On-Chip Sensor Utilizing Concatenated Micro-Ring with Enhanced Temperature Invariance Using Deep Learning

Abstract

An approach to measuring chemical concentrations using a slotted micro-ring resonator (sMRR) is proposed which is robust to spectral shifts caused by temperature variations. Two 1-D Convolutional Neural Network architectures, ResNet34 and VGG20, were trained for regression, achieving mean squared errors (MSEs) of 1.1251 × 10−4 and 1.2195 × 10−4, respectively. The models predict concentrations of water, ethanol, methanol, and propanol (0–100%) from the transmission spectra of a single-ring sMRR etched in heavily doped silicon, operating in the mid-infrared range (290–310 K). Transfer learning adapted the models for datasets with different temperature ranges, analytes (e.g., butanol), and sMRR designs, achieving comparable accuracy. Variations in accuracy across these datasets are also explored.