Reinforcement learning atomistic simulation of short-range order formation in multi-principal element alloys

Atomistic simulation of long-timescale processes such as defect diffusion and order-disorder transition is a challenging problem. These processes can involve a macroscopic timescale (millisecond or longer), which conventional methods can hardly simulate. In this work, we developed a reinforcement learning method that accelerates the kinetics simulation of defect diffusion and atomic order evolution in multi-principal element alloys. We apply the method to the short-range order (SRO) formation process in the CrCoNi alloy, where the SRO plays an essential role in the alloy's outstanding mechanical performance. The SRO formation process is simulated under different annealing temperature and vacancy concentration in the alloy. The simulation timescale is up to second level, beyond the accessible range of traditional kinetic Monte Carlo method. Incorporating extrapolation schemes, we provide numerical predictions of the short-range order parameters as a function of vacancy concentration, annealing temperature, and annealing time, which can be validated by future experiments. Our reinforcement learning method can provide a computational tool to overcome the timescale limitation of atomistic simulations.