Harnessing Physics-Inspired Machine Learning to Design Nanocluster Catalysts for Dehydrogenating Liquid Organic Hydrogen Carriers

Using liquid organic hydrogen carriers for the trans-oceanic shipment of hydrogen requires selective and low-cost dehydrogenation catalysts. Pt-based bimetallic nanoparticles have emerged as a promising class of materials that catalyse these reactions. Yet, these catalysts are limited by their high cost and poor selectivity. Machine learning methods can accelerate the discovery of catalysts meeting cost and selectivity benchmarks. The state-of-the-art machine learning methods however cannot perform this task because of challenges associated with building predictive models for large cyclic intermediates containing 20+ atoms that adsorb and react on low-symmetry active sites of bimetallic nanoparticles. Using deep learning methods is unfeasible because of the computational cost of building training datasets for such complex molecules. Hence, specific kinds of machine learning methods that are tailored for small datasets are required. Moreover, incorporating physics-based features is also essential to improve the transferability of such models to catalyst compositions outside the training set.<br/><br/>Focusing on methyl cyclohexane dehydrogenation to toluene, an industrially relevant hydrogen carrier, we introduce a machine learning approach to accelerate the design of selective and cost-effective catalysts. Using inputs to a gaussian process regression model that are inspired by physical theories, we predict the adsorption energies of large hydrocarbon intermediates and transition states that are encountered during methyl cyclohexane dehydrogenation. An active learning process is employed to minimize the size of the training set. This active learning process leverages the inherent uncertainty quantification capabilities of gaussian process regression. Across active sites of bimetallic nanoclusters having varied shapes and compositions, our model yields mean absolute errors of 0.11 – 0.25 eV for adsorption and reaction energies on test sets. Moreover, fewer than 50 datapoints per reaction intermediate are required. This performance for large hydrocarbon intermediates is superior to more data heavy deep-learning models that are built to predict the catalytic properties of less complex systems. This model is integrated with a novel lumped microkinetic model to determine the rate and selectivity of catalysts towards methyl cyclohexane dehydrogenation. The microkinetic model reveals that modifying Pt nanoclusters with IB, IIB, and post-transition elements increases dehydrogenation rates, reduces unselective reactions, and lowers Pt utilization, and in turn, the catalyst cost. These observations are consistent with experimental reports from the literature. The scalable and efficient active learning approach introduced in this work marks a significant advancement in designing catalysts for reactions involving large cyclic hydrocarbons that occur on low-symmetry active sites. The framework introduced in this work can be translated to design tailored catalysts for dehydrogenating other types of liquid organic hydrogen carriers, thus accelerating the deployment of such technologies that are vital enablers of the trans-oceanic shipment of low-carbon hydrogen.