Accelerating Inverse Design of Microwave Devices with Scattering Matrix Simulation

Inverse design (InvDes) combines simulation and optimization to algorithmically discover physical device layouts beyond conventional intuition. In practice, conventional design synthesis pipelines require hundreds or thousands of full-wave simulations and many random restarts to navigate a rugged design space—costs that become prohibitive for metallic microwave components, where accurate models demand fine spatial resolution and the objectives are strongly nonconvex. We present a microwave-focused inverse-design framework that delivers large gains in speed and memory while preserving fidelity. We build devices from a library of small building blocks, run rigorous full-wave simulations once per library element to extract its scattering parameters, then simulate candidate designs during optimization with networks of scattering matrices. This surrogate method is approximately 10000 times faster and uses 100-10000 times less memory than conventional full-wave methods, without sacrificing the fidelity needed for inverse design. Coupling the simulator to an adaptive random-search optimizer enables broad, rapid exploration of challenging design landscapes. We demonstrate the efficiency and versatility of our method by designing a range of standard microwave building blocks, including directional couplers, power splitters, phase shifters, impedance-matching networks, and demultiplexers, illustrating both the performance and generality of the approach.