More converged, less accurate? Reassessing standard choices for ab initio water using machine learning potentials

TL;DR

This study uses machine learning potentials to assess how electronic structure calculation convergence affects the accuracy of simulated water and ice properties, revealing that less converged methods may give misleadingly good results.

Contribution

It demonstrates the importance of fully converged reference calculations in evaluating electronic structure methods for water simulations, using machine learning potentials for assessment.

Findings

Widely used revPBE0-D3 setup's agreement with experiments degrades with better convergence.

Highly converged ωB97X-rV results align better with experimental data.

MP2 with triple-zeta basis set performs poorly, indicating insufficient convergence.

Abstract

Accurately simulating the properties of liquid water remains a central challenge in molecular simulations. In this work, we use machine learning potentials to investigate how the convergence settings of electronic structure calculations impact the predicted structural and dynamical properties of simulated water and ice. We evaluate the true performance of several reference methods in classical and path-integral molecular dynamics. When we compare a popular, computationally pragmatic revPBE0-D3 setup against a highly converged one, our results reveal that its widely reported experimental agreement degrades. Applying the same highly converged settings to the $\mathrm{\omega}$B97X-rV functional, we find an improved agreement with experimental results. MP2 with a triple-$\zeta$ basis set commonly used for liquid water shows poor performance, which is indicative of insufficient convergence. These findings underscore the need for fully converged reference calculations when evaluating the fundamental accuracy of electronic structure methods and developing reliable models for aqueous systems.