(290e) Deriving Descriptors and Physical Understanding from Data in Catalysis Research

Important properties of a catalytic material, such as activity and selectivity, are in general difficult to predict, in particular from first principles. The problem lies in the extreme complexity of the relation between the atomic composition of a catalyst and its catalytic properties at realistic temperatures and pressures. We demonstrate how to bridge this complexity with artificial intelligence (AI).

Single-atom metal alloy catalysts (SAACs) have recently become a very active new frontier in catalysis research. The simultaneous optimization of both facile dissociation of reactants and a balanced strength of intermediates’ binding make them highly efficient and selective for many industrially important reactions. However, discovery of new SAACs is hindered by the lack of fast yet reliable prediction of the catalytic properties of the sheer number of candidate materials. In this work, we address this problem by applying a compressed-sensing data-analytics approach SISSO [1] parameterized with density-functional inputs. Our approach is faster and more accurate than the current state-of-the-art linear relationships. Besides consistently predicting high efficiency of the experimentally studied Pd/Cu, Pt/Cu, Pd/Ag, Pt/Au, Pd/Au, Pt/Ni, Au/Ru, and Ni/Zn SAACs (the first metal is the dispersed component), we identify more than two hundred yet unreported candidates [2]. Some of these new candidates are predicted to exhibit even higher stability and efficiency than the reported ones. Our study demonstrates the importance of breaking linear relationships to avoid bias in catalysis design, as well as provides a recipe for selecting best candidate materials from hundreds of thousands of transition-metal SAACs for various applications.

Moreover, using subgroup discovery, an AI approach that discovers statistically exceptional subgroups in a dataset, we develop a strategy for identification of most important parameters of a catalytic material and competing mechanisms of a catalytic reaction. The approach is used to develop physical understanding of hydrogen activation at SAAC's [2] and address the problem of converting carbon dioxide (CO2) to fuels and other useful chemicals [3]. Large-scale conversion would alleviate detrimental effects of CO2 excessive emissions on the environment. The AI model is trained on high-throughput first-principles based data for a broad family of oxides. We demonstrate that an electron transfer to the pi*-antibonding orbital of the adsorbed molecule and the associated bending of the gas-phase linear molecule, previously proposed as the indicator of activation, are insufficient to account for the good catalytic performance of experimentally characterized oxide surfaces. Instead, our AI approach identifies the common feature of these surfaces in the binding of a molecular O atom to a surface cation, which results in a strong elongation and therefore weakening of one molecular C-O bond. This finding suggests to use the C-O bond elongation as an indicator of CO2 activation. Based on these findings, we propose a set of new promising oxide-based catalysts for CO2 conversion, and a recipe to find more.

[1] R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, and L. M. Ghiringhelli, Phys. Rev. Mater. 2, 083802 (2018).
[2] Z.-K. Han, D. Sarker, R. Ouyang, A. Mazheika, Y. Gao, and S. V. Levchenko, Nature Comm. 12, 1833 (2021).
[3] A. Mazheika, Y. Wang, R. Valero, L. M. Ghiringhelli, F. Vines, F. Illas, S. V. Levchenko, and M. Scheffler, arXiv:1912.06515 [cond-mat.mtrl-sci].