Machine Learning Interatomic Potentials: From Mechanism to Rational Design

Abstract

Machine learning interatomic potentials (MLIPs) bridge the accuracy of ab initio methods and the efficiency needed for large-scale, long-timescale atomistic simulation. This dissertation combines Bayesian active learning with deep equivariant neural network potentials and deploys the resulting workflows to decode atomic mechanisms across four domains: ion transport, phase transitions, excited-state photochemistry, and computational materials design. In solid acid electrolytes, nanosecond-scale molecular dynamics reveals a "proton slingshot" mechanism: a rotating polyanion first carries the proton, and O-H bond reorientation then extends its displacement. The combined motion enables long-range jumps, with proton concentration correlating with rotational frustration. In silicon carbide, hierarchical active learning enables simulations of up to 512,000 atoms that establish incongruent melting under pressure, resolving long-standing experimental controversies and yielding a complete pressure-temperature phase diagram. Beyond the ground state, equivariant architectures are extended to learn nonadiabatic couplings between multiple potential energy surfaces, reproducing experimental photochemical quantum yields three orders of magnitude faster than reference electronic structure methods. Finally, equivariant foundation potentials are benchmarked and fine-tuned for superprotonic conductor screening, where they overcome failure modes of bespoke models trained on small datasets and reveal anisotropic proton transport in a candidate material, corroborated by experiment. Together, these contributions show that MLIPs access qualitatively new physical regimes rather than merely accelerating existing calculations. They also establish a practical framework spanning active learning, architectural extensions, and foundation model fine-tuning for MLIP-driven materials design.