Deep Learning for Gas Sensing via Infrared Spectroscopy.

Area of Science:

• Spectroscopy
• Artificial Intelligence
• Environmental Monitoring

Background:

• Deep learning (DL) is increasingly used in spectroscopy and gas sensing.
• Identifying individual gases within multi-component mixtures from infrared (IR) spectra using DL is an underexplored area.
• Accurate gas speciation is crucial for atmospheric and industrial process monitoring.

Purpose of the Study:

• To develop and evaluate a DL model for identifying and quantifying multiple gases in mixtures using IR absorption spectra.
• To demonstrate the model's effectiveness across various atmospheric and industrial gases.
• To investigate the model's internal workings for spectral feature prioritization.

Main Methods:

• A one-dimensional deep convolutional neural network (CNN) model was designed for gas classification.
• A simulated dataset of IR absorption spectra for ten key molecules (e.g., H2O, CO2, O3, NH3) in air was generated using HITRAN data.
• The CNN model was trained on simulated spectra and tested with noisy data and synthetic experimental spectra.

Main Results:

• The DL model achieved high accuracy, ranging from 82% to 97%, in predicting gas speciation in mixtures.
• Class activation maps visualized the model's focus on specific spectral regions for classification.
• The model successfully predicted speciation for synthetic experimental mixture spectra, validating its practical applicability.

Conclusions:

• The proposed CNN model offers a generalized and effective method for multi-component gas speciation using IR spectra.
• The methodology is adaptable to different gases, spectral ranges, and spectroscopy types.
• This work represents the first application of a CNN trained on HITRAN simulations for spectral identification of gas mixtures.