Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diffusion-Based Electromagnetic Inverse Design of Scattering Structured Media

Mikhail Tsukerman, Konstantin Grotov, Pavel Ginzburg

TL;DR

This work tackles electromagnetic inverse design by learning a conditional diffusion model that maps target differential scattering cross-section profiles to metasurface geometries, specifically a $2\times2$ grid of dielectric spheres. By incorporating FiLM conditioning and a 1D U-Net, the method handles the one-to-many nature of inverse scattering and can generate diverse valid designs efficiently after training. The approach, trained on $11{,}000$ samples, achieves a best unseen-target DSCS matching error of $1.39\%$ and substantially outperforms CMA-ES in both accuracy and speed, indicating strong potential for rapid, scalable design of complex metasurfaces. These results suggest diffusion-based inverse design can accelerate exploration of metasurface architectures for applications in photonics and wireless communications, with scalable benefits as problem dimensionality increases.

Abstract

We present a conditional diffusion model for electromagnetic inverse design that generates structured media geometries directly from target differential scattering cross-section profiles, bypassing expensive iterative optimization. Our 1D U-Net architecture with Feature-wise Linear Modulation learns to map desired angular scattering patterns to 2x2 dielectric sphere structure, naturally handling the non-uniqueness of inverse problems by sampling diverse valid designs. Trained on 11,000 simulated metasurfaces, the model achieves median MPE below 19% on unseen targets (best: 1.39%), outperforming CMA-ES evolutionary optimization while reducing design time from hours to seconds. These results demonstrate that employing diffusion models is promising for advancing electromagnetic inverse design research, potentially enabling rapid exploration of complex metasurface architectures and accelerating the development of next-generation photonic and wireless communication systems. The code is publicly available at https://github.com/mikzuker/inverse_design_metasurface_generation.

Diffusion-Based Electromagnetic Inverse Design of Scattering Structured Media

TL;DR

This work tackles electromagnetic inverse design by learning a conditional diffusion model that maps target differential scattering cross-section profiles to metasurface geometries, specifically a grid of dielectric spheres. By incorporating FiLM conditioning and a 1D U-Net, the method handles the one-to-many nature of inverse scattering and can generate diverse valid designs efficiently after training. The approach, trained on samples, achieves a best unseen-target DSCS matching error of and substantially outperforms CMA-ES in both accuracy and speed, indicating strong potential for rapid, scalable design of complex metasurfaces. These results suggest diffusion-based inverse design can accelerate exploration of metasurface architectures for applications in photonics and wireless communications, with scalable benefits as problem dimensionality increases.

Abstract

We present a conditional diffusion model for electromagnetic inverse design that generates structured media geometries directly from target differential scattering cross-section profiles, bypassing expensive iterative optimization. Our 1D U-Net architecture with Feature-wise Linear Modulation learns to map desired angular scattering patterns to 2x2 dielectric sphere structure, naturally handling the non-uniqueness of inverse problems by sampling diverse valid designs. Trained on 11,000 simulated metasurfaces, the model achieves median MPE below 19% on unseen targets (best: 1.39%), outperforming CMA-ES evolutionary optimization while reducing design time from hours to seconds. These results demonstrate that employing diffusion models is promising for advancing electromagnetic inverse design research, potentially enabling rapid exploration of complex metasurface architectures and accelerating the development of next-generation photonic and wireless communication systems. The code is publicly available at https://github.com/mikzuker/inverse_design_metasurface_generation.

Paper Structure

This paper contains 15 sections, 7 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Problem
  3. Diffusion Model for Inverse Design
  4. Forward solver and dataset generation
  5. Model architecture
  6. Training and evaluation
  7. Experimental Results and Discussion
  8. Model Implementation Details
  9. Diffusion Process
  10. Reverse Process
  11. FiLM Conditioning
  12. Model Training Parameters
  13. Ground Truth vs Generated Metasurface Comparison
  14. Feasibility of Generated Structures
  15. CMA-ES Parameters Used For Comparison

Figures (4)

  • Figure 1: (A) Schematic outline of the problem; (B) Parametrization of metasurface with spheres placed in a $2 \times 2$ grid; (C) Schematic of the encoded geometry vector for a metasurface.
  • Figure 2: Architecture of the conditional diffusion model for metasurface generation.
  • Figure 3: Performance of the trained diffusion model. (A) MPE statistics over training epochs. (B) DCSC spectra comparison between ground-truth (black dashed line) and best generated metasurface sample with MPE = 1.39% (red solid line). (C) MPE distribution for 40 generated samples for selected out-of-distribution conditioning.
  • Figure 4: Comparison of the ground truth and generated metasurface.