Abstract
Terahertz (THz) spectroscopy has shown great promise in identifying and quantifying biomolecules whose vibrational modes fall within the terahertz frequency range. This work aims to develop a framework for determining sugar compositions in common chemical reactions that convert C5 and C6 sugars to higher-value products. To simulate these reaction environments, we prepared three types of solid binary mixtures as pellets: glucose-sorbitol, xylose-xylitol, and glucose-fructose, all with varying concentration ratios. Using a terahertz time-domain spectroscopy (THz-TDS) system, we acquired THz spectra of those binary mixture pellets and implemented four types of linear and nonlinear machine learning models to predict sugar compositions from the acquired spectra. Prediction results from test data sets suggest that support vector regression (SVR) shows superior performance over the rest of machine learning models in all binary mixture experiments, with the average and best root-mean-square errors (RMSE) of 5.42 and 2.83% w/w, respectively. Additionally, we investigated the capability of THz spectroscopy to differentiate molecular isomer structures by preparing sample pellets of pure d-xylose and l-xylose. Our results reveal that THz spectroscopy poses challenges in classifying enantiomerism (d-xylose and l-xylose pairs) but still shows potential for identifying functional isomerism (glucose and fructose pairs). These findings demonstrate that THz spectroscopy, combined with optimized machine learning models, offers a promising alternative to gold-standard techniques by enabling simple, rapid, and nondestructive monitoring of chemical compositions during the sugar synthesis.