Deep Learning Assisted Inverse Design of High-Power Microwave Devices

To expedite the design of high-power microwave (HPM) devices, we developed a knowledge forward network (KFN) which will be trained using a specially generated dataset obtained through the D-optimality method, allowing it to be used under small sample conditions. Subsequently, we developed a discriminator network (DN) and a generator network (GN) separately. The DN is derived from the pretrained KFN and serves the purpose of evaluating whether the designed device meets the design targets. The GN is trained with the assistance of the DN to generate devices that meet the design targets. However, due to the nonlinear interaction between an electromagnetic field and the charged particles of a high-current relativistic electron beam of HPM devices, the prediction accuracy of the pretrained KFN is limited. To address this challenge, the particle-in-cell (PIC) simulation is introduced into the inverse design process as a validation method. The method has been validated in the article, and has successfully generated several relativistic backward-wave oscillators (RBWOs) that meet the design targets.