Frost which forms on cold surfaces have negative influences in many engineering applications, e.g. air conditioner heat exchangers, aircraft wings and traffic surfaces, etc. Frost formation is particularly a critical issue as it changes thermal properties and influences heat and mass transfer processes. The frost structure is affected by humidity, temperature and flow velocity, etc. However, the relationship between frost structural properties and performance degradation rates is not fully understood. It is desired to incorporate three-dimensional (3-D) frost models in the CFD simulations to investigate the effects of frost local structure on heat and mass transfer. Experimental characterization of frost structure is difficult since it requires complex measurement technique such as in-situ µX-ray CT. Recently, we proposed a new method for 3-D reconstruction of frost structure using a replica method. However, the 3-D reconstruction size and resolution are still limited. In the present study, realistic synthetic 3-D frost structures are synthesized by generative adversarial neural network (GAN). Two approaches are taken. Firstly, 3-D training data is used to train a GAN^3D-3D network. µX-ray CT data was copped and assembled to small size patches. The GAN^3D-3D network trained with the data patches can produce new but statistically equivalent synthetic structures with any chosen size. Secondly, 3-D structure is directly reconstructed from 2-D cross-section images by GAN^2D-3D without conducting additional µX-ray CT measurement. Comparisons of morphological properties between the real and synthetic structures are conducted, which showed good agreement between the predicted and measured data.