Highly Sensitive PCF-SPR Magnetic Sensor Based on Statistical Optimization and MLP Modeling

A highly sensitive photonic crystal fiber surface plasmon resonance (PCF-SPR) magnetic sensor assisted by statistical optimization and machine learning is proposed and numerically investigated. The sensor consists of a silica-based PCF incorporating two gold-coated elliptical channels filled with magnetic fluid, enabling efficient coupling between the guided core mode and surface plasmon polaritons. Finite element method (FEM) simulations were performed over a magnetic-field range of 30–300 Oe. The air-hole diameter, pitch, and gold layer thickness were optimized using the Taguchi method and analysis of variance (ANOVA), identifying the pitch as the dominant design parameter. The optimized sensor achieves a wavelength sensitivity of 500 pm/Oe, an amplitude sensitivity of 0.0361 Oe− 1, and a resolution of 2 × 10− 2 Oe within the 30–90 Oe operating range. To accelerate sensor analysis, a multilayer perceptron (MLP) model was developed to predict the complete confinement-loss spectrum directly from the wavelength and magnetic-field intensity. The model was trained using a FEM-generated dataset and rigorously validated through a Leave-One-H-Configuration-Out (LOCO) cross-validation strategy. The MLP achieved an average R2 of 0.9878 ± 0.0101, outperforming Support Vector Regression and Random Forest models under the same validation protocol. Furthermore, robustness analysis considering ± 5