DP-GEN Machine Learning Force Fields for Ionic Liquid

The development of machine learning force fields (MLFFs) offers a promising alternative, aiming to reproduce potential energy surfaces (PES) based on DFT data. However, the quality of MLFFs largely depends on the expertise of researchers in preparing training datasets and tuning hyperparameters. Unlike traditional FFs, which are systematically built and transferable, MLFFs may face uncertainties in covering rare events during simulations, especially in systems with diverse atomic types. In this study, DeePMD was used to construct MLFFs, and the results were compared to classic and polarizable FF MD simulations. The findings reveal that incorporating non-equilibrated (nEQ) datasets enhances MLFF performance, yet discrepancies with polarizable FF results raise questions about the correctness and completeness of MLFF simulations in capturing complex behaviors.