Neural Network-Based Molecular Dynamics Simulation of Water Assisted by Active Learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Physical Chemistry

Background:

• Accurate simulation of water's complex behavior is crucial for understanding various chemical and physical processes.
• Traditional methods face challenges in capturing the nuances of water's structural and dynamic properties, especially quantum effects.

Purpose of the Study:

• To develop and validate an improved deep potential (DP) model for water simulations.
• To investigate the impact of representative sampling techniques on DP model accuracy.
• To incorporate nuclear quantum effects (NQEs) into advanced simulations for a more comprehensive understanding of water.

Main Methods:

• Combined classical molecular dynamics (MD) simulations with simulated annealing (SA) for conformational exploration.
• Utilized K-means clustering to extract representative water structures for training a deep potential (DP) model (DeePMD).
• Integrated DeePMD with centroid molecular dynamics (CMD) and path integral methods to include NQEs.

Main Results:

• The DeePMD method demonstrated high accuracy in predicting water structural properties, density, and self-diffusion coefficients, comparable to DFT-MD.
• The approach successfully reproduced the characteristic O-H stretch feature in Raman spectra.
• Representative sampling was shown to be essential for robust DP model training.

Conclusions:

• The developed DeePMD approach, enhanced with representative sampling and NQEs, offers a powerful tool for accurate water simulations.
• Incorporating NQEs is vital for precisely capturing water's spectroscopic and dynamic behaviors.
• This work advances the capability of machine learning potentials in modeling condensed matter systems.