Machine-learned interatomic potentials (MLIPs) are traditionally constrained to fulfill physical laws like rotational invariance and energy conservation. However, evidence suggests that relaxing these constraints can improve both computational efficiency and accuracy. This talk investigates the behavior of unconstrained MLIPs in the large-data limit, demonstrating that they can outperform physically constrained models in both speed and benchmark accuracy, and discusses strategies to ensure that the unconstrained nature of the model does not introduce artifacts in the simulations.
The practical applicability of models, however, depends on the quality of the data as much as its amount. This is showcased by MAD-1.6, a highly curated dataset specifically designed for robust force-field learning across the periodic table, based on the principle of Massive Atomic Diversity.
Extending the original MAD dataset to 102 elements, MAD-1.6 utilizes a highly consistent all-electron DFT workflow, at the meta-GGA level of theory, to ensure uniformity across molecules, bulk crystals, and surfaces. We demonstrate that when trained on this consistent, high-fidelity data, unconstrained models like PET-MAD-1.6 achieve exceptional accuracy and stability in practical workflows.