In this thesis, we focus on engineering design optimization problems which involve numerical simulators (finite elements, finite volumes, …) that are computationally expensive to evaluate, and a large number of design variables. This type of problem is encountered for the design of systems in many engineering fields, including the automotive domain. Bayesian optimization methods, based on surrogate models (such as Kriging for example), are particularly efficient in solving optimization problems for expensive functions, such as numerical simulators, however they do not scale well when the number of design variables (also called the problem’s dimension) is large. In particular, two aspects of Bayesian optimization are problematic in high dimension: building a globally accurate Kriging surrogate model, and correctly identifying the new observations added during the optimization loop to enrich the surrogate model. This thesis focuses on developing a new surrogate technique more adapted than classical Kriging to Bayesian optimization for high-dimensional problems. The resulting optimization algorithm is able to bypass the curse of dimensionality faced by ordinary Kriging, and to solve design optimization problems, with many variables, in an efficient manner, i.e., with few calls to the expensive numerical simulator.

 Jury: