Neural network potential from bispectrum components: A case study on crystalline silicon.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Developing accurate machine learning force fields (MLFFs) is crucial for simulating materials.
• Current methods often use limited training data, failing to capture the full potential energy surface.
• Crystalline silicon is a fundamental material requiring reliable computational models.

Purpose of the Study:

• To systematically develop and benchmark machine learning force fields for crystalline silicon.
• To address limitations of localized training sets in capturing global potential energy surface features.
• To identify optimal material descriptors and regression techniques for silicon MLFFs.

Main Methods:

• Training a silicon MLFF using randomly generated symmetrical crystal structures.
• Employing bispectrum coefficients as material descriptors.
• Utilizing neural network potential fitting and comparing various regression techniques.
• Benchmarking performance on diverse silicon datasets.

Main Results:

• Randomly generated structures improve the generalizability of MLFFs.
• Neural network potentials with bispectrum coefficients provide accurate results.
• The developed MLFF demonstrates good transferability for silicon.
• Substantial benchmarks confirm the effectiveness of the chosen methodology.

Conclusions:

• A robust methodology for developing general and transferable silicon MLFFs has been established.
• Bispectrum coefficients and neural network potentials are effective for accurate atomic-scale simulations.
• This approach enhances the reliability of machine learning models in materials science.