Uncertainty Quantification: UQTk example problems...2012/04/04 · Uncertainty Quantification:...
Transcript of Uncertainty Quantification: UQTk example problems...2012/04/04 · Uncertainty Quantification:...
Uncertainty Quantification: UQTk example problems
Bert Debusschere1 Roger Ghanem2
1Sandia National LaboratoriesLivermore, CA, [email protected]
2University of Southern CaliforniaLos Angeles, CA, USA
SIAM 2012 Conference on Uncertainty Quantification
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 1 / 81
Acknowledgement
K. Sargsyan, C. Safta, H.N. Najm— Sandia National Laboratories, CA
O.M. Knio — Duke Univ., Durham, NCA. Alexanderian — Johns Hopkins Univ., Baltimore, MD
This work was supported by:
US Department of Energy (DOE), Office of Advanced Scientific Computing Research(ASCR), Scientific Discovery through Advanced Computing (SciDAC)
DOE ASCR Applied Mathematics program.
Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly ownedsubsidiary of Lockheed Martin Corporation, for the U.S. Department of Energys National Nuclear Security Administration undercontract DE-AC04-94AL85000.
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 2 / 81
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 3 / 81
Introduction
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 4 / 81
Introduction
Uncertainty Quantification Toolkit (UQTk)
A library of C++ and Matlab functions for propagation of uncertaintythrough computational models
Mainly relies on spectral Polynomial Chaos Expansions (PCEs) forrepresenting random variables and stochastic processes
Target usage:
Rapid prototypingAlgorithmic researchTutorials / Educational
Version 1.0 released under the GNU Lesser General Public License
Downloadable from http://www.sandia.gov/UQToolkit/
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 5 / 81
Introduction
UQTk contents and development/release plans
Currently released (http://www.sandia.gov/UQToolkit/)
C++ Tools for intrusive UQ with PCEs
Under production, planned release Fall 2012
C++ Tools for non-intrusive UQMatlab tools for intrusive and non-intrusive UQKarhunen-Loeve decompositionBayesian inference toolsMany more examples and documentation
Under development
Adding support for multiwavelet based stochastic domaindecompositionsSupport for arbitrary basis types
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 6 / 81
Polynomial Chaos
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 7 / 81
Polynomial Chaos
Polynomial Chaos Expansions (PCEs)
Represent random variables with finite variance as polynomials ofstandard random variables
U(θ) ≃P∑
k=0
ukΨk(ξ)
Truncated to finite dimension n and order p
Numer of terms P + 1 = (n+p)!n!p!
ξ = (ξ1, ξ2, · · · , ξn) standard i.i.d. random variablesΨk standard orthogonal polynomialsuk spectral modes or PC coefficients
Most common standard Polynomial-Variable pairs:
(continuous) Gauss-Hermite, Legendre-Uniform(discrete) Poisson-Charlier
[Wiener, 1938; Ghanem & Spanos, 1991; Xiu & Karniadakis, 2002]
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 8 / 81
Polynomial Chaos
PCEs in the UQToolkit
// Initialize PC class
int ord = 5; // Order of PCE
int dim = 1; // Number of uncertain parameters
PCSet myPCSet("ISP",ord,dim,"LU"); // Legendre-Uniform PCEs
// Initialize PC class
int ord = 5; // Order of PCE
int dim = 1; // Number of uncertain parameters
PCSet myPCSet("NISP",ord,dim,"LU"); // Legendre-Uniform PCEs
Currently support Wiener-Hermite, Legendre-Uniform, andGamma-Laguerre (limited), Jacobi-Beta (development version)
PCSet class initializes PC basis type and pre-computes informationneeded for working with PC expansions
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 9 / 81
Polynomial Chaos
Operations on PCEs in the UQToolkit
// PC coefficients in double*
double* a = new double[npc];
double* b = new double[npc];
double* c = new double[npc];
// Initialization
a[0] = 2.0;
a[1] = 0.1;
...
// Perform some arithmetic
myPCSet.Subtract(a,b,c);
myPCSet.Prod(a,b,c);
myPCSet.Exp(a,c);
myPCSet.Log(a,c);
// PC coefficients in Arrays
Array1D<double> aa(npc,0.e0);
Array1D<double> ab(npc,0.e0);
Array1D<double> ac(npc,0.e0);
// Initialization
aa(0) = 2.0;
aa(1) = 0.1;
...
// Perform arithmetic
myPCSet.Subtract(aa,ab,ac);
myPCSet.Prod(aa,ab,ac);
myPCSet.Exp(aa,ac);
myPCSet.Log(aa,ac);
PC coefficients are either stored in double* vectors or in moreadvanced custom Array1D<double> classes
Functions can take either data type as argument
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 10 / 81
Propagation of Uncertainty
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 11 / 81
Propagation of Uncertainty
Surface Reaction Model
3 ODEs for a monomer (u),dimer (v), and inert species(w) adsorbing onto a surfaceout of gas phase.
du
dt= az − cu − 4duv
dv
dt= 2bz2 − 4duv
dw
dt= ez − fw
z = 1− u − v − w
u(0) = v(0) = w(0) = 0.0
a = 1.6 b = 20.75 c = 0.04 d = 1.0 e = 0.36 f = 0.016
0 200 400 600 800 1000Time [-]
0.0
0.2
0.4
0.6
0.8
1.0
Species Mass Fractions [-]
u v w
Oscillatory behavior for b ∈ [20.2, 21.2]
(Vigil et al., Phys. Rev. E., 1996; Makeev et al., J. Chem. Phys., 2002)
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 12 / 81
Propagation of Uncertainty
Surface Reaction Model: Intrusive Spectral Propagation(ISP) of Uncertainty
Assume PCE for uncertain parameter b and for the output variables,u, v ,w
Substitute PCEs into the governing equations
Project the governing equations onto the PC basis functions
Multiply with Ψk and take the expectation
Apply pseudo-spectral approximations where necessary
UQTk elementary operations
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 13 / 81
Propagation of Uncertainty
Surface Reaction Model: Specify PCEs for inputs andoutputs
Represent uncertain inputs with PCEs with known coefficients:
b =P∑
i=0
biΨi (ξ)
Represent all uncertain variables with PCEs with unknown coefficients:
u =P∑
i=0
uiΨi (ξ) v =P∑
i=0
viΨi (ξ) w =P∑
i=0
wiΨi (ξ) z =P∑
i=0
ziΨi (ξ)
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 14 / 81
Propagation of Uncertainty
Surface Reaction Model: Substitute PCEs into governingequations and project onto basis functions
du
dt= az − cu − 4duv
d
dt
P∑
i=0
uiΨi = a
P∑
i=0
ziΨi − c
P∑
i=0
uiΨi − 4dP∑
i=0
uiΨi
P∑
j=0
vjΨj
⟨
Ψk
d
dt
P∑
i=0
uiΨi
⟩
=
⟨
aΨk
P∑
i=0
ziΨi
⟩
−⟨
cΨk
P∑
i=0
uiΨi
⟩
−⟨
4dΨk
P∑
i=0
uiΨi
P∑
j=0
vjΨj
⟩
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 15 / 81
Propagation of Uncertainty
Surface Reaction Model: Reorganize terms
d
dtuk⟨Ψ2
k
⟩= azk
⟨Ψ2
k
⟩− cuk
⟨Ψ2
k
⟩− 4d
P∑
i=0
P∑
j=0
uivj 〈ΨiΨjΨk〉
d
dtuk = azk − cuk − 4d
P∑
i=0
P∑
j=0
uivj〈ΨiΨjΨk〉⟨Ψ2
k
⟩
d
dtuk = azk − cuk − 4d
P∑
i=0
P∑
j=0
uivjCijk
Triple products Cijk =〈ΨiΨjΨk〉〈Ψ2
k〉can be pre-computed and stored for
repeated use
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 16 / 81
Propagation of Uncertainty
Surface Reaction Model: Substitute PCEs into governingequations and project onto basis functions
dv
dt= 2bz2 − 4duv
d
dt
P∑
i=0
viΨi = 2P∑
h=0
bhΨh
P∑
i=0
ziΨi
P∑
j=0
zjΨj − 4dP∑
i=0
uiΨi
P∑
j=0
vjΨj
⟨
Ψk
d
dt
P∑
i=0
viΨi
⟩
=
⟨
2Ψk
P∑
h=0
bhΨh
P∑
i=0
ziΨi
P∑
j=0
zjΨj
⟩
−⟨
4dΨk
P∑
i=0
uiΨi
P∑
j=0
vjΨj
⟩
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 17 / 81
Propagation of Uncertainty
Surface Reaction Model: Reorganize terms
d
dtvk⟨Ψ2
k
⟩= 2
P∑
h=0
P∑
i=0
P∑
j=0
bhzizj 〈ΨhΨiΨjΨk〉 − 4d
P∑
i=0
P∑
j=0
uivj 〈ΨiΨjΨk〉
d
dtvk = 2
P∑
h=0
P∑
i=0
P∑
j=0
bhzizj〈ΨhΨiΨjΨk〉
⟨Ψ2
k
⟩ − 4dP∑
i=0
P∑
j=0
uivj〈ΨiΨjΨk〉⟨Ψ2
k
⟩
d
dtvk = 2
P∑
h=0
P∑
i=0
P∑
j=0
bhzizjDhijk − 4dP∑
i=0
P∑
j=0
uivjCijk
Pre-computing and storing the quad product Dhijk becomescumbersome
Use pseudo-spectral approach instead
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 18 / 81
Propagation of Uncertainty
Surface Reaction Model: Pseudo-Spectral approach forproducts
Introduce auxiliary variable g = z2
g = z2
f = 2bz2 = 2bg
gk =
P∑
i=0
P∑
j=0
zizjCijk
fk = 2
P∑
i=0
P∑
j=0
bigjCijk
Limits the complexity of computing product terms
Higher products can be computed by repeated use of the same binaryproduct rule
Does introduce errors if order of PCE is not large enough
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 19 / 81
Propagation of Uncertainty
Surface Reaction Model: UQTk implementation
// Build du/dt = a*z - c*u - 4.0*d*u*v
aPCSet.Multiply(z,a,dummy1); // dummy1 = a*z
aPCSet.Multiply(u,c,dummy2); // dummy2 = c*u
aPCSet.SubtractInPlace(dummy1,dummy2); // dummy1 = a*z - c*u
aPCSet.Prod(u,v,dummy2); // dummy2 = u*v
aPCSet.MultiplyInPlace(dummy2,4.e0*d); // dummy2 = 4.0*d*u*v
aPCSet.Subtract(dummy1,dummy2,dudt); // dudt = a*z - c*u - 4.0*d*u*v
All operations are replaced with their equivalent intrusive UQcounterparts
Results in a set of coupled ODEs for the PC coefficients
u, v ,w , z represent vector of PC coefficients
This set of equations is integrated to get the evolution of the PCcoefficients in time
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 20 / 81
Propagation of Uncertainty
Surface Reaction Model: Second equation implementation
// Build dv/dt = 2.0*b*z*z - 4.0*d*u*v
aPCSet.Prod(z,z,dummy1); // dummy1 = z*z
aPCSet.Prod(dummy1,b,dummy2); // dummy2 = b*z*z
aPCSet.Multiply(dummy2,2.e0,dummy1); // dummy1 = 2.0*b*z*z
aPCSet.Prod(u,v,dummy2); // dummy2 = u*v
aPCSet.MultiplyInPlace(dummy2,4.e0*d); // dummy2 = 4.0*d*u*v
aPCSet.Subtract(dummy1,dummy2,dvdt); // dvdt = 2.0*b*z*z - 4.0*d*u*v
// Build dw/dt = e*z - f*w
aPCSet.Multiply(z,e,dummy1); // dummy1 = e*z
aPCSet.Multiply(w,f,dummy2); // dummy2 = f*w
aPCSet.Subtract(dummy1,dummy2,dwdt); // dwdt = e*z - f*w
Dummy variables used where needed to build the terms in theequationsData structure is currently being enhanced to provide the operationresult as the function return value
Will allow more elegant inline replacement of operators with theirstochastic counterparts
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 21 / 81
Propagation of Uncertainty
Surface Reaction Model: ISP results
0 200 400 600 800 1000Time [-]
0.0
0.2
0.4
0.6
0.8
1.0
Species Mass Fractions [-]
u v w
Assume 0.5% uncertainty in b around nominal value
Legendre-Uniform intrusive PC
Mean and standard deviation for u, v , and w
Uncertainty grows in time
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 22 / 81
Propagation of Uncertainty
Surface Reaction Model: ISP results
0 200 400 600 800 1000Time [-]
−0.1
0.0
0.1
0.2
0.3
0.4
0.5ui
[-]
u0
u1
u2
u3
u4
u5
Modes of u
Modes decay with higher order
Amplitudes of oscillations of higher order modes grow in time
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 23 / 81
Propagation of Uncertainty
Surface Reaction Model: ISP results: PDFs
0.310 0.315 0.320 0.325 0.330 0.335 0.340 0.345 0.350u
0
50
100
150
200
250
300
350
Prob. Dens. [-]
t = 330.5t = 756.5
0.10 0.15 0.20 0.25 0.30 0.35u
0
2
4
6
8
10
12
14
16
Prob. Dens. [-]
t = 377.8t = 803.0
Pdfs of u at maximum mean (left) and maximum standard deviation(right)
Distributions get broader and multimodal as time increases
Effect of accumulating phase errors
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 24 / 81
Propagation of Uncertainty
Surface Reaction Model: Non-Intrusive Propagation ofUncertainty
Consider two approaches for Non-Intrusive Spectral Projection(NISP):
QuadratureMonte Carlo or random-sampling
Assume PCE for uncertain parameter b and the output variables
Sample input parameters at quadratrure/random sample points
Run deterministic forward model for each of the sampled inputparameters
Perform Galerkin projection uk = 〈uΨk 〉
〈Ψ2k〉
Quadrature: 〈uΨk〉 =∑Nq
i=1 wiu(bi )
Monte Carlo: 〈uΨk〉 = 1Ns
∑Ns
i=1 u(bi )
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 25 / 81
Propagation of Uncertainty
Surface Reaction Model: NISP implementation in UQTk
Quadrature:
// Get the quadrature points
int nQdpts=myPCSet.GetNQuadPoints();
double* qdpts=new double[nQdpts];
myPCSet.GetQuadPoints(qdpts);
...
// Evaluate parameter at quad pts
for(int i=0;i<nQdpts;i++){
bval[i]=myPCSet.EvalPC(b,&qdpts[i]);
}
...
// Run model for all samples
for(int i=0;i<nQdpts;i++){
u_val[i] = ...
}
// Spectral projection
myPCSet.GalerkProjection(u_val,u);
myPCSet.GalerkProjection(v_val,v);
myPCSet.GalerkProjection(w_val,w);
Monte-Carlo Sampling:
// Get the sample points
int nSamples=1000;
Array2D<double> samPts(nSamples,dim);
myPCSet.DrawSampleVar(samPts);
...
// Evaluate parameter at sample pts
for(int i=0;i<nSamples;i++){
... // select samPt from samPts
bval[i]=myPCSet.EvalPC(b,&samPt)
}
...
// Run model for all samples
for(int i=0;i<nSamples;i++){
u_val[i] = ...
}
// Spectral projection
myPCSet.GalerkProjectionMC(samPts,u_val,u);
myPCSet.GalerkProjectionMC(samPts,v_val,v);
myPCSet.GalerkProjectionMC(samPts,w_val,w);
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 26 / 81
Propagation of Uncertainty
Surface Reaction Model: NISP results
0 200 400 600 800 1000Time [-]
0.0
0.2
0.4
0.6
0.8
1.0
Species Mass Fractions [-]
NISP Quadrature
u v w
0 200 400 600 800 1000Time [-]
0.0
0.2
0.4
0.6
0.8
1.0
Species Mass Fractions [-]
NISP MC
u v w
Mean and standard deviation for u, v , and w
Quadrature approach agrees well with ISP approach using 6quadrature points
Monte Carlo sampling approach converges slowly
With a 1000 samples, results are quite different from ISP and NISP
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 27 / 81
Propagation of Uncertainty
Surface Reaction Model: Comparison ISP and NISP
0 200 400 600 800 1000Time [-]
−0.010
−0.008
−0.006
−0.004
−0.002
0.000
0.002
0.004ui
[-]
u4,ISP
u4,NISP
u5,ISP
u5,NISP
Lower order modes agree perfectly
Very small differences in higher order modes
Difference increases with time
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 28 / 81
Propagation of Uncertainty
Surface Reaction Model: Comparison ISP and NISP
0.10 0.15 0.20 0.25 0.30 0.35u
0
2
4
6
8
10
Prob. Dens. [-]
ISP, t = 803.0NISP, t = 803.0
All pdf’s based on 50K samples each and evaluated with KernelDensity Estimation (KDE)
No difference in PDFs of sampled PCEs between NISP and ISP
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 29 / 81
Propagation of Uncertainty
Surface Reaction Model: Comparison ISP, NISP, and MC
0.10 0.15 0.20 0.25 0.30 0.35u
0
2
4
6
8
10
Prob. Dens. [-]
ISP, t = 803.0NISP, t = 803.0MC, t = 803.0
All pdf’s based on 50K samples each and evaluated with KernelDensity Estimation (KDE)
Good agreement between intrusive, non-intrusive projection, andMonte Carlo sampling
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 30 / 81
Propagation of Uncertainty
ISP pros and cons
Pros:
ElegantOne time solution of system of equations for the PC coefficients fullycharacterizes uncertainty in all variables at all timesTailored solvers can (potentially) take advantage of new hardwaredevelopments
Cons:
Often requires re-write of the original codeReformulated system is factor (P+1) larger than the original systemand can be challenging to solveChallenges with increasing time-horizon for ODEs
Many efforts in the community to automate ISP
UQToolkit http://www.sandia.gov/UQToolkit/Sundance http://www.math.ttu.edu/ klong/Sundance/html/Stokhos http://trilinos.sandia.gov/packages/stokhos/...
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 31 / 81
Propagation of Uncertainty
NISP pros and cons
Pros:
Easy to use as wrappers around existing codesEmbarassingly parallel
Cons:
Most methods suffer from curse of dimensionality Nq = nNd
Many develoment efforts for smarter sampling approaches anddimensionality reduction
(Adaptive) Sparse Quadrature approachesCompressive Sensing...
Sampling methods have found very wide spread use in the community
DAKOTA http://dakota.sandia.gov/...
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 32 / 81
Uncertainty Characterization
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 33 / 81
Uncertainty Characterization
Uncertain Input Characterization
Use standard distribution
Normal distribution: often a good choice based on Central LimitTheoremLognormal: when positivity is required
Infer model parameters from data with inverse problem
Determine RV from available samples of RV
Fit standard distribution to data, e.g. MultiVariate Normal (MVN)approximationInverse Cumulative Distribution Function (CDF) mapping; Rosenblatttransformation
Dimensionality reduction for stochastic processes, Karhunen-LoeveExpansion (KLE)
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 34 / 81
Uncertainty Characterization
Bayesian Inference
Bayes formulaPosterior
︷ ︸︸ ︷
P(c|D) ∝Likelihood
︷ ︸︸ ︷
P(D|c)Prior
︷︸︸︷
P(c)
Update prior distribution/knowledge about parameter c to posteriordistribution given data D, using likelihood function L(c) ≡ P(D|c).Data D = {di}Ni=1 - measurements of some quantities of interest (QoIs).
Prior distribution P(c) is based on expert opinion/previous literature.
Likelihood function measures goodness-of-fit and is the key component thatconnects the model inputs to measured QoIs, e.g.
L(c) = P(D|c) = 1√2πσ
exp
(
−N∑
i=1
(di − fi (c))2
2σ2
)
Input parameter → output QoI functions fi (·) could be expensive or noteven available.
Usually posterior distribution is not analytically tractable:need to resort to Markov Chain Monte Carlo sampling.
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 35 / 81
Uncertainty Characterization
Markov Chain Monte Carlo
Single-site MCMC
• Set the current chain state c at an initial chain state c(0),
• Repeat for a predefined number (NMCMC ) of times,
• For k = 1, . . . ,K ,
• generate a single-site proposal c ′k from a Gaussiandistribution centered at the current chain statevalue of site ck with proposal width σk ,
• compute α = min {1,P(c′|D)/P(c|D)},• update the current chain state’s k-th element
ck = c ′k with probability α,
• End
• End
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 36 / 81
Uncertainty Characterization
Markov Chain Monte Carlo
Adaptive MCMC [Haario,2002]
• Set the current chain state c at an initial chain state c(0),
• Repeat for a predefined number (NMCMC ) of times,
• generate a proposal c′ from a multivariate Gaussiandistribution centered at the current chain state value c withproposal covariance that is learnt from previous chain states,
• compute α = min {1,P(c′|D)/P(c|D)},• update the current chain state c = c
′ with probability α,
• End
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 36 / 81
Uncertainty Characterization
Surface Reaction Model: parameter inference
500 600 700 800 900 1000
Time
0
0.1
0.2
0.3
0.4
Spec
ies
u
0.26 0.28 0.3 0.32 0.34Species u
0
0.02
0.04
0.06
0.08
0.1
Spec
ies
v
Synthetic data is generated for model outputs u,v at T = 1000.
Inferring two input parameters a and b using 10 samples on u and v
Likelihood function
P(D|a, b) = 1
2πσuσv
exp
(
−N∑
i=1
((d
(i)u − u(T ; a, b))2
2σ2u
+(d
(i)v − v(T ; a, b))2
2σ2v
)
)
Uncorrelated Gaussian noise model is assumed with standard deviationsproportional to the model value σu = 0.1u, σv = 0.1v .
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 37 / 81
Uncertainty Characterization
Surface Reaction Model: parameter inference
500 600 700 800 900 1000
Time
0
0.01
0.02
0.03
0.04
0.05
Spec
ies
v
0.26 0.28 0.3 0.32 0.34Species u
0
0.02
0.04
0.06
0.08
0.1
Spec
ies
v
Synthetic data is generated for model outputs u,v at T = 1000.
Inferring two input parameters a and b using 10 samples on u and v
Likelihood function
P(D|a, b) = 1
2πσuσv
exp
(
−N∑
i=1
((d
(i)u − u(T ; a, b))2
2σ2u
+(d
(i)v − v(T ; a, b))2
2σ2v
)
)
Uncorrelated Gaussian noise model is assumed with standard deviationsproportional to the model value σu = 0.1u, σv = 0.1v .
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 37 / 81
Uncertainty Characterization
Surface Reaction Model: parameter inference
10000 12000 14000 16000 18000 20000MCMC sample count
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Para
met
er a
10000 12000 14000 16000 18000 20000MCMC sample count
0
5
10
15
20
25
Para
met
er b
Posterior distribution based on 20000 adaptive MCMC samples.
First 4000 samples discarded
Shown only the second half of the chains
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 38 / 81
Uncertainty Characterization
Posterior on inferred parameters
0.6 0.8 1 1.2 1.4 1.6Parameter a
5
10
15
20
Para
met
er b
Width of posterior indicates the amount of uncertainty in the inferredparameters
Uncertainty can be reduced by taking more data or by reducing noisein measurements
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 39 / 81
Uncertainty Characterization
Surface Reaction Model: posterior to PC in 1d
20.73 20.74 20.75 20.76 20.77 20.78Parameter b
0
50
100
Posterior PDFPosterior PDF (PC)
The posterior describes random variable a with CDF F (·)CDF transformation F (a) = η maps random variable a touniform[0, 1] random variable η.η = Φ(ξ) maps uniform η to normal RV ξThe inverse CDF enables NISP projection
a =P∑
k=0
akΨk(ξ) ak ∝ 〈aΨk(ξ)〉 =∫
F−1(Φ(ξ))︸ ︷︷ ︸
a
Ψk(ξ)dξ
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 40 / 81
Uncertainty Characterization
Surface Reaction Model: posterior to PC in multi-d
Rosenblatt transformation maps any (not necessarily independent) setof random variables (λ1, . . . , λn) to uniform i.i.d.’s {ηi}ni=1
(Rosenblatt, 1952).
η1 = F1(λ1)
η2 = F2|1(λ2|λ1)
...
ηn = Fn|n−1,...,1(λn|λn−1, . . . , λ1)
Rosenblatt transformation is a multi-D generalization of 1D CDFmapping.
Conditional CDFs are harder to evaluate in high dimensions
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 41 / 81
Uncertainty Characterization
Projection of Rosenblatt transformed vars onto PCEs
0 0.2 0.4 0.6 0.8 1Parameter η1
0
0.2
0.4
0.6
0.8
1
Para
met
er
η 2
0.6 0.8 1 1.2 1.4 1.6Parameter a
5
10
15
20
Para
met
er b
NISP projection is enabled by inverse Rosenblatt transformation(a, b) = R
−1(ξ1, ξ2) ensures a well-defined quadrature integration
a =
P∑
k=0
akΨk(ξ) ak ∝
∫
R−1a (ξ)
︸ ︷︷ ︸
a
Ψk(ξ)dξ
b =
P∑
k=0
bkΨk(ξ) bk ∝
∫
R−1b (ξ)
︸ ︷︷ ︸
b
Ψk(ξ)dξ
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 42 / 81
Uncertainty Characterization
Surface Reaction Model: predictive confidence
0 200 400 600 800 1000Time
0
0.1
0.2
0.3
0.4
Spec
ies
u
0 200 400 600 800 1000Time
0
0.2
0.4
0.6
0.8
Spec
ies
v
Uncertainty in inferred input parameters a and b is pushed throughthe forward model
Using 2D Wiener Hermite NISP Quadrature approach
Accounts for parametric uncertainty due to data noise
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 43 / 81
Uncertainty Characterization
Multivariate Normal Approximation
Many distributions are unimodal and somewhat shaped like Gaussians
MultiVariate Normal (MVN) approximations capture the mean andcorrelation structure of the random variables
Easy to extract from a set of samples
In 1D: just compute mean and standard deviation: u = u0 + u1ξMulti-D: Cholesky factorization of covariance
# Compute mean parameter values
par_mean = numpy.mean(samples,axis=0)
# Compute the covariance
par_cov = numpy.cov(samples,rowvar=0)
# Compute the Cholesky Decomposition
chol_lower = numpy.linalg.cholesky(par_cov)
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 44 / 81
Uncertainty Characterization
MVN approximation of Bayesian posterior from MCMCsamples
1.30 1.32 1.34 1.36 1.38 1.40
S15200
5250
5300
5350
5400
CS
0.000
0.075
0.150
0.225
0.300
0.375
0.450
0.525
0.600
1.31 1.32 1.33 1.34 1.35 1.36 1.37 1.38 1.39
S1
5240
5260
5280
5300
5320
5340
5360
5380
5400
CS
Comparison of Posterior (blue) with MVN (red)
S1 = 1.351 + 0.01367ξ1
CS = 5310− 26.25ξ1 + 20.26ξ2
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 45 / 81
Uncertainty Characterization
Karhunen-Loeve (KL) Expansions
Assume stochastic process F (x , θ)
With covariance function Cov(x1, x2)
F can be written as
F (x , θ) = 〈F (x , θ)〉θ +∞∑
k=1
√
λkFk(x)ξk
Fk(x): eigenfunctions of Cov(x1, x2)
λk : corresponding eigenvalues, all positive
ξk : uncorrelated random variables, unit variance
Samples are obtained by projecting realizations of F onto Fk
Generally not independent
Special case: for Gaussian F , ξk are i.i.d. normal random variables
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 46 / 81
Uncertainty Characterization
1D Gaussian Process: Realizations
δ = 0.02
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f(x)
δ = 0.1
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f(x)
Covariance Cov(x1, x2) = exp(−(x1 − x2)2/δ2)
Sample realizations are noisier as correlation length decreases
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 47 / 81
Uncertainty Characterization
1D Gaussian Process: KL modes
δ = 0.02
0.0 0.2 0.4 0.6 0.8 1.0x
−3
−2
−1
0
1
2
3
f n
f1
f2
f3
f4
δ = 0.1
0.0 0.2 0.4 0.6 0.8 1.0x
−4
−3
−2
−1
0
1
2
3
4
5
f n
f1
f2
f3
f4
Eigenmodes of the covariance matrix
Data covariance matrix constructed from 512 Gaussian processrealizations
Higher modes are more oscillatory
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 48 / 81
Uncertainty Characterization
1D Gaussian Process: KL random variables
δ = 0.02
−4 −3 −2 −1 0 1 2 3 4ξ(θ)
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
PDF(ξ)
ξ1
ξ2
ξ3
ξ4
δ = 0.1
−4 −3 −2 −1 0 1 2 3 4ξ(θ)
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
PDF(ξ)
ξ1
ξ2
ξ3
ξ4
Random variables obtained by projecting realizations onto KL modes
Uncorrelated by construction
Also independent due to nature of Gaussian Process
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 49 / 81
Uncertainty Characterization
1D Gaussian Process: Eigenvalue spectrum
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
Eigenvalue spectrum decays more slowly as correlation lengthdecreases
More oscillatory modes needed to represent fluctuations in x
KL expansion generally is truncated after enough modes are includedto capture a specified fraction of the total variance
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 50 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
2 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
4 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
6 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
8 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
10 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
14 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
16 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
18 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
1 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
2 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
3 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
4 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
5 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
6 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
7 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
8 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
1D Gaussian Process: Reconstructed realizations
0 10 20 30 40Eigenvalue #
10-4
10-2
100
102
Eigenvalue M
agnitude
δ=0.02
δ=0.1
δ=0.2
δ=0.5
0.0 0.2 0.4 0.6 0.8 1.0x
−20
−10
0
10
20
f n
9 terms
Large scale features can be resolved with small number of modes
Smaller scale features require higher modes
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 51 / 81
Uncertainty Characterization
KL of 2D Gaussian Process
δ = 0.1
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
δ = 0.2
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
δ = 0.5
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
2D Gaussian Process with covariance:Cov(x1, x2) = exp(−||x1 − x2||2/δ2)Realizations are smoother as covariance length δ increases
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 52 / 81
Uncertainty Characterization
2D KL - Modes for δ = 0.1
√λ1f1
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ5f5
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ2f2
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ6f6
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ3f3
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ7f7
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ4f4
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ8f8
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
yDebusschere (Sandia, USC) UQTk examples SIAM UQ 2012 53 / 81
Uncertainty Characterization
2D KL - Modes for δ = 0.2
√λ1f1
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ5f5
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ2f2
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ6f6
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ3f3
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ7f7
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ4f4
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ8f8
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
yDebusschere (Sandia, USC) UQTk examples SIAM UQ 2012 54 / 81
Uncertainty Characterization
2D KL - Modes for δ = 0.5
√λ1f1
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ5f5
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ2f2
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ6f6
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ3f3
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ7f7
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ4f4
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
√λ8f8
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
yDebusschere (Sandia, USC) UQTk examples SIAM UQ 2012 55 / 81
Uncertainty Characterization
2D KL - eigenvalue spectrum
0 20 40 60Eigenvalue #
10-2
100
102
104
Eigenvalue M
agnitude
δ=0.1
δ=0.2
δ=0.5
δ = 0.14 terms 16 terms
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
32 terms 64 terms
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 56 / 81
Uncertainty Characterization
2D KL - eigenvalue spectrum
0 20 40 60Eigenvalue #
10-2
100
102
104
Eigenvalue M
agnitude
δ=0.1
δ=0.2
δ=0.5
δ = 0.24 terms 16 terms
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
32 terms 64 terms
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 57 / 81
Uncertainty Characterization
2D KL - eigenvalue spectrum
0 20 40 60Eigenvalue #
10-2
100
102
104
Eigenvalue M
agnitude
δ=0.1
δ=0.2
δ=0.5
δ = 0.54 terms 16 terms
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
32 terms 64 terms
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0y
0.0 0.2 0.4 0.6 0.8 1.0x
0.0
0.2
0.4
0.6
0.8
1.0
y
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 58 / 81
Case Study 1
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 59 / 81
Case Study 1
Uncertainty in Model Inputs
Probabilistic UQ requires specification of uncertain inputs
Require joint PDF on input space
PDF can be found given data
Typically such PDFs are not available from the literature
Summary information, e.g. nominals and bounds, is usually available
Uncertainty in computational predictions can depend strongly ondetailed structure of the missing parametric PDF
Need a procedure to reconstruct a PDF consistent with availableinformation in the absence of the raw data
“Data Free” Inference (DFI) (Berry et al., JCP 2012)
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 60 / 81
Case Study 1
The strong role of detailed input PDF structure
Simple nonlinear algebraic model (u, v) = (x2 − y2, 2xy)
Two input PDFs, p(x , y)
same nominals/boundsdifferent correlation structure
Drastically different output PDFs
different nominals and bounds
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 61 / 81
Case Study 1
Generate ignition “data” using a detailed model+noise
Ignition using a detailedchemical model formethane-air chemistry
Ignition time versus InitialTemperature
Multiplicative noise errormodel
11 data points:
di = tGRIig,i (1 + σǫi )
ǫ ∼ N(0, 1)
1000 1100 1200 1300Initial Temperature (K)
0.01
0.1
1
Igni
tion
time
(sec
)
GRI
GRI+noise
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 62 / 81
Case Study 1
Fitting with a simple chemical model
Fit a global single-stepirreversible chemical model
CH4 + 2O2 → CO2 + 2H2O
R = [CH4][O2]kf
kf = A exp(−E/RoT )
Infer 3-D parameter vector(lnA, lnE , lnσ)
Good mixing with adaptiveMCMC when start at MLE
28
30
32
34
36
lnA
10.6
10.8
lnE
0 2000 4000 6000 8000 10000Chain Step
-3-2.5
-2-1.5
-1-0.5
lnσ
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 63 / 81
Case Study 1
Bayesian Inference Posterior and Nominal Prediction
30 31 32 33 34 35
10.6
10.65
10.7
10.75
10.8
10.85
1000 1100 1200 1300Initial Temperature (K)
0.01
0.1
1
Igni
tion
time
(sec
)
GRIGRI+noiseFit Model
GRI
GRI+noise
Marginal joint posterior on(lnA, lnE ) exhibits strongcorrelation
Nominal fit model is consis-tent with the true model
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 64 / 81
Case Study 1
Correlation Slope χ and Chemical Ignition
0 0.5 1Time (sec)
0
0.05
0.1
0.15
0.2
0.25
Mas
s Fr
actio
n
Means
1000
1500
2000
2500
3000
Tem
pera
ture
(K
)CH
4
O2
CO2
H2O
T
0.46 0.462 0.464 0.466 0.468 0.47Time (sec)
0
0.01
0.02
0.03
0.04
0.05
0.06
Mas
s Fr
actio
n
CH4
O2
CO2
H2O
Standard Deviations
0
100
200
300
400
Tem
pera
ture
(K
)
T
4th Order Multiwavelet PC, Multiblock, Adaptive
σT ,max ∼ 400 K during ignition transient, χ ∼ 0.03
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 65 / 81
Case Study 1
Time evolution of Temperature PDFs in preheat stage
1300 1400 1500 1600Temperature (K)
0
0.05
0.1
0.15
Prob
abili
ty D
ensi
ty
MC
t=0.455 sec
0.459 sec
0.462 sec
0.464 sec
1300 1400 1500 1600Temperature (K)
0
0.05
0.1
0.15
Prob
abili
ty D
ensi
ty
MW
t=0.455 sec
0.459 sec
0.462 sec
0.464 sec
Similar results from MC (20K samples) and MW PC
Increased uncertainty, and long high–T PDF tails, in time
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 66 / 81
Case Study 1
Evolution of Temp. PDF – Fast Ignition Transient
0
0.01
0.02
MC
0
0.01
0.02
0
0.01
0.02
0
0.01
0.02
1000 1500 2000 2500 3000Temperature (K)
0
0.01
0.02
Den
sity
t = 0.4642 sec
0.4660 sec
0.4664 sec
0.4668 sec
0.4671 sec
0
0.01
0.02
MW
0
0.01
0.02
0
0.01
0.02
0
0.01
0.02
1000 1500 2000 2500 3000Temperature (K)
0
0.01
0.02D
ensi
ty
t = 0.4642 sec
0.4660 sec
0.4664 sec
0.4668 sec
0.4671 sec
Transition from unimodal to bimodal PDFs
Leakage of probability mass from pre-heat PDF high–T tail
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 67 / 81
Case Study 1
Time evolution of Temperature PDFs for different χ
1500 2000 2500 3000Temperature (K)
0
0.001
0.002
0.003
0.004
0.005
0.006
Prob
abili
ty D
ensi
ty
0.022875
1500 2000 2500 3000Temperature (K)
0
0.002
0.004
0.006
0.008
0.01
Prob
abili
ty D
ensi
ty
0.036675
1500 2000 2500 3000Temperature (K)
0
0.005
0.01
0.015
0.02
0.025
0.03
Prob
abili
ty D
ensi
ty
0.042425
1500 2000 2500 3000Temperature (K)
0
0.01
0.02
0.03
0.04
0.05
0.06
Prob
abili
ty D
ensi
ty
0.04325
1500 2000 2500 3000Temperature (K)
0
0.05
0.1
0.15
0.2
0.25
Prob
abili
ty D
ensi
ty
0.04395
1500 2000 2500 3000Temperature (K)
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Prob
abili
ty D
ensi
ty
0.04475
1500 2000 2500 3000Temperature (K)
0
0.002
0.004
0.006
0.008
0.01
Prob
abili
ty D
ensi
ty
0.052475
1500 2000 2500 3000Temperature (K)
0
0.001
0.002
0.003
0.004
0.005
0.006
Prob
abili
ty D
ensi
ty
0.066275
Bimodal solution PDFs for high uncertainty growth
Unimodal for low uncertainty growth, with χ ≈ 0.044
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 68 / 81
Case Study 1
Central Challenge for UQ in Chemical Kinetic Models
Need joint PDF on model parameters for forward UQ
Joint PDF structure is crucial
Joint PDF not available for chemical kinetic parameters
At best, have
Nominal parameter valuesBounds, e.g. marginal 5%, 95% quantiles
PDF can be constructed by repeating experimentsor access to original raw data
– Neither is feasible
Is there a way to construct an approximate PDF without access toraw data?
– Yes!
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 69 / 81
Case Study 1
Data Free Inference (DFI) (Berry et al., JCP, in review)
Intuition: In the absence of data, the structure of the fit model,combined with the nominals and bounds, implicitly inform thecorrelation between the parameters
Goal: Make this information explicit in the joint PDF
DFI: discover a consensus joint PDF on the parameters consistentwith given information:
– Nominal parameter values– Bounds– The fit model– The data range– ... potentially other/different constraints
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 70 / 81
Case Study 2
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 71 / 81
Case Study 2
Schlogl Model is a prototype bistable model
• ReactionsA+ 2X
a1−→
←−
a2
3X
Ba3−→
←−
a4
X .
• Propensitiesa1 = k1AX (X − 1)/2,a2 = k2X (X − 1)(X − 2)/6,a3 = k3B ,a4 = k4X .
• Nominal parametersk1A 0.03k2 0.0001k3B = λ 200k4 3.5A 105
B 2 · 105X (0) 250
0 5 10 15 200
100
200
300
400
X(t
)
0 5 10 15 200
200
400
600
800
X(t
)0 5 10 15 20
200
300
400
500
600
700
800
X(t
)
λ = 130
λ = 200
λ = 350
Time, t
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 72 / 81
Case Study 2
Polynomial Chaos expansion represents any randomvariable as a polynomial of a standard random variable
• Truncated PCE: finite dimension n and order p
X (θ) ≃P∑
k=0
ckΨk(η)
with the number of terms P + 1 = (n+p)!n!p! .
• η = (η1, · · · , ηn) standard i.i.d. r.v.Ψk standard orthogonal polynomialsck spectral modes.
• Most common standard Polynomial-Variable pairs:(continuous) Gauss-Hermite, Legendre-Uniform,(discrete) Poisson-Charlier.
[Wiener, 1938; Ghanem & Spanos, 1991; Xiu & Karniadakis, 2002]
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 73 / 81
Case Study 2
Polynomial Chaos expansion represents any randomvariable as a polynomial of a standard random variable
• Truncated PCE: finite dimension n and order p
X (θ) ≃P∑
k=0
ckΨk(η)
with the number of terms P + 1 = (n+p)!n!p! .
• η = (η1, · · · , ηn) standard i.i.d. r.v.Ψk standard orthogonal polynomialsck spectral modes.
• Most common standard Polynomial-Variable pairs:(continuous) Gauss-Hermite, Legendre-Uniform,(discrete) Poisson-Charlier.
[Wiener, 1938; Ghanem & Spanos, 1991; Xiu & Karniadakis, 2002]
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 73 / 81
Case Study 2
Polynomial Chaos expansion represents any randomvariable as a polynomial of a standard random variable
• Truncated PCE: finite dimension n and order p
X (θ) ≃P∑
k=0
ckΨk(η)
with the number of terms P + 1 = (n+p)!n!p! .
• η = (η1, · · · , ηn) standard i.i.d. r.v.Ψk standard orthogonal polynomialsck spectral modes.
• Most common standard Polynomial-Variable pairs:(continuous) Gauss-Hermite, Legendre-Uniform,(discrete) Poisson-Charlier.
[Wiener, 1938; Ghanem & Spanos, 1991; Xiu & Karniadakis, 2002]
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 73 / 81
Case Study 2
Galerkin Projection is typically needed
PC expansion: X (θ) ≃∑Pk=0 ckΨk(η) = gD(η)
Orthogonal projection: ck =〈X (θ)Ψk (η)〉
〈Ψ2k(η)〉
• Intrusive Spectral Projection (ISP)
⋆ Direct projection of governing equations⋆ Leads to deterministic equations for PC coefficients∗ No explicit governing equation for SRNs
• Non-intrusive Spectral Projection (NISP)
⋆ Sampling based⋆ No explicit evolution equation for X needed∗ Galerkin projection not well-defined for SRNs
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 74 / 81
Case Study 2
Galerkin Projection is typically needed
PC expansion: X (θ) ≃∑Pk=0 ckΨk(η) = gD(η)
Orthogonal projection: ck =〈X (θ)Ψk (η)〉
〈Ψ2k(η)〉
• Intrusive Spectral Projection (ISP)
⋆ Direct projection of governing equations⋆ Leads to deterministic equations for PC coefficients∗ No explicit governing equation for SRNs
• Non-intrusive Spectral Projection (NISP)
⋆ Sampling based⋆ No explicit evolution equation for X needed∗ Galerkin projection not well-defined for SRNs
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 74 / 81
Case Study 2
Galerkin Projection is typically needed
PC expansion: X (θ) ≃∑Pk=0 ckΨk(η) = gD(η)
Orthogonal projection: ck =〈X (θ)Ψk (η)〉
〈Ψ2k(η)〉
• Intrusive Spectral Projection (ISP)
⋆ Direct projection of governing equations⋆ Leads to deterministic equations for PC coefficients∗ No explicit governing equation for SRNs
• Non-intrusive Spectral Projection (NISP)
⋆ Sampling based⋆ No explicit evolution equation for X needed∗ Galerkin projection not well-defined for SRNs
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 74 / 81
Case Study 2
Karhunen-Loeve decomposition reduces stochastic processto a finite number of random variables
• KL decomposition:
X (t, θ) = X (t) +∞∑
n=1
ξn(θ)√
λnfn(t)
• Uncorrelated, zero-mean KL variables:
〈ξn〉 = 0, 〈ξnξm〉 = δnm
• SSA(continuum) ←→ KL(discrete)
X (t)←→ ξ = (ξ1, ξ2, . . . )
0 2 4 6 8 10 12 14 16 18 20
Time, t-300
-200
-100
0
100
KL
mod
es,
λi1/
2 fi(t
)
0 1 2 3 4 5 6 7 8 9n
1000
10000
1e+05
1e+06
Eig
enva
lues
, λ n
-2 -1 0 1 2
ξ1
-4
-2
0
2
ξ 2
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 75 / 81
Case Study 2
K-L decomposition captures each realization
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 76 / 81
Case Study 2
K-L decomposition captures each realization
0 2 4 6 8 10 12 14 16 18 20Time, t
0
100
200
300
400
500
600
700
X(t
) L=10L=30L=60X(t)
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 76 / 81
Case Study 2
PC expansion of a random vector
ξ =P∑
k=0
ckΨk(η)
Galerkin projection
ck =〈ξΨk(η)〉〈Ψ2
k(η)〉
is not well-defined,since ξ and η do not belong to the same stochastic space.
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 77 / 81
Case Study 2
PC expansion of a random vector
ξ =P∑
k=0
ckΨk(η)
Galerkin projection
ck =〈ξΨk(η)〉〈Ψ2
k(η)〉
is not well-defined,since ξ and η do not belong to the same stochastic space.
Need a map ξ ↔ η.
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 77 / 81
Case Study 2
Rosenblatt transformation
• Rosenblatt transformation maps any (not necessarilyindependent) set of random variables (ξ1, . . . , ξn) to uniformi.i.d.’s {ηi}ni=1 (Rosenblatt, 1952).
η1 = F1(ξ1)
η2 = F2|1(ξ2|ξ1)η3 = F3|2,1(ξ3|ξ2, ξ1)...
ηn = Fn|n−1,...,1(ξn|ξn−1, . . . , ξ1)
• Inverse Rosenblatt transformation ξ = R−1(η) ensures awell-defined quadrature integration
〈ξiΨk(η)〉 =∫
R−1(η)iΨk(η)dη
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 78 / 81
Case Study 2
KL+PC+Data Partitioning represent the dynamics of a bimodal process
0 5 10 15 20
Time, t0
100
200
300
400
500
600
700X
KL
PC(t
)
KL-PC representation, 5 KL modes, 3rd PC order
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 79 / 81
Advanced topics
Outline
1 Introduction to the UQ Toolkit (UQTk)
2 Spectral Polynomial Chaos Expansions (PCEs)
3 Propagation of Uncertainty through Computational Models
4 Characterization of Input Uncertainty
5 Case Study 1: Chemical Mechanism and Input Correlations
6 Case Study 2: Representation of Non-Gaussian Process with PCE
7 Advanced Topics
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 80 / 81
Advanced topics
Advanced Topics
Sensitivity analysis
Domain decomposition methods; multiwavelets
Adaptive Sparse Quadrature
Stochastic preconditioning (time rescaling)
Data Free Inference (DFI)
Bayesian Compressive Sensing (BCS)
PCEs with random coefficients versus Gaussian Processes
Model uncertainty, comparison, selection
... Stay Tuned ...
Debusschere (Sandia, USC) UQTk examples SIAM UQ 2012 81 / 81