Interpretable prediction of photonic crystal fiber characteristics using Kolmogorov-Arnold network

This work presents a neural network framework for photonic crystal fiber (PCF) modeling based on the Kolmogorov-Arnold network (KAN) architecture, which integrates data-driven learning with symbolic interpretability. The model is trained to predict four PCF properties-effective refractive index, effective mode area, dispersion, and confinement loss-from the structural parameters. In addition to achieving high numerical accuracy across all targets (R²>0.98), the KAN framework yields analytic, closed-form symbolic expressions that explicitly reveal the functional dependencies learned by the network. These compact formulas identify the normalized air-hole diameter as the dominant structural parameter while capturing subtle wavelength-dependent nonlinearities. The extracted expressions align well with established physical trends, confirming that the framework provides reliable and interpretable insight into modal behavior. Owing to its transparency and generality, this approach offers a valuable tool not only for PCF design and optimization but also for broader photonics applications requiring efficient and explainable modeling.