Recent experiments show that the grain boundaries (GBs) of copper nanoparticles (NPs) lead to an outstanding performance in reducing CO2 and CO to alcohol products. We report here multiscale simulations that simulate experimental synthesis conditions to predict the structure of a 10 nm Cu NP (158 555 atoms). To identify active sites, we first predict the CO binding at a large number of sites and select four exhibiting CO binding stronger than the (211) step surface. Then, we predict the formation energy of the *OCCOH intermediate as a descriptor for C-C coupling, identifying two active sites, both of which have an under-coordinated surface square site adjacent to a subsurface stacking fault. We then propose a periodic Cu surface (4 by 4 supercell) with a similar site that substantially decreases the formation energy of *OCCOH, by 0.14 eV.