AutoLens

AI-powered automated optical lens design using differentiable optimization and deep learning

AutoLens is an automated lens design project that uses gradient backpropagation and curriculum learning to optimize optical systems from scratch. Built on top of the DeepLens differentiable optics framework, AutoLens delivers optimization capabilities that go far beyond what commercial lens design software can offer.

Key Features

• End-to-end differentiable ray tracing with PyTorch
• Curriculum learning for ab initio (from-scratch) lens design
• Wide-angle, full-frame, and cellphone lens optimization
• Aspherical and spherical surface support
• GPU-accelerated via CUDA
• Compatible with the DeepLens framework

Getting Started

# Clone the repository
git clone https://github.com/singer-yang/AutoLens.git
cd AutoLens

# Create and activate the conda environment
conda env create -f environment.yml
conda activate autolens

# Run the automated lens design script
python 1_autolens.py

# (Optional) Refine an existing lens design
python 2_lens_optim.py

Lens Design Examples

Example 1: Wide-Angle Lens

• Field of View: 80°
• F-number: F/2.0
• Focal Length: 4.55 mm

Example 2: Full-Frame Lens

• Field of View: Full-frame (43.3 mm diagonal)
• F-number: F/2.0
• Focal Length: 50 mm

Example 3: Multiple Designs from Random Seeds

20 randomized automated lens design results for FoV 80°, F/2.0, 4.55 mm focal length — demonstrating the diversity of solutions the optimizer can discover.

Example 4: Ultra Design Targets

AutoLens supports challenging ultra-specification targets such as F/1.2 high-speed camera lenses and 120° wide-angle cellphone lenses. The figure below shows 8 diverse optimized designs across these demanding specifications.

Project Structure

AutoLens/
├── 1_autolens.py        # Automated design from scratch (curriculum learning)
├── 2_lens_optim.py      # Refine an existing lens with Adam optimizer
├── 0_hello_deeplens.py  # Introductory example
├── configs/             # Lens configuration files
├── deeplens/            # Core ray-tracing engine (DeepLens submodule)
├── datasets/            # Example lens files
└── assets/              # Images, animations, and other static assets

Citation

If you find this repository helpful, please cite our paper:

@article{yang2024curriculum,
  title={Curriculum learning for ab initio deep learned refractive optics},
  author={Yang, Xinge and Fu, Qiang and Heidrich, Wolfgang},
  journal={Nature Communications},
  year={2024},
  publisher={Nature Publishing Group}
}

License

This project is licensed under the Apache License 2.0.