Machine-learning-enabled exploitation of gas-sensing descriptors: A case study of five pristine metal oxides

Machine learning, a new research paradigm, is promising for device design, material discovery, and theoretical exploration, overcoming the drawbacks of experimental approaches. In this study, we used machine learning to explore the gas-sensing reactions of metal oxides. The training data were experimental results. Three types of features - metal oxide and gas molecule properties and operating conditions - were extracted as inputs for the machine learning algorithms. The genetic algorithm-adjusted artificial neural network achieved the best performance, and the neural network test was performed using two new experimental datasets. Further, Shapley Additive exPlanations were employed to understand the findings and explore key descriptors for gas sensing. The feature importance was ranked, and the first six features were selected as key descriptors for gas sensing using metal oxides. Furthermore, theoretical investigations were conducted, and the correlations between features and outcomes were analyzed. We found that the active site for gas sensing on pristine metal oxides comprises ionic oxygen loosely bonded to the metal sites. This study provides an example of the use of machine learning for material discovery. We expect that this study will encourage more studies on machine learning for gas sensing.