Neural Network Water Model Based on the MB-Pol Many-Body Potential

Area of Science:

• Computational Chemistry
• Materials Science
• Physical Chemistry

Background:

• The MB-pol many-body potential accurately models water properties but is computationally expensive.
• High computational cost limits the application of accurate potentials in large-scale simulations.

Purpose of the Study:

• Develop a computationally efficient deep potential neural network (DPMD) model for water, based on the MB-pol potential.
• Evaluate the accuracy and transferability of the DPMD model for various water phases and conditions.

Main Methods:

• Trained a DPMD model using configurations from the MB-pol potential.
• Investigated the impact of training set composition (liquid vs. cluster configurations) on model performance.
• Assessed the model's ability to predict liquid, vapor-liquid equilibrium, and supercooled water properties.

Main Results:

• A DPMD model trained primarily on liquid configurations accurately described the bulk liquid phase but underestimated vapor-liquid coexistence densities.
• Incorporating cluster configurations into the training set significantly improved the prediction of vapor coexistence densities.
• The DPMD model demonstrated good accuracy for supercooled liquid water densities, despite these conditions not being in the training set.

Conclusions:

• Neural network models, like DPMD, can achieve both accuracy and transferability for simulating water.
• The composition of the training set is crucial for developing reliable and versatile neural network potentials.
• Careful construction of representative training data is essential for the success of DPMD models in materials simulations.