Uncertainty-Aware Deep Ensembles for Robust and Reliable Chemical Sensor Arrays

Selective detection of sulfur-containing gases such as hydrogen sulfide (H₂S), methyl mercaptan (CH₃SH), and dimethyl sulfide ((CH₃)₂S) is important for halitosis-related breath and environmental monitoring. However, conventional metal-oxide chemiresistive sensors show high cross-reactivity toward these chemically similar species, which limits their practical selectivity. To address this, we present a deep-ensemble-assisted electronic nose based on catalytic metal nanoparticles-decorated metal oxide nanofiber arrays. A 15-channel multi-sensor array was fabricated by anchoring diverse metal catalysts (Pt, Pd, Ir, and Co) onto metal oxide nanofiber scaffolds (SnO₂, Co₃O₄, and WO₃) through an intense pulsed light based photothermal process. Based on comprehensive datasets collected from a 15-channel multi-sensor array under varying gas species, concentrations, operating temperatures, and humidity levels, we developed a deep-ensemble learning framework. This architecture simultaneously performs gas-species classification and concentration quantification while providing predictive uncertainty, thereby demonstrating a proof-of-concept reliability-aware platform for VSC sensing with potential relevance to future breath-related and environmental monitoring applications.