Application of machine learning methods to analyze the presence of harmful impurities in the atmosphere based on spectral data

The work is devoted to the development of an integrated approach to the analysis of the presence of harmful impurities in the atmospheric air. This approach involves: 1) using the measurement results of terahertz absorption spectra of air containing harmful impurities; 2) creating and using a neural network to analyze the data obtained. Sets of model absorption spectra of a gas mixture with different qualitative and quantitative compositions are generated to train the neural network. The application of a neural network to model sets of spectra demonstrated the identification of six gas components with concentrations up to 0.01 ppm. The neural network has achieved 90-95% accuracy in gas detection. A series of experiments were conducted for real gases, showing the sensitivity of the THz spectroscopy method to low concentrations of gases in atmosphere.