MOCLIP: A Foundation Model for Large-Scale Nanophotonic Inverse Design

Overview

MOCLIP (Metasurface Optics Contrastive Learning Pretrained) is a foundation model for metasurface optics design, trained using contrastive learning on a large-scale experimentally generated dataset of silicon-on-glass metasurfaces. The model was pretrained on 466,537 unique geometry–spectrum pairs, enabling joint representation learning of metasurface geometries and their corresponding polarization-resolved transmission spectra.

Details on the dataset generation and MOCLIP architecture can be found in the following arxiv preprint: MOCLIP: A Foundation Model for Large-Scale Nanophotonic Inverse Design.

Alongside the pretrained model, this repository provides a 10k randomly sampled subset of the full dataset for reproducibility, experimentation, and benchmarking.

Getting Started

Hardware Requirements

MOCLIP was trained and evaluated on an NVIDIA RTX 4090 GPU with CUDA support. While the model can technically run on a CPU, performance will be prohibitively slow for most practical use cases. We strongly recommend using an NVIDIA GPU with CUDA support for inference.

Software Requirements

• Ubuntu 24.04 LTS (tested; macOS and Windows are expected to work but are not officially validated)
• Python ≥ 3.9
• PyTorch ≥ 2.4 (with CUDA 12.1 support)
• NumPy ≥ 1.26
• TorchVision ≥ 0.19
• PyYAML ≥ 0.2.5

Installation

1. Clone the repository:

   git clone https://github.com/RodionovSA/MOCLIP_release.git
   cd MOCLIP_release

2. Install PyTorch with CUDA support (see pytorch.org for the command matching your CUDA version). For CUDA 12.1:

   pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

3. Install the remaining dependencies:

   pip install numpy>=1.26 pyyaml>=0.2.5 matplotlib

How to Use

Detailed usage examples are provided in notebooks/MOCLIP_example.ipynb.

Citation

If you use this repository, model, or dataset in your work, please cite:

@article{rodionov2025moclip,
  title   = {MOCLIP: A Foundation Model for Large-Scale Nanophotonic Inverse Design},
  author  = {Rodionov, S. and Burguete-Lopez, A. and Makarenko, M. and Wang, Q. and Getman, F. and Fratalocchi, A.},
  journal = {arXiv preprint arXiv:2511.18980},
  year    = {2025}
}