Generation of Stochastic Interconnect Responses via Gaussian Process Latent Variable Models

We introduce a novel generative model for stochastic device responses using limited available data. This model is oblivious to any varying design parameters or their distribution and only requires a small set of "training" responses. Using this model, new responses are efficiently generated whose distribution closely matches that of the real data, e.g., for use in Monte-Carlo-like analyses. The modeling methodology consists of a vector fitting step, where device responses are represented by a rational model, followed by the optimization of a Gaussian process latent variable model. Passivity is guaranteed by a posteriori discarding of nonpassive responses. The novel model is shown to considerably outperform a previous generative model, as evidenced by comparing accuracies of distribution estimation for the case of differential-to-common mode conversion in two coupled microstrip lines.