Accurately distinguishing gases with nearly identical molecular structures─such as nitric oxide (NO) and nitrogen dioxide (NO2)─remains challenging for conventional sensors. We report a palm-sized (5 cm × 5 cm) electronic nose that integrates an ultralow-power microelectro-mechanical systems (MEMS) sensor array with a spatiotemporal deep-learning model (STNet), for trace-level detection and quantification of NO and NO2. The array contains nine carbon-based nanocomposite sensors monolithically fabricated on a 3 mm × 3 mm chip; each sensor operates at room temperature, consumes <2 mW, and achieves detection limits below 0.5 ppm for both gases. STNet combines an enhanced Transformer encoder with a temporal convolutional network, simultaneously capturing intersensor correlations and long-range temporal dependencies. Evaluated on laboratory-generated data sets, the system reduces misclassification rates by up to 50% and improves concentration-prediction accuracy by 25% relative to state-of-the-art CNN and LSTM baselines. Powered and controlled by a smartphone running the embedded STNet model, the device delivers on-site analysis with subsecond latency. By uniting highly selective sensing hardware with efficient edge-level inference, this platform overcomes long-standing limitations in selectivity, portability, and power consumption, offering a scalable solution for environmental monitoring, industrial process control, and medical diagnostics.
Wang et al. (2025) studied this question.