A Continuous Action Space Tree search for INverse desiGn (CASTING) Framework for Materials Discovery

Abstract

Fast and accurate prediction of optimal crystal structure, topology, and microstructures is important for accelerating the design and discovery of new materials. A challenge lies in the exorbitantly large structural and compositional space presented by the various elements and their combinations. Speed, accuracy, and scalability are three desirables for any inverse design tool to sample efficiently across such a vast space. While traditional global optimization approaches (e.g., evolutionary algorithm, random sampling based) have demonstrated the ability to predict new crystal structures that can be used as super-hard materials, semiconductors, and photovoltaic materials to name a few, it is highly desirable to develop approaches that converge faster to the solution, have better solution quality, and are scalable to high dimensionality. Reinforcement learning (RL) approaches are emerging as powerful design tools capable of addressing these issues but primarily operate in discrete action space. In this work, we introduce CASTING, which is an RL-based scalable framework for crystal structure, topology, and potentially microstructure prediction. CASTING employs an RL-based continuous search space decision tree (MCTS -Monte Carlo Tree Search) algorithm with three important modifications (i) a modified rewards scheme for improved search space exploration (ii) a 'windowing' or 'funneling' scheme for improved exploitation and (iii) adaptive sampling during playouts for efficient and scalable search. Using a set of representative examples ranging from metals such as Ag to covalent systems such as C and multicomponent systems (graphane, boron nitride, and complex correlated oxides), we demonstrate the accuracy, the speed of convergence, and the scalability of CASTING to discover new metastable crystal structures and phases that meet the target objective.