(306e) Microkinetic Modeling and Design of Catalytic Materials with Fine-Tuned Machine-Learning Interatomic Potentials

The in-silico modeling and design of catalyst materials must tackle the extreme complexity of chemical reactions at catalytic surfaces. This makes the direct application of density-functional theory (DFT) methods computationally prohibitive, especially when targeting a wide combinatory space of elements of the periodic table. This problem can be addressed with machine learning (ML) techniques, which can significantly reduce the number of DFT calculations required to find minima and saddle points of potential energy surfaces.

In this contribution, we apply DFT and ML to study CO2 hydrogenation reactions on a wide range of metallic catalysts, including single-atom alloys (SAAs), i.e., materials made of single metal atoms dispersed into another metal, able to break the Brønsted-Evans-Polanyi relations. We use graph neural networks (GNNs) interatomic potentials from the Open Catalyst project [1] and DFT to calculate the adsorption energies and activation energies of the reaction mechanisms, and we fine-tune the pre-trained GNN models within an active learning loop. Lastly, we apply structure-dependent mean-field microkinetic modeling [2] to calculate the catalytic performances (activity and selectivity) of the materials, accounting for the different active sites of catalyst nanoparticles.

The application of the framework to the CO2 hydrogenation process allows us to rationalize how reaction mechanisms and catalytic performances change with the catalyst composition, paving the way toward the design and nano-engineering of the catalytic materials. The methodology that we propose can be applied to a wide range of systems in heterogeneous catalysis.

1. L. Chanussot, et al., ACS Catalysis, 11, 10, 6059–6072, 2021.
2. R. Cheula, M. Maestri, Catalysis Today 387, 159-171, 2022.