Inverse Design of a Microstrip Meander Line Slow Wave Structure with XGBoost and Neural Network

Abstract

We present a new machine learning (ML) deep learning (DL) synthesis algorithm for the design of a microstrip meander line (MML) slow wave structure (SWS). Exact numerical simulation data are used in the training of our network as a form of supervised learning. The learning results show that the training mean squared error is as low as 5.23 × 10−2 when using 900 sets of data. When the desired performance is reached, workable geometry parameters can be obtained by this algorithm. A D-band MML SWS with 20 GHz bandwidth at 160 GHz center frequency is then designed using the auto-design neural network (ADNN). A cold test shows that its phase velocity varies by 0.005 c, and the transmission rate of a 50-period SWS is greater than −5 dB with the reflectivity below −15 dB when the frequency is from 150 to 170 GHz. Particle-in-cell (PIC) simulation also illustrates that a maximum power of 3.2 W is reached at 160 GHz with 34.66 dB gain and output power greater than 1 W from 152 to 168 GHz.