Uncertainty Quantification with Invertible Neural Networks for Signal Integrity Applications

We present a machine learning based tool to quantify uncertainty for prediction problems regarding signal integrity. Harnessing invertible neural networks, we convert the inverse posterior distribution given by the network to address uncertainty in frequency responses as a function of design space parameters. As an example, we consider a differential plated-through-hole via in package core and predict S-parameters from its geometrical properties. Results show 3.3 % normalized mean squared error when compared with responses from a fullwave EM simulator.