A GPU-accelerated 2D TM FDTD electromagnetic simulator built with PyTorch/CUDA.
Designed as a high-performance alternative to Meep for microwave imaging, MIMO radar, and inverse scattering.
Measured results (Tesla T4 GPU, Colab free tier):
| Grid | Kaggle 2×T4 (per GPU) | Colab T4 | Meep CPU | GPU / Meep |
|---|---|---|---|---|
| 64² | 7 Mcells/s | 6 Mcells/s | 14 Mcells/s | 0.5× |
| 128² | 24 Mcells/s | 25 Mcells/s | 20 Mcells/s | 1.2× |
| 256² | 102 Mcells/s | 94 Mcells/s | 15 Mcells/s | 6.8× |
| 512² | 384 Mcells/s | 350 Mcells/s | 16 Mcells/s | 24× |
| 1024² | 1,481 Mcells/s | — | — | 92× (est.) |
At 512² grid: 24× faster than Meep on Kaggle T4.
At 1024² grid: 1,481 Mcells/s — 92× faster than Meep (Kaggle 2×T4).
Combined 2×T4: 2,871 Mcells/s = 179× faster than Meep CPU.
CPU-only comparison (all scenarios, laptop):
Gaussian pulse hitting a glass slab (εᵣ=4). Analytical prediction: R=11.1%, T=88.9%. Simulation confirms both values.
5 GHz sinusoidal source between PEC walls. Guided mode propagates since f=5 GHz > f_c=2.5 GHz. Mode pattern clearly visible.
Plane wave scattering from a dielectric cylinder (εᵣ=9, r=15mm). Backscatter/forward ratio = 1.53. 8 detectors on ring record scattered field in all directions.
1 GHz pulse penetrating a concrete wall (εᵣ=6, σ=0.05 S/m) and scattering off a hidden conductive target. Return echo visible at step 500. Wall loss ≈ −9.2 dB.
Two coherent sources at 2.4 GHz creating standing wave pattern. Fringe spacing = λ/2 = 62mm. Time-averaged |Hz|² reveals the interference structure.
1 GHz pulse through skin (δ=13mm) → fat (δ=71mm) → muscle (δ=12mm). Progressive attenuation log-curve matches analytical skin depths.
8-element Uniform Linear Array at 10 GHz. Cylinder target at 120° from boresight. Angular resolution ≈ λ/(N·d) = 14.3°. Delay-and-sum beamforming angular spectrum shown.
16-antenna MIMO array, 1 GHz. Four samples: healthy brain + 3 clot positions (left hemisphere, right hemisphere, frontal lobe). DAS backprojection localizes frontal clot to within 2mm.
200×200 grid, elliptical breast phantom: skin/adipose/fibroglandular/tumor layers. 16 TX antennas, 1 GHz. DAS localization error: 5.4mm (tumor radius=12mm). Breast imaging advantage: Adipose→Tumor contrast Δεᵣ=50 vs brain-clot Δεᵣ=15 (3.3× stronger signal).
16-antenna circular array, 150×150 grid. DAS reconstruction correctly localizes 12mm tumor. Material map shows skull/brain/tumor clearly.
- GPU-accelerated FDTD via PyTorch tensor ops — no Python loops over spatial indices
- 2D TM mode: {Ex, Ey, Hz} with correct Yee-grid staggering and leapfrog
- Lossy dispersive materials — per-cell εᵣ, σ (skull, brain, adipose, tumor)
- First-order Mur absorbing boundary conditions
- CPML coefficient scaffold (dual-staggered E/H, ready for Phase 2)
- MIMO circular array imaging — delay-and-sum backprojection
- Batch simulation support (multiple simultaneous TX runs)
- Real-time visualization and MP4/GIF animation export
- 21-test physics verification suite
- Click the badge → Runtime → T4 GPU → Run all cells
git clone https://github.com/shahzaibshazoo/waveforge.git
cd waveforge
pip install torch numpy matplotlib pytestpython examples/basic_2d_wave.py # free-space pulse
python examples/02_dielectric_slab.py # reflection/transmission
python examples/03_waveguide.py # guided mode
python examples/04_scattering_cylinder.py # radar cross section
python examples/05_through_wall_radar.py # through-wall detection
python examples/06_multipath_interference.py # WiFi interference
python examples/07_tissue_penetration.py # bio-tissue attenuation
python examples/08_ev_radar_ula.py # automotive radar
python examples/09_brain_clot_dataset_sample.py # brain clot dataset
python examples/10_breast_tumor_mimo.py # breast tumor imaging
python examples/brain_mimo_imaging.py # full brain MIMOwaveforge/
├── src/core/
│ ├── grid.py # YeeGrid, CFL enforcement
│ ├── fields.py # SoA field tensor container (batch-ready)
│ ├── materials.py # Per-cell εᵣ, σ → Ca/Cb FDTD coefficients
│ ├── sources.py # Gaussian, Ricker, Sinusoidal sources
│ ├── boundaries.py # Mur ABC + dual-staggered CPML scaffold
│ └── fdtd2d.py # 2D TM FDTD engine
├── src/visualization/
│ └── plot2d.py # Field snapshots + FuncAnimation
├── examples/ # 10 simulation scenarios
├── tests/ # 21 physics tests
├── benchmarks/ # GPU vs CPU vs Meep comparisons
├── notebooks/ # Colab GPU benchmark notebook
└── assets/ # Plots and simulation outputs
pytest tests/test_fdtd2d.py -v # 21/21 passed- CPML full psi field updates
- TFSF plane wave source
- Near-to-far field transform
- 3D extension
- Differentiable physics (adjoint method)
- 10K brain/breast dataset generator
MIT