Universal Interatomic Potential and Simulation of Kinetics

Electrochemical interfaces are chemically and structurally so complex [Advanced Materials 34 (2022) 2108252; Energy & Environmental Science 14 (2021) 4882; Advanced Materials 33 (2021) 2100404 ] that they typically evade simple models. I will describe the recent development of a universal neural interatomic potential (UNIP) that covers 96 elements on the periodic table, from Hydrogen to Curium. More than two thousand GPU years were used to generate the ab initio training data guided by active learning. Diverse test simulations have shown this universal potential has outstanding performance, with energy error significantly less than the chemical accuracy (43 meV/atom) for even chemically very complex systems. Going from a few hundred atoms in DFT to up to 50,000 atoms with UNIP, one can study realistic microstructures such as curved interfaces, realistic phase transformations, plastic deformation and damage evolution, electrochemical interfaces, etc. A reinforcement learning (RL) technique to guide long-timescale simulation is also introduced. [J Materiomics 9 (2023) 447; Advanced Science 11 (2024) 2304122]