Insights into Supported Subnanometer Catalysts Exposed to CO via Machine-Learning-Enabled Multiscale Modeling

Abstract

Subnanometer catalysts offer high noble metal utilization and superior performance for several reactions. However, understanding their structures and properties on an atomic scale under working conditions is challenging due to the large configurational space. Here, we introduce an efficient multiscale framework to predict their stability exposed to an adsorbate. The framework integrates a comprehensive toolset including density functional theory (DFT) calculations, cluster expansion, machine learning, and structure optimization. The end-to-end machine-learning workflow guides DFT data generation and enables significant computational acceleration. We demonstrate the approach for CO-adsorbed Pd n (n = 1− 55) clusters on CeO 2 (111). Simulation results reveal that CO can facilitate restructuring by stabilizing smaller planar structures and bilayer structures of specific intermediate sizes, consistent with experimental reports. Metal−support interactions, preferential CO adsorption, and metal nuclearity and structure control catalyst stability. The framework allows automatic discovery of stable catalyst structures and a systematic strategy to exploit properties in the subnanometer scale.

Deducing subnanometer cluster size and shape distributions of heterogeneous supported catalysts

“…This can be rationalized by CO preferentially adsorbing on lower wavenumber bridge and threefold sites on the Pd 20 /CeO 2 cluster, while predominantly occupying higher wavenumber atop and bridge sites on Pd 10 /CeO 2 . The preferential adsorption on threefold and bridge sites on larger supported Pd clusters has also been observed in the literature 28 . Thus, we choose to operate in the saturated CO coverage regime for the remainder of our work due to increase in the number of spectroscopic peaks as compared to at differential coverages.…”

“…To tackle these barriers, we determine the ensemble of low-energy metal/adsorbate configurations for each cluster size at a given temperature and CO partial pressure using a cluster genetic algorithm coupled with a Grand Canonical Monte Carlo (GCMC) algorithm 28 . To achieve this, one needs to develop Hamiltonians describing the metal-support, metal-metal, metal-adsorbate, and adsorbate-adsorbate (lateral) interactions using DFT and machine learning.…”

“…We employ a machine-learned Hamiltonian to describe the energy of bare Pd n /CeO 2 and a cluster genetic algorithm to predict low-energy structures 51 . We also developed a second machine-learned Hamiltonian to describe CO adsorption on Pd n /CeO 2 clusters at arbitrary surface coverages that accounted for lateral interactions 28 . Both Hamiltonians were trained using DFT data.…”

Data‐Driven Machine Learning for Understanding Surface Structures of Heterogeneous Catalysts

“…c) Surface phase diagram for Ag(111) exposed to O 2 atmosphere, generated via GCMC. Reprinted with permission [83] . Copyright 2022, American Chemical Society.…”

“…More than 6000 structures were sampled to obtain the surface phase diagram for oxidation (Figure 6c), and random forests (RF) regression was used to determine the structural features that govern surface stability. Wang et al [83] . performed this ab initio GCMC simulation on CO‐adsorbed Pt clusters supported on CeO 2 (111) surface (Figure 6d), and applied a cluster genetic algorithm (CGA) [84] to predict low‐energy structures to obtain the surface phase diagram as a function of temperature and CO pressure.…”