Bayesian optimization for material discovery processes with noise

Abstract

An augmented Bayesian optimization approach is presented for materials discovery with noisy and unreliable measurements. A challenging non-Gaussian, non-sub-Gaussian noise process is used as a case study for the discovery of additives for the promotion of nucleation of polyethylene crystals. NEMD (non-equilibrium molecular dynamics) data are used to validate and characterize the statistical outcomes of the candidate additives and the Bayesian optimization performance. The discovered candidates show nearly optimal performance for silicon for the class of tetrahedrally coordinated crystals and a material similar to graphene but more compliant for the class of hexagonally coordinated crystals. The Bayesian approach using a noise-augmented acquisition function and batched sampling shows a sub-σ level of median accuracy and an improved robustness against noise.