Reinforcement Learning Accelerated Atomistic Surface Reconstruction in Semi-Grand Canonical Ensemble

Identifying the structure of surfaces and interfaces is crucial for understanding active sites in target reactions and determining defect concentrations as a function of thermodynamic variables for applications in catalysis and electronics. Recent advancements in integrating machine learning with first-principles calculations have demonstrated successful cases of surface reconstruction. Building on these efforts, we present a deep reinforcement learning approach to facilitate atomistic simulations of surface reconstruction. Using SrTiO₃ as a model system, we formulate a new reward function that accounts for both thermodynamic and kinetic factors pathways of atomistic growth on surfaces within a semi-grand canonical ensemble. We employ double Q-learning for the deep Q-network (DQN). Efficient data sampling strategies were used for the development of neural network interatomic potentials (NN-IP) for SrTiO₃. The trained DQN enables faster acquisition of stable surfaces under given thermodynamic conditions compared to Markov chain Monte Carlo (MCMC) simulations.