Machine Learning Methods for Classification of Organic and Inorganic Compounds Raman Spectra

The study presents a comparative analysis of the machine learning methods effectiveness for classifying Raman spectra to enable automated identification of organic and inorganic compounds. A dataset contains about 2000 spectra of 20 organic and inorganic compounds, obtained using a 785 nm laser source, was compiled for the research. The experimental setup included a laser, optical elements for signal shaping and filtering, and a diffraction gratings spectrometer for data acquisition. Prior to model training, baseline correction and normalization of spectra to the maximum value were performed. The classification algorithms employed were logistic regression, support vector machines, random forest, gradient boosting, k-nearest neighbors (k-NN), as well as a combination of k-NN with dimensionality reduction via principal component analysis. Test experiments performance was evaluated using receiver operating characteristic (ROC) analysis and the area under the curve (AUC) metric was calculated. An analysis of algorithm parameters, runtime, and spectral data processing specifics was conducted, enabling a comprehensive characterization of each method for the given dataset. The implementation of machine learning methods for the identification of organic and inorganic compounds with a signal-to-noise ratio of about 8 and an AUC value of at least 0.95 for binary classification is shown.