Improve the performance of machine-learning potentials by optimizing descriptors.

Area of Science:

• Computational materials science
• Machine learning in physics
• Atomistic simulations

Background:

• Machine-learning (ML) potentials offer a computationally efficient alternative to density functional theory (DFT) for atomic simulations.
• The performance of ML potentials heavily relies on the quality of descriptors used to represent atomic environments.
• Descriptor optimization is crucial for enhancing the accuracy and applicability of ML potentials.

Purpose of the Study:

• To implement a differentiable descriptor optimization method for machine-learning potentials.
• To investigate the advantages of optimized descriptors, particularly for small training datasets.
• To validate the improved predictive capabilities of ML potentials with optimized descriptors.

Main Methods:

• Implemented a differentiable approach for descriptor representation in ML potentials.
• Trained and evaluated ML potentials with and without descriptor optimization using aluminum as a case study.
• Compared the accuracy of ML potentials against first-principles calculations and dynamical simulations.

Main Results:

• ML potentials with optimized descriptors demonstrated superior performance, especially with limited training data.
• The optimized potentials accurately predicted energies and forces, matching first-principles calculations.
• Statistical results from dynamical simulations were effectively reproduced by the trained potentials.

Conclusions:

• Descriptor optimization is a key strategy for improving machine-learning interatomic potentials.
• The developed method enhances the accuracy and expands the application scope of ML potentials in materials science.
• This approach offers a pathway to more reliable and efficient atomic simulations.