Automated Design of Broadband Folded-Waveguide Slow-Wave Structures for Traveling-Wave Tubes via Deep Reinforcement Learning

In this article, deep reinforcement learning (DRL) is utilized to explore high-performance folded-waveguide (FWG) slow-wave structures (SWSs) for broadband application in traveling-wave tubes (TWTs). DRL combines deep learning (DL) for perception and reinforcement learning (RL) for decision-making to form a complete automated system from input data to output action control. The deep ${Q}$ -network (DQN), a typical model of DRL is employed to learn how to design FWG SWSs like human beings, and to apply the learned design experience to meet different application requirements. To verify the effectiveness of this method concretely, a DQN is trained in ${G}$ -band to explore the general design policy of FWG SWSs for realizing broadband beam-wave interaction, and it will be exploited to search for well-structured FWG SWSs in the frequency band centered at 850 GHz. Using CST to predict the performance of the structure designed by the DQN, the simulation results show that the reflection coefficient ${S}_{{11}}$ is less than −20.3 dB and the transmission loss is less than 1.95 dB/mm from 750 to 1000 GHz. With a pencil electron beam of 20 kV and 10 mA, the hot simulation results show that the 3-dB bandwidth with the output power greater than 1.6 W and the gain greater than 26 dB can reach 110 GHz from 800 to 910 GHz. Such a design method will facilitate researchers to obtain a practical FWG structure at specified operating conditions, and help to get a deeper insight into the design process from the perspective of artificial intelligence (AI).