Library-Based LAMMPS Implementation of High-Dimensional Neural Network Potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Accurate atomic interaction potentials are crucial for molecular dynamics (MD) simulations.
• Traditional empirical force fields struggle with complex bonding situations.
• Electronic structure calculations provide high accuracy but are computationally expensive.

Purpose of the Study:

• To develop an efficient C++ library for implementing neural network potentials (NNPs).
• To enable large-scale reactive molecular dynamics simulations with accurate atomic interactions.
• To integrate NNPs into the LAMMPS molecular dynamics package.

Main Methods:

• Developed a C++ library for efficient neural network potential-energy and force evaluations.
• Implemented the neural network potential within the LAMMPS molecular dynamics package.
• Utilized liquid water as a test system to demonstrate performance.

Main Results:

• The C++ library achieves very high efficiency in evaluating neural network potentials.
• The implementation allows for accurate representation of atomic interactions.
• Demonstrated successful application in molecular dynamics simulations of liquid water.

Conclusions:

• The developed library and LAMMPS implementation significantly enhance the feasibility of large-scale reactive MD simulations.
• Neural network potentials offer a computationally efficient alternative to traditional methods for complex systems.
• This work paves the way for more accurate and extensive simulations in materials science and chemistry.