Deep Learning for Gas Sensing via Infrared Spectroscopy.

Deep learning methods, a powerful form of artificial intelligence, have been applied in a number of spectroscopy and gas sensing applications. However, the speciation of multi-component gas mixtures from infrared (IR) absorption spectra using deep learning remains to be explored. Here, we propose a one-dimensional deep convolutional neural network gas classification model for the identification of small molecules of interest based on IR absorption spectra in flexible user-defined frequency ranges. The molecules considered include ten that are of interest in the atmosphere or in industrial and environmental processes: water vapor, carbon dioxide, ozone, nitrous oxide, carbon monoxide, methane, nitric oxide, sulfur dioxide, nitrogen dioxide, and ammonia. A simulated dataset of IR absorption spectra for mixtures of these molecules diluted in air was generated and used to train a deep learning model. The model was tested against simulated spectra containing noise and was found to provide speciation predictions with accuracy from 82 to 97%. The internal operation of the model was investigated using class activation maps that illustrate how the model prioritizes spectral information for classification. Finally, the model was demonstrated for the prediction of speciation for two synthetic experimental mixture spectra. The proposed model and the dataset generation strategies are generalized and can be implemented for other gases, different frequency ranges, and spectroscopy types. The multi-component speciation method developed herein is the first application of a convolutional neural network model, trained on HITRAN-based simulations, for spectral identification.