Performance Analysis of Machine Learning Models for Angular Interrogation of Surface Plasmon Resonance

Abstract

Surface plasmon resonance (SPR) paves the way for several cutting-edge sensing technologies well-known for being label-free and real-time monitoring. The angular scanning technique, one of the most common SPR applications, was performed by illuminating the SPR-based sensor with multiple incident angles of a single-wavelength laser beam. For refractive index sensing, the optical reflectance is absorbed in a specific angle, known as a plasmonic angle, which can be observed as a dark band when captured using a camera. Various methods have been proposed to locate the plasmonic position based on the detected image. This manuscript presented an analysis of the performance of machine learning on the identification of plasmonic angles based on the reflectance spectra for refractive index sensing. The reflectance curves are generated using Fresnel equations and the transfer matrix method with shot noise. After training and validating, the rational quadratic gaussian process regression model provides the most accurate model for predicting the plasmonic angle positions. The model can predict the plasmonic angles accurately for all studied refractive indices with a root mean square error of $3.83 \times 10^{\mathbf{-4}}$ RIU. Furthermore, the analysis of noise performance illustrated that a low number of photons could significantly degrade the model’s accuracy and precision. The theoretical performance can be achieved at the photon energy level of 8.14 pJ.