FDTDX is an efficient open-source Python package for the inverse design of three-dimensional photonic nanostructures using the Finite-Difference Time-Domain (FDTD) method. Built on JAX, it provides native GPU support and automatic differentiation capabilities, making it ideal for large-scale 3D design in nanophotonics.

• High Performance: GPU-accelerated FDTD simulations with multi-GPU scaling capabilities
• Memory Efficient: Leverages time-reversibility in Maxwell's equations for efficient gradient computation
• Automatic Differentiation: Built-in gradient-based optimization for complex 3D structures
• User-Friendly API: Intuitive positioning and sizing of objects in absolute or relative coordinates
• Large-Scale Design: Capable of handling simulations with billions of grid cells
• Open Source: Freely available for research and development

Visit our documentation for:

• Detailed API reference
• Tutorial guides
• Best practices

Also check out our whitepaper for some examples and a comparison to other popular FDTD-frameworks.

Install FDTDX using pip:

pip install fdtdx  # Basic CPU-Installation
pip install fdtdx[cuda12]  # GPU-Acceleration (Highly Recommended!)
pip install fdtdx[rocm]   # AMD-GPU (only python<=3.12)

For development installation, see the contributing guidelines!

# The following lines often lead to better memory usage in JAX
# when using multiple GPU.
export XLA_PYTHON_CLIENT_ALLOCATOR="platform"
export XLA_PYTHON_CLIENT_PREALLOCATE="false"
export NCCL_LL128_BUFFSIZE="-2"
export NCCL_LL_BUFFSIZE="-2"
export NCCL_PROTO="SIMPLE,LL,LL128"

If you find this repository helpful for you work, please consider citing:

@article{schubert2025quantized,
  title={Quantized inverse design for photonic integrated circuits},
  author={Schubert, Frederik and Mahlau, Yannik and Bethmann, Konrad and Hartmann, Fabian and Caspary, Reinhard and Munderloh, Marco and Ostermann, J{\"o}rn and Rosenhahn, Bodo},
  journal={ACS omega},
  volume={10},
  number={5},
  pages={5080--5086},
  year={2025},
  publisher={ACS Publications}
}

This project was developed at the Institute of Information Processing at Leibniz University Hannover, Germany and sponsored by the cluster of excellence PhoenixD (Photonics, Optics, Engineering, Innovation across Disciplines).