Inverse design of a novel multiport power divider based on hybrid neural network

In this study, we propose an inverse design approach based on a neural network for a novel multiport power divider (MP-PD) with complex geometry. The inverse design approach is obtaining geometry from the desired physical performance to address the challenge of conventional methods. We develop a hybrid neural network model for this inverse design. The backbone architecture incorporates a bidirectional long short-term memory module, a multihead self-attention module, and convolutional modules. This hybrid neural network is employed to capture the feature of physical performance and learn the relationship between the geometric structure of the proposed MP-PD and its corresponding physical performance. Consider the design of the power divider as an end-to-end methodology that directly maps design requirements to optimal geometric parameters. The neural network transfers the designed process into multiple-input-multiple-output. We adopt the network model to successfully predict 20 geometric parameters of MP-PDs for two distinct operating frequencies. The two operating frequencies are those utilized in real engineering applications, which are 3.5 GHz in the 5G band and 2.45 GHz in the trackside communication band. The predicted MP-PD improves the return loss and bandwidth by 8.05 dB and 0.25 GHz, respectively, over the desired performance. The experiments and comparisons demonstrate the effectiveness and accuracy of our inverse design approach. The efficiency and flexibility of design are also significantly improved by the hybrid neural network model.