ReLMM: Reinforcement Learning Optimizes Feature Selection in Modeling Materials

A challenge to materials discovery is the identification of the physical features that are most correlated to a given target material property without redundancy. Such variables necessarily comprise the optimal search domain in subsequent material design. Here, we introduce a reinforcement learning-based material model (ReLMM) as a tool for analyzing a given database in identifying a minimal or near minimal subset of physical features for the design of a material with a given target property. We aim for minimality in the selected subset with respect to its size─smaller being better─ while maintaining the desired accuracy of the prediction. We have shown, using synthetic multiscale data sets, that ReLMM can identify the relative importance of features, and thus help identify which should be selected across scales. In the context of semiconducting materials, ReLMM can be used to improve the prediction of the band gap by identifying which features should be selected in model building. For metal halide perovskites, ReLMM was seen to find a near minimal data set at least as well as, if not better than, state-of-the-art feature selection tools such as LASSO and XGBoost. We also found that our domain-science oriented approach can be used to uncover the hierarchical structure of a material from a database consisting of molecular-scale, mesoscale and device-scale features and labels in complementarity with an earlier hierarchical model called NestedAE.