CO₂ Isotopologue Quantification Using Direct Frequency Comb Spectroscopy and Machine Learning

The estimation of molecular concentrations in gas mixtures has applications in healthcare, environmental monitoring, and food quality assessment. Direct frequency comb spectroscopy provides benefits as a sensing platform in these application areas due to its accuracy, species selectivity and capacity for broadband analysis and miniaturization. One challenge associated with direct frequency comb spectroscopy is the accurate estimation of the concentration levels of molecules with overlapping absorption features. Traditional analysis methods, such as curve fitting, do not scale well to mixtures containing many different molecules. Here, we use machine learning methods to estimate the concentration level of two isotopologues of carbon dioxide: ¹²C¹⁶O₂ and ¹³C¹⁶O₂. We compare several machine learning methods for analyzing both transmission spectra and the image output of a virtually imaged phased array spectrometer. Additionally, we evaluate two data augmentation methods to improve the model's generalization ability. Our optimal approach, the transformer method with shifted spectra data augmentation, outperforms curve fitting estimates with a mean absolute error of 14.7 ppm for ¹²C¹⁶O₂ and 0.3 ppm for ¹³C¹⁶O₂, both 3 orders of magnitude smaller than the equivalent curve fitting values.