Journal of Materials Informatics, Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 27, 2024

The application of machine learning methods for predicting potential energy surface and physical properties within materials science has garnered significant attention. Among recent advancements, Kolmogorov-Arnold Networks (KANs) have emerged as a promising alternative to traditional Multi-Layer Perceptrons. This study evaluates the impact substituting Perceptrons with KANs four established frameworks: Allegro, Neural Equivariant Interatomic Potentials, Higher Order Message Passing Network (MACE), Edge-Based Tensor Prediction Graph Network. Our results demonstrate that integration enhances prediction accuracies, especially complex datasets such HfO2 structures. Notably, using exclusively in output block achieves most improvements, improving accuracy computational efficiency. Furthermore, employing facilitates faster inference improved efficiency relative utilizing throughout entire model. selection optimal basis functions depends on specific problem. strong enhancing potentials material property predictions. Additionally, proposed methodology offers generalizable framework can be applied other ML architectures.

Language: Английский