Related Experiment Video
Updated: Jul 17, 2026
10:27
Contrast-Matching Detergent in Small-Angle Neutron Scattering Experiments for Membrane Protein Structural Analysis and Ab Initio Modeling
Published on: October 21, 2018
12.9K
Guidelines for creating artificial neural network empirical interatomic potential from first-principles molecular
Kohei Shimamura1, Shogo Fukushima2, Akihide Koura2
1Graduate School of System Informatics, Kobe University, Kobe 657-8501, Japan.
The Journal of Chemical Physics
|October 3, 2019
Summary
We developed guidelines for creating artificial neural network (ANN) potentials from limited first-principles molecular dynamics (FPMD) data. This enables long-timescale simulations, significantly accelerating computations.
Area of Science:
- Computational Materials Science
- Artificial Intelligence in Chemistry
- Condensed Matter Physics
Background:
- First-principles molecular dynamics (FPMD) simulations offer high accuracy but are computationally expensive, limiting system size and simulation time.
- Existing FPMD methods are constrained to hundreds of atoms and picoseconds, hindering the study of long-timescale phenomena.
- There is a need for efficient methods to extend the timescale of molecular dynamics (MD) simulations without sacrificing accuracy.
Purpose of the Study:
- To provide guidelines for developing artificial neural network empirical interatomic potentials (ANN potentials) using limited FPMD data.
- To enable long-timescale MD simulations under specific temperature and pressure conditions using ANN potentials.
- To demonstrate the effectiveness of ANN potentials by accelerating MD simulations and reproducing experimental properties.
Main Methods:
- Trained ANN potentials using FPMD data, focusing on the convergence of the radial distribution function (g(r)).
- Monitored and minimized pressure errors during ANN training using total energy and atomic forces from FPMD.
- Incorporated additional FPMD data for near-atom interactions to enhance potential robustness.
- Applied the developed guidelines to create ANN potentials for α-Ag2Se.
Main Results:
- ANN potentials for α-Ag2Se successfully reproduced g(r) and mean square displacements compared to FPMD.
- Long-timescale properties, such as specific heat, were accurately predicted by the ANN potential, matching FPMD and experimental values.
- MD simulations using the ANN potential achieved over 10^4 acceleration compared to FPMD simulations.
- The proposed guidelines simplify ANN potential creation and enable the use of previously generated FPMD data.
Conclusions:
- The developed guidelines facilitate the creation of accurate ANN potentials from limited FPMD data.
- ANN potentials enable significant acceleration of MD simulations, making long-timescale studies feasible.
- This approach enhances the utility of existing FPMD datasets and advances computational materials science.
