Cellulose I crystallinity estimation using a combination of infrared spectroscopy and machine learning approaches

This study presents a novel approach for estimating the crystallinity index (CrI) of cellulose I materials by integrating infrared (IR) spectroscopy with machine learning techniques. Microcrystalline cellulose (MCC) was used as a crystalline standard, while ball-milled MCC, xylan, and lignin powders served as amorphous references to prepare mixture samples covering a CrI range from 0.0 % to 82.2 %. IR spectra of these samples were collected and preprocessed with second-derivative transformation and vector normalization. Random forest (RF) models were employed for spectral variable selection based on mean decrease in impurity (MDI), identifying the 1400-900 cm^-1 spectral region as highly informative for CrI prediction. Subsequent modeling with multilayer perceptron (MLP) using this selected region achieved superior predictive performance with an R² of 0.959 and RMSE of 0.052. Comparative analysis with partial least squares regression (PLSR), decision trees (DT), and RF confirmed the robustness and accuracy of the MLP model. The optimized model was further validated on diverse cellulose materials, including lignocellulosic pulps and cellulose nanomaterials, demonstrating strong agreement with X-ray diffraction measurements. This work provides a set of crystalline and amorphous standards for calibration and offers a rapid and reliable method for cellulose crystallinity estimation, potentially benefiting cellulose science.