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

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Materials Science

Background:

• Growing demand for alternative energy sources due to fossil fuel depletion.
• Environmental concerns driving research into cleaner energy technologies.
• Direct coal liquefaction as a key area for coal conversion.

Purpose of the Study:

• To develop a rapid and accurate method for predicting reaction energy barriers in direct coal liquefaction.
• To overcome the computational limitations of traditional Density Functional Theory (DFT) methods.
• To facilitate high-throughput screening of hydrogen-donor solvents and coal-radical interactions.

Main Methods:

• Development of a reaction-pathway-informed machine learning (ML) framework.
• Integration of physicochemical descriptors for coal-derived radicals and hydrogen-donor molecules.
• Utilizing Morgan fingerprints and dimensionality reduction for structural descriptors.
• Evaluation of machine learning algorithms including eXtreme Gradient Boosting (XGBoost) and Gradient Boosting Regression (GBR).

Main Results:

• Baseline ML model achieved a coefficient of determination (R ²) of 0.925.
• Incorporation of structural descriptors significantly improved predictive performance.
• XGBoost and GBR models demonstrated high accuracy, with R ² values exceeding 0.99 on the training set and 0.933/0.947 on the test set.
• The developed framework offers a data-driven approach for predicting reaction barriers.

Conclusions:

• The ML framework provides a computationally efficient alternative to DFT for predicting reaction barriers.
• This approach accelerates the understanding of reaction mechanisms in direct coal liquefaction.
• The findings guide the optimization of coal conversion processes and the rational design of hydrogen-donor solvents.