Global Potential Energy Surface Exploration Using Generative Stochastic Surface Walking (Gen-SSW)

Abstract Global potential energy surface (PES) exploration offers a unique way to predict the structure of material from first principles. Here, we develop Gen-SSW method to enable autonomous and fast global PES exploration, which combines a general-purpose SE(3)-equivariant 3D diffusion model, LASP-Gen, with stochastic surface walking (SSW) global optimization. Contrary to the conventional SSW global PES exploration that requires a long serial trajectory to overcome high barriers between large structural basins, Gen-SSW establishes an autonomous, massively parallel global optimization framework, which realizes global PES exploration via divide-and-conquer: many short SSW trajectories are performed based on the seed structures batch-generated from LASP-Gen. By applying Gen-SSW across complex materials, we show that (i) LASP-Gen can efficiently yield a structural pool with huge structure diversity and reasonable local chemical bonding patterns, and (ii) By docking LASP-Gen with SSW, Gen-SSW can identify efficiently low-energy polymorphs of materials, as exampled by ionic materials (TiO2, LiFePO4), amorphous alloys (Ge2Sb2Te5), strong covalent materials (boron) and molecular crystals (phenol crystal). In particular, we find a new boron allotropy, namely α̃-B, that is only 6 meV/atom above α-boron but 17 meV/atom more stable than γ-boron, and suggest a synthetical route for making this new phase via solid–solid phase transition.