AI and the World's Hard Problems · Chemistry & materials · arXiv · 2026-10-19
Training Dataset Matters for Machine Learning Interatomic Potentials: Water Simulation as an Example
Using water as an example, this paper shows that machine-learning interatomic potentials (MLIPs) depend on training datasets that cover the relevant thermodynamic conditions and phases. Insufficiently sampled datasets led to wrong predictions of density, melting temperature and nuclear quantum effects. Comprehensive datasets computed with the same electronic structure method gave consistent predictions regardless of MLIP architecture, with densities in quantitative agreement with reference calculations.