Selective Sensing of Hydrogen and Ammonia Using a Single CMOS-Compatible Sensor and Transfer Learning Methods

The development of high-performance gas sensors is crucial for ensuring safety and efficiency in the emerging hydrogen economy, particularly for detecting hydrogen (H₂) and ammonia (NH₃), which are essential for hydrogen storage, transportation, and energy applications. Hydrogen is highly flammable, with a lower explosive limit of 4%, while ammonia is toxic and can cause severe health hazards; thus, their early and accurate detection is critical to prevent accidents and ensure safe handling. However, most hydrogen sensors exhibit cross-sensitivity to ammonia, making it challenging to distinguish between the two gases. Additionally, blends of ammonia and hydrogen are considered as alternative fuels to achieve zero-carbon emissions. Detecting them in mixture form is essential, as the flammability and toxicity limits of the mixture differ from those of the individual gases, requiring precise monitoring for safety, process optimization, and efficient fuel utilization. In this study, we employ palladium (Pd) nanoparticle-decorated electrostatically formed nanowire (Pd-EFN) sensor for the selective detection of H₂, NH₃, and their mixtures at low concentrations. The EFN sensor, a multiple-gate depletion-mode field-effect transistor (FET) fabricated using complementary metal-oxide-semiconductor (CMOS)-compatible processes, provides unique multigate electrostatic control, enabling enhanced sensitivity and selectivity. Experimental results demonstrate a highly reversible response, with distinct "electrostatic fingerprints" observed across different back-gate voltages, allowing for improved gas differentiation. Using supervised machine learning techniques including Linear and Kernel Support Vector Machine, AdaBoost, Gradient Boosting, Extra Trees, Random Forest, Decision Tree, Linear Discriminant Analysis, and K-Nearest Neighbors, we achieved up to 94% classification accuracy in distinguishing H₂ vs NH₃ and H₂ vs (NH₃ + H₂), respectively. Additionally, adopting a transfer learning approach using the VGG-19 neural network and leveraging sensor response maps as inputs, further improved accuracy to approximately 97 and 96%, respectively. Furthermore, the ability to discern the individual gases and the mixture (H₂/NH₃/(NH₃ + H₂)) was improved from 77 to 87% with the use of transfer learning. The ability to selectively identify individual gases and their mixtures using a single sensor with high accuracy, without the need for sensor arrays, paves the way for advanced, miniaturized, and cost-effective gas sensing platforms, demonstrating potential for real-world applications in hydrogen safety and environmental monitoring.