Gradient-based training and pruning of radial basis function networks with an application in materials physics.

Area of Science:

• Physics
• Materials Science
• Computer Science

Background:

• Machine learning models require interpretability and robustness in scientific applications.
• Existing techniques may lack efficiency or scalability for complex datasets.

Purpose of the Study:

• To introduce a fully gradient-based training technique for radial basis function networks.
• To develop efficient model pruning criteria for continuous and binary data.
• To enhance the interpretability of machine learning models in material physics.

Main Methods:

• Developed a novel, fully gradient-based approach for training radial basis function networks.
• Derived closed-form optimization criteria for pruning models.
• Implemented an efficient and scalable open-source solution.
• Applied the technique to a material physics problem involving atomic migration.

Main Results:

• Achieved compact and interpretable pruned models from larger, complex ones.
• Demonstrated the technique's effectiveness on real-world material physics data.
• Visualizations revealed key atomic configurations influencing migration processes.

Conclusions:

• The proposed method offers an interpretable and robust machine learning solution for scientific applications.
• Pruned models provide valuable insights into material behavior.
• Findings can guide the development of improved machine learning descriptors for materials science.