Bayesian Optimization for Simulation-Based Design of Multi-Model Systems

We address the problem of simulation-based design using multiple interconnected expensive simulation models, each modeling a different subsystem. Our goal is to find the globally optimal design with minimal model evaluation costs. To our knowledge, the best existing approach is to treat the whole system as a single expensive model and apply an existing Bayesian optimization (BO) algorithm. This approach is likely inefficient due to the need to evaluate all the component models in each iteration. We propose a multi-model BO approach that dynamically and selectively evaluates one component model per iteration based on linked emulators for uncertainty quantification and the system knowledge gradient (KG) as acquisition function. Building on this, we resolve problems with constraints and feedback couplings that often occur in real complex engineering design by penalizing the objective emulator and reformulating the original problem into a decoupled one. The superior efficiency of our approach is demonstrated through solving an analytical problem and a multidisciplinary design problem of electronic packaging optimization.