Machine-learned acceleration for molecular dynamics in CASTEP.

Area of Science:

• Computational Materials Science
• Materials Modeling
• Machine Learning Applications

Background:

• Machine learning (ML) interatomic potentials offer promise for materials modeling but often require significant expertise for new systems.
• Established density-functional theory (DFT) packages are widely used but can be computationally intensive.
• Integrating ML potentials with first-principles methods is crucial for advancing materials discovery.

Purpose of the Study:

• To present a computational methodology for combining the CASTEP simulation software with on-the-fly fitting and evaluation of ML interatomic potentials.
• To establish a framework for systematically improving ML potential accuracy by regular comparison with DFT reference data.
• To demonstrate the practical application of this integrated approach in materials simulations.

Main Methods:

• Developed a computational framework to couple CASTEP, a first-principles simulation package, with ML interatomic potential models.
• Implemented an on-the-fly fitting and evaluation scheme for ML potentials during simulations.
• Utilized regular checks against DFT reference data to ensure and quantify the accuracy of evolving ML models.

Main Results:

• Successfully integrated ML interatomic potential development with CASTEP simulations.
• Demonstrated the effectiveness of using DFT reference data for real-time ML model validation and improvement.
• Applied the methodology to perform high-temperature molecular-dynamics simulations of carbon nanostructures with validated ML potentials.

Conclusions:

• The presented computational methodology facilitates the use of ML interatomic potentials in materials modeling by ensuring accuracy through DFT validation.
• This approach lowers the barrier to entry for utilizing ML potentials in complex simulations.
• The freely available code supports academic research in computational materials science and ML applications.