High-Throughput Machine-Learned Force-Fields Employing Workflow for Heterocatalyst Screening

Heterogeneous catalysis consumes significant energy resources and would thus benefit from the discovery of new catalytically active materials that would lower the energy consumption and therefore reduce the cost of the produced chemicals. The adsorption energies (AEs) of reactants and intermediates of a given chemical reaction are a good descriptor for a catalyst’s activity [1] and could be used for as proxies for catalyst discovery. Modern computing infrastructure and density functional theory (DFT) facilitate the calculation of AEs for many materials with satisfying speed and accuracy. Still, DFT calculations are too expensive for high-throughput scans of thousands of materials when multiple binding sites and different materials facets should be considered for a more sophisticated description of materials activity. Machine learned force-fields (MLFFs) providing offer a solution to this accuracy-efficiency dilemma, since they provide close to DFT accuracy for a fraction of the computational cost.<br/>In this work, we will present our current workflow for obtaining the relevant AEs for CO₂ thermoreduction to methanol. For a list of candidate metals and their alloys, we start with bulk geometries from the Materials Project database [2] with the python API. We then use a trained MLFF from the Open Catalyst Project [3] to calculate the surface stability for all facets with Miller indices {-2, -1, 0, 1, 2}. We pick the most stable cuts for each facet. Subsequently, we identify all possible high symmetry binding sites on those facets and predict AEs for the reaction key intermediates: *H, *OH, *OCHO and *OCH₃ [4]. To validate the MLFF, we compare the AEs for more than 100 materials to results from selected single-point DFT calculations, to estimate the mean absolute error for each material and adsorbate. The last step is important for checking the accuracy of used MLFFs. We find that for more than 80% of materials, the MLFF is accurate to within 0.25 eV for the AEs. Materials with distinctly worser accuracy (above 0.5 eV), e.g. MnCo, MnGa, FeCo, In₃Ru, are flagged and removed from our discovery pipeline. We believe that our workflow can easily be used to speed-up the search for new catalysts, when good accuracy, complexity of materials structure and its surfaces is crucial for a good activity estimation.<br/><br/>[1] J. K. Nørskov et. al. <b>J. Catal. 209</b>, 275–278 (2002)<br/>[2] A. Jain et al. <b>APL Mater. 1</b>, 011002 (2013); https://materialsproject.org/<br/>[3] I L. Chanussot et al. <b>ACS Catal. 11</b>, 6059-6072 (2021); R. Tran et al. <b>ACS Catal. 13</b>, 3066-3084 (2023); https://opencatalystproject.org/ .<br/>[4] P. Amann et al. <b>Science 376</b>, 603–608 (2022).