A machine-learning potential-based generative algorithm for on-lattice crystal structure prediction

We propose a crystal structure prediction method based on a novel structure generation algorithm and on-lattice machine-learning interatomic potentials. Our algorithm generates atomic configurations with arbitrary supercells by assigning atomic species to sites on the given lattice and evaluates their energy using machine-learned potentials. We demonstrate two advantages of this approach. Firstly, our structure generation algorithm conducts intelligent configurational space sampling, focusing on low-energy structures and reducing computational costs. Secondly, the use of machine-learning interatomic potentials significantly reduces the number of DFT calculations. We demonstrate the efficiency of our method by constructing the convex hull of binary Nb-W, ternary Mo-Ta-W and quaternary Nb-Mo-Ta-W systems. We identify new stable structures not present in the AFLOW database, which we employ as our baseline. Due to the computational efficiency of our method, we anticipate that it can pave the way for the efficient high-throughput discovery of multicomponent materials.