A multi-agentic framework for real-time, autonomous freeform metasurface design

Publication out now in Science Advances

This repository hosts the resources accompanying our study, A multi-agentic framework for real-time, autonomous freeform metasurface design.

Overview

We present MetaChat, a multi-agentic computer-aided design framework, which combines agency with millisecond-speed deep learning surrogate solvers to automate and accelerate photonics design. MetaChat is capable of performing complex freeform design tasks in nearly real-time, as opposed to the days-to-weeks required by the manual use of conventional computing methods and resources.

Near real-time, multi-objective, multi-wavelength autonomous metasurface design is enabled by two key contributions:

• Agentic Iterative Monologue (AIM): A novel agentic system designed to seamlessly automate multiple-agent collaboration, human-designer interaction, and computational tools
• FiLM WaveY-Net: A semi-general fullwave surrogate solver, which supports conditional fullwave modeling—enabling simulations with variable conditions, including source angle, wavelength, material, and device topology—while maintaining high fidelity to the governing physics

Repository structure

• metachat-aim/: Source code for the AIM agentic design stack.
• film-waveynet/: Source code for the FiLM WaveY-Net surrogate solver, including scripts for training and inference (pretrained weights downloadable via Zenodo, training and validation data downloadable via Stanford Digital Repository; see below).
• web-app/: Code for an example web app, which includes a frontend and a GPU server backend, to run MetaChat.

Data availability

All data used for training and validation in the study and referenced by the code here (dielectric structures, sources, Ex, Ey, and Hz fields) can be downloaded from the Stanford Digital Repository. Further information can be found on the Metanet Page. The pretrained best_model.pt checkpoint is hosted on Zenodo.

Reference

The following BibTeX entry can be used to cite MetaChat, this code, and data:

@article{lupoiu2025multiagentic,
	title = {A multi-agentic framework for real-time, autonomous freeform metasurface design},
	volume = {11},
	url = {https://www.science.org/doi/full/10.1126/sciadv.adx8006},
	doi = {10.1126/sciadv.adx8006},
	language = {en},
	number = {44},
	journal = {Science Advances},
	author = {Lupoiu, Robert and Shao, Yixuan and Dai, Tianxiang and Mao, Chenkai and Edée, Kofi and Fan, Jonathan A.},
	year = {2025},
}

Contact

Corresponding author: jonfan@stanford.edu

If you have any questions or need help setting up either AIM or FilM WaveY-Net, don't hesitate to reach out!