Summary:

Accurately determining transition state energies is a major bottleneck in the computational screening of catalytic materials. In this work, we applied a machine learning model based on Gaussian process regression and a graph kernel to predict transition state energies efficiently. We applied the model to study the reverse water-gas shift reaction (the hydrogenation of CO₂ to CO and water) on metals and single-atom alloy catalysts, demonstrating a substantial improvement over conventional linear scaling relations. We also estimated uncertainties with model ensembles and showed how these uncertainties propagate to the catalytic activity predictions through the construction of ensembles of microkinetic models. Then, we employed the trained model to screen additional materials and identify promising catalyst candidates. This work underscores the potential of combining advanced graph-based machine learning with atomistic simulations and microkinetic modeling for catalyst discovery, offering a robust and scalable framework for the study of complex catalytic reactions.

View paper