Integrated Chemiresistor Arrays for VOC Identification in Air Through Machine Learning and Image Analysis

Chemiresistor-based sensors face challenges detecting chemicals in the presence of moisture. In this study, we identify chemical scents in room air by combining multi-chemistry sensor arrays with machine learning and image analysis. State-of-the art pico-drop technology was used to deposit a diverse set of monolayer coated gold nanoparticles with micron-level precision onto microfabricated electrodes to form chemiresistor arrays. Chemical scents of volatile organic compounds at 0.10, 0.15, 0.25, and 0.40 P/Psat partial pressures in dry air and room air were sampled by the sensor array. Unique chemical-fingerprints were obtained from the distinct chemical interactions between the volatiles and the sensing materials in both dry and humid air conditions. Machine learning classification models were employed for chemical scent identification using both direct sensor responses and image analysis of data radar plots. Conventional random forest, subspace K-nearest neighbors, and neural network models achieved 95-97%, 97-100%, and 90-97% classification accuracy, respectively, for chemical scents at the four partial pressures, both in the absence and presence of room air humidity. A new approach using image analysis based convolutional neural network models achieved 92-100% and 96-100% classification accuracy for the same scents in the absence and presence of moisture, respectively. A combined class of all four concentrations achieved 95-97% accuracy for chemical scents both in dry and room air using conventional machine learning algorithms as well as image analysis models. Results demonstrate that arrays of diverse sensing materials combined with artificial intelligence have potential to improve sensor technologies for a wide range of chemicals in real world environments.