Neural network techniques for fast parametric modeling of vias on multilayered circuit packages

This paper provides an overview of recent advances of neural network techniques for fast and parametric modeling of vias on the multilayered circuit packages. First, we review a space-mapping neural network technique for broadband and completely parametric modeling of vias. This technique exploits the merits of space-mapping technology and incorporates an equivalent circuit into the model structure. The neural network is trained to learn the multi-dimensional mapping between the geometrical variables and the values of independent circuit elements in the equivalent circuit. Once trained with the EM data, this model provides accurate and fast prediction of the EM behavior of vias with geometry parameters as variables. We also review a combined neural networks and transfer functions technique for via modeling. This technique is capable of providing accurate simulation models even if an equivalent circuit is not available. It retains the EM level accuracy and reduces CPU time significantly compared to EM simulator. Experiments in comparison with measurement data and EM simulations are included to demonstrate the merits of these neural network techniques.