CO₂ Isotopologue Quantification Using Direct Frequency Comb Spectroscopy and Machine Learning.

Area of Science:

• Spectroscopy
• Machine Learning
• Environmental Monitoring

Background:

• Direct frequency comb spectroscopy offers accurate, selective, and broadband gas analysis.
• Estimating concentrations in complex gas mixtures with overlapping absorption features is challenging.
• Traditional curve fitting methods struggle with complex mixtures.

Purpose of the Study:

• To apply machine learning for precise estimation of carbon dioxide isotopologue concentrations.
• To compare different machine learning models and data augmentation techniques.
• To overcome limitations of traditional methods in spectral analysis.

Main Methods:

• Utilized machine learning models (including transformers) to analyze spectral data.
• Investigated analysis of both transmission spectra and virtually imaged phased array spectrometer outputs.
• Employed data augmentation techniques, specifically shifted spectra, to enhance model generalization.

Main Results:

• The transformer method with shifted spectra data augmentation achieved high accuracy.
• Achieved mean absolute errors of 14.7 ppm for ¹²C¹⁶O₂ and 0.3 ppm for ¹³C¹⁶O₂.
• Outperformed traditional curve fitting by 3 orders of magnitude.

Conclusions:

• Machine learning, particularly the transformer model with data augmentation, offers a superior approach for spectral analysis.
• This method accurately quantifies molecular concentrations in complex gas mixtures.
• Advanced spectral analysis holds promise for healthcare, environmental monitoring, and food quality.