Theoretical Catalyst Screening of Multielement Alloy Catalysts for Ammonia Synthesis Using Machine Learning Potential and Generative Artificial Intelligence

In catalyst design, it is important to use multielement alloys to search for highly reactive atomic compositions and configurations. We developed a rapid calculation method for finding highly reactive surfaces by generating candidates of atomic configurations with high activity using a generative adversarial network (GAN). The evaluation of the reaction energy and turnover frequency (TOF) for each catalyst candidate is necessary; however, this is computationally costly with conventional methods such as density functional theory (DFT). Here, this was accelerated by the universal neural network potential (UNNP). By iterating the generation and evaluation processes, we explored atomic configurations that improve the reactivity for the ammonia synthesis of an alloy catalyst comprising Pd, Rh, and Ru. The microkinetic analysis, considering the six elementary reactions in the NH3 synthesis, was performed on more than 600 catalyst candidates. The generated ternary alloy surface structures exhibited higher TOFs than their binary and monometallic counterparts. The highly reactive configurations exhibited stronger N adsorption on the Ru–Ru bridge site on the step of the surface surrounded by Pd atoms. This method can be employed for an efficient computational catalyst screening, with detailed atomistic structure of catalysts and thermodynamics or kinetics properties.