Surface segregation machine-learned with inexpensive numerical fingerprint for the design of alloy catalysts

Abstract

Metal alloy catalysts have been known to enhance many essential reactions. Surface-adsorbate interaction of an alloy catalyst depends on the composition and nanostructure of the surface, which can be changed by a phenomenon called surface segregation. Thus, thermodynamic information on the surface segregation is critical to tune the surface-adsorbate interaction, and thus the reactivity of alloy catalysts. Collected 1,366 density functional theory-calculated segregation energies (Esegr) from literature were featurized by 19-dimensional inexpensive numerical fingerprint, which represents facet-, site-, and elemental-dependencies. Deep neural network (DNN) model was constructed based on the data. Reasonable interpolative and extrapolative prediction by the DNN was clearly demonstrated using principal component (PC) analysis. On top of that, impact of each feature on the Esegr was analyzed by an explainable artificial intelligence approach, giving useful insights for alloy catalyst design based on surface segregation. As an example of the applications, the DNN model was used to explain the formation thermodynamics of site-selective nanosegregated Pt alloy catalyst, and was also used to perform elemental screening to find impurity-host metal combinations feasible to form the nanosegregated alloy catalyst. In addition, the guidance for data collection to improve the DNN model was given by analyzing PC space and performing the weight analysis of PCs.