ML-assisted training workflow for range-separated water force field

Introduction

This water forcefield describes the potential energy surface of water with the nonbonding interaction predicted by multipolar polarizable model (implemented in DMFF.ADMP) in the long range and machine learning model (EANN) in the short range and bonding interaction calculated by MB-pol. Further details please refer to our paper.

In this repository, the folder training_workflow contains the relevant scripts for training EANN model, including training ML-DFT model (DeePKS), transfer learning (TL), active learning (AL), and ensemble knowledge distillation (EKD). The folder final_PES contains a MD simulation case with the final water model benchmarked in our paper. The folder md_benchmarks contains the cases for bulk-phase properties benchmarks. The folder ccsdt_data contains the CCSD(T)/CBS energy data of all dimers, trimers, tetramers, pentamers, and octamers used for the training/validation of the EANN model.

Installation

• DMFF - https://github.com/deepmodeling/DMFF

• i-PI - pip install i-pi

Example

Run MD simulations using our water PES with the following commands.

cd final_PES
sbatch sub.sh