Seven machine-learned water models test how density functional choice shapes water's behaviour

The work

Technical · from the record

Seven machine learning force fields for liquid water were trained, one for each density functional theory exchange-correlation description, using VASP's on-the-fly learning scheme in a cell of 64 water molecules. Each force field then drove molecular dynamics of 512 water molecules, from which the authors computed the six-dimensional molecular pair correlation function, three-body angle and tetrahedral order distributions, hydrogen bond counts, excess entropy, viscosity and the self-diffusion coefficient, with the classical SPC/E model as a reference. Comparing these quantities with published experimental values, the authors report that functionals without dispersion corrections give over-structured water, more negative excess entropy and suppressed diffusion, that orientational excess entropy varies linearly with the translational component, and that RPBE-D3 agrees most closely with experiment across density, radial distribution function, excess entropy, viscosity and self-diffusivity among the seven functionals. They attribute the similarity between RPBE-D3 and SPC/E to comparable Born effective charges and a shared long-range dispersion form.

How AI was used

A separate machine learning force field was fitted for each of seven exchange-correlation descriptions of water (PBE, PBE-D3, PBE-TS, RPBE, RPBE-D3, R2SCAN+rVV10, vdW-DF-cx) using the on-the-fly training scheme in VASP, in which ab initio reference configurations are generated and the kernel-based potential refitted during an NpT molecular dynamics run of 64 water molecules over 50,000 steps of 1 fs along a 100 to 400 K temperature ramp, with a kernel basis capped at 6,000 functions, element-reduced descriptors, energies scaled to isolated-atom references and the hydrogen mass set to 8 amu during training only. The trained force fields were then used in place of explicit ab initio dynamics to run 1 ns NpT equilibration at 1 bar and 300 K to fix the density, followed by NVT production trajectories of 512 water molecules at 0.5 fs time steps. All structural, thermodynamic and transport analysis — the six-dimensional pair correlation function, marginal orientational distribution functions, three-body angles, tetrahedral order parameter, excess entropy via the Lazaridis-Karplus route, Green-Kubo viscosity and Einstein-relation diffusion with the Yeh-Hummer finite-size correction — was conventional post-processing of these trajectories rather than further model use, and the SPC/E reference was simulated in LAMMPS.