Evaluation of Machine-Learning Interatomic Potentials for Liquid Water Systems

Abstract

This report benchmarks the performance of several machine learning interatomic potentials (MLIPs) in simulating thermodynamic and transport properties of water. Classical force fields are efficient but often fail to capture complex interactions such as many-body effects and nuclear quantum effects. Quantum-scale methods, such as ab initio molecular dynamics, can account for these phenomena with high accuracy, but their computational cost limits their use to small systems and short timescales. MLIPs offer a practical compromise by learning interatomic potentials from quantum mechanical data, enabling simulations that are both accurate and scalable.

We evaluate two MLIPs, DeePMD and Allegro, trained on SCAN-metaGGA multiphase data, and assess their ability to reproduce key properties such as self-diffusivity and isothermal compressibility. The models yielded self-diffusivity values of 0.82 × 10−9 m2 /s and 0.62 × 10−9 m2 /s, both of which are significantly lower than the experimental value of 2.30 × 10−9 m2 /s. This discrepancy is attributed to overly strong hydrogen bonding from the SCAN meta-GGA functional, consistent with prior studies. We also observed that training on multi-phase datasets caused the models to produce structural features characteristic of a mixture of phases, as revealed by the radial distribution functions. While DeePMD produced values closer to experiment, it exhibited unphysical trends, including decreasing diffusivity with increasing system size and decreasing isothermal compressibility with increasing temperature.

These results illustrate both the promise and limitations of MLIPs in capturing the complex behavior of water and suggest that care must be taken in dataset construction and model validation.