Planification d'exp\'eriences s\'equentielle dans un contexte de m\'eta-mod\'elisation multi-fid\'elit\'e

TL;DR

This paper develops a multi-fidelity co-kriging approach for sequential experimental design to efficiently predict outputs of complex computer codes by leveraging multiple code levels and optimizing point selection based on variance contributions.

Contribution

It introduces a novel sequential design strategy for multi-fidelity co-kriging models, including methods to select the code level for simulation based on variance contributions.

Findings

Effective multi-fidelity surrogate modeling with co-kriging.

Strategies for sequentially adding points to improve model accuracy.

Method for choosing the optimal code level during experiments.

Abstract

Large computer codes are widely used in engineering to study physical systems. Nevertheless, simulations can sometimes be time-consuming. In this case, an approximation of the code input/output relation is made using a metamodel. Actually, a computer code can often be run at different levels of complexity and a hierarchy of levels of code can hence be obtained. For example, it can be a finite element model with a more or less fine mesh. The aim of our research is to study the use of several levels of a code to predict the output of a costly computer code. The presented multi-stage metamodel is a particular case of co-kriging which is a well-known geostatistical method. We first describe the construction of the co-kriging model and we focus then on a sequential experimental design strategy. Indeed, one of the strengths of co-kriging is that it provides through the predictive co-kriging variance an estimation of the model error at each point of the input space parameter. Therefore, to improve the surrogate model we can sequentially add points in the training set at locations where the predictive variances are the largest ones. Nonetheless, in a multi-fidelity framework, we also have to choose which level of code we have to run. We present here different strategies to choose this level. They are based on an original result which gives the contribution of each code on the co-kriging variance.