Inverse Design of Metasurface based Absorbers using Physics Guided Conditional Diffusion Models

Technical Summary: Inverse Design of Metasurface based Absorbers using Physics Guided Conditional Diffusion Models

Problem Statement
The inverse design of metasurfaces—mapping specific electromagnetic (EM) responses to physical geometries—is a challenging task due to the high dimensionality, nonlinearity, and non-uniqueness of the design space. Conventional methodologies rely on iterative optimization routines coupled with full-wave simulations, which are computationally intensive and time-consuming, often requiring months to converge on a solution. Furthermore, existing deep learning approaches, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), face significant limitations. GANs often suffer from training instability and mode collapse, making consistent conditioning on continuous performance metrics difficult. VAEs frequently produce blurred reconstructions and lack diversity. Additionally, many generative models produce designs with fine or irregular features that are impractical for fabrication, and they often fail to strictly adhere to target spectral constraints during the generation process.

Methodology
The authors propose a Physics-Guided Conditional Diffusion Framework to address these challenges. The core of the approach is a diffusion model trained to generate 3-channel colored images representing metasurface configurations (including geometry, material resistivity, and layering) based on target reflection characteristics and material properties.

Key technical components include:

1. Conditional Generation with FiLM: To integrate conditioning information (target reflection spectra $S_{11}$ and material parameters) effectively, the authors employ Feature-wise Linear Modulation (FiLM). Unlike simple concatenation, FiLM modulates intermediate feature maps within the denoising U-Net using learned affine transformations ($\gamma$ and $\beta$). This mechanism significantly improves conditional alignment and spectral controllability throughout the diffusion denoising process.
2. Physics-Aware Regularization: To enforce adherence to physical laws and target spectra, a pre-trained Surrogate EM Simulator (SES) is embedded into the training pipeline. The SES predicts the reflection spectrum of the generated geometry, and a spectral consistency loss ($L_{spec}$) is calculated as the Mean Squared Error (MSE) between the predicted and target spectra. This loss is added to the standard diffusion loss, guiding the model to produce physically consistent designs.
3. Classifier-Free Guidance (CFG): The model utilizes CFG to balance conditional and unconditional generation, allowing for stronger conditioning during inference by combining conditional and unconditional noise predictions.
4. Dataset and Encoding: The model is trained on a custom dataset of 39,664 samples comprising 30 distinct classes of meta-atom geometries (single and dual-layer) across various commercial substrate materials. The dataset addresses imbalance through stratified splitting and a weighted sampling strategy, ensuring the model learns diverse spectral characteristics, including rare ultra-wideband responses. The meta-atoms are encoded as 3-channel images: Channel 1 for geometric patterns, Channel 2 for resistivity-weighted patterns, and Channel 3 for layer configuration.

Key Contributions

• Novel Framework: Introduction of a diffusion-based inverse design framework specifically tailored for metasurface absorbers, integrating FiLM for high-fidelity conditioning and a surrogate simulator for physics-based regularization.
• Improved Conditioning: Demonstration that FiLM-based conditioning outperforms standard concatenation methods in aligning generated geometries with continuous spectral targets.
• Design Diversity: The model generates multiple geometrically distinct patterns for the same spectral condition, providing engineers with diverse design alternatives rather than a single solution.
• Physical Realizability: By incorporating material properties as inputs rather than outputs and enforcing fabrication constraints during dataset creation, the model ensures generated designs are manufacturable.

Results
The proposed framework was evaluated across the 2–18 GHz frequency range:

• Accuracy: The model achieved an average spectral Mean Squared Error (MSE) of 0.0006 and a Band Alignment Accuracy (BAA) of 0.958 between target and generated spectra.
• Efficiency: The model generates a suitable design in approximately 30 seconds, a significant reduction compared to conventional iterative approaches which can take several months.
• Ablation Studies: Comparative analysis showed that the proposed model (Diffusion + SES + FiLM) significantly outperformed baseline diffusion models, conditional WGAN-GP, and VAEs in terms of spectral accuracy and the percentage of valid designs (100% valid for the proposed model vs. 30-60% for others).
• Experimental Validation: A planar Radar Absorbing Structure (RAS) based on a generated meta-atom was fabricated and measured. The experimental results showed greater than 90% power absorption within the target frequency range (8.5–11 GHz), closely matching both the target and simulated spectra.

Significance
The paper claims that this framework provides an efficient and scalable solution for metasurface inverse design, effectively reducing reliance on computationally expensive EM simulations. By combining generative AI with physics-based constraints, the approach achieves high conditional fidelity while maintaining the ability to produce diverse, manufacturable designs. The successful experimental validation confirms the practical realizability of the method. The authors note that future work could extend this approach to oblique angles of incidence and polarization-dependent responses, further enhancing its applicability to real-world EM design problems.