PhysBO: Learning Where the Physics is Right for Bayesian Optimization

Bayesian optimization is widely used to optimize expensive scientific and engineering systems, where cheap physical models are often available as prior knowledge. Existing methods assume this prior knowledge has a single global level of reliability. In practice, simplified models are only valid locally: a reduced reaction mechanism may fail once neglected reactions become active, while empirical correlations often extrapolate poorly. We introduce $\texttt{PhysBO}$, a Bayesian optimization framework that treats low- and high-fidelity predictions as competing explanations whose significance varies across the design space, yielding a surrogate that adaptively decides where physical knowledge should be trusted. We further derive an acquisition function that explicitly values experiments resolving uncertainty over which local explanation is correct, recovering the classical model-discrimination objective from Bayesian experimental design as an exploration term. We illustrate $\texttt{PhysBO}$ on a chemical engineering optimization problem, where it achieves lower simple regret over existing physics-informed and vanilla BO methods.