Machine Learning Based Uncertainty Quantification of Extrapolated Design Space and Frequency Response for RF Structures

Current deterministic machine learning models provide point estimates for predictions without any metric quantifying its inaccuracy for test inputs. In this paper, we focus on uncertainty analysis for a recently developed machine learning model used for design space and frequency response extrapolation using variational inference. This information equips the designer to identify how well the model performs for a given test input and hence identify if further training is required. We also explain here how much data is enough to train this model well. We discuss these approaches for a 5th order interdigital bandpass filter at 28GHz.