High-sensitivity grated D-shaped Au–SnSe₂ SPR fiber-optic sensor for temperature detection using Bayesian regularized neural networks

Abstract

Conventional optical-fiber temperature sensors often suffer from limited sensitivity, narrow dynamic range, and reliance on manual spectral interpretation without intelligent predictive integration. To address these limitations, we propose a grated D-shaped fiber-based SPR sensor for temperature sensing application. We used gold (Au) and tin selenide (SnSe₂) as the sensing layers. The analyte we have chosen is ethanol–chloroform solution with 50–50% and 60–40%, respectively. We varied the temperature (T) from 0 to 100 °C with 10 °C intervals. This work makes two main contributions: First, it optimizes the structure of the Au/SnSe₂ grating-assisted D-shaped configuration and analyzes its practical feasibility, including a study of how a ± 2% fabrication tolerance affects performance. Second, it introduces a machine learning framework that compares with various models: Bayesian Regularized Artificial Neural Network (BRANN), Kolmogorov–Arnold Network (KAN), Gaussian Process Regression (GPR), XGBoost, CatBoost etc., to automatically and accurately predict resonance wavelength (RW) and temperature sensitivity (TS) for the range 0–100 ⁰C. We got the best performance with the BRANN model with mean absolute percentage error (MAPE) of 0.009985%. We achieved a maximum TS of around 12 nm/°C for both concentration variations, manually. The maximum value of the figure of merit (FOM) was 0.132°C⁻¹ for 50–50% solution and 0.137 °C⁻¹ for 60–40% solution, respectively. It provides a promising platform for high-sensitivity temperature monitoring in applications requiring compact size, high immunity to electromagnetic interference, and optical remote sensing.