Quantifying the Limits of Methane Activation in Cu-exchanged Zeolites using Reactive and Interpretable Machine Learning based Potentials

Copper-based zeolites have been widely explored as promising catalysts for the methane valorization reaction to form methanol. These studies are motivated by the hope of finding an elusive ‘Goldilocks’ topology or an active site that shows high methanol selectivity at reasonable methane conversions. As large-scale screening studies with density functional theory (DFT) remain challenging for zeolite catalysts, we now show that a reactive and interpretable machine learning-based potential (rMLP), developed using multistage active learning algorithm and a curriculum-based training (CBT) approach can be used to overcome this bottleneck. Our rMLP approach replaces expensive DFT-based NEB calculations without appreciable accuracy loss. We calculate methane activation barriers for all possible [CuOCu]2+ sites across 52 zeolites with an MAE of 0.07 eV versus DFT. By comparing with known experimental measurements, our approach establishes the limits of methane activation performance across 52 zeolite topologies. Finally, we show that our curriculum-based training (CBT) approach, which relies on several different types of calculations, gradually “teaches” the model about different relevant parts of the PES. This progressive training approach has important implications for the interpretability of emerging machine learning-based approaches.