Optics-informed deep learning training for classifying gas absorption spectra on the edge

In recent years, the analysis of gases in the air has been crucial, as sensing and monitoring are essential for understanding air quality. However, air monitoring is challenging since it requires battery powered components due to the necessity of deploying sensors and processors in a distributed manner outside with limited power resources. This fueled the development of miniaturized silicon photonic spectrometers, which are able to extract gas absorption spectra with minimum power consumption. However, such analog devices are sensitive to fabrication imperfections, environmental conditions, and temperature, making their use impractical outside of laboratory conditions. Recent approaches apply spatial heterodyne Fourier transform (SHFT) spectrometers enabling the passive calibration with Deep Learning techniques. Although deep learning models are known for their ability to learn the underlying behaviour from noisy data, their complexity hinders their application when lower power consumption is required. To this end, neuromorphic photonics promises several advantages in the deployment of neural networks, enabling high-speed operations with low power consumption. Integrating neural networks into photonic hardware leads to additional training challenges, due to the easily saturated photonic activation functions. In this work, we propose an optics-informed training pipeline for classifying gas absorption spectra utilizing photonic neural networks, taking into account their transfer functions. We provide an extensive experimental study that shows the potential ability of photonic neural networks to be deployed in such scenarios, demonstrating the usage of various photonic activation functions.