MEDIA INHERENTLY STATISTICAL DIFFUSION IN MICROSTRUCTURES FINAL COARSE VMS AFFINE DERIVE COARSE-TODEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID MAP CORRECTION FORMULATION CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory STOCHASTIC WEAK FORM U u : u L2 (W; H 1 ( D)), u u g V v : v L2 (W; H 1 ( D)), v 0 Weak formulation : Find u U such that, for all v V (u,t , v) a(u, v) ( f , v); a(u, v) : ( K u, v) VMS hypothesis Exact solution u u u C F Coarse solution U U C U F Subgrid solution V V C V F FINAL COARSE VMS AFFINE DERIVE COARSE-TODEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID MAP CORRECTION FORMULATION CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory EXPLICIT SUBGRID MODELLING: IDEA DERIVE THE WEAK FORMULATION FOR THE GOVERNING EQUATIONS COARSE WEAK FORM PROJECT THE WEAK FORMULATION ON COARSE AND FINE SCALES FINE (SUBGRID) WEAK FORM COARSE-TO-SUBGRID MAP EFFECT OF COARSE SOLUTION ON SUBGRID SOLUTION AFFINE CORRECTION SUBGRID DYNAMICS THAT ARE INDEPENDENT OF THE COARSE SCALE LOCALIZATION, SOLUTION OF SUBGRID EQUATIONS NUMERICALLY FINAL COARSE WEAK FORMULATION THAT ACCOUNTS FOR THE SUBGRID SCALE EFFECTS CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SCALE PROJECTION OF WEAK FORM Find u U and u U such that, for all F F and v V Projection of weak form on coarse scale C C F F v V C C (u,Ct , vC ) (u,Ft , vC ) a(uC , vC ) a(u F , vC ) ( f , vC ) Projection of weak form on subgrid scale (u,Ct , v F ) (u,Ft , v F ) a(uC , v F ) a(u F , v F ) ( f , v F ) u F uˆ F u F 0 EXACT SUBGRID SOLUTION COARSE-TO-SUBGRID MAP SUBGRID AFFINE CORRECTION FINAL COARSE VMS AFFINE DERIVE COARSE-TODEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID MAP CORRECTION FORMULATION CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SPLITTING THE SUBGRID SCALE WEAK FORM Subgrid scale weak form (u , v ) (u , v ) a(u , v ) a(u , v ) ( f , v ) C ,t F F ,t F C F F F F Coarse-to-subgrid map F F C F F F ˆ ˆ (u , v ) (u,t , v ) a(u , v ) a(u , v ) 0 C ,t F Subgrid affine correction (u , v ) a(u , v ) ( f , v ) F0 ,t CORNELL U N I V E R S I T Y F F0 F F Materials Process Design and Control Laboratory NATURE OF MULTISCALE DYNAMICS ASSUMPTIONS: 1 1 NUMERICAL ALGORITHM FOR SOLUTION OF THE MULTISCALE PDE t SUBGRID TIME STEP t COARSE TIME STEP A(t ) B (t ) ũC ūC Coarse solution field at start of time step ûF t CORNELL U N I V E R S I T Y Coarse solution field at end of time step t Materials Process Design and Control Laboratory REPRESENTING COARSE SOLUTION COARSE MESH ELEMENT D( e ) RANDOM FIELD DEFINED OVER THE ELEMENT u C ( x, t , ) FINITE ELEMENT PIECEWISE POLYNOMIAL REPRESENTATION C u 1 (t, ) ( x) USE GPCE TO REPRESENT THE RANDOM COEFFICIENTS C u 1 s0 s (t) s ( ) ( x) nbf nbf PC Given the coefficients u s (t ) , the coarse scale solution is completely defined C CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory COARSE-TO-SUBGRID MAP COARSE MESH ELEMENT D( e ) ANY INFORMATION FROM COARSE TO SUBGRID SOLUTION CAN BE PASSED ONLY THROUGH uCs (t ) uˆ ( x, t , ) 1 s 0 uCs (t )Fs ( x, t , ) F COARSE-TOSUBGRID MAP nbf PC INFORMATION BASIS FUNCTIONS THAT ACCOUNT FROM COARSE FOR FINE SCALE SCALE EFFECTS FINAL COARSE VMS AFFINE DERIVE COARSE-TODEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID MAP CORRECTION FORMULATION CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SOLVING FOR THE COARSE-TO-SUBGRID MAP START WITH THE WEAK FORM (u,Ct , v F ) (uˆ,Ft , v F ) a(uC , v F ) a(uˆ F , v F ) 0 APPLY THE MODELS FOR COARSE SOLUTION AND THE C2S MAP u C s ( s Fs ) ,t , v F K uCs ( s Fs ) , v F 0 AFTER SOME ASSUMPTIONS ON TIME STEPPING ( s Fs ),t , v F K ( s Fs ), v F 0 THIS IS DEFINED OVER EACH ELEMENT, IN EACH COARSE TIME STEP 0 CORNELL U N I V E R S I T Y t t Materials Process Design and Control Laboratory BCs FOR THE COARSE-TO-SUBGRID MAP ( s Fs ),t , v F K ( s Fs ), v F 0 INTRODUCE A SUBSTITUTION s s Fs F F , , v K , v 0 s t s CONSIDER AN ELEMENT x3 x4 F F , v K , v 0 s ,t s , s ,t ( x , t ', ) s ( ) x1 s s CORNELL U N I V E R S I T Y x2 s ,t ( x, 0, ) s ,t ( x, tn , ) Materials Process Design and Control Laboratory SOLVING FOR SUBGRID AFFINE CORRECTION START WITH THE WEAK FORM (u,Ft 0 , v F ) a(u F 0 , v F ) ( f , v F ) CONSIDER AN ELEMENT x4 x3 WHAT DOES AFFINE CORRECTION MODEL? – EFFECTS OF SOURCES ON SUBGRID SCALE – EFFECTS OF INITIAL CONDITIONS x1 uF0 0 x2 IN A DIFFUSIVE EQUATION, THE EFFECT OF INITIAL CONDITIONS DECAY WITH TIME. WE CHOOSE A CUT-OFF To reduce cut-off effects and to increase efficiency, we can use the quasistatic subgrid assumption s ,t u,Ft 0 0 FINAL COARSE VMS AFFINE DERIVE COARSE-TODEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID MAP CORRECTION FORMULATION CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory MODIFIED COARSE SCALE FORMULATION We can substitute the subgrid results in the coarse scale variational formulation to obtain the following C C C C C C u , , v u , , v u K , v s t s s s t s s ( f , v C ) K u F 0 , v C u F 0 , t , v C We notice that the affine correction term appears as an antidiffusive correction Often, the last term involves computations at fine scale time steps and hence is ignored FINAL COARSE VMS AFFINE DERIVE COARSE-TODEFINE PROBLEM WEAK FORM HYPOTHESIS SUBGRID MAP CORRECTION FORMULATION CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory DIFFUSION IN A RANDOM MICROSTRUCTURE – A MIXTURE MODEL IS USED AS AN EXAMPLE OF GENERATING A HETEROGENEOUS DISTRIBUTION OF CONDUCTIVITY – WE ASSUME THAT THE DIFFUSION COEFFICIENTS OF INDIVIDUAL CONSTITUENTS ARE NOT KNOWN EXACTLY k* ( ) k*0 k*1 ( ) THE INTENSITY OF THE GRAY-SCALE IMAGE IS MAPPED TO THE CONCENTRATIONS DARKEST DENOTES PHASE LIGHTEST DENOTES PHASE k ( x, ) (k ( ) k ( )) I ( x) k ( ) u (x, 0, ) 0 CORNELL U N I V E R S I T Y u |( x 1) 0, u |( x 0) 1, k u |( y 0,1) 0 n Materials Process Design and Control Laboratory RESULTS AT TIME = 0.05 FIRST ORDER GPCE COEFF SECOND ORDER GPCE COEFF FULLY RESOLVED GPCE SIMULATION RECONSTRUCTED FINE SCALE SOLUTION (VMS) MEAN CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory RESULTS AT TIME = 0.2 FIRST ORDER GPCE COEFF SECOND ORDER GPCE COEFF FULLY RESOLVED GPCE SIMULATION RECONSTRUCTED FINE SCALE SOLUTION (VMS) MEAN CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory HIGHER ORDER TERMS AT TIME = 0.2 FOURTH ORDER GPCE COEFF FIFTH ORDER GPCE COEFF FULLY RESOLVED GPCE SIMULATION RECONSTRUCTED FINE SCALE SOLUTION (VMS) THIRD ORDER GPCE COEFF CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SUPPORT-SPACE – STOCHASTIC GALERKIN f ( ( )) - Joint probability density function of the inputs A { ( ) : f ( ( )) 0} - The input support-space denotes the regions where input joint PDF is strictly positive Triangulation of the supportspace Any function X ( ( )) can be represented as a piecewise polynomial on the triangulated support-space h X ( ( )) - Function to be approximated X h ( ( )) - Piecewise polynomial approximation over support-space L2 convergence – (mean-square) (X Error in approximation is penalized severely in high input joint PDF regions. We use importance based refinement of grid to avoid this CORNELL U N I V E R S I T Y h ( ( )) X ( ( )))2 f ( ( ))d Chq 1 A h = mesh diameter for the support-space discretization q = Order of interpolation Materials Process Design and Control Laboratory IMPLEMENTATION OF SUPPORT-SPACE A stochastic process W ( x, t , ) can be interpreted as a random variable at each spatial point Two-level grid approach A D Support-space grid Spatial domain • Mesh dense in regions of high input joint PDF Spatial grid • There is finite element interpolation at both spatial and random levels D( e ) ( x, ) A( e ') Element • Each spatial location handles an underlying support-space grid • Highly OOP structure nbf nbf ' nbf i 1 j 1 i 1 f ( x, ) fi ( ) Ni ( x) fi j N 'j ( ) Ni ( x) CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory CAPTURING UNSTABLE EQUILIBRIUM • Computational details – 1600 bilinear elements for spatial grid • Time of simulation – 1.5 nondimensional units Cold wall c 0 1 0.75 Insulated Y Insulated 0.5 • Rayleigh number – uniformly distributed random variable between 1530 and 1870 (10% fluctuation about 1700) • Prandtl number – 6.95 • Time stepping – 0.002 nondimensional units 0.25 0 0 0.25 0.5 0.75 1 X Hot wall h 1 • Support-space grid – Onedimensional with ten linear elements • Simulation about the critical Rayleigh number – conduction below, convection above • Both GPCE and support-space methods are used separately for addressing the problem • Failure of Generalized polynomial chaos approach • Support-space method – evaluation and results against a deterministic simulation CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory FAILURE OF THE GPCE 9.2E-07 5.7E-06 1.0E-05 1.5E-05 2.0E-05 2.5E-05 3.0E-05 3.4E-05 1 -6.4E-08 4.3E-07 1 1.4E-06 1.9E-06 2.4E-06 2.9E-06 3.4E-06 Y-vel X-vel 0.75 0.75 0.5 Y Y Mean X- and Yvelocities determined by GPCE yields extremely low values !! (Gibbs effect) 0.25 0 0.5 0.25 0 0.25 0.5 0.75 0 1 X -5.0E-03 -3.6E-03 -2.1E-03 -7.1E-04 7.1E-04 1 2.1E-03 3.6E-03 0.5 0.75 1 1.6E-03 2.8E-03 4.0E-03 5.2E-03 Y-vel 0.75 Y Y 0.5 0.5 0.25 0.25 0 0 0.25 0.5 X CORNELL 0.25 -3.2E-03 -2.0E-03 -8.0E-04 3.9E-04 1 5.0E-03 0.75 X- and Yvelocities obtained from a deterministic simulation with Ra = 1870 (the upper limit) 0 X X-vel U N I V E R S I T Y 9.3E-07 0.75 1 0 0 0.25 0.5 0.75 1 X Materials Process Design and Control Laboratory PREDICTION BY SUPPORT-SPACE METHOD -5.0E-03 -3.6E-03 -2.1E-03 -7.1E-04 7.1E-04 1 2.1E-03 3.6E-03 5.0E-03 -3.2E-03 -2.0E-03 -7.4E-04 4.9E-04 1 2.9E-03 4.2E-03 5.4E-03 Y-vel X-vel 0.75 Y Mean X- and Yvelocities determined by support-space method at a realization Ra=1870 Y 0.75 0.5 0.25 0.25 0 0.5 0 0 0.25 0.5 0.75 1 0 0.25 -5.0E-03 -3.6E-03 -2.1E-03 -7.1E-04 7.1E-04 1 2.1E-03 3.6E-03 1 1.6E-03 2.8E-03 4.0E-03 5.2E-03 0.75 Y Y 0.5 0.5 0.25 0.25 0 0 0.25 0.5 X CORNELL 0.75 Y-vel 0.75 X- and Yvelocities obtained from a deterministic simulation with Ra = 1870 (the upper limit) -3.2E-03 -2.0E-03 -8.0E-04 3.9E-04 1 5.0E-03 X-vel 0.5 X X U N I V E R S I T Y 1.7E-03 0.75 1 0 0 0.25 0.5 0.75 1 X Materials Process Design and Control Laboratory SPARSE GRID COLLOCATION If the number of random inputs is large (dimension D ~ 10 or higher), the number of grid points to represent an output on the support-space mesh increases exponentially GPCE for very high dimensions yields highly coupled equations and ill-conditioned systems (relative magnitude of coefficients can be drastically different) Instead of relying on piecewise interpolation, series representations, can we choose collocation points that still ensure accurate interpolations of the output (solution) CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SMOLYAK ALGORITHM: SPARSE GRIDS Full tensor product grid: 289 points Example of using sparse grids to build interpolating functions: Discontinuous functions Sparse Grid: 65 points Left to right: Improving interpolation depth Number of points required to construct For an error around 2x10-2: interpolating functions reduces combinatorially. Required number of points using sparse grids 3300 Reduction more significant as the number of Required number of points using full tensor products: 32769 dimensions increases CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SMOLYAK ALGORITHM LET OUR BASIC 1D INTERPOLATION SCHEME BE SUMMARIZED AS Ui( f ) a xi X i x i f ( xi ) IN MULTIPLE DIMENSIONS, THIS CAN BE WRITTEN AS (U i1 U id )( f ) xi1 X i1 xid X id (axi1 axid ) f (x i1 , , x id ) TO REDUCE THE NUMBER OF SUPPORT NODES WHILE MAINTAINING ACCURACY WITHIN A LOGARITHMIC FACTOR, WE USE SMOLYAK METHOD U 0 0, i U i U i 1 , i i1 Aq ,d ( f ) Aq 1,d ( f ) (i1 id id )( f ) i q IDEA IS TO CONSTRUCT AN EXPANDING SUBSPACE OF COLLOCATION POINTS THAT CAN REPRESENT PROGRESSIVELY HIGHER ORDER POLYNOMIALS IN MULTIPLE DIMENSIONS A FEW FAMOUS SPARSE QUADRATURE SCHEMES ARE AS FOLLOWS: CLENSHAW CURTIS SCHEME, MAXIMUM-NORM BASED SPARSE GRID AND CHEBYSHEV-GAUSS SCHEME CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SMOLYAK ALGORITHM Extensively used in statistical mechanics Uni-variate interpolation Provides a way to construct interpolation functions based on minimal number of points Ui( f ) a xi X i x i f ( xi ) Multi-variate interpolation (U i1 Univariate interpolations to multivariate (U i1 U id )( f ) interpolations xi1 X i1 U id )( f ) xid X id ( axi1 (axi1 axid ) f (x , , x id ) xi1 X i1 xid iX id 1 Smolyak interpolation U 0 0, i U i U i 1 , Some degradation in accuracy i i1 Aq ,d ( f ) Aq 1,d ( f ) (i1 id id )( f ) i q Maximal reduction when the function is assumed to be smooth CORNELL U N I V E R S I T Y D = 10 ORDER CC FE 3 1581 1000 4 8801 10000 5 41625 100000 Materials Process Design and Control Laboratory SPARSE GRID COLLOCATION METHOD Solution Methodology PREPROCESSING Compute list of collocation points based on number of stochastic dimensions, N and level of interpolation, q Compute the weighted integrals of all the interpolations functions across the stochastic space (wi) Solve the deterministic problem defined by each set of collocated points Use any validated deterministic solution procedure. Completely non intrusive POSTPROCESSING 0.301 0.260 0.301 0.260 0.220 0.220 0.180 0.180 0.140 0.140 0.100 0.100 0.060 0.060 0.020 0.020 Compute moments and other statistics with simple operations of the deterministic data at the collocated points and the preprocessed list of weights Std deviation of temperature: Natural convection CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory USING THE COLLOCATION METHOD FOR HIGHER DIMENSIONS 1. Flow through heterogeneous random media Alloy solidification, thermal insulation, petroleum prospecting Look at natural convection through a realistic sample of heterogeneous material Square cavity with free fluid in the middle part of the domain. The porosity of the material is taken from experimental data1 Left wall kept heated, right wall cooled Numerical solution procedure for the deterministic procedure is a fractional time stepping method 1. Reconstruction of random media using Monte Carlo methods, Manwat and Hilfer, Physical Review E. 59 (1999) CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory FLOW THROUGH HETEROGENEOUS RANDOM MEDIA Experimental correlation for the porosity of the sandstone. Material: Sandstone Eigen spectrum is peaked. Requires large dimensions to accurately represent the stochastic space Simulated with N= 8 Number of collocation points is 11561 (level 4 interpolation) 15 1 0.9 0.8 0.7 0.6 0.5 Eigen spectrum 0.4 0.3 Eigenvalue 10 Numerically computed 5 0.2 0.1 0 0 10 20 30 40 50 60 0 5 10 15 20 Index CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory FLOW THROUGH HETEROGENEOUS RANDOM MEDIA Temperature Snapshots at a few collocation points Temperature 9.1 6.4 3.7 1.0 -1.7 -4.4 -7.1 -9.8 0.9 0.8 0.7 0.6 0.4 0.3 0.2 0.1 14.2 10.1 6.1 2.0 -2.0 -6.1 -10.1 -14.1 CORNELL U N I V E R S I T Y 12.1 8.6 5.1 1.7 -1.8 -5.3 -8.7 -12.2 7.0 4.4 1.8 -0.8 -3.4 -6.0 -8.6 -11.2 Streamlines Temperature 0.9 0.8 0.7 0.6 0.4 0.3 0.2 0.1 Y velocity 0.94 0.81 0.69 0.56 0.44 0.31 0.19 0.06 y-Velocity 0.9 0.8 0.7 0.6 0.4 0.3 0.2 0.1 FIRST MOMENT SECOND MOMENT 0.097 0.084 0.071 0.058 0.045 0.032 0.019 0.006 Y velocity 5.056 4.382 3.708 3.034 2.359 1.685 1.011 0.337 Materials Process Design and Control Laboratory USING THE COLLOCATION METHOD FOR HIGHER DIMENSIONS 2. Flow over rough surfaces Thermal transport across rough surfaces, heat exchangers Look at natural convection through a realistic roughness profile Rectangular cavity filled with fluid. Lower surface is rough. Roughness auto correlation function from experimental data2 T (y) = -0.5 Lower surface maintained at a higher temperature Rayleigh-Benard instability causes convection T (y) = 0.5 y = f(x,ω) Numerical solution procedure for the deterministic procedure is a fractional time stepping method 2. H. Li, K. E. Torrance, An experimental study of the correlation between surface roughness and light scattering for rough metallic surfaces, Advanced Characterization Techniques for Optics, Semiconductors, and Nanotechnologies II, CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory NATURAL CONVECTION ON ROUGH SURFACES Experimental ACF Experimental correlation for the surface roughness 1 Eigen spectrum is peaked. Requires large dimensions to accurately represent the stochastic space 0.8 0.6 V2 Simulated with N= 20 (Represents 94% of the spectrum) 0.4 0.2 Number of collocation points is 11561 (level 4 interpolation) 0.44 0.44 0.31 0.31 0.19 0.19 0.06 0.06 -0.06 -0.06 -0.19 -0.19 -0.31 -0.31 -0.44 -0.44 0.44 0.31 0.19 0.06 -0.06 -0.19 -0.31 -0.44 CORNELL U N I V E R S I T Y 0 0.5 1 1.5 2 V1 Numerically computed Eigen spectrum Sample realizations of temperature at collocation points 16 12 Eigenvalue 0.44 0.31 0.19 0.06 -0.06 -0.19 -0.31 -0.44 0 8 4 0 5 10 15 20 Index Materials Process Design and Control Laboratory NATURAL CONVECTION ON ROUGH SURFACES FIRST MOMENT SECOND MOMENT Temperature Temperature 0.17 0.14 0.12 0.10 0.08 0.06 0.03 0.01 0.44 0.31 0.19 0.06 -0.06 -0.19 -0.31 -0.44 Streamlines Y Velocity 7.63 6.62 5.60 4.58 3.56 2.54 1.53 0.51 Roughness causes improved thermal transport due to enhanced nonlinearities Results in thermal plumes Can look to tailor material surfaces to achieve specific thermal transport CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory Statistical characterization of microstructures Can we compute statistical response of a class of microstructures subjected to applied loads based on limited experimental information? Features of a microstructure Grain size (in 3D, grain volume) When a specimen is manufactured, the microstructures at a sample point will not be the same always. Orientation Distribution Function Rodrigues’ representation FCC fundamental region CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory Technique employed Maximum entropy (MAXENT): The probability distribution that maximizes entropy and satisfies the given (experimental/simulation-based) information is the least-biased estimate that can be made. Problem formulation Numerical implementation Limited microstructures computed using phase field simulations We employ the voronoi cell tessellation technique for representing microstructures. Extract features of the microstructure Geometrical: grain size Texture: ODFs Conjugate Compute a PDF of microstructures Entropy gradient Compute bounds on macroscopic properties microstructure feature constraints Meshing a statistical class of microstructures using CUBIT CORNELL U N I V E R S I T Y features of microstructure, I Given information about microstructures. We use grain size and texture features Materials Process Design and Control Laboratory Statistical class of 3D Aluminium polycrystals Three statistical Aluminium polycrystal samples generated using phase field simulations Comparison of grain size distributions between a phase field simulation from the representative class and a MaxEnt sample First four statistical moments of grain sizes (volumes) Probability mass function 0.25 Grain volume distribution using phase field simulations pmf reconstructed using MaxEnt 0.2 0.15 0.1 0.05 0 0 CORNELL U N I V E R S I T Y 2000 4000 6000 8000 100001200014000160001800020000 Grain volume (voxels) Materials Process Design and Control Laboratory ODF reconstruction using MAXENT Input ODF 0.35 Grain size distribution of a microstructural sample. Comparison with the MaxEnt distribution Probability mass function 0.3 0.25 Reconstructed samples using MAXENT 0.2 0.15 Represent ation in FrankRodrigues space Rcorr=0.9644 KL=0.0383 0.1 0.05 0 0 5000 10000 15000 Grain volume (voxels) 20000 25000 A microstructural specimen computed from the MaxEnt distribution CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory Statistical variation of properties Homogenization scheme: First order stress averaging Scheme employing Hill’s criterion 60 How to mesh a microstructure? We employ hexahedral elements using Cubit software Mean std Equivalent stress (MPa) 50 Mean stress-strain curve 40 30 Aluminium polycrystal with rate-independent strain hardening. Pure tensile test. 20 10 0 0 1 2 Equivalent strain CORNELL U N I V E R S I T Y 3 Statistical variation of homogenized stressstrain curves. -4 x 10 Materials Process Design and Control Laboratory HETEROGENEOUS DIFFUSION a) Two phase materials b) Micro-emulsions, c) porous media, d) ceramics e) Polycrystals f) Foams, blends - To apply physical processes on these heterogeneous systems - worst case scenarios - variations on physical properties Different morphology, anisotropy Aim: To develop a procedure to predict statistics of properties of heterogeneous materials undergoing certain phenomena CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory OVERVIEW OF METHODOLOGY Given certain properties P1, P2, .. Pn, that the structure satisfies. STEP 1 These properties are usually statistical: Volume fraction, 2 Point correlation, auto correlation Reconstruct realizations of the structure satisfying the properties. Monte Carlo, Gaussian Random Fields, Stochastic optimization ect STEP 2 Solve the heterogeneous property problem in the reduced stochastic space for computing property variations. Collocation schemes CORNELL U N I V E R S I T Y STEP 3 Construct a reduced stochastic model from the data. This model must be able to approximate the class of structures. KL expansions, FFT and other transforms, Autoregressive models, ARMA models Materials Process Design and Control Laboratory EXAMPLE: THERMAL DIFFUSION THROUGH TUNGSTEN—SILVER MATRIX MC-Potts model, generate microstructures database. Apply the KL transform Z First 9 eigen values are enough X Tungsten-silver composite image1 15 Z 10 5 0 0 Y 1.5807E-02 1.4753E-02 1.3699E-02 1.2645E-02 1.1592E-02 1.0538E-02 9.4840E-03 8.4303E-03 7.3765E-03 6.3227E-03 5.2689E-03 4.2151E-03 3.1613E-03 2.1076E-03 1.0538E-03 0 5 5 Y 10 10 15 X 15 1. S. Umekawa, R. Kotfila and O.D. Sherby, Elastic properties of a tungsten-silver composite above and below the melting point of silver, J. Mech. Phys. Solids 13 (1965) CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory REDUCED MODEL FOR THE STRUCTURE Represent any microstructure as a linear combination of the eigen-images I = Iavg + I1a1 + I2a2+ I3a3 + … + Inan =a1 + ..+ an +a2 Image I belongs to the class of structures? It must satisfy certain conditions a) Its volume fraction must equal the specified volume fraction b) Volume fraction at every pixel must be between 0 and 1 c) It should satisfy higher order statistics Thus the n tuple (a1,a2,..,an) must further satisfy some constraints. CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory REDUCED MODEL FOR THE STRUCTURE Constraints on the coefficients 15 a v 10 i i v i 5 a I ( j ) 1, j 1: NPixels a I ( j ) 0, j 1: NPixels i i 0 i -5 i i i -10 10 15 15 10 20 -15 -10 -5 0 5 10 15 5 Construct the Convex Hull of the set of linear inequalities. This is the allowable set of coefficients. This represents the space of allowable microstructures 0 -5 In this space all the structures are equiprobable. This represents a stochastic space in (n-1) dimensions. Actually a plane in n dimensions, Call this the ‘material plane’ CORNELL U N I V E R S I T Y -10 -15 11 11.5 12 12.5 13 13.5 14 14.5 15 15.5 16 Materials Process Design and Control Laboratory PHYSICAL PROBLEM UNDER CONSIDERATION Structure size 20x20x20 μm Tungsten Silver Matrix T= -0.5 T= 0.5 Heterogeneous property is the thermal diffusivity. Tungsten: ρ 19250 kg/m3 k 174 W/mK c 130 J/kgK Left wall maintained at -0.5 Silver: ρ 10490 kg/m3 Right wall maintained at +0.5 k 430 W/mK All other surfaces insulated c 235 J/kgK Diffusivity ratio αAg/αW = 2.5 CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory COLLOCATION SCHEME: SAMPLE REALIZATIONS First column: conductivity Second column: Temperature 15 10 5 0 -5 -10 -15 11 CORNELL U N I V E R S I T Y 11.5 12 12.5 13 13.5 14 14.5 15 15.5 16 Materials Process Design and Control Laboratory MEAN STATISTICS Temperature isosurfaces Mean temperature: No variations closer to the surfaces, significant variations inside Mean distribution of silver CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory SECOND ORDER STATISTICS Temperature slice Property slice Left, isosurface of temperature deviation 0 0 5 5 y 10 10 15 15 CORNELL U N I V E R S I T Y x 0 0 5 5 y 10 10 15 Right, isosurface of properties x 15 Materials Process Design and Control Laboratory HIGHER ORDER STATISTICS CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory Future research directions Algorithms to address the curse-of-dimensionality Adaptivity in the support space, adaptive sparse-grid quadrature rules, SPDE model reduction, etc. Stochastic multiscale advection-diffusion-reaction Stochastic multiscale modeling in materials Information-theoretic algorithms for coupling statistics across length scales Robust design techniques Interface stochastic and statistical (Bayesian) computation CORNELL U N I V E R S I T Y Materials Process Design and Control Laboratory