Machine Learning Study of the Surface Reconstructions of Cu$_{2}$O(111) Surface

The atomic structure of the most stable reconstructed surface of cuprous oxide (Cu$_{2}$O)(111) surface has been a longstanding topic of debate. In this study, we develop on-the-fly machine-learned force fields (MLFFs) to systematically investigate the various reconstructions of the Cu$_{2}$O(111) surface under stoichiometric as well as O- and Cu-deficient or rich conditions, focusing on both ($\sqrt{3}$$\times$$\sqrt{3}$)R30° and (2$\times$2) supercells. By utilizing parallel tempering simulations supported by MLFFs, we confirm that the previously described nanopyramidal and Cu-deficient (1$\times$1) structures are the lowest energy structures from moderately to strong oxidizing conditions. In addition, we identify two promising nanopyramidal reconstructions at highly reducing conditions, a stoichiometric and a Cu-rich one. Surface energy calculations performed using spin-polarized PBE, PBE+U, r$^{2}$SCAN, and HSE06 functionals show that the previously known Cu-deficient configuration and nanopyramidal configurations are at the convex hull (and, thus, equilibrium structures) for all functionals, whereas the stability of the other structures depends on the functional and is therefore uncertain. Our findings demonstrate that on-the-fly trained MLFFs provide a simple, efficient, and rapid approach to explore the complex surface reconstructions commonly encountered in experimental studies, and also enhance our understanding of the stability of Cu$_{2}$O(111) surfaces.