Deep-Learning-Enabled Inverse Design of Electromagnetic Structures

Abstract

The design of electromagnetic (EM) structures and their co-design with related circuits plays an important role in the development of high-frequency circuits and systems. Current design methodologies rely on a bottom-up approach to build electromagnetic structures out of a set of structures whose function is known previously, such as transmission lines. This thesis discusses the use of deep learning to enable a top-down approach to design of EM structures, known as inverse design, where structures are synthesized from a set of desired operating parameters. This has the potential to greatly speed up EM structure design, as well as to find designs closer to the overall optimal in the design space. In this thesis, a tandem neural network approach is used for inverse synthesis of matching networks, a type of EM structure. Tandem neural networks have been shown to be effective in previous works on inverse design for other applications. The work here also serves as a first step in the creation of an integrated codebase/tool for EM structure and circuit co-design in the industry standard Cadence Virtuoso software using the Python programming language.