Physics-aware multifidelity Bayesian optimization: A generalized formulation

• DOI: 10.1016/j.compstruc.2024.107302
• Publication Date: 2024-02-07T22:18:56Z
• Authors: Francesco Di Fiore; Laura Mainini
• Keywords: Bayesian Optimization; Engineering optimization

The adoption of high-fidelity models for many-query optimization problems is majorly limited by the significant computational cost required their evaluation at every query. Multifidelity Bayesian methods (MFBO) allow to include costly responses a sub-selection queries only, and use fast lower-fidelity accelerate process. State-of-the-art rely on purely data-driven search do not explicit information about physical context. This paper acknowledges that prior knowledge domains engineering can be leveraged these searches, proposes generalized formulation MFBO embed form domain awareness during procedure. In particular, we formalize bias as multifidelity acquisition function captures structure domain. permits partially alleviate from learning properties on-the-fly, sensitively enhances management multiple sources information. method allows efficiently simulations guide while containing overall expense. Our physics-aware presented illustrated two classes frequently met in science engineering, namely design health monitoring problems.