Machine Learning for automatic identification of new minor species

One of the main difficulties to analyze modern spectroscopic datasets is due to the large amount of data. For example, in atmospheric transmittance spectroscopy, the solar occultation channel (SO) of the NOMAD instrument onboard the ESA ExoMars2016 satellite called Trace Gas Orbiter (TGO) had produced $\sim$10 millions of spectra in 20000 acquisition sequences since the beginning of the mission in April 2018 until 15 January 2020. Other datasets are even larger with $\sim$billions of spectra for OMEGA onboard Mars Express or CRISM onboard Mars Reconnaissance Orbiter. Usually, new lines are discovered after a long iterative process of model fitting and manual residual analysis. Here we propose a new method based on unsupervised machine learning, to automatically detect new minor species. Although precise quantification is out of scope, this tool can also be used to quickly summarize the dataset, by giving few endmembers ("source") and their abundances. We approximate the dataset non-linearity by a linear mixture of abundance and source spectra (endmembers). We used unsupervised source separation in form of non-negative matrix factorization to estimate those quantities. Several methods are tested on synthetic and simulation data. Our approach is dedicated to detect minor species spectra rather than precisely quantifying them. On synthetic example, this approach is able to detect chemical compounds present in form of 100 hidden spectra out of $10^4$, at 1.5 times the noise level. Results on simulated spectra of NOMAD-SO targeting CH$_{4}$ show that detection limits goes in the range of 100-500 ppt in favorable conditions. Results on real martian data from NOMAD-SO show that CO$_{2}$ and H$_{2}$O are present, as expected, but CH$_{4}$ is absent. Nevertheless, we confirm a set of new unexpected lines in the database, attributed by ACS instrument Team to the CO$_{2}$ magnetic dipole.
Comment: 26 pages, 10 figures