GPU Fourier optics in Python: build an optical path, propagate complex fields, and inspect the result.
中文 · Quickstart · API · Examples · Visual gallery
A synthetic Gaussian beam broadens as it propagates. The red curves show the paraxial prediction for the beam radius; the background is the computed intensity. Reproduce this figure · Parameters, checks and environment.
| Area | Implemented building blocks |
|---|---|
| Scalar optics | Cached angular-spectrum propagation, Fresnel and Rigorous modes; lenses, apertures, gratings and spatial or momentum filters |
| Vector optics | Three-component fields, transverse spectral projection, Jones elements and Richards–Wolf focusing |
| Thin films | Transfer-matrix calculations and a multilayer momentum-space filter |
| Workflows | Field containers, parameter sweeps, coherent or incoherent spectral synthesis |
| Inspection | Cross sections, longitudinal cuts, intensity and phase; physical regression tests |
Use this for controlled Fourier-optics models. The scalar Rigorous mode uses the exact homogeneous-medium angular-spectrum dispersion; its name does not imply a full-vector Maxwell or arbitrary-geometry solver. See physics guidance and the source modules for each model's assumptions.
Python 3.10+, an NVIDIA GPU with a CUDA 12 compatible driver, and the declared CuPy dependencies are required for GPU propagation.
python -m venv .venv
# Windows:
.venv\Scripts\activate
# Linux/macOS: source .venv/bin/activate
python -m pip install -r requirements.txt
python -m pip install -e .
python examples/public_gaussian.pyThe last command writes three PNG/SVG figures and a JSON manifest to docs/showcase/. It requires no downloaded beam, material or experimental data. On Windows, import optical_system before using CuPy kernels so the repository's CUDA DLL discovery runs. Dependency installation is shown for a fresh environment; the gallery was generated with the existing environment.
A minimal field propagation:
import numpy as np
from optical_system.system import OpticalSystem
import cupy as cp
from propagation.propagator import AngularSpectrumPropagator
x = cp.linspace(-64, 64, 512, endpoint=False) # micrometres
X, Y = cp.meshgrid(x, x)
u0 = cp.exp(-(X**2 + Y**2) / 8.0**2)
propagator = AngularSpectrumPropagator(x, x, wavelength=0.532, mode="Fresnel")
u1 = propagator.propagate(u0, z=np.pi * 8.0**2 / 0.532)
intensity = cp.asnumpy(cp.abs(u1)**2)All lengths must share one unit. OpticalSystem normalizes its initial field to unit discrete squared norm; the minimal low-level example above preserves the input amplitude.
flowchart LR
A["Coordinates + wavelength + initial field"] --> B["Optical elements / filters"]
B --> C["Scalar or vector propagation"]
C --> D["Complex field / Field container"]
D --> E["Cuts, metrics and saved figures"]
E --> F["Sampling and physical checks"]
| Parameter | Value |
|---|---|
| Wavelength / input amplitude waist | 0.532 µm / 8 µm |
| Grid / sample spacing | 512 × 512 / 0.25 µm |
| Periodic FFT window | [−64, 64) µm on both axes |
| Propagation | Fresnel, 101 planes, 0 ≤ z ≤ 2zᵣ; zᵣ = πw₀²/λ ≈ 377.94 µm |
| Intensity normalization | Input on-axis peak = 1 for every displayed plane |
| Beam radius | w = 2√⟨x²⟩, weighted by the complete 2D intensity |
 |  |
| Beam spreading and peak reduction at fixed input normalization. | Fresnel FFT results compared with the analytic paraxial radius. |
This is a homogeneous scalar, monochromatic, paraxial example (w₀ ≫ λ). FFT propagation assumes a periodic window; the beam must remain well inside it. Refine the grid and enlarge the window when changing the parameters. The comparison checks this specific discretization against its paraxial model; it does not measure error against a general Maxwell solution.
The recorded run used Python 3.11.9, MyOptics 0.5.1, NumPy 2.1.2, Matplotlib 3.9.2, CuPy 13.3.0 and an NVIDIA RTX 4060. Its maximum relative radius error was 1.17 × 10⁻¹² and discrete power drift 4.44 × 10⁻¹⁶. Full run manifest records parameters and asset hashes.
- Supported examples: vector focusing, q-plates, multilayers and broadband workflows.
- Quickstart, documentation home, benchmarks and changelog.
examples/legacy/contains historical scripts that are not maintained as current runnable examples.
python -m pytest tests/GPU tests are marked and skip when CUDA is unavailable. A passing suite only covers the tested models and parameter cases.
The package metadata currently declares Proprietary and this repository has no open-source license grant. Public source visibility does not grant general redistribution or reuse rights.