Condensed phase properties and transferable neural network potentials

Transferable neural network potentials (NNP) are undergoing rapid development. Many practical applications of NNPs focus on single molecules; e.g., using NNPs as a fast replacement for quantum chemical methods for dihedral angle scans in force field development. Similarly, the reference data on which most portable NNPs have been trained are single molecules. As NNPs are beginning to be used to simulate more complex systems, such as solute-solvent simulations, the question arises whether the current generation of transferable NNPs is accurate enough to reproduce condensed phase properties, which in most cases are outside the training domain of the models. Here we present a first analysis of how well two transferable NNPs (ANI-2x, MACE-OFF23-(S/M)) perform in predicting properties such as density, heat of vaporization, heat capacity, and isothermal compressibility of several pure liquids (water, methanol, acetone, benzene, n-hexane at room temperature, and N-methylacetamide at 100C). In addition, we look at selected pair correlation functions and the mean square displacement. Currently, each of the models has specific weaknesses, and seemingly small flaws lead to poor performance when applied to condensed phase simulations. This suggests that the reproduction of condensed phase properties should be considered in the training/validation of transferable NNPs, as is standard in the development of classical mechanical force fields.