Abstract
<jats:p>Herein we report the melting points of ice exhibited by some recently published foundation Neural Network Potentials (NNPs) using the direct coexistence method: SO3LR, Orb v3, MACE-MP, FENNIX-Bio1, NEP89 and PET-MAD as well as the first generation model ANI-2x. Using our previously published LICH-TEST algorithm, we classify the structural motifs adopted by individual molecules over time, finding that most models correctly exhibit ice growth and shrinkage via interfacial hexagonal structures. However, the observed melting temperatures vary from the experimental value significantly, with the lowest and highest nearly 150 K apart. The growth rates of ice below the melting points were also found to vary significantly, in some cases exceeding classical water models by one decade. SO3LR was the only NNP exhibiting an accurate value, and represents the best trade-off of speed and accuracy for the simulation of ice nucleation in (bio)organic systems. Disconcertingly the MACE, Orb and ANI models overestimate the melting point to such a degree that liquid water is effectively under deep supercooling when simulated at standard conditions. By comparing variants of the latter two potentials, we infer that an accurate description of dispersion interactions during training and/or evaluation improves the water density isobar and leads to slightly better melting temperatures, albeit with a substantial speed penalty if added during inference. We conclude with some general recommendations for training the next generation of foundation models, in order to improve their description of ice-water interactions.</jats:p>