Liquid water simulation using hydrogen bond corrected SCAN and neural network potentials.

Accurately reproducing the structure of liquid water with ab initio molecular dynamics (AIMD) simulation is a crucial first step on the path towards accurately predicting the properties of liquid solutions without relying on experiment. Density functional theory (DFT) is normally used to approximate the forces in these simulations. However, no DFT functional has been shown to give an entirely satisfactory description of the structure of liquid water. Here, I propose a simple correction to the strongly constrained and appropriately normalised (SCAN) DFT functional, that corrects the strength of the hydrogen bonding interaction with a simple exponential potential fitted to dimer energy calculations. The resulting SCAN-CH functional provides an excellent description of the structure of liquid water. Long time scale NPT simulations are enabled by the use of neural network potentials, which demonstrate that the simulations are well converged and that the density of water is also more accurately reproduced with this method.