Refinement and performance benchmark for range-separated water force field.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Previous work developed a coupled cluster with singles and doubles and perturbative triples (CCSD(T)) level range-separated water force field combining physics-driven and machine learning models.
• Expensive CCSD(T)/CBS calculations resulted in limited quantum mechanical (QM) data and missing force labels, causing training instability.
• Instability in bulk phase simulation is a universal problem in training machine learning potentials, especially at the CCSD(T) level.

Purpose of the Study:

• To overcome limitations in QM data and force labels for training machine learning water models.
• To develop and validate a new training workflow for stable and accurate water force fields.
• To improve the simulation of bulk properties using machine learning potentials.

Main Methods:

• Implemented an active learning protocol for comprehensive sampling across temperatures and densities.
• Employed an intermediate force label technique using a machine learning density functional.
• Utilized an ensemble knowledge distillation method for model stabilization.

Main Results:

• The new training workflow significantly stabilized the water model.
• Achieved sub-chemical accuracy for both cluster energies and experimental properties.
• Demonstrated state-of-the-art performance in benchmarks for densities, radial distribution functions, dielectric constants, diffusivity, and infrared spectra.

Conclusions:

• The developed training protocol effectively overcomes limitations in data and labels for machine learning potentials.
• The stabilized water model exhibits high accuracy and reliability for various physical properties.
• This approach proves effective for creating robust and accurate machine learning potentials in computational chemistry.