A Critical Assessment of Neural Network Potentials for Water and the Role of Nuclear Quantum Effects through the Van Hove Correlation Function

High accuracy studies of the properties of liquid water should combine the accuracies achievable by ab initio molecular dynamics (AIMD) with the long time and large length scales achievable by classical MD. Here, we assess two different neural network potential (NNP) based modeling strategies within classical MD in predicting the spatiotemporal correlations in liquid water i.e., the Van Hove correlation function (VHF). In principle, the NNP’s can deliver ab initio quality results at significantly larger time and length scales than would be directly accessible using ab initio methods. We apply the DeepMD [1] and NequIP [2] approaches and critically assess their efficacy, particularly regarding the size of training set and accuracy of the final predictions. By varying the training sets and using path integral approaches, we analyze the role of nuclear quantum effect on the VHF. These results are also contrasted with recent inelastic X-ray scattering data [3].

[1] Wang, Han, et al. "DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics." Computer Physics Communications 228 (2018): 178-184.

[2] Batzner, Simon, et al. "Se (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials." arXiv preprint arXiv:2101.03164 (2021).

[3] Iwashita, Takuya, et al. "Seeing real-space dynamics of liquid water through inelastic x-ray scattering." Science advances 3.12 (2017): e1603079.