Neuromorphic photonic neural networks for on-chip gas spectra classification

The growing concern about air pollution, especially in recent years, has increased the interest in developing systems that perform air monitoring, track air quality, and detect air pollutant gases, like carbon dioxide and methane. Advanced gas detection technologies, including miniaturized silicon photonic spectrometers, are essential for identifying unhealthy living conditions, especially in urban environments, and enable us to monitor air effectively, reduce harmful emissions, and enhance overall environmental quality. These cost-effective and small-sized devices detect specific absorption features of different gases in the atmosphere and have great potential to be used in sensor devices. However, these spectrometers are highly sensitive to fabrication errors and environmental conditions, which can affect their accuracy. Deep learning models have shown promising results in classifying spectra, since they can work with noisy data. Still, their complexity and high resource demands make them impractical for deployment in edge-based environmental monitoring systems. At the same time, photonic neural networks enable high-speed operations with low power consumption and promise several benefits for analyzing gas absorption spectra and detecting the corresponding gas, taking advantage of neuromorphic photonic hardware accelerators. However, training a neuromorphic photonic neural network is not trivial and presents additional training challenges. To this end, in this work, we propose a pipeline for implementing photonic neural networks to classify gas absorption spectra and detect different gases in the environment, considering the limitations of the neuromorphic photonic hardware and spectrometer imperfections. The experimental results indicate the effectiveness of the proposed method, as the real-hardware-integrated model achieves results comparable to those of the software model. • Pipeline for implementing photonic neural network to classify gas absorption spectra. • Training and deploying photonic hardware-aware neuromorphic models. • Utilizing photonic neuromorphic hardware for practical and low-power gas sensing.