(389g) Hierarchical Screening Aided By Machine Learning and Genetic Algorithm to Optimize Catalysts for Plasma-Assisted NH3 Synthesis

Plasma-assisted catalysis is emerging as a “green” alternative to decarbonize and decentralize NH₃ synthesis. However, this kind of process needs to improve its energy yield to become economically feasible. Here, we present a hierarchical screening computational pipeline we developed to search for catalyst compositions that could potentially convert N2/H2 to NH3 more efficiently. In our screening, we use the negative of the (DFT) calculated reaction energy for the HNNH2* + H(g) = HNNH3* step as the ultimate catalytic descriptor, as the above was found in earlier work [1] to robustly correlate with NH3 yields in plasma reactors. However, to reduce computational cost by reducing the search space on which the ultimate descriptor is calculated, we exploited the correlation between the above descriptor and catalyst nitrophobicity. Namely, we set up a first stage where a genetic algorithm (GA) optimizes catalyst composition to maximize nitrophobicity, with the latter being characterized by the N* binding energy, which is calculated by an artificial neural network (ANN) with a mean absolute error around 15 kJ/mol. The latter ANN uses the (single component) density of states (DOS) of the atoms conforming as input and was trained using the data made available at Catalysis-Hub.org [2]. In a second stage, the presumed most nitrophobic compositions from the first stage are verified by through DFT calculated N* binding energies, and the most nitrophobic compositions are taken to the final screening stage, where the ultimate catalytic (reaction energy) descriptor is calculated via DFT. Binary compositions such as Co3Sn, Ag3Pd, Ag3Pb, and “high entropy” compositions such as Cu2Hg2Ni3PbTl2, emerged as compositions with a better value of the reaction energy descriptor than was found in metals previously tested experimentally.

1. Liu et al. ACS Sustainable Chem. Eng.2022, 10, 6, 2034
2. Winther et al. Sci Data 2019, 6, 75