Graph-based machine learning model for adsorption energy prediction of CO2 hydrogenation intermediates on metal oxide catalysts
Year: 2026
Type: preprint
Abstract: Mixed metal oxide catalysts are highly active for CO2 hydrogenation to methanol and typically more selective than classical Cu-based catalysts. Adsorption energies of key reaction intermediates can serve as activity descriptors for high-throughput screening of metal oxide-dopant combinations, but this requires numerous computationally demanding density functional theory (DFT) calculations. Machine-learning interatomic potentials (MLIPs) and adsorption energy prediction models offer faster alternatives, yet current architectures struggle with the structural and electronic complexity of mixed metal oxides introduced by dopants, oxygen vacancies, and diverse adsorption motifs. In this work, a graph-based Gaussian process regression (GPR) model is developed, combining graph representations of adsorption structures with node-level electronic descriptors. Trained on 638 adsorption structures of CO and all formate-pathway intermediates on pristine and Zn-doped ZrO2 and HfO2, as well as pristine In2O3 surfaces, the model achieves a test-set mean absolute error (MAE) of 0.16 eV. Leave-one-feature-out analysis identifies the d-band center, surface work function, and adsorbate LUMO energy as the dominant descriptors governing the prediction accuracy. Coupling the model with UMA small v1.2-D3 establishes an efficient workflow for rapid structural relaxation and accurate adsorption energy prediction, enabling high-throughput screening of mixed metal oxide catalysts for CO2 hydrogenation to methanol.
Source: ChemRxiv
Institutions: TU Wien, Universitat de Barcelona
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220868799