Theoretical Catalyst\nScreening of Multielement Alloy\nCatalysts for Ammonia Synthesis Using Machine Learning Potential and\nGenerative Artificial Intelligence

In catalyst design,\nit is important to use multielement alloys\nto search for highly reactive atomic compositions and configurations.\nWe developed a rapid calculation method for finding highly reactive\nsurfaces by generating candidates of atomic configurations with high\nactivity using a generative adversarial network (GAN). The evaluation\nof the reaction energy and turnover frequency (TOF) for each catalyst\ncandidate is necessary; however, this is computationally costly with\nconventional methods such as density functional theory (DFT). Here,\nthis was accelerated by the universal neural network potential (UNNP).\nBy iterating the generation and evaluation processes, we explored\natomic configurations that improve the reactivity for the ammonia\nsynthesis of an alloy catalyst comprising Pd, Rh, and Ru. The microkinetic\nanalysis, considering the six elementary reactions in the NH<sub>3</sub> synthesis, was performed on more than 600 catalyst candidates. The\ngenerated ternary alloy surface structures exhibited higher TOFs than\ntheir binary and monometallic counterparts. The highly reactive configurations\nexhibited stronger N adsorption on the Ru–Ru bridge site on\nthe step of the surface surrounded by Pd atoms. This method can be\nemployed for an efficient computational catalyst screening, with detailed\natomistic structure of catalysts and thermodynamics or kinetics properties.