Deep Learning Models to Develop V-Doped SnO ₂ Nanofilms to Distinguish between CO and H ₂ for Applications as Sensors

Traditional machine learning approaches for gas sensing often rely on complex feature engineering and offline processing, limiting their direct deployment on resource-constrained edge devices. In this work, we present a hardware-aware Edge AI framework to address the selectivity challenge of vanadium-doped tin oxide (V-doped SnO 2 ) sensors. While the synthesized sensors exhibited high sensitivity, they showed indistinguishable response magnitudes to CO and H 2, causing significant amplitude ambiguity. To overcome this, a lightweight physics-informed Multilayer Perceptron (MLP) model was developed to extract distinct temporal kinetic fingerprints from raw response segments. Notably, the model incorporates an exponential activation function that aligns with the intrinsic Langmuir-type adsorption kinetics, enabling nearly 100% classification accuracy and R 2 > 0.99 with minimal architectural complexity. Practical deployment on a Raspberry Pi 5 validated the system’s extreme efficiency, demonstrating millisecond inference latency, a memory footprint of <5 KB, and negligible accuracy loss (<1%) after INT8 quantization. This work establishes a viable pathway for integrating nanomaterial science with tinyML, enabling intelligent, low-power, and autonomous monitoring systems.