TITLE: Conditional Latent Diffusion for Electromagnetic Inverse Design

ABSTRACT: This project tackles an electromagnetic (EMX) inverse design problem: reconstructing binary, physically valid layouts from frequency-domain S-parameter measurements. The task is difficult because the forward simulator is expensive and non-differentiable. We evaluate diffusion-based inverse methods on a progression of problems: a toy linear inverse task, electrical resistance tomography (ERT) with a differentiable forward model, and ultimately EMX. We applied a family of Diffusion Posterior Sampling (DPS) techniques to enforce measurement consistency. Although these methods work on the toy problem, they were unreliable for ERT despite available gradients, showing strong sensitivity to guidance strength and producing unstable reconstructions. These issues worsened for EMX, where gradients are unavailable. We added a zeroth-order gradient approximation to run DPS end-to-end, but reconstruction became prohibitively slow. We therefore shifted to conditional diffusion, embedding measurements directly into the generative model, which was far more stable and effective. Using conditional latent diffusion further improved efficiency and produced high-quality EMX reconstructions with better measurement consistency at much lower cost. We will report the success of the proposed scheme across multiple design topologies.

MAJOR ADVISOR: Alireza Aghasi
COMMITTEE: Arun Natarajan
COMMITTEE: Jinsub Kim