(216h) Machine-Learning Model Development for Electrocatalyst Discovery

Zheng Li, and Hongliang Xin*

*hxin@vt.edu

Session: Catalysis and Reaction Engineering Division- Electrocatalysis and Photoelectrocatalysis

Transition metal alloys, metal oxides and molecular complexes have shown great promise for catalyzing many chemical reactions. For a given type of catalyst, its surface reactivity can be tailored by varying geometric characteristics, metal ligands, organic functional groups and extrinsic factors (e.g., solvation, support) in the vicinity of active sites [1]. With hierarchical complexities in catalyst design, a priori estimation of chemical reactivity of surface atoms is attractive. However, it is very time-consuming and costly, if at all possible, to search for highly optimized catalyst structures by the empirical testing and/or quantum-chemical calculations. Therefore, we tackle this problem by developing the predictive, data-driven machine-learning-augmented models (e.g., artificial neural network, kernel ridge regression) that can unravel the complex, non-linear interactions between the electronic structure properties with reactivity descriptors, thus enabling fast and accurate predictions of surface reactivity of a broad range of material space without resorting to the expensive quantum-chemical calculations.

Since the efficiency and accuracy of machine-learning models are strongly dependent on the input features, we develop various types of data representation of catalyst surfaces. For metal alloys and metal oxides, we engineer numerical representation of surface metal atoms by using geometry and electronic structure based primary features [2-4] including the local electronegativity and the d-band features that are dependent on the surroundings of an adsorption site, together with the intrinsic properties of active metal atoms including the electronegativity, ionic potential, electron affinity, etc. For the description of molecular complexes, we apply the Coulomb matrix [5], which is a simple matrix form crystal structure representation based on the molecular complex structure coordination and atoms physical properties. Another type of representation method for molecular complexes is based on the similarity indexing of molecular complexes with the associating functional groups and structure components. The machine-learning-augmented models exhibit outstanding prediction power in screening for the bimetallic, metal oxide and molecular complexes surfaces with desired adsorption properties. Compared with the traditional high-throughput computational and experimental trial-and-error approach, the machine-learning models have great potential in accelerating the discovery of new catalytic materials with desired properties. Sensitivity analysis of the machine-learning models can shed light on the nature of chemical bonding and its governing factors, providing chemical insights into catalytic processes and fundamental knowledge base of molecule-catalyst interactions.

[1] Vojvodic, Aleksandra, and Jens K. Nørskov. 2015. “New Design Paradigm for Heterogeneous Catalysts.” National Science Review 2 (2): 140–49.

[2] 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.

[3] 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.

[4] Z. Li, S. Wang, W. Chin, and H. Xin, Rapid Screening of Bifunctional Alloy Catalysts Enabled by Machine Learning, Submitted to Journal of Materials Chemistry A. (2017).

[5] M. Rupp, A. Tkatchenko, K.-R. Müller, and O.A. von Lilienfeld, Fast and Accurate Modeling of Molecular Atomization Energies with Machine Learning, Phys. Rev. Lett. 108 (2012) 58301.