The catalytic activity of a catalyst is fundamentally governed by the atomic arrangement on its surface under working conditions,1 a feature that remains challenging to characterise using conventional electron microscopy or standard density functional theory (DFT) calculations.2 In this study, using acetic acid ketonisation on aluminium-doped magnetite as a model system,3-4 we demonstrate the potential of neural network-based machine learning to predict dopant-induced surface geometries, which can help optimise the optimal doping concentration of aluminium.
We investigated aluminium-doped magnetite for the ketonisation of acetic acid to acetone, a reaction that can be used for bio-oil upgrading. Experimental tests on samples with 0, 3, 6, and 17 at% Al revealed that the 17 at% sample performed best, but the optimal doping level remained unclear. Using Monte Carlo simulations and a neural network potential, we modelled the magnetite (111) surface across a concentration range (0, 1, 3, 6, 10, 13, and 17 at%). The simulations reveal that Al dopants increasingly aggregate into clusters (connected by oxygen), with the total number of clusters peaking at ~13 at%. Crucially, the experimentally measured acetone production rates correlate linearly with this simulated cluster count, predicting that the true catalytic optimum is 13 at% Al—a finding not directly accessible from experiment alone. Experimental validation for the predicted optimum sample is underway. Moreover, all cluster types inhibit the formation of oxygen vacancies, thereby enhancing catalyst stability.
Overall, this work underscores the synergistic power of machine learning and DFT in resolving catalyst surface structures under realistic conditions and in elucidating their impact on catalytic performance. These insights offer valuable guidance for the rational design and optimisation of catalysts.