Using New Generation Neural Network Potentials to Benchmark Ice-Water Equilibria

Ice and mixed phase clouds in the earth's atmosphere form predominantly through heterogeneous nucleation on seed particles, such as mineral dust and organics. Determining the atomistic ice-nucleation mechanism on these particles is challenging for experiments and simulations. When simulating ice nucleation using Molecular Dynamics (MD), one typically relies on classical empirical potentials (force fields) to describe interactions between atoms in the particle surface and water. However, due to the large number of different materials ice-nucleating particles can consist of, accurate classical empiric potentials are not available for all systems, leading to heavy computational costs for creating and testing new ones.  

In recent years, foundation neural network potentials (NNPs), trained on large sets of quantum chemical data, aim to enable simulations of any system, thus circumventing the issue of creating new potentials. These NNPs would ideally combine accuracies of Density Functional Theory (DFT) with simulation speeds of classical MD. To determine the viability of using foundation models in MD simulations of heterogenous ice nucleation, we have benchmarked the ice-water equilibria of four NNPs: SO3LR (Kabylda et al. 2025), Orb v3 (Rhodes et al. 2025), Fennix-Bio1 (Plé et al. 2025) and ANI-2x (Devereux et al. 2020). We determined their melting points and, where not available in the literature, the water density isobars they exhibit in the temperature range 250-300 K. We have used the coexistence method: A system initially containing hexagonal ice and liquid water is simulated in the NPT ensemble, and the melting point is determined as the temperature at which the number of ice-like water molecules (counted using the classification algorithm LICH-TEST) does not change over time.  

The SO3LR potential was the only one of the four displaying a melting point close to 273 K. Fennix-Bio1 underestimated the melting point by 20 K, while both Orb v3 and ANI-2x overestimated it by over 75 K. By comparing variants of the latter two models, we can infer that inclusion of dispersion interactions during either training or inference improves the water density isobar, which in turn leads to a more accurate melting point. In addition, we find that while the NNPs are in theory reactive models, no Grotthus-like proton transfers were observed in the simulations.  

Kabylda et al.: Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields, ChemRxiv, 2025. 

Rhodes et al.: Orb-v3: atomistic simulation at scale, https://arxiv.org/abs/2504.06231, 2025.

Plé et al.: A Foundation Model for Accurate Atomistic Simulations in Drug Design, ChemRxiv, 2025. 

Devereux et al.: Extending the Applicability of the ANI Deep Learning Molecular Potential to Sulfur and Halogens, Journal of Chemical Theory and Computation, 16, 4192–4202, 2020.