A neural network potential based on pairwise resolved atomic forces and energies.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Molecular simulations are crucial for molecular and materials design.
• Machine learning (ML)-based potential energy functions promise efficient simulations at quantum chemical accuracy.
• Previous work introduced PairF-Net, an ML approach using a pairwise interatomic scheme for force prediction.

Purpose of the Study:

• To enhance the PairF-Net model by incorporating energy conservation.
• To couple the ML model with a molecular mechanical (MM) environment using OpenMM.
• To evaluate the performance of the updated PairF-Net for both gas-phase and aqueous solution simulations.

Main Methods:

• Developed an updated PairF-Net model with intrinsic energy conservation.
• Coupled the ML model to the OpenMM package for hybrid ML/MM simulations.
• Utilized the rMD17 dataset for gas-phase validation and introduced the rMD17-aq dataset for aqueous solution benchmarking.

Main Results:

• The updated PairF-Net demonstrated good agreement with the rMD17 dataset for energy and force predictions in the gas phase.
• ML models trained on gas-phase data successfully predicted forces for molecules in aqueous solution via hybrid ML/MM simulations.
• The model accurately reproduced molecular energy, atomic forces, and dynamical distributions for aqueous solutions using the new rMD17-aq dataset.

Conclusions:

• The enhanced PairF-Net model effectively simulates molecular systems with energy conservation.
• Hybrid ML/MM simulations enable accurate predictions for molecules in aqueous solution.
• The developed model and dataset advance the application of ML potentials in complex chemical environments.