Photonic Sensors and Machine Learning

Résumé fourni par la source

Chapter 14 explores advanced optimization techniques for machine learning (ML) models tailored for surface plasmon resonance (SPR)-based sensors. These techniques are specifically designed for resource-constrained embedded devices, thereby enabling real-time inference. SPR sensors, renowned for their high sensitivity and specificity, are widely used in biosensing, environmental monitoring, and industrial process control. Recent chip-level innovations have led to compact, portable SPR devices, but the computational demands of sophisticated models—particularly Gaussian process regression (GPR)—can exceed the limited processor, memory, and energy budgets of embedded platforms. In our previous work, we improved the SPR sensor performance by 10–12% in terms of the sensor&s;s figure of merit (FOM) by optimizing key parameters such as light wavelength (λ) and metal layer thickness ( d m ) using the GPR ML model. However, the high computational load of that GPR model may make it difficult for direct deployment on resource-limited SPR sensors, particularly when built on chip. Recognizing that on-device ML can offer reduced latency, enhanced privacy, lower bandwidth usage, and autonomous operation, this work introduces a kernel-optimized GPR framework specifically tailored for real-time inference on edge devices. By refining the radial basis function (RBF) and the rational quadratic (RQ) kernel parameters, we balance model complexity against predictive accuracy. Among our optimized configurations, one achieves an R 2 value of 0.86 and sufficiently low mean squared error (MSE) to meet real-time requirements. Our ML experiments confirm that on-device GPR inference significantly lowers latency, preserves data privacy, and enables autonomous decision-making without extensive offloading to external servers. This advancement opens up real-time SPR sensing scenarios in medical diagnostics, environmental monitoring, and industrial control despite strict memory and power constraints. Ultimately, our results illustrate how kernel refinements and model compression can transform GPR into a resource-efficient approach for portable SPR sensors, paving the way for self-contained, intelligent solutions in diverse high-impact applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.