Ultralow Prediction Error in SPR-Based PCF Sensor Using Kolmogorov–Arnold and Artificial Neural Networks

This paper presents the application of machine learning (ML) techniques such as Kolmogorov-Arnold Networks (KAN) and Artificial Neural Networks (ANN) for predicting confinement loss (CL) in photonic crystal fiber (PCF) sensors based on the surface plasmon resonance (SPR) concept. A thin gold (Au) coating is incorporated into the PCF structure to induce plasmonic effects. Confinement loss is analyzed through simulations performed in COMSOL Multiphysics for analytes with refractive indices in the range of 1.30- 1.40. In this study, a novel KAN model is introduced alongside a conventional ANN to predict confinement loss with significantly improved precision. The results demonstrate that the KAN model achieves superior predictive performance, yielding a mean absolute percentage error (MAPE) of 0.02193, compared to 0.04327 for the ANN and prior studies. Furthermore, the KAN model attains an outstanding coefficient of determination (R2 score) of 0.9996 and a remarkably low R2 score adjusted of 0.999, which indicate improved performance in ML-based PCF-SPR sensor design. This work thus provides an effective framework for developing high-performance PCF-SPR sensors using advanced ML algorithms, with promising applications in biosensing and chemical diagnostics.