On the Effectiveness of Bayesian AutoML methods for Physics Emulators