Communication Dans Un Congrès Année : 2023

Multi-Fidelity Sequential Bayesian Optimization and Reliability Assessment Method for the Design of Complex Systems

Résumé

Solving a Reliability-Based Design Optimization (RBDO) problem is a challenging task, especially in the presence of complex systems that are modeled using computationally expensive numerical solvers. This topic has been extensively investigated in the literature. Amongst the proposed strategies, the Sequential Optimization and Reliability Assessment (SORA) method implements a decoupled resolution of the RBDO problem by sequentially iterating between deterministic optimization and reliability analysis. The idea of SORA is to solve, at each iteration, an equivalent deterministic optimization of the RBDO problem. This is achieved in a region that is considered as feasible, by shifting the constraints by a distance derived from the target failure probability. Once the deterministic optimization is performed, a reliability analysis phase is achieved by solving another optimization problem, in order to update the shifting vector for the next deterministic optimization problem. In this paper, we propose to solve the optimization problems in SORA with a Bayesian approach through an active learning strategy based on multi-fidelity surrogate models. First, the surrogate models are built in an augmented space where design variables and uncertain variables are combined. Second, the computational cost of SORA is reduced by taking advantage of lower fidelity solvers to build the surrogates. The merits of this innovative framework are demonstrated on an analytical test case as a well as on a realistic test case involving the optimization of a sounding rocket subject to a probability constraint on the target altitude to reach.

Fichier principal
Vignette du fichier
DTIS2024-224-U23_19599-Publiée.pdf (676.15 Ko) Télécharger le fichier
Origine Fichiers éditeurs autorisés sur une archive ouverte
Licence

Dates et versions

hal-04635030 , version 1 (04-07-2024)

Licence

Identifiants

Citer

Romain Espoeys, Loïc Brevault, Mathieu Balesdent, Sophie Ricci, Paul Mycek. Multi-Fidelity Sequential Bayesian Optimization and Reliability Assessment Method for the Design of Complex Systems. UNCECOMP 2023 - 5th International Conference on Uncertainty Quantification in Computational Science and Engineering, Jun 2023, Athènes, Greece. ⟨10.7712/120223.10329.19599⟩. ⟨hal-04635030⟩
158 Consultations
306 Téléchargements

Altmetric

Partager

  • More