Bayesian Optimisation of a Metasurface using a Penalised Objective Function

Kilian Bihannic (1, 2) , Jérémy Omer (1) , Renaud Loison (3) , Guillaume Reille (2)

1 IRMAR - Institut de Recherche Mathématique de Rennes
2 Thales DMS France, SAS
3 IETR - Institut d'Électronique et des Technologies du numéRique

This study formulates the design of a metasurface as an unconstrained optimisation problem. The objective function is assumed to be expensive to evaluate and the performance of the optimisation process is assessed by the number of objective function evaluations. This characteristic of the problem motivates the use of a bayesian optimisation strategy called Efficient Global Optimisation (EGO). An undesirable modeling property of a natural objective function is solved by jointly minimising a necessary condition of optimality. We show numerically that penalising the objective improves the speed and robustness of the optimisation process.

Mots clés: Bayesian Optimisation; Metasurface; Penalisation; Convolution

Domaines: Recherche opérationnelle [math.OC]; Combinatoire [math.CO]

Kilian Bihannic, Jérémy Omer, Renaud Loison, Guillaume Reille. Bayesian Optimisation of a Metasurface using a Penalised Objective Function. 23ème congrès annuel de la Société Française de Recherche Opérationnelle et d'Aide à la Décision, INSA Lyon, Feb 2022, Villeurbanne - Lyon, France. ⟨hal-03595256⟩