Reinforcement learning in crystal structure prediction

Crystal Structure Prediction (CSP) is a fundamental computational problem in materials science. Basin-hopping is a prominent CSP method that combines global Monte Carlo sampling to search over candidate trial structures with local energy minimisation of these candidates. The sampling uses a stochastic policy to randomly choose which action (such as a swap of atoms) will be used to transform the current structure into the next. Typically hand-tuned for a specific system before the run starts, such a policy is simply a fixed discrete probability distribution of possible actions, which does not depend on the current structure and does not adapt during a CSP run. We show that reinforcement learning (RL) can generate a dynamic policy that both depends on the current structure and improves on the fly during the CSP run. We demonstrate the efficacy of our approach on two CSP codes, FUSE and MC-EMMA. Specifically, we show that, when applied to the autonomous exploration of a phase field to identify the accessible crystal structures, RL can save up to 46% of the computation time.