Séminaire scientifique - Sylvain DUBREUIL (ONERA)

Sylvain Dubreuil, ingénieur-chercheur au Département traitement de l’information et systèmes à l'ONERA de Toulouse, sera présent sur le campus d'Aix-en-Provence et donnera un séminaire scientifique vendredi 27 novembre à 10h30 intitulé "Multidisciplinary Design Analysis and Optimization formulation based on disciplinary surrogate models".

Pour les aixois, il sera possible d'assister à la présentation en salle des Conseils, ceux à distance pourront y assister via le lien Teams ci-dessous.

Résumé :

Multidisciplinary Design Analysis and Optimization (MDAO) deals with the modelling of complex systems, which are generally solved using partitioned approaches. Formally, a multidisciplinary analysis (MDA) can be formulated as a nonlinear system of equations in which each block of equations is handled by a specific solver (i.e., a disciplinary solver). Coupling between the solvers to find the equilibrium state of the system is achieved through an iterative algorithm, such as the nonlinear Gauss–Seidel algorithm. In a multi-query parametric context, such as optimization or uncertainty quantification, the computational cost associated with repeatedly solving the MDA can be prohibitive, making surrogate modelling an attractive option. This work focuses on the construction of disciplinary Gaussian process (GP) surrogate models, as opposed to treating the entire MDA as a single black-box function. Approximating only the computationally expensive disciplinary solvers using disciplinary GP surrogate models offers several advantages, including a more flexible framework for easily integrating high-fidelity sources of information and better management of computational resources. However, this approach also presents specific challenges arising from the nonlinear coupling between disciplinary GPs. These include acquiring training data when the ranges of the coupling variables are unknown, sampling the probability distribution of the fixed-point solution, performing sensitivity analysis of this distribution, and dealing with the potentially high dimensionality of the coupling variables. These questions have been investigated at ONERA and ISAE-SUPAERO through two PhD theses (G. Berthelin and I. Da Costa Cardoso). The main achievements and remaining challenges will be presented in this talk. This work is carried out in collaboration with N. Bartoli and T. Lefebvre at ONERA, and C. Gogu at ISAE-SUPAERO.

A voir également...

https://teams.microsoft.com/meet/3697029998057205?p=onYdzXwGmCDzb1DnV3