Projet SAMOURAI (ANR-20-CE46-0013)


Summary of the project

Metamodeling and optimization under uncertainty for complex simulators

Towards frugal simulation methods for robust design applications
Increasing the efficiency of simulation-based industrial processes requires improvements in uncertainty quantification and numerical optimization. A major difficulty arises from the “black box” nature of the simulator, a direct consequence of the growing complexity of the applications targeted. The main objectives of the project, driven by application needs, are to develop efficient optimization methods based on metamodels, pushing back their current limits of performance and applicability. The ambition is to meet the following four challenges:
(i) Work package 1: design metamodels adapted to large-scale problems (typically around 100 input variables) in the context of a limited budget of simulations (around 500);
(ii) Work package 2: adapt sequential enrichment strategies to large-scale problems in the context of optimal and reliable design ;
(iii) Work package 3: develop metamodeling and optimization methods involving non-continuous input variables such as discrete or categorical variables;
(iv) Work package 4: take into account the presence of simulator failures by learning the associated hidden constraints and integrating them into the optimization procedure.

Variable selection, Gaussian process models and active learning adapted to expensive simulators for optimization under uncertainty
To build efficiently predictive metamodels from a limited set of simulations, the study of new sensitivity indices to identify influential variables, metamodels adapted to high dimensions and a robust optimization method for their hyperparameters were proposed. The second objective of the project was to identify the set of controlled input values of a simulator enabling to reach a target risk level for an output of interest also dependent on uncertain variables, using an adapted enrichment criterion. Problems with a low probability of failure were solved by a sequence of problems associated with a progressively reduced failure region. For the third objective, metamodels with a kernel adapted to sets of vectors were developed and coupled to a Bayesian optimization method with an enrichment criterion optimized by a specific evolutionary algorithm. Finally, an active learning method was developed to predict the probability of simulation failure, and coupling strategies were used to take these hidden constraints into account in “black box” optimization methods.

The project has resulted in a dozen of papers and around 30 oral communications in international journals and conferences, as well as the sharing of code whose sources are available on public repositories (CRAN, GitHub or GitLab).
In addition, the results are being used for applications in the field of renewable and low-carbon energies and air transport to reduce CO2 emissions: in the reliable and optimal design of wind turbines and engine turbine blades, for the optimal layout of wind farms and the analysis of uncertainties in thermal-hydraulic models for nuclear safety.

The Samourai project is a collaborative research-enterprise project (PRCE) coordinated by IFPEN. It also involves CEA, CentraleSupélec, EDF R&D, EMSE, Safran Tech and Polytechnique Montréal.

Publications of the project

WP1

Journal articles (published or preprint)

Oral communications and posters

Source codes

WP2

Journal articles (published or preprint)

Oral communications and posters

Source codes

WP3

Journal articles (published or preprint)

Oral communications and posters

WP4

Journal articles (published or preprint)

Oral communications and posters

Source codes