Feature engineering of machine-learning chemisorption models for catalyst design

We integrate machine-learning algorithms into the descriptor-based design approach for rapid screening of transition-metal catalysts. By engineering numerical representation of surface metal atoms using easily accessible features such as the local electronegativity and the effective coordination number that are dependent on the surroundings of an adsorption site, together with the intrinsic properties of active metal atoms including the electronegativity, ionic potential, and electron affinity, the machine-learning model optimized with similar to 250 ab initio adsorption energies on bimetallic alloys can capture complex, non-linear adsorbate/substrate interactions with the root mean squared errors (RMSE) similar to 0.12 eV. We applied the model to search for {100}-terminated multimetallic copper (Cu) catalysts for electrochemical CO2 reduction where the *CO adsorption energy represents an important efficiency metric. Compared with the traditional high-throughput computational and experimental trial-and-error approach, the machine-learning chemisorption models have great potential in accelerating the discovery of interesting catalytic materials. As the complexity of catalyst structures increases, new features will be needed to learn under-lying correlations and avoid introducing significant errors on top of the average DFT prediction errors expected with standard semi-local generalized gradient approximation (GGA) functionals. (C) 2016 Elsevier B.V. All rights reserved.