Numerical techniques for infrared spectra analysis of organic and inorganic volatile compounds for biomedical applications

We consider the possibility of using a combination of machine and deep learning in the spectral analysis for multicomponent gas mixtures. The experimental setup consists of a quantum cascade laser with a tuning range of 5.3 to 12.8 mu m, a peak power up to 150 mW, and a Herriot astigmatic gas cell with an optical path of up to 76 m. We used acetone, ethanol, methanol, acetaldehyde, and ethylene as test substances for the described techniques. For detection and clustering biomarkers, we used machine learning methods such as t-distributed stochastic neighbor embedding, principal component analysis, and classification methods such as decision tree, k-nearest neighbors, logistic regression, and support vector machines. We used a shallow convolutional neural network (CNN) based on TensorFlow (Google) and Keras for spectral analysis of gas mixtures. We modeled IR spectra of pure substances using an NIST database as training and validation sets. Then we used experimental spectra as a test set. We showed that logistic regression gives us the best result for pure substances' classification. Next, we modeled gas mixtures from synthetic IR spectra for CNN as training and validation sets. We showed that neural networks trained on synthetic spectra can recognize synthetic gas mixtures and experimental individual gaseous substances. We suggest using machine learning methods for pure substances clustering and classification and CNN for gas mixture components identification. We experimentally obtained minimum detectable concentration at ppm level for pure substances. Finally, we estimated that detection limits for the described experimental setup and numerical techniques occurs at levels around 10 to 50 ppb. (C) 2021 Society of Photo-Optical Instrumentation Engineers (SPIE)