Novel Discontinuity Modeling with Machine Learning and Application to Microwave Test Fixture De-embedding

Abstract

Novel approach of modeling with recurrent neural networks is proposed and tested for de-embedding of discontinuities (e.g. connectors, probes, etc.) in microwave test fixtures. Very high numbers of parameters of the test fixture are used for generating the model to cover the actual hardware and measurement variation. Excellent results are demonstrated when de-embedding with DUT, compared to the well-known TRL method. It is shown that by taking advantage of the high dimensional regression capability of machine learning, one can take into consideration and mitigate the uncertainty of the discontinuity, as well as of the substrate structure on which the test fixture is built on, arises from variation of calibration kits and repeated measurements. The machine learning approach is very successful in addressing the issues of time consuming, error prone, and computational stability of the very widely used measurement technique of TRL.