Machine Learning for Automating Millimeter-Wave Directional Coupler Designs

A computational approach to automate mm-wave Directional Couplers can greatly improve upon the current techniques that rely on trial-and-error. To build such an automated framework, we propose a Machine Learning technique that, within a computational time of seconds, can fully automate the physical realizations of mm-wave Directional Couplers over a wide range of electrical specifications and metal options. Our proposed technique includes the training of fully connected neural networks that accurately learn the physical-electrical relationships of coupled lines and includes the multi-head gradient-based optimizers that optimally compute the physical dimensions for given design specifications. By demonstrating the automated designs of mm-wave Directional Couplers with numerous coupling factors and dielectric thicknesses, we verify the effectiveness of our proposed computational method.