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

Abstract

In our study, we combined classical molecular dynamics (MD) simulations with the simulated annealing (SA) method to explore the conformational landscape of water molecules. By using the K-means clustering method, we processed the MD simulation data to extract representative samples of water molecular structures used to train a deep potential (DP) model. Our DeePMD method showed accuracy in predicting water structural properties compared to DFT-MD results. Meanwhile, this approach achieves a balanced prediction of water density and self-diffusion coefficients compared with earlier DeePMD simulations. These results highlight the essential role of representative sampling techniques in training the DP model. Furthermore, we demonstrated the effectiveness of combining the DeePMD simulation with the centroid molecular dynamics (CMD) approach, which incorporates nuclear quantum effects (NQEs). This approach successfully reproduced the shoulder feature at 3250 cm^-1 in the Raman spectra for the O-H stretch. Incorporating the path integral method into the DeePMD simulations underscores the importance of considering NQEs in understanding water molecules' structural and dynamic behaviors.