Dall-EM: Generative AI with Diffusion Models for New Design Space Discovery and Target-to-Electromagnetic Structure Synthesis

Design of RF and mmWave circuits and electromagnetic (EM) structures is typically a designer experience-driven process, aided with time and resource-intensive simulations, iterative design methods, and ad-hoc optimization. The starting points of this process are some preselected parameterized templates that can limit the design space and achievable performance and functionalities. While there are templates that exist for simpler EM structures operable over a narrow range of frequencies (such as simple matching networks, symmetrical power dividers, etc.), for many structures that require more complex spectral responses, there are no set templates. This paper presents a generative AI approach toward rapid synthesis of arbitrary shaped EM structures with designer scattering parameters (S-parameters) utilizing a directed diffusion approach. While these models have been extremely successful in generating complex images, we show for the first time the diffusion model for synthesis of EM structures (represented as images), allowing search space of arbitrary shaped structures. Compared to traditional genetic algorithms, we demonstrate design convergence in seconds. Even compared with prior works with predictive AI models, generative AI approach reduces design time by at least ~10 times. Since diffusion models are known to unearth a much richer design space compared to generative adversarial networks and variational encoders, we believe that this can open up a new dimension for generative AI approaches towards EM and circuit synthesis.