Machine Learning-Enabled Screening and Experimental Validation of Ionic Liquids for High-efficient Cellulose Dissolution

Abstract

Ionic liquids (ILs) are widely utilized for cellulose dissolution due to their high structural tunability and exceptional capacity to disrupt hydrogen bond networks. However, the hard-to-predict cellulose solubility presents a significant challenge and available models for this task remain limited. In this study, based on an up-to-date dataset of 329 cellulose solubility data points across 195 ILs, we first systematically evaluate the performance of COSMO-RS for cellulose-in-IL solubility prediction, which is selected as a representative of mechanism-driven method and a baseline for the further development of data-driven methods. Subsequently, five statistical algorithms combined with two types of IL representations (namely Morgan fingerprints and RDKit descriptors), are rigorously trained and tested for the modelling of cellulose-in-IL solubility following the classical machine learning workflow. In addition, the Transformer-Convolutional Neural Networks is also extended on this task as an assessment of advanced deep learning framework. The best-performing model is employed to guide a large-scale screening of 1224 potential ILs. As a validation, five previously untested ILs are taken to conduct temperature-dependent solubility measurements. This study provides a valuable tool for guiding IL-based cellulose preprocessing toward related material discovery.