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

We propose a method for crystal structure prediction based on a new structure generation algorithm and on-lattice machine learning interatomic potentials. Our algorithm generates the atomic configurations assigning atomic species to sites of the given lattice, and uses cluster expansion or low-rank potential to evaluate their energy. We demonstrate two benefits of such approach. First, our structure generation algorithm offers a ``smart'' configurational space sampling, targeting low-energy structures which significantly reduces computational costs. Second, the application of machine learning interatomic potentials significantly reduces the number of DFT calculations. We discuss how our algorithm resembles the latent diffusion models for image generation. We demonstrate the efficiency of our method by constructing the convex hull of Nb-Mo-Ta-W system, including binary and ternary Nb-W and Mo-Ta-W subsystems. We found new binary, ternary, and quaternary stable structures that are not reported in the AFLOW database which we choose as our baseline. Due to the computational efficiency of our method we anticipate that it can pave the way towards efficient high-throughput discovery of multicomponent materials.