Safe Bayesian optimization with simulation-informed Gaussian process for the constraints

In safety-critical industrial scenarios, algorithms must efficiently find optimal input parameters that ensure safe outcomes. Safe Bayesian optimization is a viable solution that guarantees constraint fulfillment during optimization when the constraints are continuous. However, in some industrial contexts, constraints are only known to be fulfilled or not. Consequently, learning these constraints becomes a classification problem, and the theoretical guarantees of safe Bayesian optimization do not apply in this scenario. This paper addresses this limitation by introducing an enhanced version of safe Bayesian optimization, incorporating a simulation-informed Gaussian Process (GP) for handling classification constraints. We applied this novel approach to optimize the parameters of a computational model for the Isotope Separator On-Line (ISOL) at the MYRRHA facility. The results revealed a significant reduction in constraint violations by approximately 87% compared to learning the true classification constraints using the Laplace approximation.