A Data-Efficient Training Model for Signal Integrity Analysis based on Transfer Learning

The signal integrity analysis is an essential part of electronic design, while direct high speed signal analysis becomes a time-consuming work. As machine learning exhibits high performance in communication fields in recent years, deep neural network(DNN) is thought to be a key tool to predict eye diagram metrics. However, DNN based signal integrity analysis faces two challenges: demands for amounts of labelled data and long training period. In this paper, we propose a signal integrity analysis model based on transfer learning. The model makes full use of a trained network and trains networks for various channel environments. To achieve the same predicting accuracy, 64% labelled data are utilized for training compared to DNN. The application of hyper-parameters in the neural network improves the prediction accuracy of the eye diagram width and height by 42.7% and 49.24% compared to current methods in few shot signal integrity analysis.