(582ba) Feature Engineering of Machine-Learning Models for Metal Oxides

Zheng Li, and Hongliang Xin*

Department of Chemical Engineering, Virginia Tech, Blacksburg, VA, 24061

*hxin@vt.edu

Session: Catalysis and Reaction Engineering Division – Poster Sessions

Perovskites with the ABO₃ type structure have attracted enormous research efforts in recent years due to their high activity for catalyzing oxygen evolution in (photo)-electrochemical systems. The thermodynamically and kinetically unfavorable oxygen evolution reaction (OER) is considered as a bottleneck for many energy conversion processes such as metal-air batteries and photocatalytic water splitting cells [1]. We aim to design cost-effective metal oxides for catalyzing the sluggish OER. It is, however, very time-consuming and costly to search for highly optimized metal oxides by high-throughput experiments and/or quantum-chemical calculations due to the vast materials space of composition and geometry. In this poster, we extend our previous study [2-3] on metal alloy catalysts to metal oxides by developing a machine-learning-augmented chemisorption model that captures adsorption properties of active sites on metal oxides.

As the chosen of appropriate data representation method is the most critical ingredient for the success of machine learning models, we applied and analyzed various types of data representation methods to describe the metal oxide catalyst surfaces. For instance, we explore electronic based features such as the B-site metal e_g band filling, d-band upper edge, oxidation state, d-orbital splitting, ligand strength, and easily accessible structural based features such as the coordination number, together with the intrinsic properties of metal atoms. By using both electronic based features and structural features, the machine-learning models (e.g., artificial neural network, kernel ridge regression) optimized with available ab initio adsorption energies on perovskite (001) surfaces can capture adsorption energies of *O, *OH, and *OOH intermediates with the root mean squared errors (RMSE) smaller than the DFT-GGA calculation error of 0.2 eV. To shed light on the underlying factors that govern the adsorbate/catalyst interaction, we also employ a feature importance analysis (e.g., exploratory analysis, principle component analysis) to the feature datasets. Compared with the traditional high-throughput computational and experimental trial-and-error approach, the machine-learning chemisorption models exhibit great potential in accelerating the discovery of high-performance oxide materials for electrocatalysis.

[1] J. Suntivich, K.J. May, H.A. Gasteiger, J.B. Goodenough, and Y. Shao-Horn, A perovskite oxide optimized for oxygen evolution catalysis from molecular orbital principles, Science. 334 (2011) 1383–1385.

[2] X. Ma, Z. Li, L.E.K. Achenie, and H. Xin, Machine-Learning-Augmented Chemisorption Model for CO2 Electroreduction Catalyst Screening, J. Phys. Chem. Lett. 6 (2015) 3528–3533.

[3] Z. Li, X. Ma, and H. Xin, Feature engineering of machine-learning chemisorption models for catalyst design, Catalysis Today. 280, Part 2 (2017) 232–238.