Advanced applications of Raman spectroscopy and machine learning techniques

Abstract

Raman spectroscopy is considered a great analytical tool in many fields. The obtained spectra can provide valuable information, including sample composition, purity, heterogeneity, etc. Many data analysis techniques have been utilized to extract the information of research interest and build prediction models from complicated Raman spectroscopic data without bias from analysts. As a novel branch of data analysis techniques, machine learning techniques benefit from great accuracy and excellent handling of complicated data. The possible applications of combining various machine learning models and spectroscopic data are endless; one such application being the determination of material properties. Raman spectroscopy and machine learning techniques have been utilized in the water content determination of ionic liquids (ILs). ILs are considered novel “green solvents” for various applications. The water content in ILs greatly influences their properties and application efficiency. Traditionally, the water content is determined through the Karl Fischer titration method; this method is accurate but hard to apply to real-time water content monitoring because of its destructive nature. Raman spectroscopy as an inherently non-destructive technique is especially qualified for real-time water content monitoring. Combined with various multivariate regression analysis methods, Raman spectroscopy has been utilized to determine the water content of ILs at varying temperatures, providing a potentially fast and non-destructive operando method to monitor the water content in ILs when the temperature may be simultaneously changed. ILs and their novel analogs, deep eutectic solvents (DESs), are quite popular in the field of separations due to their high degree of tunability and excellent solubility. The infinite dilution activity coefficients (IDACs) of molecular solutes in ILs and DESs are key solvent selection characteristics in separations. This characteristic reveals the interaction between the solute to be separated and solvent used as separation media. A factorization-machine-based neural network machine learning modeling method was utilized to predict the IDACs of molecular solutes in ILs and DESs from their chemical information. The proposed model showed robust IDAC predictability for molecular solutes and ILs from various families, and the molecular solute IDAC prediction in DESs was achieved for the first time together with ILs. Besides applications in novel solvents, Raman spectroscopy and machine learning techniques have been applied to improve current analysis methods in forensic science. As one of the most popular cosmetic products worldwide, lipsticks are considered useful forensic evidence in crime scenes. The interpretation and comparison of forensic evidence are important since they reflect the potential connection between people and the crime scene. Thus, the differentiation of lipstick samples is worthy of investigation. Combined with Raman spectroscopy, a series of machine learning classifiers, such as linear discriminant analysis, have been utilized to develop a more reliable and efficient lipstick sample differentiation method than the traditional visual inspection forensic method. The excellent classification accuracy of the developed method revealed its potential in the on-site inspections for lipstick samples.