Invertible Neural Networks for High-Speed Channel Design Parameter Distribution Estimation

High-speed channels have complex topologies with large number of design and process variables. Design and analysis of such channels, for given specifications has been a challenge for the signal integrity community for a while. In this work, we present the use of a unique architecture of invertible neural networks (INN) for predicting the joint posterior distributions in the channel design space, conditioned on the channel passing given specifications in eye-height and eye-width. To provide a proof-of-concept, we perform a full-factorial channel simulation on a 16 Gbps differential pair channel with five variables to calculate the joint posterior distributions for the channel parameters. We then compare them to the ones obtained by the INN. Such an inverse design technique gives a fast channel model when used in the forward direction and a joint posterior distribution estimator when used in the reverse direction for the same cost of training. Channel parameter correlations that are identified based on the estimated joint posterior distributions are also shown.