Kolmogorov-Arnold Network Made Learning Physics Laws Simple.

Area of Science:

• Materials Science
• Machine Learning
• Computational Physics

Background:

• Contrastive learning is increasingly used in physical systems due to its cross-modal and scalable nature.
• Kolmogorov-Arnold Networks (KANs) offer a new paradigm in neural network architectures.

Purpose of the Study:

• To develop a novel contrastive learning framework, KCCP, integrating CLIP and KAN principles.
• To establish robust correlations between crystal structures and their physical properties.
• To evaluate the performance of KANs against MLPs for this task.

Main Methods:

• Developed Kolmogorov-Arnold Contrastive Crystal Property Pretraining (KCCP).
• Integrated principles of CLIP (Contrastive Language-Image Pre-training) and KANs.
• Conducted comparative analysis between KANs and Multilayer Perceptrons (MLPs).

Main Results:

• KANs demonstrated significantly superior performance compared to MLPs.
• KANs achieved higher accuracy and faster convergence speeds in predicting crystal properties.
• KCCP successfully established correlations between crystal structures and physical properties.

Conclusions:

• KCCP offers a promising approach for cross-data structural and cross-modal physical models.
• KANs represent a powerful alternative to MLPs in machine learning applications for physical systems.
• This work extends contrastive learning capabilities to the domain of physical systems.