Sharp Increase in Catalytic Selectivity in Acetylene Semihydrogenation on Pd Achieved by a Machine Learning Simulation-Guided Experiment

Pd (metal) is of key value as a heterogeneous hydrogenation catalyst for its high activity and stability. It, however, fails in selective acetylene hydrogenation: at high H-2 pressures and 100% conversion, the dominant product is ethane, not the desirable ethene. Despite decades' efforts, even the structure of the catalyst remains to be veiled by the in situ formation of unknown PdHx and, controversially, PdCx phases. Here, by combining our recently developed machine learning potential global optimization, microkinetic simulation, and catalysis experiment, we resolve a Pd4H3 phase formed under hydrogenation reaction conditions (e.g., 298 K and p(H-2) > 0.1 atm), in which the exposed Pd4H3(100) open surface is the most responsible for catalyzing the deep hydrogenation to ethane at high H-2 pressures. This finding is rooted in the thermodynamics phase diagram for the Pd-H bulk and surfaces from millions of structure candidates explored by stochastic surface walking (SSW) global optimization and the lowest-energy pathways for the hydrogenation on different surfaces. Guided by the theoretical prediction, Pd catalysts with a large particle size (26 nm) dominated by the close-packed (111) surface are synthesized and tested for selective acetylene hydrogenation in comparison with that of the commercial Pd/C catalyst (particle size similar to 2 nm). We show that simple nanostructure engineering improves markedly the selectivity by 16 times, from 4.5% for the commercial Pd catalyst to 76% for our designed Pd catalysts at 100% acetylene conversion and 293 K, showing great promise for machine learning-guided catalyst design. General guidelines to further improve catalyst selectivity are proposed.