<p>We propose a novel paradigm, which is a blend of generative modelling, deep learning surrogates, as well as reinforcement learning (RL), to automatically design high-efficiency terahertz (THz) nano-antennas in communication systems. The classical design method is highly reliant on manual intuition and trial and error simulation which are computationally expensive and do not necessarily discover non-intuitive geometries. We use a variational autoencoder (VAE) to sample over a broad space of antenna designs, a convolutional neural network (CNN) surrogate to predict antenna electromagnetic performance and an RL agent to optimize design given fabrication constraints. Antenna geometries are geometries coded into a latent space by the VAE, enabling the geometry to be sampled and reconstructed effectively, and the CNN surrogate utilizes this latent code to predict performance as fast as possible, eliminating the costly full-wave simulations. The RL agent also explores the latent space in order to maximize a rewarding choice on the basis of bandwidth, efficiency and manufacturability. Moreover, the framework is dynamically tailored to suit the design of the antennas to suit the range of operating frequencies of photonic THz sources, in order to be compatible with existing communication modules. The outcomes of the experiment demonstrate that the proposed method makes it possible to achieve the work of the antennas with operation in the 2-plus THz range and efficiencies up to 80% that is superior to the performance of the traditional constructions such as bow-tie or slot antennas. Such a combination of generative models and RL does not just allow the design process to go faster, but exposes geometries that are counter-intuitive to design; thereby enabling optimization of data in THz device engineering. This paper is a step forward to closed loop and automated design systems in the next generation technologies of communication.</p>