Design And Modeling of Oxide Catalysts by machine LEarning and atomistic Simulations

The DAMOCLES project aims to apply data-driven modeling (i.e., molecular simulations combined with machine learning) to study and tailor metal oxide catalysts for CO2 hydrogenation processes (reverse water-gas shift, CO2 methanation, and CO2 to methanol), with the final goal of screening oxides in search of new and better catalytic materials. The starting point of the project is the newly-released OC22 oxide data set (from Meta FAIR and Ulissi's group) comprising approximately 50k adsorption energies of relevant molecules on multi-component oxide surfaces spanning 52 elements of the periodic table. State-of-the-art machine learning models (e.g., Gaussian process regression, and graph neural networks), will be applied for the prediction of relevant adsorption energies not included in the sparsely labeled OC22 dataset. New density-functional theory (DFT) calculations will target the adsorption energy of molecules not considered in the data set and activation energies of the elementary reactions of the CO2 hydrogenation paths. The catalytic performances (i.e., activity and selectivity) of the oxide surfaces in OC22 will be evaluated through microkinetic modeling. Active learning will be applied to iteratively improve the model predictions with additional DFT calculations targeting the parameters selected with sensitivity analysis and uncertainty quantification. The overarching goal of the DAMOCLES project is to understand the effects of the structure and composition of the catalyst surface in the processes of CO2 hydrogenation and identify new promising catalytic materials. This insight can guide experimental researchers in the synthesis of oxide materials with improved catalytic performances, thus helping the development of economically sustainable processes for the transformation of waste CO2 into useful chemicals and fuels.