Description
Date depot: 23 janvier 2023
Titre: Fighting High Dimensionality in Bayesian Optimization for Applications in Mechanical Design
Directrice de thèse:
Carola DOERR (LIP6)
Directeur de thèse:
Fabian DUDDECK (Université Technologique de Münich (TUM))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Recent advances in innovative manufacturing technologies and computer-aided tools have focused attention on the prototyping phase of mechanical structures, where artificial intelligence can help automate workflows and limit human intervention. Since mechanical designs are usually characterized by very expensive numerical simulations, surrogate-based optimization techniques play a crucial role in their optimization. Unfortunately, the number of parameters that define a component under study can be very large. This poses a challenge for surrogate-based optimization, as the number of training samples that are needed to build reliable models increases exponentially with the dimensionality of the problem. Appropriate techniques to handle this \emph{curse of dimensionality} are therefore at high demand.
In this PhD project, we aim at investigating new dimensionality reduction methods to improve the performance of Bayesian optimization for high-dimensional mechanical applications. Of particular interest are unsupervised learning techniques from the field of geometric deep learning, e.g., autoencoders, as well as the construction of low-dimensional manifolds via linear and nonlinear mappings.
Doctorant.e: Santoni Maria Laura