Machine Learning Prediction of Molar Mass Distribution of Polymers Forming Fibrous Structures

Abstract Molecular weight parameters, such as molar mass distribution (MMD), are core characteristics of polymer materials. While MMD is routinely measured by gel permeation chromatography (GPC), the procedure takes time and resources and requires specific, relatively costly instruments. Polysaccharides such as cellulose and chitin are particularly difficult to dissolve. Herein, we report on the potential of combining hyperspectral imaging (HSI) with machine learning (ML) to predict, at different length scales, the MMDs of these sparingly soluble fibrous polymers from visible-range HS images (420−730 nm) of the solid substrate. A Gaussian process regression model was trained on principal component representations of the HS cubes against MMDs measured by GPC. HCl-vapor hydrolysis served as a controlled degradation pathway that altered the degree of polymerization without detectable morphological change. Although samples appear identical to the eye and under electron microscopy, HSI captures systematic optical differences that can be empirically related to GPC-measured MMD within the investigated hydrolyzed samples. Predicted distributions matched the GPC reference, with curve RMSEs of 0.009−0.027 (native cellulose paper), 0.017−0.050 (cellulose nanopaper), and 0.016−0.029 (chitin nanopaper). The principal features of the bimodal MMD of cellulose nanopaper and the leveling-off degree of polymerization were both recovered. The workflow uses a microscope, a tunable bandpass filter, and a digital camera, completing HSI acquisition and prediction in a few minutes per sample. This approach provides a non-destructive, solvent-free complement to traditional analytical techniques. The presented case study is a proof-of-concept, calibrated against GPC reference data for the specific materials and protocols studied here. Expanding to other polymer systems would require establishing a similar chromatographic reference for each new system.