Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Hubert Beck, Ondrej Marsalek

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 $\mathrmω$B97X-rV functional, we find an improved agreement with experimental results. MP2 with a triple-$ζ$ 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.

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

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 B97X-rV functional, we find an improved agreement with experimental results. MP2 with a triple- 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.
Paper Structure (20 sections, 5 equations, 18 figures)

This paper contains 20 sections, 5 equations, 18 figures.

Table of Contents

  1. Introduction
  2. Computational Details
  3. Results
  4. Model validation
  5. MD Simulations
  6. Discussion
  7. Conclusions
  8. Adapting the Training Dataset
  9. Complete Test Sets
  10. Role of Basis Sets and Pseudopotentials
  11. Radial Distribution Function
  12. Covalent and Hydrogen Bonds
  13. Pressure--Density Curve
  14. Diffusion Coefficient
  15. Hydrogen Bond Lifetimes
  16. ...and 5 more sections

Figures (18)

  • Figure 1: Top panels: RMSEs for the 4 different C-NNPs on test sets of liquid water, ice and a slab with a liquid-vacuum interface for both classical (left) and quantum (right) structures. Bottom panels: Distribution of force disagreements along an $NVT$ trajectory of liquid water using classical (left) and path integral (right) MD. The main panels use a logarithmic scale for the y-axis, while the insets shows the distribution for low disagreements on a linear scale.
  • Figure 2: The oxygen-oxygen RDF for the 4 C-NNPs. In each panel, the result from the classical MD simulation is shown by the solid line in the respective colour, the PIMD result is shown by the black dashed line and the experimental reference Skinner2013/10.1063/1.4790861 is indicated by the gray-shaded area.
  • Figure 3: Analysis of the structure of covalent and hydrogen bonds. Top panel: The distribution and means of lengths of covalent oxygen-hydrogen bonds for classical (solid lines/bars) and PI (dashed lines/bars) MD run with the 4 different C-NNPs. The main plot shows the distributions of bond lengths and the inset the mean bond lengths. Middle panel: The distributions of the proton sharing coordinate. Bottom panel: The distributions of the hydrogen bond angle. The inset compares the half width at half maximum of the different methods.
  • Figure 4: The pressure--density curve of liquid water (top) and ice $\mathrm{I_h}$ (bottom) for the 4 different C-NNPs. The black curve in the top panel and the black diamond in the bottom panel shows the experimental reference for each system Grindley1971/10.1063/1.1675455Sanz2004/10.1103/PhysRevLett.92.255701. The error bars indicate statistical errors obtained by block averaging. As in the other plots, solid lines are used for classical MD and dashed lines for PIMD.
  • Figure 5: Comparison of the diffusion coefficients for the different C-NNPs in classical and quantum MD. The pale colours at the end of each bar signals the magnitude of the finite size corrections added to the calculated values. The gray vertical area indicates the experimental reference value including its confidence interval Holz2000/10.1039/b005319h.
  • ...and 13 more figures