Inverse Design of Folded Waveguide SWSs for Application in TWTs Based on Transfer Learning of Deep Neural Network

This article reports on the design and demonstration of a practical bidirectional fully connected deep neural network (BFC-DNN) for folded waveguide (FWG) slow wave structures (SWSs) in multiple frequency bands, which can be used to speed up the design process of FWG-SWS traveling-wave tube (TWT) with high performance in different frequency bands. The BFC-DNN is first trained to inverse design FWG-SWS using exact numerical simulation results in a form of supervised learning, which shows that the training loss is lower than 0.05. The simulation results of CST demonstrate that the transmission of the structure designed by the BFC-DNN is higher than - 0.1 dB at 34 GHz, with a phase velocity of 0.267 c. Based on the transfer learning, the pretrained BFC-DNN model can be fine-tuned from the Ka-band with a smaller dataset at 850 GHz, and the inverse design of an 850-GHz central frequency FWG-SWS can be successfully generated. The simulation results show that the output power of the structure designed by the BFC-DNN is 2.86 W and the gain reaches 33.6 dB, where the input power is 1.25 mW at 850 GHz.

Automated summary

This article presents a bidirectional fully connected deep neural network (BFC-DNN) for practical design of folded waveguide slow wave structures (SWSs) in multiple frequency bands. The BFC-DNN is trained using exact numerical simulation results in a supervised learning manner, and demonstrates high performance in different frequency bands.