Summer school CEA-EDF-INRIA 2021

Multi-fidelity, multi-level, model selection/aggregation: how the presence of several versions of a code can improve the prediction of complex phenomena ?

Location: Paris Jussieu

Dates: 14-18 June 2021

Scientific organizers : Guillaume Perrin (Université Gustave Eiffel), Merlin Keller (EDF R&D), Cédric Durantin (CEA/DAM)

Numerical simulation has become an essential tool for predicting the behaviour of complex
systems. This simulation is generally based on computer codes, of which there may be several
versions, modelling the same system but with a complexity, precision and calculation cost that
may be significantly different. For example, a version modeled in three dimensions is a priori more
accurate than another corresponding to a simplification in one or two dimensions, but this version
is also much more demanding in computing time.
As a general rule, the availability of a new version of a code is accompanied by a progressive
decline of previous versions. However, the presence of several versions of physical or numerical
models, with comparable or clearly hierarchical precisions, is a great asset. And their joint use can
often make it possible to predict complex phenomena with increased precision, for a given
simulation cost.

The aim of this summer school is to answer the following two questions:

1. How to exploit different versions of a code with a hierarchy of accuracies?

2. How to exploit different competitive versions of a code (no clear hierarchy of precision
between the different versions)?

Multi-level and multi-fidelity approaches will be introduced in order to answer the first question,
while Bayesian model averaging (BMA) approaches and model selection techniques will provide
leads to answer the second question.


    • Introduction to statistical learning - Guillaume Perrin (Université Gustave Eiffel)

    • Multilevel Monte Carlo methods for Uncertainty Quantification - Fabio Nobile (Ecole Polytechnique Fédérale de Lausanne)

    • Multifidelity surrogate modeling - Claire Cannamela and Baptiste Kerleguer (CEA/DAM)

    • Introduction to Bayesian calibration and Bayesian model averaging - Paola Cinnella (Sorbonne Université)

    • Model/estimator selection - Sylvain Arlot and Mélina Gallopin (Université Paris Saclay)