Abstract
The evolution of fiber optic surface plasmon resonance (FO-SPR) sensors has significantly advanced the field of sensing, driving innovations in biomedical diagnostics, environmental monitoring, and industrial automation. However, traditional machine learning approaches used to enhance these sensors often suffer from high prediction errors, restricting their impact. In this work, we present a sensor-focused knowledge distillation framework aimed at improving the prediction error and efficiency of Figure of Merit (FOM) prediction, an essential parameter for FO-SPR sensor optimization. The proposed method leverages an ensemble of fully connected neural networks (FNNs) as teacher models to extract nonlinear relationships between structural parameters and sensor performance. These models generate soft labels that encapsulate intermodel uncertainty and latent patterns. A compact student model is then trained using a dual-loss function that integrates both soft teacher outputs and ground-truth data. The resulting distilled model outperforms existing methods (e.g., GPR), achieving high performance (MSE: 8223.022, RMSE: 90.681, MAE: 50.712, and R2: 0.98). Residual analysis validates its robustness under varying conditions. This approach enables real-time, data-driven optimization of photonic sensor design, particularly in selecting optimal wavelengths and material thicknesses, thus representing a scalable solution for next-generation intelligent sensor systems.

This publication has 27 references indexed in Scilit: