224G Package Interconnect Design Study - Based on Artificial Neural Network Modeling Approach

Abstract

a new modeling methodology for package interconnect study based on abstraction from physical model to circuit model to Artificial Neural Network (ANN) is proposed here. The mapping between circuit model and physical model is done through S-parameter correlation and the training data for ANN is generated through sweeping variables of the circuit model. The trained network is used to explore design space of package interconnect and to predict its electrical performance. The methodology is applied to 224G PAM4 package interconnect design study. 15 variables are defined for the circuit model and 74K data points are generated for training the neural network. The trained network gives good predictions for the differential insertion loss of package interconnect with a testing Mean Squared Error (MSE) of 0.004 at 56GHz and prediction results can be generated instantly for any variable combinations defined in the circuit model. The proposed modeling methodology has demonstrated effectiveness and good accuracy.