École Nationale Supérieure de
Techniques Avancées

Abstract

Simulation tools developed in recent years still have a limited impact on mechanical design. While each element or sub-systems model are gradually refined, the overall optimization of these systems dealt with methodological, organizational and IT difficulties. In this context, a clearly scientific problem is computation time. Indeed, computation time is too important to allow optimization within a industrial limited timeframe. To remove this lock, we place ourselves in the context of multi-fidelity datas which are powered by more or less accurate numerical simulations. In fact, it is pointless to have accurate simulations far from the optimum design. To use these surrogate models, we must obviously have in a nonlinear context an effective simulator for generating these more or less accurate simulations. The crux of the matter is that trends of constraints and objectives functions are preserved even for a lower accuracy. It appears that reduced order methods are naturally suited to multi-fidelity optimization calculations using low fidelity data. The major contribution of this internship is to propose a multi-fidelity optimization strategy based on the two major ingredients mentioned above: multi-fidelity surrogate models and reduced order models in a context of non-linear structure computation. The scientific contribution of this work is to apply the use of scale models in the multi-fidelity surrogate models and illustrate the construction of an associated prototype software. This proof-of-concept is based on an academic case of auxetic bimaterial structure.