Quantification of cellulose II using infrared spectroscopy: Machine learning approaches

In cellulose-containing materials, the cellulose II allomorph is often of commercial interest, as it results from regenerated and mercerized fibers (e.g., lyocell and viscose), whereas cellulose III and IV are primarily of scientific interest. This paper presents a machine learning approach for quantifying cellulose II in cellulose-based materials using infrared spectroscopy. Second-derivative preprocessing significantly enhanced the ability to detect cellulose II-related features. Feature importance analysis revealed that the 1400-900 cm^-1 region was highly informative, representing key vibrational modes associated with crystalline and amorphous cellulose, as well as the coexistence of cellulose I and II. By focusing on this spectral range, model robustness was improved while computational cost was reduced. Four regression models-random forest, decision tree, partial least squares regression, and multilayer perceptron-were developed and compared. Among them, the best performance was achieved with the random forest model trained on second-derivative spectra in the selected region (1400-900 cm^-1), achieving an R² of 0.994, RMSE of 0.022, and MRE of 0.052. The best-performing model was also applied to evaluate NaOH-treated wood pulp, CNC, and lyocell. The results of this study demonstrate that the combination of infrared spectroscopy and machine learning provides a promising alternative to conventional X-ray diffraction for the rapid and reliable quantification of cellulose II content.