Component analysis of gas mixtures by THz spectra based on machine learning with training on mixtures or individual components data

This study analyzes different approaches to identifying components in multicomponent mixtures spectra in terahertz spectroscopy using machine learning techniques. A dataset of absorption spectra for mixtures of eight gases, including H₂O, O₃, O₂, CH₄, SO₂, NH₃, NO, and H₂S, using the HITRAN database was generated. The spectra were generated under various conditions, including temperature and pressure. Support Vector Machine and Catboost machine learning algorithms was employed to analyze the absorption spectra of gas mixtures. The accuracy of detection was compared depending on whether training was done on individual component spectra or on mixture spectra. Limits of each approach were determined and accuracy of individual components detection dependencies were established on the substance percentage, pressure and number of components. It is shown that the approach to training on mixtures allows for more accurate identification of the components of mixtures containing a larger number of gases. The results demonstrate the effectiveness of these methods in accurately identifying components, even in complex mixtures containing multiple components with overlapping spectral lines.