Topology Optimization of Microstrip Lines Using Twin Deep Neural Networks for Performance Prediction and Accuracy Evaluation

抄録

This paper presents topology optimization of microstrip lines using twin deep neural networks (DNNs) for prediction of scattering parameters and its accuracy evaluation. Topology optimization can be accelerated by using a DNN that acts as a surrogate model for time-consuming EM simulations. However, if the prediction accuracy of the DNN for performance prediction is not high enough, the optimization will fail due to misleading caused by prediction errors. To reduce the risk of optimization failure, the present method introduced an additional DNN to evaluate the accuracy of the performance prediction. The proposed method is shown to be effective in avoiding misleading and speeding up the optimization process through numerical and experimental results.