Machine-learning for discovery of descriptors for gas-sensing: A case study of doped metal oxides

Conventionally, gas sensors are studied based on functional materials, case by case, using experimental methods. In this study, 872 datasets with 34 features of doped oxides, extracted from the literature, were used to analyze the key features of gas-sensing reactions and understand gas-sensing mechanisms from a global perspective using a genetic algorithm-optimized artificial neural network. Shapley additive explanations were employed to determine the importance and relationships of the features. Based on the physical meaning of the important features, the characteristics of doped oxides and gas molecules are described in terms of sensitivity, measurement range, and selectivity. The reactivity of gas molecules has a strong impact on the sensitivity of doped oxides; sensors with low sensitivity usually have a wide measuring range. The selectivity of doped oxides depends on a suitable doping ratio, oxygen configurations, and the electrophilicity of gas molecules. The best characteristics of the doped oxides in gas sensing were predicted using a generative adversarial net. This study contributes toward improving gas-sensing materials.