nati-nissan/MetaMamba

---
license: mit
viewer: false
language:
- en
tags:
- electromagnetics
- metamaterials
- metasurface
- inverse-design
- generative-ai
- mamba
- state-space-models
- physics-informed-ai
- scattering-parameters
- huygens-metasurface
- fine-tuning
pretty_name: "MetaMamba: Multilayered Huygens' Metasurface Design Dataset"
size_categories:
- 100K<n<1M
task_categories:
- tabular-regression
---

# MetaMamba: Multilayered Huygens' Metasurface Design Dataset

This dataset accompanies the two-part article:

> **Harnessing Selective State Space Models to Enhance Semianalytical Design of Fabrication-Ready Multilayered Huygens' Metasurfaces**
>
> - **Part I** — *Field-based Semianalytical Synthesis*
> - [](https://arxiv.org/abs/2603.03837)
> - **Part II** — *Generative Inverse Design (MetaMamba)*
> - [](https://arxiv.org/abs/2603.03877)

**GitHub Repository (code + processing scripts):** https://github.com/nati-nissan/MetaMamba

---

## Overview

This dataset supports forward surrogate training and CST calibration for the MetaMamba pipeline — a generative inverse design framework for multilayer transmissive Huygens Metasurface (HMS) unit cells.

The design problem: given a target scattering response (transmission efficiency |T|² and phase φ), recover the five-layer Jerusalem-Cross (JC) patch geometry **W** = (W₁, …, W₅) that realizes it. This is an inherently one-to-many inverse problem, addressed by the AR-Mamba generator described in Part II.

Two data sources are provided, covering both a single design frequency and a broadband frequency range:

| Source | Description |
|--------|-------------|
| **LAYERS (SA)** | Large-scale semi-analytical data generated by the LAYERS solver (introduced in Part I). Fast, physics-based, used for surrogate pretraining. |
| **CST Microwave Studio** | High-fidelity full-wave simulations under periodic Floquet boundary conditions. Small but accurate; used for surrogate calibration. |

---

## Files

| File | Source | Frequency | Samples | Role in pipeline |
|------|--------|-----------|---------|-----------------|
| `sa_dataset_20ghz.csv` | LAYERS (SA) | 20 GHz | ~524,000 | Forward surrogate **pretraining** |
| `cst_dataset_20ghz.csv` | CST | 20 GHz | 1,080 | Forward surrogate **calibration** |
| `sa_freq_resp_18_to_22.csv` | LAYERS (SA) | 18–22 GHz | ~65,000 | Broadband surrogate **pretraining** |
| `cst_freq_resp_18_to_22.csv` | CST | 18–22 GHz | 1,080 | Broadband surrogate **calibration** |

> The CST datasets reuse the same 1,080 unit-cell geometries for both single-frequency and broadband calibration. Each full-wave simulation inherently yields the broadband response, so no additional CST budget is required for the broadband regime.

---

## Column Schema

### Single-frequency files (`*_20ghz.csv`)

| Column | Description | Unit |
|--------|-------------|------|
| `W1` … `W5` | JC patch leg lengths (design parameters) | mil |
| `t_square` | Transmission power efficiency \|T\|² at 20 GHz | (0 to 1) |
| `t_pha` | Transmission phase φ at 20 GHz | degrees (−180° to 180°) |

### Frequency-response files (`*_freq_resp_18_to_22.csv`)

| Column | Description | Unit |
|--------|-------------|------|
| `W1` … `W5` | JC patch leg lengths | mil |
| `eff_18.0` … `eff_22.0` | Transmission power efficiency \|T\|² sampled across [18, 22] GHz | (0 to 1) |
| `pha_18.0` … `pha_22.0` | Transmission phase φ sampled across [18, 22] GHz | degrees (−180° to 180°) |

> **Data processing** (normalization, train/val/test splits) is handled by scripts in the GitHub repository. Raw files are provided here as-is.

---

## Citation

If you use this dataset, please cite both parts of the paper:

```bibtex
@misc{marcus2026harnessingselectivestatespace,
title={Harnessing Selective State Space Models to Enhance Semianalytical Design
of Fabrication-Ready Multilayered Huygens' Metasurfaces:
Part I - Field-based Semianalytical Synthesis},
author={Sherman W. Marcus and Natanel Nissan and Vinay K. Killamsetty
and Ravi Yadav and Dan Raviv and Raja Giryes and Ariel Epstein},
year={2026},
eprint={2603.03837},
archivePrefix={arXiv},
primaryClass={physics.app-ph},
url={https://arxiv.org/abs/2603.03837},
}

@misc{nissan2026harnessingselectivestatespace,
title={Harnessing Selective State Space Models to Enhance Semianalytical Design
of Fabrication-Ready Multilayered Huygens' Metasurfaces:
Part II - Generative Inverse Design (MetaMamba)},
author={Natanel Nissan and Sherman W. Marcus and Dan Raviv
and Raja Giryes and Ariel Epstein},
year={2026},
eprint={2603.03877},
archivePrefix={arXiv},
primaryClass={physics.app-ph},
url={https://arxiv.org/abs/2603.03877},
}
```

---

## License

This dataset is released under the [MIT License](https://opensource.org/licenses/MIT).