Predicting the Solubility of Lignin via Machine Learning

Lignin is a highly promising renewable
resource, but its practical
application faces challenges due to its polydispersity and variability
in solubility. This study utilized real-world characterization data
(gel permeation chromatography (GPC) and HSQC NMR) to construct the
molecular structures of 100 lignins of varying molecular weights.
We used a machine learning (ML) approach, combining structural features
with quantum chemical information, to predict the solubilities of
these lignins in various solvents (calculated using COSMOtherm software).
The machine learning model demonstrated high accuracy (R2 values of 0.987, 0.892, and 0.970, respectively), demonstrating
its effectiveness in predicting lignin solubility based on structure
and solvent properties. Furthermore, SHAP analysis elucidated the
influence of individual molecular features on solubility predictions,
contributing to our understanding of how the lignin structure influences
solubility. This study provides valuable insights into the selection
of highly soluble green solvents and the preparation of monodisperse
lignin.