Generative RF Design

The Inverse Design Problem

Traditional antenna design is forward: choose a structure, simulate it, iterate. Inverse design reverses this: specify the desired electromagnetic behavior, let an algorithm find the structure that produces it. This is mathematically harder—the mapping from behavior to structure is non-unique and non-convex—but generative models make it tractable by learning a prior over plausible designs from large training corpora of simulated structures.

Diffusion Models for RF Topology

We adapt score-based diffusion models to the domain of antenna geometry. A forward noising process gradually corrupts known antenna structures into noise; the reverse denoising process learns to generate plausible structures from noise, guided by EM performance objectives as conditioning signals. This allows sampling novel topologies conditioned on target S11 response, gain pattern, or polarization purity—producing designs that meet specifications without any human-guided iteration.

Closing the Loop with Simulation

Generative models require large training datasets. We build automated pipelines that generate candidate RF structures, simulate them via HFSS and openEMS, and label the results. These simulation outputs feed back into the model's training distribution, creating a self-improving design loop: each generation of designs is better than the last, because the model learns from a progressively richer dataset of what works and what does not.