SIMPLE-NN: An efficient package for training and executing neural- network interatomic potentials

K Lee and D Yoo and W Jeong and S Han, COMPUTER PHYSICS COMMUNICATIONS, 242, 95-103 (2019).

DOI: 10.1016/j.cpc.2019.04.014

The molecular dynamics (MD) simulation is a favored method in materials science for understanding and predicting material properties from atomistic motions. In classical MD simulations, the interaction between atoms is described by an empirical interatomic potential, so the reliability of the simulation hinges on the accuracy of the underlying potential. Recently, machine learning (ML) based interatomic potentials are gaining attention as they can reproduce potential energy surfaces (PES) of ab initio calculations, with a much lower computational cost. Therefore, an efficient code for training ML potentials and inferencing PES in new configurations would widen the application range of MD simulations. Here, we announce an open-source package, SNU Interatomic Machine-learning PotentiaL packagE-version Neural Network (SIMPLE-NN) that generates and utilizes the ML potential based on the artificial neural network with the Behler-Parrinello type symmetry function as descriptors for the chemical environments. SIMPLE-NN uses the Atomic Simulation Environment (ASE) package and Google Tensorflow for high expandability and efficient training, and also supports the in-house code for quasi-Newton method. Notably, the package features a weighting scheme based on the Gaussian density function (GDF), which significantly improves accuracy and reliability of ML potentials by resolving sampling bias that exists in typical training sets. For MD simulations, SIMPLE-NN interfaces with the LAMMPS package. We demonstrate the performance and usage of SIMPLE-NN with examples of SiO2. Program summary Program Title: SIMPLE-NN Program Files doi: http: //dx.doi.org/10.17632/pjv2yr7pvr.1 Licensing provisions: GPLv3 Programming language: Python/C++ Nature of problem: Inferencing the potential energy surface for the given system with accuracy comparable to ab initio methods but with much lower computational costs. Solution method: Calculate descriptor vectors that encode local chemical environment. High-dimensional neural network is used to predict the total energy from the descriptor vectors. The trained neural network can be used for molecular dynamics simulations. (C) 2019 Elsevier B.V. All rights reserved.

Return to Publications page