A comparative analysis of ten machine learning models for performance prediction in a numerically simulated photonic crystal fibre-based surface plasmon resonance sensor

This study proposes a novel, highly sensitive surface plasmon resonance (SPR) sensor based on a hexagonal photonic crystal fibre (PCF) sensor, tailored for analytes' refractive index (RI) detection from 1.30 to 1.43. The sensor’s optical characteristics are analysed using finite element method (FEM) simulations. The proposed design achieves a peak wavelength sensitivity of 29,000 nm/RIU and an amplitude sensitivity of 2,653.68 RIU⁻¹. Additionally, a high resolution of 3.45 × 10⁻⁶ RIU underscores its effectiveness in capturing minute variations in refractive index. We conducted a comparative evaluation of ten machine learning algorithms for predicting confinement loss and amplitude sensitivity. Our results show that ensemble methods, particularly Extra Trees, Random Forest, Gradient Boosting, and XGBoost, achieve exceptionally high prediction accuracy for confinement loss (R2 > 0.999), while LightGBM outperforms other models for amplitude sensitivity prediction (R2 = 0.9448). The proposed sensor is stable and suited for detecting analytes for food safety, and adulteration monitoring applications.