Differential Via Modeling Using Multilayer Perceptron-Sequential (MLP-SEQ) Neural Network

In this paper, an encoder-decoder structured multi-layer perceptron-sequential (MLP-SEQ) networks are proposed to model high-speed differential vias for estimating differential insertion loss (IL) and return loss (RL). Sequential neural networks including recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) are introduced as the decoder NN to treat frequency responses as sequences. The proposed models are validated by finite element method (FEM) simulation results. The accuracy and training times of MLP-RNN, MLP-LSTM, and MLP-GRU models are compared and analyzed. Based on the MLP-LSTM model, various design of experiments (DoEs) are conducted to enhance the reproducibility and reliability of the proposed model. In addition, to further improve accuracy, various methods to enhance long-term memory of the encoder's output feature node are investigated.