Application of Deep Learning for High-speed Differential Via TDR Impedance Fast Prediction

A deep neural network (DNN) model is developed in this paper for fast prediction of time-domain reflectometer (TDR) impedance for differential vias in high-speed printed circuit boards (PCBs). Unlike traditional empirical linear modeling approaches, the DNN model more accurately maps the nonlinearity between via geometrical parameters and differential impedance. How to select neural network type, training functions and how to select an efficient set of training data are discussed in the paper. Good correlations between the predicted impedances and target values prove the accuracy and reliability of the DNN model. The calculation time for a single data point is reduced to milliseconds, so that the design efficiency of high-speed differential via design is significantly increased.