We present a new approach for the generation of stable structures of nanoclusters using deep learning methods. Our method consists in constructing an artificial potential energy surface, with local minima corresponding to the most stable structures and which is much smoother than "real" potential in the intermediate regions of the configuration space. To build the surface, graph convolutional networks are used. The method can extrapolates the potential surface to cases of structures with larger number of atoms than was used in training. Thus, having a sufficient number of low-energy structures in the training set, the method allows to generate new candidates for the ground-state structures, including ones with larger number of atoms. We applied the approach to silica clusters $(SiO_2)_n$ and for the first time found the stable structures with n=28...51. The method is universal and does not depend on the atomic composition and number of atoms.