IEEEReflectarrays (RAs) have been attracting considerable interest in the recent years due to their appealing features, in particular, a possibility of realizing pencil-beam radiation patterns, as in the phased arrays, but without the necessity of incorporating the feeding networks. These characteristics make them attractive solutions, among others, for satellite communications or mobile radar antennas. Notwithstanding, available microstrip implementations are inherently narrow-band, and heavily affected by conductor and surface wave losses. RAs based on grounded dielectric layers offer improved performance and flexibility in terms of shaping the phase reflection response. In either case, a large number of variables (induced by the need for independent adjustment of individual unit cell geometries), and the necessity of handling several requirements, make the design process of reflectarrays a challenging endeavor. In particular, RA optimization is extremely expensive when conducted at the level of EM simulation models, otherwise necessary to ensure reliability. A practical solution is surrogate-assisted design with the metamodels constructed for the RA unit elements. Unfortunately, conventional modeling methods require large numbers of training data samples to render accurate surrogates, which turns detrimental to the optimization process efficiency. This work proposes an alternative approach with the unit element representations constructed using deep learning with automated adjustment of the model architecture. As a result, design-ready surrogates can be constructed using only a few hundred samples, and the total RA optimization cost is reduced to only a handful of equivalent EM analyses of the entire array. Our approach is validated using an RA incorporating 3D pyramidal-shaped elements and favorably compared to benchmark techniques. Experimental verification of the obtained design is discussed as well.