Computational design of nanoalloy catalysts from DFT, genetic algorithms and machine learning

Nanoalloys have been found to display highly promising catalytic properties for a range of chemical and electrochemical reactions, e.g. Pt-based nanoalloys for the oxygen reduction reaction (ORR) in fuel cells and Au-Cu and Cu-Ni nanoalloys for electro-reduction of CO2 (CO2RR) into fuels and chemicals.

The activity, selectivity and stability of the catalysts depends specifically on the composition, structure and ordering of the nanoalloys, and the ability to predict and optimize these properties in silico holds great potential. Here, we combine interatomic potentials and density functional theory (DFT) level calculations with genetic algorithms (GA) and machine learning (ML) to accelerate the search for the compositions, which provide the most stable and catalytically active nanoalloys.

We will present a number of examples for bi- and tri-metallic nanoalloys and core-shell particles, e.g. displaying improved catalytic activity over pure platinum for ORR, and the identification of novel mixed core mixed shell particles for CO2RR. A traditional GA allows the search to be performed for optimisation of a single variable, e.g. stability or activity, but we also present a new ML-GA approach, which allows for a multi-objective optimisation to be performed, where two or more variables can be searched simultaneously.

The MLGA yields a substantial reduction in the number of minimizations that need to be performed to find a range of structures of interest. Finally, a new and computationally fast approach to identify and correct for systematic DFT-errors is presented. The approach, which utilizes an ensemble of different exchange-correlation functionals, is demonstrated for electro-reduction of CO2.

Here, the identification and correction for errors associated with C=O double bonds leads to the identification of the preferred reaction pathways and the design of suitable electro-catalysts for production of formic acid.