Adaptive sampling methods via machine learning for materials screening

Abstract

High-throughput virtual screening by using a combination of first-principles calculations and Bayesian optimization (BO) has attracted much attention as a method for efficient material exploration. The purpose of the virtual screening is often to search for the materials whose properties meet a certain target criterion, while the conventional BO aims to find the global extremum. Some recent works use the conventional BO by converting target properties for such motivation. On the other hand, an adaptive sampling method, where the acquisition function is based on the probability that a data point achieves a target property within a specific range, is suggested previously [Kishio et al., Chemom. Intell. Lab. Syst. 127, 70 (2013)]. In this paper, we demonstrate that such adaptive sampling is effective for the exploration of the materials whose properties meet target criteria. We conducted simulations of material exploration using an in-house database constructed by first-principles calculations and compared the performance of the adaptive sampling and conventional BO approaches. Furthermore, we evaluate and discuss the performance of acquisition functions extended to multi-objective problems for material exploration, considering multiple-target properties simultaneously.