High-throughput screening of bimetallic catalysts enabled by machine learning

We present a holistic machine-learning framework for rapid screening of bimetallic catalysts with the aid of the descriptor-based kinetic analysis. A catalyst database, which contains the adsorption energies of *CO and *OH on {111}-terminated model alloy surfaces and fingerprint features of active sites from density functional theory calculations with the semi-local generalized gradient approximation (GGA), is established and used in optimizing the structural and weight parameters of artificial neural networks. The fingerprint descriptors, rooted at the d-band chemisorption theory and its recent developments, include the sp-band and d-band characteristics of an adsorption site together with tabulated properties of host-metal atoms. Using methanol electro-oxidation as the model reaction, the machine-learning model trained with the existing dataset of similar to 1000 idealized alloy surfaces can capture complex, non-linear adsorbate/metal interactions with the RMSE similar to 0.2 eV and shows predictive power in exploring the immense chemical space of bimetallic catalysts. Feature importance analysis sheds light on the underlying factors that govern the adsorbate/metal interactions and provides the physical origin of bimetallics in breaking energy-scaling constraints of *CO and *OH, the two most commonly used reactivity descriptors in heterogeneous catalysis.