Machine learning and optical emission spectroscopy

Applying machine learning to plasma spectra has potential to expand access to plasma conditions as well as applications in e.g. gas sensors or clinical breath analysis. We have recently demonstrated CH$_{\mathrm{4}}$ detection at 1~ppm using a He low-temperature RF plasma jet and low-cost spectrometer. This identification occurs without dependence on carbon-based spectral lines. However, the high number of data variables (wavelengths) poses a severe challenge for machine learning and to progress to fully realistic gas environments requires greater understanding of the relationship between plasma, spectral features and model development. To this end we investigate the PLS-DA algorithm family which can provide feedback as to the significance of particular variables to a model's success or failure. Here we report models capable of differentiating different molecular species (CH$_{\mathrm{4}}$, H$_{\mathrm{2}}$ and C$_{\mathrm{2}}$H$_{\mathrm{2}})$ down to 1ppm. We use the Variable Importance on Projection (VIP) scores as a primary feature reduction technique to select the most influential wavelengths, which include mainly the major lines (He, H$_{\mathrm{x}})$ as well as impurity lines with C$_{\mathrm{x}}$H$_{\mathrm{y}}$ species having limited effect.