Machine Learning Enables Accurate Molecular Quantum Dynamics

Valerii Andreichev on Feb 23, 2026 submitted an arXiv preprint describing machine-learning methods to construct smooth potential energy surfaces for molecular simulations. The paper emphasizes transfer learning that requires a minimal number of expensive high-level electronic-structure training points and applies semiclassical approximations, notably perturbatively corrected instanton theory, to capture tunnelling and anharmonicity. This reduces cost for high-accuracy chemical reaction dynamics.