With the development of Internet of Things technology, various sensors are under intense development. Electrostatically formed nanowire (EFN) gas sensors are multigate Si sensors based on CMOS technology and have the unique advantages of ultralow power consumption and very large-scale integration (VLSI) compatibility for mass production. In order to achieve selectivity, machine learning is required to accurately identify the detected gas. In this work, we introduce automatic learning technology, by which the common algorithms are sorted and applied to the EFN gas sensor. The advantages and disadvantages of the top four tree-based model algorithms are discussed, and the unilateral training models are ensembled to further improve the accuracy of the algorithm. The analyses of two groups of experiments show that the CatBoost algorithm has the highest evaluation index. In addition, the feature importance of the classification is analyzed from the physical meaning of electrostatically formed nanowire dimensions, paving the way for model fusion and mechanism exploration.
Keywords: automatic learning; electrostatically formed nanowires; machine learning; selectivity; sensor; volatile organic compounds.