Acceleration of crystal structure relaxation with deep reinforcement learning

Abstract We introduce a Deep Reinforcement Learning (DRL) model for crystal structure relaxation and compare different types of neural network architectures and reinforcement learning algorithms for this purpose. Numerical experiments are conducted on Al-Fe structures, with potential energy surfaces generated using EAM potentials. We examine the influence of parameter settings on model performance and benchmark the best-performing models against classical optimization algorithms. Additionally, the model’s capacity to generalize learned interaction patterns from smaller atomic systems to more complex systems is assessed. The results demonstrate the potential of DRL models to enhance the efficiency of structure relaxation compared to classical optimizers.