A support vector machines framework for identification of infrared spectra

In prior work (Chowdhury, M.A.Z., Rice, T.E. & Oehlschlaeger, M.A., Appl. Phys. B 127, 34 (2021)), we found support vector machines (SVM) to be adept at learning patterns from spectral data within a THz frequency range (7.33-11 cm(-1)) for the purposes of gas-phase speciation. Here, we implement SVM, in a one-versus-rest framework, for the classification of infrared spectra in a broad frequency range (400-4000 cm(-1) or 2.5-25 mu m) for 34 gas-phase compounds at pressures ranging from 0.1 to 1 atm and for absorber mole fractions from 1 ppm to 1 (pure gases). Within the SVM framework, hyperparameters for the classifier were optimized to choose an optimum kernel for the SVM and acceptable soft margin constant to minimize misclassifications. The framework is tested using cross-validation strategies to determine the dependence of performance on variation in pressure and absorber concentration. Validation was carried out by considering experimental absorption spectra, from the literature, in three random trials, where the combined experimental classification accuracy was greater than 90%. A simulated spectral dataset containing artificial noise was used to further evaluate the SVM classifier in studies where the frequency range and resolution were varied, to better interrogate the capabilities of the SVM framework.