TorchRDIT

TorchRDIT is an advanced software package designed for the inverse design of meta-optics devices, utilizing an eigendecomposition-free implementation of Rigorous Diffraction Interface Theory (R-DIT). It provides a GPU-accelerated and fully differentiable framework powered by PyTorch, enabling the efficient optimization of photonic structures.

This framework achieves up to 16.2× speedup compared to traditional inverse design methods based on Rigorous Coupled-Wave Analysis (RCWA). By integrating differentiable R-DIT with topology optimization techniques and neural networks (e.g., SIREN), TorchRDIT facilitates the design of complex meta-optics devices, including:

• Parameter-constrained and free-form meta-atoms.
• Reconfigurable photonic structures using optical phase-change materials.
• High-performance meta-lenses and beam deflectors.

With its focus on computational efficiency and flexibility, TorchRDIT lays the foundation for next-generation metasurface design workflows, offering capabilities for the creation of intricate and multifunctional photonic devices.

TorchRDIT is used for the papers:

• Eigendecomposition-free inverse design of meta-optics devices
• Inverse Design of Photonic Structures Using Automatic Differentiable Rigorous Diffraction Interface Theory
• A 3D-Printed Millimeter-Wave Free-Form Metasurface Based on Automatic Differentiable Inverse Design
• Differentiable Inverse Design of Free- form Meta-optics Using Multiplicative Filter Network

Features

The EM algorithms used in this package are Rigorous Diffraction Interface Theory (R-DIT) and Rigorous Coupled Wave Analysis (RCWA) . The solver is developed on an automatic differentiable framework (PyTorch) and is fully differentiable for more advanced applications.

Key Capabilities

• Eigendecomposition-free R-DIT: Up to 16.2× speedup over traditional RCWA
• Full differentiability: Seamless integration with gradient-based optimization
• GPU acceleration: CUDA and MPS backend support
• GDS Export/Import: Industry-standard format for fabrication workflows
  • Export photonic masks to GDS format
  • Import masks from GDS vertex data
  • Support for complex geometries with holes
  • Batch processing for multiple designs
• Material handling: Dispersive materials with automatic fitting
• Flexible geometry: Arbitrary unit cells and lattice systems

Fast Fourier Factorization

TorchRDIT’s FFF implementation follows the tangent-field techniques popularized by fmmax and S4, but is written entirely in PyTorch (torchrdit/src/torchrdit/vector.py). When is_use_FFF=True, the solver automatically instantiates the tangent generator, yet you can also invoke it explicitly via compute_tangent_field:

from torchrdit.vector import compute_tangent_field

tx, ty = compute_tangent_field(
    mask.float(),
    XO=solver.XO,
    YO=solver.YO,
    lattice_t1=solver.lattice_t1,
    lattice_t2=solver.lattice_t2,
    harmonics=tuple(int(k) for k in solver.harmonics),
    scheme="POL",  # POL, NORMAL, JONES, or JONES_DIRECT
)

Choose between the four built-in schemes—POL, NORMAL, JONES, JONES_DIRECT—to match the field representation expected by your Fourier factorization workflow while keeping gradients intact.

For background on the algorithmic formulation see:

• M. F. Schubert and A. M. Hammond, "Fourier modal method for inverse design of metasurface-enhanced micro-LEDs," Opt. Express 31, 42945 (2023).
• V. Liu and S. Fan, "S4 : A free electromagnetic solver for layered periodic structures," Computer Physics Communications 183, 2233–2244 (2012).

Installation

You can install the package directly from PyPI:

pip install torchrdit

Dependencies

• torch>=2.5.1
• numpy>=2.2.0
• scipy>=1.15.3
• scikit-image>=0.25.2
• matplotlib>=3.10.3
• tqdm>=4.67.1
• pyyaml>=6.0.2
• ruff>=0.12.2
• gdstk>=0.9.60

Note: The gdstk library may require additional system dependencies:

• zlib - compression library
• qhull - computational geometry library

These are typically available through your system's package manager (e.g., apt-get install libz-dev libqhull-dev on Ubuntu).

Usage

TorchRDIT includes several example files to demonstrate its capabilities:

R-DIT Algorithm Examples

• example_gmrf_rdit.py: GMRF with hexagonal unit cells using R-DIT algorithm

Dispersive Materials Examples

• example_gmrf_dispersive.py: GMRF with dispersive materials and permittivity fitting

Design Optimization Examples

• example_gmrf_variable_optimize.py: Demonstration of parameter optimization using automatic differentiation

GDS Export/Import Examples

• example_gds_export.py: Comprehensive GDS functionality demonstration

Quick GDS Example

from torchrdit.shapes import ShapeGenerator
from torchrdit.gds import mask_to_gds, gds_to_mask

# Create shape and export to GDS
shape_gen = ShapeGenerator(X, Y, grids)
mask = shape_gen.generate_circle_mask(center=(0, 0), radius=0.5)
mask_to_gds(mask, shape_gen.get_layout(), "DEVICE", "output.gds")

# Import from GDS
reconstructed = gds_to_mask("output.json", shape_gen)

For more detailed documentation, please visit our Wiki.

Related Works

• Eigendecomposition-free inverse design of meta-optics devices
• Inverse Design of Photonic Structures Using Automatic Differentiable Rigorous Diffraction Interface Theory

Citation

The package contains the work from multiple publicatinos. Please consider to cithe the following paper for the R-DIT solver:

Yi Huang, Ziwei Zhu, Yunxi Dong, Hong Tang, Bowen Zheng, Viktor A. Podolskiy, and Hualiang Zhang, "Eigendecomposition-free inverse design of meta-optics devices," Opt. Express 32, 13986-13997 (2024)

@article{Huang:24,
author = {Yi Huang and Ziwei Zhu and Yunxi Dong and Hong Tang and Bowen Zheng and Viktor A. Podolskiy and Hualiang Zhang},
journal = {Opt. Express},
keywords = {Diffraction theory; Finite-difference time-domain method; Inverse design; Machine learning; Neural networks; Refractive index},
number = {8},
pages = {13986--13997},
publisher = {Optica Publishing Group},
title = {Eigendecomposition-free inverse design of meta-optics devices},
volume = {32},
month = {Apr},
year = {2024},
url = {https://opg.optica.org/oe/abstract.cfm?URI=oe-32-8-13986},
doi = {10.1364/OE.514347},
abstract = {The inverse design of meta-optics has received much attention in recent years. In this paper, we propose a GPU-friendly inverse design framework based on improved eigendecomposition-free rigorous diffraction interface theory, which offers up to 16.2\&\#x2009;\&\#x00D7; speedup over the traditional inverse design based on rigorous coupled-wave analysis. We further improve the framework\&\#x2019;s flexibility by introducing a hybrid parameterization combining neural-implicit and traditional shape optimization. We demonstrate the effectiveness of our framework through intricate tasks, including the inverse design of reconfigurable free-form meta-atoms.},
}

License

This project is licensed under the GPL-3.0 License. See the LICENSE file for details.

Feedback and Support

If you have any questions, issues, or suggestions, please open an issue or email us at yi_huang at student dot uml dot edu.