Universal Descriptors of Propane Dehydrogenation Activity at Atomically Dispersed Metal–X Sites (X = O, S, C, and N): Machine Learning-Powered Inverse Catalyst Design

Abstract

Atomically dispersed metal–X (M–X) sites (X = O, S, C, and N) have shown great promise as active centers for propane dehydrogenation (PDH). In this work, the formation energies of adsorbed H at the X site (E_H@X) and coadsorbed H&H at the M–X site (E_H&H@M–X) are identified as two universal descriptors of the PDH activity at M–X sites by establishing their scaling relations with the rate-determining state energies. The derived volcano-shaped activity map is capable of providing a rational interpretation of experimentally reported catalysts. To rapidly predict E_H@X and E_H&H@M–X without DFT calculations, motif-specific features, including intrinsic elemental properties and the number of atoms characterizing the coordination environments, are constructed to train two extra-trees regression models to perform a multitiered virtual screening of 1,659,373 potential catalysts. Pt₁Co₁–Ga₂O₃ is found to show a higher activity than previously known Ir₁–Ga₂O₃, because the substrate imposes a specific geometrical arrangement of the surface Pt, Co, Ga, and O atoms, which strengthens the binding of the rate-determining transition state through Lewis acid–base interactions and therefore imparts specific function to the highly active Pt–O site for PDH. These findings provide a general strategy for the inverse catalyst design by combining DFT-based microkinetic analysis with machine-learning algorithms.