Deep Learning-Based Modeling and Synthesis of 3D Electromagnetic Structures

Abstract

To meet and improve system specifications in developing technologies, the Integrated Circuits (IC) industry began exploring 3-D IC design because it addresses the limitations of traditional 2-D chip design, offering enhanced performance, reduced power consumption, and increased integration density. However, the design and optimization of electromagnetic (EM) structures are critical processes for functionality, efficiency, labor, and costs. Traditional design methods are extensive, challenging, and limiting for particular circuit topologies optimal performance. Hence, a robust and nontraditional method is necessary to explore this complex design space efficiently and design an optimal EM structure. Recent research, conducted for planar EM structures from the Integrated Microsystems Research Lab (IMRL) at Princeton University, presents a reverse-model for circuit design, improving design time and creating a new design space to potentially reach a global optimal EM structure. A deep learning-based methodology provides a design procedure led by desired performance metrics, such as scattering parameters, rather than interfacing various electrical elements to execute a task, to then optimize through parametric sweeps. While research following this method developed concrete results, the next step is to utilize this framework to multilayer structures, adding depth and complexity to the traditional design methodologies. The development of 3-D requires techniques that handle tedious and laborious design tasks to be effective and reliable. This thesis presents the first steps towards a feasible 3-D EM synthesizer through MATLAB.