Electronic noses (e-noses) based on metal-oxide semiconductor (MOX) sensors have been extensively studied for gas detection and pattern recognition, yet they remain isolated instruments with no interoperability between devices. A garlic clove measured on two different devices produces completely different raw voltage readings, because sensor outputs depend on device-specific constants: supply voltage Vcc, load resistance RL, baseline resistance R₀, and environmental conditions. The same garlic measured on the same device on different days produces different readings, because MOX sensors drift with temperature, humidity, and age. We introduce a modular feature framework for digital olfaction: a taxonomy of feature categories that extracts information from MOX sensor time-series along five dimensions — device-agnostic, absolute, temporal, health, and hardware. We formalize the widely-used device-agnostic normalization Rₛ / R₀ and prove it cancels both Vcc and RL completely, enabling theoretical interoperability across any analog MOX circuit regardless of supply voltage or load resistor. We further prove mathematically that this normalization does not cancel differences in the sensor-specific sensitivity constants a and b in the power-law Rₛ / R₀ = a * Cᵇ, which vary across different MOX sensor models—and even across units of the same model due to manufacturing tolerances. Consequently, zero-shot cross-device transfer between any two independently manufactured devices is impossible without calibration; we derive the minimum calibration requirement and outline this as future work. We validate the framework on two independent datasets: (1) session-invariance of 88. 5% classification accuracy on held-out measurement sessions across 50 food substances (SmellNet dataset), significantly above chance level (2%, t = 60. 78, p < 0. 000001) ; (2) long-term drift stability on the UCI Gas Sensor Array Drift Dataset with mean intra/inter separation ratio of 1. 18 across 36 months of real sensor aging. Ablation studies on the device-agnostic feature group show that cross-channel selectivity ratios are the most discriminative component, while per-channel features are individually redundant. In a baseline comparison, the framework outperforms learned representations, including a contrastive 1D-CNN (81. 8%) and the ScentFormer transformer (53. 0%). An informal zero-shot cross-device test between a 3-sensor OpenSmell device and the SmellNet device yields accuracy near chance (11–19%), consistent with the mathematical limits we derive. The framework includes a standardized recording protocol, a principled sensor array design guide, and a discussion of chemical information boundaries. We explicitly acknowledge the fundamental limits of MOX sensors and outline directions for future sensor technologies. All code, data, hardware designs, and documentation are open-source.
Praise James (Tue,) studied this question.