Detecting trace methane levels with plasma optical emission spectroscopy and supervised machine learning

Trace methane detection in the parts per million range is reported using a novel detection scheme based on optical emission spectra from low temperature atmospheric pressure microplasmas. These bright low-cost plasma sources were operated under non-equilibrium conditions, producing spectra with a complex and variable sensitivity to trace levels of added gases. A data-driven machine learning approach based on partial least squares discriminant analysis was implemented for CH4 concentrations up to 100 ppm in He, to provide binary classification of samples above or below a threshold of 2 ppm. With a low-resolution spectrometer and a custom spectral alignment procedure, a prediction accuracy of 98% was achieved, demonstrating the power of machine learning with otherwise prohibitively complex spectral analysis. This work establishes proof of principle for low cost and high-resolution trace gas detection with the potential for field deployment and autonomous remote monitoring.