Machine learning-assisted performance prediction and SHAP-based analysis of a multilayer SPR sensor for arsenic detection

Groundwater arsenic contamination poses a serious environmental and public health challenge, necessitating the development of rapid, sensitive, and reliable optical sensing technologies for water quality monitoring. This work presents a physics-assisted machine learning framework for performance prediction, explainable analysis, and robustness evaluation of a multilayer surface plasmon resonance (SPR) sensor employing an S-FPL53/MgF 2 /Cu/Ni/ZnSe structure operating in the Kretschmann configuration. A comprehensive dataset was generated using the transfer matrix method (TMM) by varying the multilayer thicknesses within fabrication-feasible ranges and was subsequently used to train and evaluate multiple regression models, including Linear Regression, Random Forest, Gradient Boosting, and XGBoost. Comparative analysis based on RMSE, MAE, and R 2 demonstrates that ensemble learning models substantially outperform conventional regression approaches, with XGBoost achieving the highest prediction accuracy ( R 2 = 0.9949 ). SHapley Additive exPlanations (SHAP) and Spearman correlation analysis reveal that the Cu and Ni layers are the dominant structural parameters governing sensor performance, which is further supported by plasmonic damping and evanescent-field confinement theory. Finite-element-method (FEM) simulations confirm the accuracy of the TMM model through excellent agreement in resonance characteristics, while Monte Carlo analysis under 5% and 10% Gaussian fabrication uncertainties demonstrates the robustness of both the optimized sensor and the trained machine learning surrogate models. The proposed framework provides an efficient and interpretable approach for the design, optimization, and reliability assessment of high-performance multilayer SPR sensors for arsenic detection and can be readily adapted to other refractive-index-based optical sensing applications.