Parameter Prediction of SIR Dual Band Bandpass Filter Based on Convolutional Neural Network

To solve the problems of long modeling and simulation optimization time and poor design performance of traditional dual band bandpass filters, a convolutional neural network is adopted to predict the parameters of the filter and achieve optimization design in this paper. Firstly, an SIR microstrip dual band bandpass filter was designed based on a three section microstrip stepped impedance resonator unit. On this basis, the S-parameter simulation results of the microstrip filter were used as inputs, and the sensitive structural parameters of the filter were used as outputs. The S-parameter dataset obtained from HFSS electromagnetic simulation was used to train and optimize the model parameters of the convolutional neural network, achieving the prediction of the structural parameters of the filter. Finally, the performance of the filter with the predicted structural parameters was verified by simulation. The results show that when the center frequencies of the microstrip SIR dual band bandpass filter are 3GHz and 4.35GHz, respectively, the in band S11 is less than -20dB and S21 is greater than -0.15dB, which matches the expected filter design specifications well. From the perspective of simulation optimization time, the convolutional neural network model takes 0.1 hours to train in a single session, which is 0.4 hours less than the fully connected neural network optimization method. However, the traditional HFSS simulation debugging optimization method has uncertain time consumption.