Integrating High-Throughput DFT and Machine Learning for Predictive Design of Hydrogen-Donor Solvents in Coal...

The depletion of conventional oil and gas resources, together with increasing environmental concerns, has renewed interest in direct coal liquefaction, where accurate prediction of reaction energy barriers is critical for mechanistic understanding, process optimization, and the screening of hydrogen-donor solvents. However, the substantial computational cost of density functional theory (DFT) calculations restricts its application in high-throughput exploration of radical-solvent combinations. Here, we develop a reaction-pathway-informed machine learning framework for rapid and accurate energy barrier prediction. By integrating physicochemical descriptors of coal-derived radicals and hydrogen-donor molecules, the baseline model achieves a coefficient of determination (R ²) of 0.925. Further incorporation of structurally resolved descriptors generated from Morgan fingerprints via dimensionality reduction significantly improves predictive performance. Among the evaluated algorithms, eXtreme Gradient Boosting (XGBoost) and Gradient Boosting Regression (GBR) show superior accuracy, achieving R ² values exceeding 0.99 on the training set and 0.933 and 0.947, respectively, on the independent test set. This approach provides a promising data-driven framework for reaction barrier prediction within the investigated hydrogen-donor solvent and coal-radical systems, offering valuable guidance for process optimization and the rational design of high-performance hydrogen-donor solvents in coal conversion technologies.