| On-chip sensing offers a promising solution due to its compactness and capability for real-time environmental monitoring. Silicon photonic SHFT spectrometers enable compact and low-power gas sensing, but their interferometric outputs are highly sensitive to fabrication tolerances and temperature shifts, which distort measured responses and degrade downstream gas classification. In this work, we present a compensation framework based on denoising autoencoders that reconstruct distorted interferograms into stable reference domains: (i) a fixed-temperature reference for thermal drift removal and (ii) an ideal-chip reference for fabrication error removal, enabling the modeling and correction of faults in the edge devices’ outputs. By modeling and correcting these variations, the proposed approach restores device stability, leading to improved prediction accuracy of pollutant detection models. Experimental results confirm that the proposed compensation strategy enhances air pollutant detection performance, enabling robust and reliable on-chip environmental detection across diverse conditions compared to baseline models. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.