Abstract:Recent developments in many-body potential energy representation via deep learning have brought new hopes to addressing the accuracy-versus-efficiency dilemma in molecular simulations. Here we describe DeePMD-kit, a package written in Python/C++ that has been designed to minimize the effort required to build deep learning based representation of potential energy and force field and to perform molecular dynamics. Potential applications of DeePMD-kit span from finite molecules to extended systems and from metallic systems to chemically bonded systems. DeePMD-kit is interfaced with TensorFlow, one of the most popular deep learning frameworks, making the training process highly automatic and efficient. On the other end, DeePMD-kit is interfaced with high-performance classical molecular dynamics and quantum (path-integral) molecular dynamics packages, i.e., LAMMPS and the i-PI, respectively. Thus, upon training, the potential energy and force field models can be used to perform efficient molecular simulations for different purposes. As an example of the many potential applications of the package, we use DeePMD-kit to learn the interatomic potential energy and forces of a water model using data obtained from density functional theory. We demonstrate that the resulted molecular dynamics model reproduces accurately the structural information contained in the original model.

What problem does this paper attempt to address?

The paper primarily addresses the conflict between accuracy and efficiency in the modeling of potential energy surfaces (PES) in molecular simulations. Specifically, traditional ab initio molecular dynamics (AIMD) offers high accuracy but is limited by the enormous computational cost required for density functional theory (DFT) calculations of potential energy and forces. This restricts its application to systems of only a few hundred to a few thousand atoms and to time scales of about 100 picoseconds. On the other hand, empirical force fields (FFs) can perform larger-scale and longer-time simulations, but their accuracy and transferability are often questioned, and the parameter fitting process is usually cumbersome and somewhat subjective. To address these issues, the researchers proposed a deep learning-based method to construct the potential energy representation of molecular systems, namely Deep Potential for Molecular Dynamics (DeePMD). DeePMD-kit is a software package designed to implement this method. It uses deep neural networks to learn the representation of the potential energy surface from DFT data and can perform molecular dynamics simulations at speeds close to those of empirical force fields while maintaining DFT-level accuracy. The main contributions of DeePMD-kit are as follows: 1. **Software Implementation**: Provides an implementation of the DeePMD method, i.e., DeePMD-kit, which can integrate with TensorFlow for rapid training, testing, and evaluation of potentials and forces. It can also integrate with classical molecular dynamics and path integral molecular dynamics software packages such as LAMMPS and i-PI for corresponding simulations. 2. **Theoretical Framework**: Details the theoretical framework of the DeePMD method, including how to construct atomic environment descriptors, how to compute atomic energies through neural networks, and how to calculate forces and stress tensors. 3. **Software Architecture**: Describes the software architecture and workflow of DeePMD-kit, including data preparation, model training, model testing, and performing molecular dynamics simulations. 4. **Example Validation**: Demonstrates the performance of DeePMD-kit through an example of a liquid water system containing 64 water molecules. The results show that the trained DeePMD model has high accuracy in predicting energies and forces and can reproduce the structural information from the original DFT model. In summary, DeePMD-kit provides a powerful tool for achieving high-accuracy and efficient molecular dynamics simulations, promising to advance the field of molecular simulations further.

DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics

DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics

DeePMD-kit v2: A software package for deep potential models

Learning DeePMD-Kit: A Guide to Building Deep Potential Models

86 PFLOPS Deep Potential Molecular Dynamics simulation of 100 million atoms with ab initio accuracy

Deep Potential Molecular Dynamics: a scalable model with the accuracy of quantum mechanics

DeePKS-kit: A package for developing machine learning-based chemically accurate energy and density functional models

Many-body interactions and deep neural network potentials for water

DeePKS + ABACUS as a Bridge between Expensive Quantum Mechanical Models and Machine Learning Potentials

DeePKS+ABACUS as a Bridge between Expensive Quantum Mechanical Models and Machine Learning Potentials

Efficient Deep Molecular Dynamic Model Training on Heterogeneous System

Training One DeePMD Model in Minutes: a Step Towards Online Learning.

DeePCG: Constructing coarse-grained models via deep neural networks

Adaptive coupling of a deep neural network potential to a classical force field

DP-GEN: A concurrent learning platform for the generation of reliable deep learning based potential energy models

Exciting DeePMD: Learning excited state energies, forces, and non-adiabatic couplings

Deepks: A Comprehensive Data-Driven Approach Toward Chemically Accurate Density Functional Theory