Towards understanding temperature modulated SMOX gas sensor arrays for outdoor air quality applications

Monitoring of pollutant gases such as ozone has become of paramount importance due to the impact of air quality on both health and environment. Gas sensors based on Semiconducting Metal Oxides (SMOXs) are cost and size efficient and offer excellent sensitivity. Their selectivity can be greatly improved by operating them in a cyclic temperature modulation mode combined with advanced chemometric models. In this work, we present a convolutional neural network (CNN) architecture that is trained on the resistance readout of a single commercial multi-pixel SMOX gas sensor. The trained models outperform other regressors, not only in the prediction of pollutant gases over a wide range of concentrations in gas mixtures under dynamic laboratory conditions, but also in quantifying the ozone concentration in real outdoor air. A SHapley Additive exPlanations (SHAP) analysis is carried out to interpret and compare the different models and their extracted features of the temperature cycle. • Temperature modulated SMOX sensors increase the chemometric performance. • Pollutant gas concentrations can be predicted in laboratory and real conditions. • CNN models are suitable for the prediction and robust in terms of generalization. • Depending on the test conditions, relevant temperature cycle features are different.