Overview
We present MetaChat, a multi-agentic computer-aided design framework that combines agency with millisecond-speed deep-learning surrogate solvers to automate and accelerate photonics design. MetaChat performs complex freeform design tasks in nearly real-time, rather than the days-to-weeks required with conventional methods.
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.
The latest Stanford nanophotonics benchmark dataset is available for download below. The web scraped Materials Expert Agent database materials.db is available for download via Zenodo.
FiLM WaveY-Net
A semi-general full-wave surrogate solver that supports conditional full-wave modeling—enabling simulations with variable conditions such as source angle, wavelength, material, and device topology—while maintaining high fidelity to the governing physics.
Training and validation data for FiLM WaveY-Net (dielectric structures, sources, and the Ex, Ey, Hz fields) are hosted on the Stanford Digital Repository. The pretrained best_model.pt checkpoint is hosted on Zenodo.
Reference
The following BibTeX entry can be used to cite this work:
@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