Understanding nitrogen adsorption in compositionally disordered oxides through descriptor-based machine learning

초록

The adsorption strength of atomic nitrogen has long been employed as a central activity descriptor in ammonia synthesis and decomposition catalysis. However, on compositionally disordered multicomponent oxide surfaces, where local environments and adsorption-induced structural rearrangements vary significantly, the factors associated with variations in nitrogen adsorption energetics remain unclear. In this study, nitrogen adsorption on rocksalt-type multicomponent oxide surfaces composed of equimolar Ni, Mg, Cu, and Zn cations was investigated using density functional theory (DFT) calculations combined with machine learning (ML) analysis. A diverse set of adsorption configurations revealed a broad distribution of adsorption energies arising from local compositional heterogeneity. To rationalize these variations, physically motivated descriptors derived solely from pre-adsorption surface properties, capturing electronic, compositional, and geometric characteristics of adsorption sites, were constructed. Correlation and feature importance analyses indicated that, among the descriptors considered in the present models, electronic and compositional descriptors associated with the adsorption-site metals provided greater predictive contributions than the selected geometric descriptors. The evaluated ML approaches showed comparable predictive performance across diverse local environments. Overall, the results demonstrate that pre-adsorption electronic and compositional descriptors retain meaningful predictive utility for nitrogen adsorption energies across configurations exhibiting different degrees of adsorption-induced structural relaxation. © 2026 Elsevier B.V.

키워드