Abstract
This work describes deep-learning-enabled inverse design of multi-port electromagnetic (EM) structures co-designed with circuits that can enable the synthesis of novel high-frequency on-chip passives and circuits with designer scattering parameters in a rapid and automated fashion. The design of EM structures for high-frequency circuits typically starts from a pre-selected topology of unit functional elements that are subsequently optimized for the desired scattering parameters through time- consuming parameter sweeps, ad hoc optimization algorithms, or prior expert experience. In the space of all possible manufac- turable EM structures, there is no reason to believe that these (and therefore, the co-designed circuits) will be close to being “globally” optimal. Inverse design attempts to take a top-down approach to synthesis of EM structures and circuits by efficiently searching in a vastly larger design space of nearly arbitrary distributed structures for the desirable scattering parameters. To allow search of this design space, we need to eliminate time- and resource-intensive EM simulations in iterative search algorithms. To this end, we demonstrate a deep-learning-based forward model that captures accurately the scattering parameters of any arbitrary planar EM structure on chip and demonstration of rapid synthesis of millimeter-wave (mmWave) EM structures utilizing the aforementioned model. As a proof of concept, we syn- thesize a broadband, low-loss but seemingly arbitra