(538f) An Efficient Method for Predicting Adsorption Energies and Catalytic Performance | AIChE

(538f) An Efficient Method for Predicting Adsorption Energies and Catalytic Performance


Through careful experimental and computational studies over the past few decades, significant progress has been made in developing mechanistic understanding and design principles for many important catalytic reactions on transition metal surfaces. However, even when design principles exists, applying them to design new catalysts is challenging, largely due to the large phase space of alloy surfaces. High-throughput screening using existing computational or experimental methods is difficult and expensive. The goal of this work is to develop an efficient, general model for predicting adsorption energies without the need for quantum chemical simulations, which will be useful in initial screening for a wide variety of reactions.

By combining physical insight with modern machine-learning techniques, we have developed a method for efficiently predicting adsorption energies for a variety of intermediates (including those containing C, H, N, O, and S) on a large phase space of metal alloy surfaces, including alloys of nearly all of the d-block metals. We leverage physical insight to improve generality, and machine learning to improve accuracy. Our methodology uses previously developed expressions for predicting adsorption energies from the electronic structure parameters of a surface.[1] To predict these electronic structure parameters, we use physically motivated features primarily based on tight-binding and electron donation, and employ a robust feature selection process to create low-bias linear models as well as higher bias, higher accuracy nonlinear models. This allows high prediction accuracy as well as error estimation. This methodology was developed and trained using data generated using standard DFT calculations, but requires no quantum chemical simulations when applied for screening.

We apply this method to screen a large number of surfaces for several important reactions, including the oxygen reduction reaction, ammonia synthesis, and CO hydrogenation. We take descriptors from several previous studies where adsorption energies have been used to predict catalytic performance, and predict values for these descriptors on a large number of surfaces, allowing simultaneous high-throughput screening for multiple reactions.

[1] Montemore, M. M.; Medlin, J. W. J. Am. Chem. Soc. 2014, 136 (26), 9272–9275.