Design optimization of high-sensitivity PCF-SPR biosensor using machine learning and explainable AI

Abstract

Photonic crystal fiber based surface plasmon resonance (PCF-SPR) biosensors are sophisticated optical sensing platforms that enable precise detection of minute refractive index (RI) variations for various applications. This study introduces a highly sensitive, low-loss, and simply designed PCF-SPR biosensor for label-free analyte detection, operating across a broad RI range of 1.31 to 1.42. In addition to conventional methods, machine learning (ML) regression techniques were integrated to predict key optical properties, while explainable AI (XAI) methods, particularly Shapley Additive exPlanations (SHAP), were used to analyze model outputs and identify the most influential design parameters. This hybrid approach significantly accelerates sensor optimization, reduces computational costs, and improves design efficiency compared to conventional methods. The proposed biosensor achieves impressive performance metrics, including a maximum wavelength sensitivity of 125,000 nm/RIU, amplitude sensitivity of −1422.34 RIU ⁻ ¹, resolution of 8 × 10 ⁻ ⁷ RIU, and a figure of merit (FOM) of 2112.15. ML models demonstrated high predictive accuracy for effective index, confinement loss, and amplitude sensitivity. SHAP analysis revealed that wavelength, analyte refractive index, gold thickness, and pitch are the most critical factors influencing sensor performance. The combination of a simple yet efficient design and advanced ML-driven optimization makes this biosensor a promising candidate for high-precision medical diagnostics, particularly cancer cell detection, and chemical sensing applications.