Machine learning for gas-phase species spectral identification

Abstract

Rotational and vibrational absorption spectra are well known to provide distinct and complex fingerprints for molecules. Machine learning (ML), and specifically deep learning (DL) methods, are capable of recognizing and generalizing patterns, such as those found in molecular absorption spectra. In this work, ML-based frameworks are developed for the classification of gas-phase molecular absorption spectra for volatile organic compounds (VOCs), halogenated compounds, sulfides, nitrogen-containing compounds, and others of industrial and environmental relevance. When coupled with absorption gas sensors, these frameworks provide automated gas detection. Methods have been developed for the identification of dozens of gaseous molecules based on absorption spectral fingerprints in both the terahertz (THz) and infrared (IR) frequency regions. Models have all been trained on simulated datasets, such that the inconsistent availability of prior experiment does not limit the development of classifiers. Trained models have been validated against simulations, with and without noise, and experiments from our laboratory and the literature. ML and DL classification models have been developed for classifying speciation associated with pure and mixture spectra at average accuracies from 90 to 99%. The DL models have been further examined by gradient- weighted class activation mapping, which provides a visual and interpretable explanation for the models’ internal decision making processes, aiding in gas sensor design. The dataset generation strategy and ML/DL approaches are generalized and can be extrapolated to other spectroscopy types, frequency ranges, and sensors.