Machine learning-supported solvent design for lignin-first biorefineries and lignin upgrading

Abstract

Solvent selection is a difficult task for lignin-first biorefineries and lignin upgrading as the solvent must satisfy multiple complex technical requirements, while remaining extremely stable to allow recycling. High lignin solubility is a common selection criterion, but the ideal solvent for lignin-first biorefineries also requires non-reactivity towards acids and stabilising reagents encountered in the reaction liquor. To facilitate the search for promising solvents, we developed a computational solvent design framework. The framework consists of a graph-based genetic algorithm for molecular design wherein a graph neural network is used for lignin solubility predictions. Based on these predictions, the genetic algorithm iteratively optimises the molecular structures, inspired by evolutionary strategies, such as selection, cross-over, and mutation. The developed framework designed numerous solvents with high potential for application in lignin-first biorefineries and lignin upgrading. For these solvents, experiments confirmed solubilities between 20 and 60 wt.% across different types of lignin. Notably, several solvents were stable under typical biorefinery process conditions. Furthermore, the explainability of graph neural networks enabled us to link the lignin solubility predictions with structural features of the solvents, providing a clear rationale for solvent selection.