Neural space mapping approaches to EM-based statistical analysis
Description
Computationally efficient neural space mapping methods for highly accurate electromagnetics-based statistical analysis and yield estimation are described and contrasted in this presentation. Statistical analysis of RF and microwave circuits is realized after a regular space-mapping optimization process, around the space-mapped nominal solutions. Comparisons are presented between different strategies for developing neural networks that implement suitable (input and output) mappings. We first consider developing an input space mapping to enhance available coarse models for efficient statistical analysis. This is realized in a linear fashion (using a Broyden-based algorithm) as well as in a nonlinear manner using a neural network to implement the mapping. The corresponding neural space mapping is “trained” using reduced sets of EM data. We illustrate how both of these input space mapping techniques can lead to inaccurate yield prediction for some problems. Finally we consider the combination of a linear input mapping with a non-linear (neural) output mapping to develop more accurate and equally efficient surrogate models. This strategy allows the transformation of conventional equivalent circuit models into accurate vehicles for EM-based statistical analysis and design. The design and statistical analysis of synthetic problems and microstrip circuits, using commercially available CAD tools, illustrate the techniques.ITESO, A.C.