Diffusion-Based Electromagnetic Inverse Design of Scattering Structured Media

Conference: NeurIPS 2025 arXiv: 2511.05357 Code: Available (https://github.com/mikzuker/inverse_design_metasurface_generation) Area: Image Generation / Diffusion Models / Electromagnetic Inverse Design Keywords: Conditional Diffusion Model, Metasurface Inverse Design, Electromagnetic Scattering, FiLM Conditioning, Meta-structure Generation

TL;DR

This paper proposes a conditional diffusion model-based framework for electromagnetic inverse design that directly generates dielectric-sphere metasurface geometries from target differential scattering cross sections (DSCS), bypassing costly iterative optimization. The approach naturally handles the non-uniqueness of the inverse problem and outperforms CMA-ES evolutionary optimization while being orders of magnitude faster.

Background & Motivation

• Value of metasurfaces: Engineered metasurfaces enable precise control of electromagnetic waves, with applications in high-resolution imaging, compact optical devices, and next-generation wireless communications.
• Challenges in inverse design:
• Nonlinear boundary conditions make the structure–response mapping highly complex.
• The design space is high-dimensional.
• One-to-many problem: A single scattering profile can correspond to multiple distinct geometries.
• Bottlenecks of traditional methods: Topology optimization and genetic algorithms rely on iterative simulations, incurring high computational costs and requiring expert tuning (CMA-ES requires 15–20 hours per optimization run).
• Advantages of generative models: Generative models can sample from the distribution of feasible designs, naturally handling the one-to-many characteristic; diffusion models offer stable training and diverse outputs.

Method

Overall Architecture

An end-to-end conditional generation pipeline:

1. Forward simulator (SMUTHI): Computes DSCS for a \(2\times2\) array of dielectric spheres → generates 11,000 training samples.
2. Conditional diffusion model: Learns the inverse mapping from DSCS profiles to metasurface geometries.
3. Inference: Given a target DSCS, samples multiple candidate structures.

Metasurface Parameterization

• A virtual square substrate is divided into an \(N \times N\) grid (here \(N=2\), i.e., 4 cells).
• Each cell contains one dielectric sphere described by 3 parameters: center \((x, y)\) and radius \(r\).
• Encoded as a 1D vector \(\in \mathbb{R}^{3N^2} = \mathbb{R}^{12}\), with all parameters normalized to \([0,1]\).
• Conditional input: DSCS values at 10 polar angles.

Key Designs

1. 1D U-Net Denoising Network

• A one-dimensional U-Net architecture processes the 12-dimensional geometry vector.
• Channel configuration: {16, 32, 64, 128, 128, 64, 32, 16}.
• DDPM framework with 1,000 denoising steps.

2. FiLM Conditioning Mechanism

Feature-wise Linear Modulation (FiLM) is used to inject the target DSCS into the network:

\[\text{FiLM}(F_{i,c}) = \gamma_{i,c} \cdot F_{i,c} + \beta_{i,c}\]

• \(\gamma\) and \(\beta\) are produced by a two-layer network from the 10-dimensional DSCS condition vector.
• FiLM transformations are applied at each layer of the U-Net, enabling the model to adapt to specific electromagnetic response requirements.

3. Noise Schedule

A cosine noise schedule (Nichol & Dhariwal, 2021) is employed:

\[\bar{\alpha}_t = \frac{f(t/T)}{f(0)}, \quad f(\tau) = \cos^2\left(\frac{\tau + s}{1 + s} \cdot \frac{\pi}{2}\right)\]

4. Forward Simulator

• The SMUTHI package (based on the T-matrix method) is used for efficient computation of electromagnetic scattering by spherical objects.
• A fixed refractive index of \(n = 2\) is used as a hyperparameter.
• 11,000 unique samples with corresponding 10-angle DSCS values are generated.

Loss & Training

• Loss function: Standard DDPM denoising loss, minimizing the L2 distance between predicted and true noise:

\[\mathcal{L} = \mathbb{E}_{t, y_0, \epsilon}\left[\|\epsilon - \epsilon_\theta(y_t, t)\|^2\right]\]

• Learning rate \(4 \times 10^{-6}\), batch size 16, trained for 116 epochs.
• EMA decay coefficient 0.995.
• Evaluation metric: Mean Percentage Error (MPE) between the DSCS of generated structures and the target DSCS.

Key Experimental Results

Main Results: Generation Quality and Generalization

Metric	Value
Best single-sample MPE	1.39%
Median MPE over 40 samples	18.91%
Interquartile range	Compact, no significant outliers

• Tested on unseen metasurface configurations (out-of-distribution generalization).
• The best sample nearly perfectly reproduces the target DSCS profile.
• Different samples correspond to distinct geometries yet yield consistent scattering responses—validating the model's natural handling of inverse problem non-uniqueness.

Comparison with CMA-ES Evolutionary Optimization

Metric	Diffusion Model	CMA-ES
MPE	~3%	~5%
One-time training cost	6 hours	—
Per-design inference time	Seconds	15–20 hours
Forward simulation calls	11,000 (one-time)	~105,000 per run
Multi-problem advantage	Amortized cost decreases	Re-optimized each time

Key distinction: The diffusion model's computational cost is one-time; after training, each generation requires no additional simulation calls. CMA-ES requires a full optimization for every new problem instance.

Ablation Study

• Checkpoints are saved every 1,000 training steps; all three MPE statistics (mean, median, standard deviation) show consistent decrease throughout training.
• Stable convergence within 116 epochs demonstrates that the model successfully learns the physics–geometry mapping.

Key Findings

1. Generation quality surpasses optimization: Diffusion model MPE (~3%) outperforms CMA-ES (~5%).
2. Orders-of-magnitude speedup: Post-training inference takes only seconds, versus tens of hours for CMA-ES.
3. Natural handling of non-uniqueness: Different samples yield geometrically distinct yet scattering-equivalent designs.
4. Out-of-distribution generalization: High-quality designs are generated for unseen target DSCS profiles.

Highlights & Insights

• Amortization argument for computational efficiency: Forward simulation cost is 11K calls (one-time) versus 105K calls per problem for CMA-ES—an overwhelming advantage when solving multiple inverse design problems.
• Implicit learning of physical constraints: The model learns the geometry–scattering mapping from data, without explicitly embedding Maxwell's equations.
• Non-uniqueness as diversity: The "ill-posed" nature of the inverse problem becomes an advantage in the generative modeling framework—providing multiple candidate design solutions.
• Practical feasibility: Generated structures are realizable under RF laboratory conditions (~10 GHz, ~30 cm scale).

Limitations & Future Work

1. Scale limitation: Only a \(2\times2\) grid (12-dimensional parameters) is validated; extension to larger grids (e.g., \(4\times4\), \(8\times8\)) remains to be explored.
2. Limited conditioning information: Only 10-angle DSCS values are used as conditions; denser angular sampling may improve accuracy.
3. Fixed refractive index: The refractive index is fixed as a hyperparameter rather than treated as a designable variable.
4. High median MPE (18.91%): While the best samples are excellent, overall average quality has room for improvement.
5. Small training dataset (11K): Scaling behavior with larger datasets and more complex structures has not been explored.
6. Caution in comparison with CMA-ES: The two approaches represent fundamentally different methodologies; a comprehensive comparison requires additional evaluation dimensions.

Related Work & Insights

• An et al. (2019): Used GANs to generate multifunctional metasurfaces, but training instability was a concern.
• Pahlavani et al. (2022): Applied VAEs to generate 3D-printed mechanical metamaterials, demonstrating the potential of generative models for inverse design.
• Bastek et al. (2022): Used deep learning to invert structure–property mappings of truss metamaterials.
• FiLM (Perez et al., 2017): A general-purpose conditioning layer for visual reasoning, here adapted for physics-based conditioning.
• Insight: The application of diffusion models to physical inverse design remains at an early stage, with potential extensions to photonic crystals, antenna arrays, and acoustic metamaterials.

Rating

Dimension	Score	Comment
Novelty	★★★☆☆	The method itself is standard DDPM+FiLM; innovation lies in the application domain.
Technical Depth	★★★☆☆	Architecture is concise and effective, but theoretical contributions are limited.
Experimental Thoroughness	★★★☆☆	Feasibility is demonstrated, but scale is small and comparisons are limited.
Value	★★★★☆	Practical demand for electromagnetic inverse design is clear; speedup is substantial.
Writing Quality	★★★★☆	Concise and clear; problem formulation is well-defined and figures are well-presented.

Related Papers

• [NeurIPS 2025] What We Don't C: Manifold Disentanglement for Structured Discovery
• [ICLR 2026] RIDER: 3D RNA Inverse Design with Reinforcement Learning-Guided Diffusion
• [NeurIPS 2025] A Gradient Flow Approach to Solving Inverse Problems with Latent Diffusion Models
• [NeurIPS 2025] Where and How to Perturb: On the Design of Perturbation Guidance in Diffusion and Flow Models
• [NeurIPS 2025] Schrödinger Bridge Matching for Tree-Structured Costs and Entropic Wasserstein Barycentres