Kolmogorov–Arnold\nNetwork Made Learning Physics\nLaws Simple

In recent years, contrastive learning has gained widespread\nadoption\nin machine learning applications to physical systems primarily due\nto its distinctive cross-modal capabilities and scalability. Building\non the foundation of Kolmogorov–Arnold Networks (KANs) [Liu, Z. et al. Kan: Kolmogorov-arnold\nnetworks. arXiv 2024, 2404.19756], we introduce a novel contrastive\nlearning framework, Kolmogorov–Arnold Contrastive Crystal Property\nPretraining (KCCP), which integrates the principles of CLIP and KAN\nto establish robust correlations between crystal structures and their\nphysical properties. During the training process, we conducted a comparative\nanalysis between Multilayer Perceptron (MLP) and KAN, revealing that\nKAN significantly outperforms MLP in both accuracy and convergence\nspeed for this task. By extending the capabilities of contrastive\nlearning to the realm of physical systems, KCCP offers a promising\napproach for constructing cross-data structural and cross-modal physical\nmodels, representing an area of considerable potential.