Surrogate Modeling with Complex-valued Neural Nets and its Application to Design of sub-THz Patch Antenna-in-Package
In this paper, we propose a surrogate model for both forward and inverse modeling with complex-valued neural networks. The complex domain offers the benefits of higher functionality and better representation. To that end, we propose a deep complex dense network by introducing complex dense blocks built with fully-connected layers that support complex operations. We further propose an inverse optimization objective that minimizes the modeling error while optimizing the design space parameters that achieve the target specifications. We apply our proposed approach for the design of a sub-THz patch antenna-in-package operating at 140 GHz frequency band. This surrogate model has the capability of reducing design cycle time and it gives the designer a quick prototype.