See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/352993637

Recent developments on surrogate models for stochastic simulators

Presentation · June 2021

DOI: 10.3929/ethz-b-000492981

CITATIONS READS

0 533

1 author:

Bruno Sudret
ETH Zurich
416 PUBLICATIONS 14,762 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Probabilistic Dam Breach Modeling View project

Rice Crop Growth Stage Estimation using High-Frequency PolSAR Data View project

All content following this page was uploaded by Bruno Sudret on 05 July 2021.

The user has requested enhancement of the downloaded file.

Recent Developments on Surrogate Models for Stochastic Simulators

Bruno Sudret

Chair of Risk, Safety and Uncertainty Quantification, ETH Zurich

4th Int. Conf. on Uncertainty Quantification in Computational Sciences and Engineering, (UNCECOMP’2021),
Athens (Greece) – June 30, 2021
How to cite?

This presentation is a plenary lecture given at the 4th International Conference on Uncertainty
Quantification in Computational Sciences and Engineering on June 30, 2021.

How to cite ?
Sudret, B., Recent Developments on Surrogate Models for Stochastic Simulators,
4th International Conference on Uncertainty Quantification in Computational Sciences and
Engineering, streamed from Athens (Greece), June 30th, 2021.

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 1 / 42
The SAMOS project

• The SAMOS project (“SurrogAte Modelling for stOchastic Simulators”) is funded by the Swiss National
Science Foundation under Grant # 175524 (May 2018 - April 2022).

Xujia Zhu
Use of generalized lambda distributions and sparse polynomial chaos ex-
pansions

Nora Lüthen
UNCECOMP’2021 paper U18995: “Surrogating stochastic simulators using
Karhunen-Loève expansion, sparse PCE and advanced statistical modelling”

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 2 / 42
Computational models in engineering

Computational models in science and engineering are used:

• To explore the design space (“virtual prototypes”)
• To optimize the system (e.g. minimize the mass) under performance constraints
• To assess its robustness w.r.t. uncertainty and its reliability
• Together with experimental data for calibration purposes

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 3 / 42
Computational models: the abstract viewpoint

A computational model may be seen as a black box program that computes quantities of interest (QoI)
(a.k.a. model responses) as a function of input parameters

Vector of input
Computational Model response
parameters
model M y = M(x) ∈ RQ
x ∈ RM

Allows for purely data-driven approaches based on {X , M(X )}

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 4 / 42
Deterministic vs. stochastic simulators

Deterministic simulators Stochastic simulators

4 4

PDF of the ouput

3 3
Output value

Stochastic
2 simulator 2

*
Deterministic

lue
simulator

va
an
1 1

Me
*
*

0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Input parameters Input parameters

Output Md (x) is a real number Output Ms (x) is a random variable

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 5 / 42
Example: wind turbine simulation

Wind generation Aero-servo-elastic simulator

Mean speed
Drag
25
Lift
Input parameters

Wind speed (m/s)

20 pitch control
Mean speed
Turbulence intensity yaw control
15
Shear exponent
Air density 10 Turbulence
Inclination angle
5
0 200 400 600
Time (s)
Wave

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 6 / 42
Example: epidemiology
Terminology
• St : number of susceptible individuals at time t
• It : number of infected individuals at time t
• Rt : number of recovered individuals at time t

System dynamics
• Susceptible individuals can get infected due to
close contact with infected individuals (S → I )
• Infected individuals can recover and becomes
immune to future infections (I → R)
• Random contact and recovery modelled with
Poisson processes

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 7 / 42
Example: mathematical finance

Geometrical Brownian motion

dSt = µ St dt + σ St dWt

• St : stock price, Wt : Wiener process

• µ: drift, σ : volatility

Asian option
• The payoff (of a call option) is contingent on the average price of the underlying asset
Z t
1
C = max {AT − K, 0} , with At = Su du.
t 0

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 8 / 42
Outline

Introduction and literature

Generalized lambda distributions

Generalized lambda/PCE models

With replications
Without replications

Application examples
Analytical example
Stochastic SIR model
Wind turbine application

Conclusions & outlook

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 8 / 42
Formal definition

• A (scalar) stochastic simulator Ms is a mapping:

Ms : DX × Ω → R
(x, ω) 7→ Ms (x, ω)

where DX is the input parameters space and {Ω, F, P} is a probability space

• When fixing x = x0 , the output is a random variable Y |X = x0 ≡ Ms (x0 , ω)
• When fixing the seed ω = ω0 we get a deterministic simulator x 7→ Ms (x, ω0 ) (a.k.a. trajectory)

Latent variables
Ms can be seen as a deterministic function M of input parameters x and latent variables Z :
Ms (x, ω) = M (x, Z(ω))

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 9 / 42
Computational costs induced by stochastic simulators

• Replications are needed to estimate the PDF of Y |X = x

• Many runs must be carried out by varying X for uncertainty propagation, sensitivity analysis,
optimization, etc.
• Realistic simulators (e.g. for wind turbine design) are costly

Need for surrogate models

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 10 / 42
Surrogate models (deterministic simulators)

A surrogate model M̃ (a.k.a. emulator) is an approximation of the original computational model M:

M̃d (x) ≈ Md (x) in some sense

with the following features:

• It assumes some regularity of the model M and some general functional properties

• It is built from a limited set of runs of the original model M called the experimental design

X = x(i) , i = 1, . . . , N

Simulated data

• It is fast to evaluate!

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 11 / 42
Surrogate models for uncertainty quantification

Name Shape Parameters

X
Polynomial chaos expansions M̃(x) = cα Ψα (x) cα
α∈A !
R
X M
Y (i) (i)
Low-rank tensor approximations M̃(x) = bl vl (xi ) bl , zk,l
l=1 i=1
Kriging (a.k.a Gaussian processes) M̃(x) = β T · f (x) + Z(x, ω) 2
β , σZ ,θ
m
X
Support vector machines M̃(x) = ci K(xi , x) + b c, b
i=1
(Deep) Neural networks M̃(x) = fn (· · · f2 (b2 + f1 (b1 + w1 · x) · w2 )) w, b

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 12 / 42
Ingredients for building a surrogate model

• Smartly post-process the data {X , M(X )} through a learning algorithm

Name Learning method

Polynomial chaos expansions sparse grid integration, least-squares,

compressive sensing

Low-rank tensor approximations alternate least squares

Kriging maximum likelihood, Bayesian inference

Support vector machines quadratic programming

h
2 i
• Validate the surrogate model, e.g. estimate the generalization error ε = E M(X) − M̃(X)

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 13 / 42
Back to stochastic simulators
Our goal is to develop emulators that are:
• Data-driven (i.e. non-intrusive w.r.t. to the original simulator)
• General-purpose: no restrictive assumption (e.g. Gaussian) on the family of the output distribution is
made
• Able to tackle the full distribution of Y |X = x, but also quantities of interest (e.g. mean, variance,
quantiles)
• Providing a representation of Y |X = x easy to sample from

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 14 / 42
Existing approaches

The literature on stochastic simulators is both old and new:

• Replication-based approaches

• Gaussian models

• Estimation of the conditional distribution

• Latent variable models 1

• Random field representations 2

1
Yan & Perdikaris (2019), Comput. Mech.

Zhu & Sudret, EMI/PMC Conference, Columbia University, 2021

2
Azzi et al. (2019, 2020), Int. J. Uncertainty Quantification; Lüthen et al. , UNCECOMP’2021

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 15 / 42
Replication-based approaches

Main idea

• Estimate distributions/QoIs based on replications

• Treat the estimated parameters as outputs from a
deterministic simulator and apply standard surrogate
models

Literature
• Stochastic Kriging: Ankenman et al. (2006) Stochastic Kriging for simulation metamodeling, Oper. Res.

• Quantile Kriging: Plumlee & Tuo (2014) Building accurate emulators for stochastic simulations via quantile Kriging, Technometrics

• Kernel density estimation: Moutoussamy et al. (2015) Emulators for stochastic simulation codes, ESAIM: Math. Model. Num. Anal.

• Generalized lambda model: Zhu & Sudret (2020) Replication-based emulation of the response distribution of stochastic simulators
using generalized lambda distributions, Int. J. Uncertainty Quantification

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 16 / 42
Assuming normality: Kriging models

Main idea
• Response distributions are normal
• Mean function µ(x) and log-variance function log(V (x)) are modeled by Gaussian processes

Literature
• Full Bayesian setup: Goldberg et al. (1997) Regression with input-dependent noise: a Gaussian process treatment, NIPS10

• Iterative fitting: Marrel et al. (2012) Global sensitivity analysis of stochastic computer models with joint metamodels, Stat. Comput.

• Maximum likelihood: Binois et al. (2018) Practical heteroscedastic Gaussian process modeling for large simulation experiments, J.
Comput. Graph. Stat.

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 17 / 42
Conditional distribution estimation
Main approaches
• Estimate the joint distribution of (X, Y ) by kernel smoothing, then compute the conditional PDF by:

f (x, y)
f (y | x) =
f (x)
• Use parametric models to represent the conditional distribution directly

Literature
• Kernel smoothing: Hall et al. (2004) Cross-validation and the estimation of conditional probability densities, J. Amer. Stat. Assoc.

• Vine copula: Kraus & Czado (2017) D-vine copula based quantile regression, Comput. Stat. Data Anal.

• Generalized lambda model: Zhu & Sudret (2021) Emulation of stochastic simulators using generalized lambda models, Submitted to
SIAM/ASA J. Unc. Quant.

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 18 / 42
Outline

Introduction and literature

Generalized lambda distributions

Generalized lambda/PCE models

Application examples

Conclusions & outlook

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 18 / 42
Definition

• The Freimer-Mudholkar-Kollia-Lin (FMKL) lambda distribution is defined through its quantile function
Q(u; λ) by 4 parameters
 
1 uλ3 − 1 (1 − u)λ4 − 1
Q(u; λ) = λ1 + −
λ2 λ3 λ4

where:
– λ1 is the location parameter
– λ2 > 0 is the scale parameter
– λ3 , λ4 are shape parameters

• The PDF is obtained by:

1 λ2 −1
fY (y; λ) = = λ −1 with u = Q (y; λ)
Q0 (u; λ) u 3 + (1 − u)λ4 −1

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 19 / 42
Properties

• GLDs approximate well unimodal PDFs (bell-, U-shaped, bounded and unbounded)
• λ3 and λ4 control the shape and boundedness
( (
−∞, λ3 ≤ 0 +∞, λ4 ≤ 0
Bl (λ) = 1
, Bu (λ) = 1
λ1 − λ2 λ3
, λ3 > 0 λ1 + λ2 λ4
, λ4 > 0

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 20 / 42
Summary chart
Moments of order

• Blue points: infinite

support
k do not exist

• Red points: finite support,

with PDF = 0 at the bound
• Green points: finite
support, with PDF 6= 0 at
the bound

Moments of order k do not exist

Zhu & Sudret (2020), Replication-based emulation of the response distribution of stochastic simulators using generalized lambda distributions, Int. J. Uncertainty Quantification

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 20 / 42
Outline

Introduction and literature

Generalized lambda distributions

Generalized lambda/PCE models

With replications
Without replications

Application examples

Conclusions & outlook

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 20 / 42
Problem statement
Let us consider a stochastic simulator MS and the following experimental design with replications

• Experimental design of size N in the X -space: X = x(1) , x(2) , . . . , x(N )

• R replications for each x(i) ∈ X : Y (i) = y (i,1) , y (i,2) , . . . , y (i,R)

Two approaches
Zhu & Sudret (2020) Replication-based emulation of the response distribution of stochastic simulators using generalized lambda distributions, Int. J. Uncertainty Quantification

• Infer-and-fit: infer a lambda distribution for each point x(i) of the experimental design, then fit a sparse
polynomial chaos expansion to the parameters λ
• Joint inference: improve the previous results by maximum likelihood optimization

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 21 / 42
Local inference of lambda distributions (“Infer”)
(i) (i)
 (i,1) (i,R)
Estimate λ based on Y = y , ... ,y
• Maximum likelihood estimation
R
!
(i)
X λ2
λ̂ = arg max log
λ
r=1
uλi,r3 −1 + (1 − ui,r )λ4 −1

where !
−1 (i,r)

(i,r) 1 uλi,r3 − 1 (1 − ui,r )λ4 − 1
ui,r = Q y ;λ , y = Q(ui,r ; λ) = λ1 + −
λ2 λ3 λ4

• Nonlinear equation that can be solved numerically (Q is a monotonically increasing function)

Polynomial chaos expansions

n (1) (N )
o
• Fit 4 PCEs from the lambda data points Λ = λ̂ , . . . , λ̂

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 22 / 42
Polynomial chaos expansions in a nutshell
Ghanem & Spanos (1991; 2003); Xiu & Karniadakis (2002); Soize & Ghanem (2004)

• We assume here for simplicity that the input parameters are independent with
Xi ∼ fXi , i = 1, . . . , M
• PCE is also applicable in the general case using an isoprobabilistic transform X 7→ Ξ

The polynomial chaos expansion for a deterministic simulator X 7→ Md (X) reads:

M
X Y
Md (X) = aα Ψα (X) Ψα (X) = Pα(i)
i
(Xi )
α∈NM i=1

where:
• Ψα (X) are basis functions (multivariate orthonormal polynomials)
• aα are coefficients to be computed (coordinates)
Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 23 / 42
Polynomial chaos expansions
Truncation schemes q=1

• “Full basis” of degree p: 6

5
4

α2
( M
) 3
X 2

α ∈ Ap,M = α ∈ NM ; αi ≤ p 1
0
0 1 2 3 4 5 6
i=1 α1

q = 0.5

• q -norm truncation: 6
5
4

α2

M
! q1  3
  2
p,q,M M def
X
αiq
1
α∈A = α ∈ N ; kαkq = ≤p , 0<q<1 0
  0 1 2 3 4 5 6
α1
i=1

Sparse expansion: A only contains relevant basis functions taken out of a candidate basis:
• Least-angle regression
• Orthogonal matching pursuit, subspace pursuit
• etc.
Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 24 / 42
Fitting polynomial chaos expansions (“Fit”)




• Gather the data from the “Fit”-step: x(1) , λ(1) , x(2) , λ(2) , . . . , x(N ) , λ(N )
• Represent the distribution parameters λ by polynomial chaos expansions
X
λk (x) ≈ λPC
k (x; a) = ak,α ψα (x) for k = 1, 3, 4
α∈Ak
!
PC
X
λ2 (x) ≈ λ2 (x; a) = exp a2,α ψα (x)
α∈A2

• PCE coefficients are calibrated by the sparse regression method Hybrid-LARS

Blatman & Sudret (2011) Adaptive sparse polynomial chaos expansion based on Least Angle Regression, J. Comput. Phys.

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 25 / 42
Improvement: joint fitting

Rationale
• The results of the Infer-and-fit algorithm depends on the accuracy of the local inference
• Idea: build a global model for the joint distribution of inputs and outputs:

fX,Y (x, y) = fY |X (y | x) · fX (x)

where the conditional PDF is represented by a lambda model:

GLD
(x, y; a) = fYGLD PC


fX,Y |X y; λ (x; a) · fX (x)

Procedure
Find the optimal PCE coefficients a that minimize the Kullback-Leibler divergence between fX,Y and
GLD
fX,Y :
GLD


â = arg min DKL fX,Y || fX,Y ( . ; a)
a

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 26 / 42
Corresponding loss function

Solution
After some basic algebra, the minimization problem reduces to:

â = arg min EX,Y log fˆY |X Y ; λPC (X; a)



a

Estimator
• The sparse PCE bases for λPC (x) are taken from the “Infer-and-Fit” approach

• The expectation w.r.t. X and Y is computed from the experimental design X = x(1) , . . . , x(N ) ,
with R replications in each point (outputs y (i,r) ):
N R
1 XX
log fˆY |X y (i,r) ; λPC (x(i) ; a)


â = arg min
a NR
i=1 r=1

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 27 / 42
Illustration of the algorithm

Toy example

• Analytical lambda model

λ1 (x) = 50x3 − 75x2 + 35x − 4
λ2 (x) = exp(−3x2 + 3x − 1)
λ3 (x) = −0.2 + 0.7x
λ4 (x) = 0.4 − 0.6x

• Experimental design of N = 20, replications R = 40

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 28 / 42
Illustration of the algorithm

Joint fitting

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 29 / 42
Outline

Introduction and literature

Generalized lambda distributions

Generalized lambda/PCE models

With replications
Without replications

Application examples

Conclusions & outlook

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 29 / 42
Introduction
Goal
Propose a framework to build a stochastic emulator without replications (... that can also be applied if
there is replicated data)
Rationale
The joint fitting estimator does not require replications!

â = arg min EX,Y log fˆY |X Y ; λPC (X; a)



a

With Without
Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 30 / 42
Procedure
Ingredients
• An experimental design X = x(1) , . . . , x(N ) and model evaluations y (i) def

= M(x(i) , ωi )
• Pre-selected PC bases for the four polynomial chaos expansions of λi (x), i = 1, . . . , 4

Selection of PCE bases

• PCE models for the mean and variance of the model output, built using the feasible generalized
least-square method
• Use the PCE basis of µ(x) (resp. log σ 2 (x)) for λ1 (resp. λ2 )
• PCE of degree 1 for λ3 and λ4 (it is assumed that the shape of the response distribution does not vary
nonlinearly with x)

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 31 / 42
Outline

Introduction and literature

Generalized lambda distributions

Generalized lambda/PCE models

Application examples
Analytical example
Stochastic SIR model
Wind turbine application

Conclusions & outlook

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 31 / 42
Error metrics for stochastic emulators
Wasserstein distance
• It is the L2 distance between the quantile functions:
Z 1
def
2
d2WS (Y, Ŷ ) = kQY − Q̂Y k2L2 = QY (u) − Q̂Y (u) du
0

Normalized Wasserstein distance



EX d2WS Y (X), Y GLaM (X)
ε=
Var [Y ]

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 32 / 42
Geometric Brownian motion

dSt = x1 St dt + x2 St dWt

• St : stock price, Wt : Wiener process

• x1 : drift, x2 : volatility
• The analytical distribution of St exists (Itô’s calculus)

x1 − x22 /2 t, x2 t



St (x)/S0 ∼ LN

Setup
• X1 ∼ U (0, 0.1), X2 ∼ U (0.1, 0.4)


• Y = S1 for t = 1 is of interest, i.e. Y (x) ∼ LN x1 − x22 /2, x2
• Experimental design: X are generated using the Latin hypercube sampling
• No replications
Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 33 / 42
PDF predictions (ED with N = 500)

x = (0.07, 0.12) x = (0.04, 0.21)

x = (0.05, 0.3) x = (0.02, 0.33)

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 34 / 42
Convergence study
• Experimental design of size 250, 500, 1000, 2000, 4000


• 50 independent runs for each scenario EX d2WS Y (X), Y GLaM (X)
ε=
• Normalized Wasserstein distance as a performance indicator Var [Y ]
10 ° -.. -.... �...-. ....- - ....- ....- - .... - ....- -.... �..- ....- - ....- ....- - .... - ....- -.... - ...�.... - ....- ....- - .... - ....- -.... - ...�. ....- - ....- ----�
GLaM
KCDE ..
...... .. + .. ...... ...... ...... ...... ...... ..... :. ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... .. het GP
+
,
T
...... .. .. ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ............ ...... ...... ...... ...... ...... ...... ...... ...... ..
+ ----- Normal
------
.. --
............,, ... ...... ...... ...... + , ... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ..

� -.....
=::: ;:: �:: j:: : : :
�- �-...............
�- �- �- �- � �- ._�-... __-
...�
:•• :• :•� :� :� : � ·� \.; \� :� :_:: :_:::. ..-
• -=, �- �- �- �- �- �- �- �- �- -� � T- �-�-�-�-�-�-�-�-�-�-�-� - - -
+ ..• :_:. :_:. :_:. :_:. :_:. :: : :: : :: :
. ::: . ----
11111111!1. �- �- �- �- �- �- �- �- �- �· - - -- �

1
... .. ...1.. .. ...... ...... ...... ...... ...... .. " ........ � ..... ......--
..... --. -c-:;a;:-,
±
...... ...... ...... ...... ...... .. . ·+· ...... ..
...... ...... ...... ...... ...... ......
1 -- ----
... ...... ...... ...... ...... ...... ...... ...... . ...... ...... ...... .... �............ ...... .... � .......... .... 1 .... ...... ...... ...... ...... ...... ·=t=· ...... ..
...... � �
:

-@ . .
1 1 -- �-----

···. l ...... ...... ...... ..... . .... . .... l ...... ...... ...... ...... ...... ......
--
� 1
_l_ .... ...... . : . .� .._ .••. •..••. • �.-:.� •.

••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •••••• •r•••••• •••••• •••••• •••••• •••••• ••••••···
... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ...... ......
: 1
�.�<22•••••• ••
. ...... ...... ...... ...... � ............ ...... ...... ..
+ · ·• •
.. .... .... .... .... .... .... .... .... .... .... .... .... .... .... .... ...1 .... .... .... .... .... .... .... 1.... .... .... .... .... .... .... .
1 1
..1.... ..1....
1
..1....

250 500 1000 2000 4000

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 35 / 42
Outline

Introduction and literature

Generalized lambda distributions

Generalized lambda/PCE models

Application examples
Analytical example
Stochastic SIR model
Wind turbine application

Conclusions & outlook

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 35 / 42
Stochastic SIR model
• Mt = St + It + Rt : total population
• St : number of susceptible individuals at time t
• It : number of infected individuals at time t
• Rt : number of recovered individuals at time t

Setup
• Total population Mt = 2000
• The initial condition: S0 ∼ U(1300, 1800),
I0 ∼ U(20, 200)
• Y (x): total number of infected individuals
during the outbreak (without counting I0 )

Binois et al. (2018) Practical heteroscedastic Gaussian process modeling for large simulation experiments, J. Comput. Graph. Stat.

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 36 / 42
PDF predictions (ED with N = 500)

x = (1364, 61) x = (1500, 120)

x = (1600, 135) x = (1714, 165)

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 37 / 42
Convergence study
• Experimental design of size 250, 500, 1000, 2000, 4000


• 50 independent runs for each scenario EX d2WS Y (X), Y GLaM (X)
ε=
• Normalized Wasserstein distance as a performance indicator Var [Y ]

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 38 / 42
Outline

Introduction and literature

Generalized lambda distributions

Generalized lambda/PCE models

Application examples
Analytical example
Stochastic SIR model
Wind turbine application

Conclusions & outlook

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 38 / 42
Wind turbine application
• Five input variables:
– Mean speed U , turbulence intensity I and shear exponent α are uniformly
distributed in the following domain Mean speed
Drag
Lift

pitch control

yaw control

Turbulence

Wave

– Air density ρ ∼ U (0.8, 1.4), inclination angle β ∼ U (−10, 10)

• Model output: maximum flapwise bending moment Y = Mb

Simulation
• 485 training points with 50 replications
• 120 test points with 500 replications

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 39 / 42
PDF predictions
The normalized Wasserstein distance is ε = 0.013

x = (8.3, 2.2, −0.21, 0.98, −7.2) x = (11.8, 2.9, 0.29, 0.97, 0.21)

x = (9.1, 1.0, 0.47, 1.3, 4.0) x = (4.0, 1.4, 0.52, 1.2, 3.5)

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 40 / 42
One word on sensitivity analysis

Due to the stochasticity in the simulator MS (x) ≡ Md (x, z), various Sobol’ indices can be defined:
• Classical indices based on partial variances:

Var [E [Y |X u ]]
Su = u = {i1 , . . . , ik } ⊂ {1, . . . , M }
Var [Y ]
• QoI-based indices, corresponding to a particular feature of Y (x), e.g. mean, variance, quantile

Var [E [QoI(X)|X u ]] e.g.

SuQoI = , QoI(x) = µY (x), σY2 (x), qα (x)
Var [QoI(X)]
• Trajectory-based indices that are conditioned on the latent variables Z :

Var [Y | X u , Z]
Sutraj (Z) =
Var [Y | Z]

See details in: Zhu & Sudret, Global sensitivity analysis for stochastic simulators based on generalized lambda surrogate models,
Reliab. Eng. Sys. Safety (2021)

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 41 / 42
Conclusions

• Stochastic simulators are used in many fields of applied sciences and engineering

• Building general-purpose emulators is necessary for optimization, sensitivity analysis, etc.

• We propose a framework based on generalized lambda distributions and polynomial chaos expansions

• Replications are not mandatory ... but can be used !

• General-purpose algorithms are available using bricks from UQLab

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 42 / 42
 www.uqlab.com

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 42 / 42
UQLab: The Uncertainty Quantification Software

• free access to academia

• 3,700 registered users Country # Users
United States 590
• 1,450+ active users from 92 countries
China 512
http://www.uqlab.com France 345
Switzerland 292
Germany 275
• The cloud version of UQLab, accessible via United Kingdom 157
an API (SaaS) Italy 148
• Available with python bindings for beta Brazil 130

testing India 123

Canada 88
https://uqpylab.uq-cloud.io/ As of June 15, 2021

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 42 / 42
UQWorld: the community of UQ https://uqworld.org/

Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 42 / 42
Questions ?

The Uncertainty Quantification

Software
www.uqlab.com

The Uncertainty Quantification

Community
www.uqworld.org

Chair of Risk, Safety & Uncertainty Quantification

www.rsuq.ethz.ch

View publication stats Surrogate models for stochastic simulators UNCECOMP’2021 – June 30, 2021 B. Sudret 42 / 42