The lowest-energy structures of AgCu nanoalloys are searched for by global optimization algorithms for sizes 100 and 200 atoms depending on composition. Even though the AgCu system is very weakly miscible in macroscopic samples, the mixing energy for these nanoalloys turns out to be clearly negative for both sizes, a result which is attributed to the stabilization of non-crystalline Cu@Ag core-shell structures at the nanoscale. The mixing energy is a quantity nowadays unknown in its functional form, so that its prediction may take advantage of machine learning techniques. A support vector regressor is then implemented to successfully predict the mixing energy of AgCu nanoalloys of both sizes. Moreover, with the help of unsupervised learning algorithms, it is shown that the automatic classification of such nanoalloys into different physically meaningful structural families is indeed possible. Finally, thanks to the harmonic superposition approximation, the temperature-dependent probabilities of such structural families are calculated.