More Converged, Less Accurate? Reassessing Standard Choices for Ab Initio Water Using Machine Learning Potentials.

Area of Science:

• Computational chemistry
• Materials science
• Physical chemistry

Background:

• Simulating liquid water properties is crucial for understanding chemical and physical processes.
• Electronic structure calculations are fundamental for accurate molecular simulations.
• Convergence settings in calculations significantly influence simulation outcomes.

Purpose of the Study:

• To investigate the impact of electronic structure calculation convergence on simulated water and ice properties.
• To evaluate the performance of different computational methods under varying convergence levels.
• To identify reliable computational strategies for aqueous systems.

Main Methods:

• Machine learning potentials were employed for molecular dynamics simulations.
• Classical and path-integral molecular dynamics were used.
• Several electronic structure methods, including revPBE0-D3, ωB97X-rV, and MP2, were assessed at different convergence levels.

Main Results:

• A computationally pragmatic revPBE0-D3 setup showed degraded experimental agreement compared to a highly converged calculation.
• The ωB97X-rV functional demonstrated improved agreement with experimental data when using converged settings.
• MP2 with a triple-ζ basis set exhibited poor performance, indicating insufficient convergence.

Conclusions:

• Fully converged reference calculations are essential for assessing the accuracy of electronic structure methods.
• Developing reliable models for aqueous systems necessitates attention to calculation convergence.
• The choice of electronic structure method and convergence criteria critically impacts the simulation of water properties.