Abstract
The present simulation work highlights how to apply the hybrid machine learning (HML) approach for the classification of fundamental core (FC) and surface plasmon polariton (SPP) modes of surface plasmon resonance (SPR) driven photonic crystal fibre (PCF) sensor loaded separately with analytes of refractive index (RI) variation from 1.37 to 1.40 in steps 0.01. The simulation results are obtained from COMSOL Multiphysics 6.0 software in addition to Python and MATLAB software. The core region of the sensor is made up of fused silica, which is surrounded by two types of cladding air holes. The plasmonic layer of gold is selected for excitation of surface plasmons, which is an essential requirement for accurate detection of various samples under consideration. The unwanted reflections from the outer surface of the design are protracted by perfectly matched layer (PML). For implementation of ML approach, the simulated data set containing real and imaginary parts of effective refractive indices (neff) of FC and SPP modes corresponding to both x- and y-polarization cases of the optimized structure loaded with sample of RI 1.38, operating between the wavelength range 0.60μm to 1.0μm in steps of 0.005, μm is created. This dataset is further utilized by the proposed HML model consisted of central neural network (CNN), XG- Boost and support vector machine (SVM). The stratified 5-fold cross-validation technique is used for training, validation, and testing purposes necessary for mode classification. The performance of the proposed sensor is comprehensively examined with the help of 6 statistical performance metrics (1) accuracy, (2) precision, (3) recall, (4) specificity, (5) F1-Score(F1-S) and (7) Matthews Correlation Coefficient (MCC). The proposed SPR-driven PCF biosensor loaded with analyte of RI 1.39 exhibits the highest sensitivity of 13500 nm/RIU. The implementation of the HML model yields an accuracy of 99.69 ± 0.62%, F1-score of 99.81 ± 0.37% and MCC value of 98.97 ± 2.06%, indicating its excellent mode classification performance. This idea of implementing the HML model for classification of modes of SPR-driven PCF sensor may be further extended in the design and development of future biosensors for sensing of various biological or chemical samples of RI variation from 1.37 to 1.40.