Comprehensive Survey of Open-Source RCWA / FDTD Optical Solvers
Survey date: 2026-02-11 | For COMPASS project reference
1. RCWA Solvers
1.1 Solvers Already Integrated in COMPASS
| Solver | Language | GPU | AD | License | ⭐ | Status | Notes |
|---|---|---|---|---|---|---|---|
| torcwa | Python/PyTorch | CUDA | ✅ PyTorch | LGPL | ~171 | ⚠️ Low activity (2023~) | S-matrix, suitable for metasurface inverse design. Development stalled |
| grcwa | Python/autograd | ❌ | ✅ autograd | GPL | ~94 | ❌ Discontinued (2020~) | CPU only. Topology optimization. Effectively abandoned |
| meent | Python (NumPy/JAX/PyTorch) | JAX/PyTorch | ✅ JAX/PyTorch | MIT | ~112 | ✅ Active | 3 backends supported, best for ML integration. Developed by Korean company (KC-ML2) |
Assessment: meent is the best in terms of licensing (MIT), maintenance, and flexibility. grcwa needs reassessment for long-term integration viability.
1.2 Integration Candidates (High Priority)
| Solver | Language | GPU | AD | License | ⭐ | Status | Key Strengths |
|---|---|---|---|---|---|---|---|
| fmmax | Python/JAX | JAX | ✅ | MIT | ~137 | ✅ Active (Meta) | State-of-the-art vector FMM (Li inverse rule extension), Brillouin zone integration, batching support. AR/VR optics |
| torchrdit | Python/PyTorch | CUDA | ✅ | GPL-3.0 | ~11 | ✅ Active | No eigenvalue decomposition required (R-DIT), 16.2x speedup over conventional RCWA, GDS import/export |
| S4 (phoebe-p fork) | C++/Lua/Python | ❌ | ❌ | GPL-2.0 | ~166 | ⚠️ Fork active | RCWA reference implementation, Li inverse rule, Solcore/RayFlare integration. 25-year history |
fmmax details:
- Developed by Meta Reality Labs for AR/VR diffractive optics design
- 4 vector FMM formulations (Pol, Normal, Jones, Jones-direct) → best convergence
- JAX JIT compilation + batching for simultaneous computation of multiple configurations
- Anisotropic and magnetic material support
torchrdit details:
- R-DIT (Rigorous Diffraction Interface Theory) eliminates eigenvalue decomposition, the traditional RCWA bottleneck
- PyTorch-based → similar interface to existing torcwa code possible
- Fast Fourier Factorization (POL, NORMAL, JONES, JONES_DIRECT)
- Automatic permittivity fitting for dispersive materials
1.3 Special Purpose / Reference
| Solver | Language | GPU | License | ⭐ | Key Features |
|---|---|---|---|---|---|
| inkstone | Python/NumPy | ❌ | AGPL-3.0 | ~63 | Tensor permittivity/permeability (anisotropic, magneto-optic, gyromagnetic). Partial recomputation optimization |
| nannos | Python | ✅ | GPL-3.0 | ~20 | Multiple FMM formulations, AD, GPU acceleration. GitLab hosted |
| rcwa_tf | Python/TensorFlow | CUDA | BSD-3 | ~51 | TF-based, Lorentzian broadening (gradient stabilization), batch optimization |
| rcwa (edmundsj) | Python | ❌ | MIT | ~134 | Built-in refractiveindex.info DB, TMM+RCWA, ellipsometry |
| EMUstack | Fortran/Python | ❌ | GPL-3.0 | ~28 | Hybrid 2D-FEM + scattering matrix. Strong for metallic/plasmonic structures |
| MESH | C++ | ❌(MPI) | GPL-3.0 | ~33 | RCWA + thermal radiation transfer (near-field/far-field). Stanford Fan Group |
| RETICOLO | MATLAB | ❌ | Freeware | N/A | Industry standard (Zeiss, Intel, Apple, Samsung). V10 (2025.01) anisotropy support. 25-year history |
| rcwa4d | Python | ❌ | MIT | ~40 | Incommensurate periodicity (twisted bilayer/moiré structures). Stanford Fan Group |
| RCWA.jl | Julia | CUDA | GPL-3.0 | ~46 | S-matrix + ETM, eigenvalue-free algorithm, CUDA 5x speedup |
| EMpy | Python | ❌ | MIT | ~219 | TMM + RCWA + mode solver. Most GitHub stars |
2. FDTD Solvers
2.1 Solvers Already Integrated in COMPASS
| Solver | Language | GPU | AD | License | ⭐ | Status | Notes |
|---|---|---|---|---|---|---|---|
| Meep | C++/Python | ❌ | ✅ (adjoint) | GPL-2.0+ | ~1,500 | ✅ Active | Most mature open-source FDTD. MPI parallelization. Lack of GPU support is a drawback |
| flaport/fdtd | Python/PyTorch | CUDA | ✅ PyTorch | MIT | ~650 | ⚠️ Low activity | For education/prototyping. Simple API. Lacks advanced material models |
| fdtdz | C++/CUDA/JAX | CUDA | ✅ JAX | MIT | ~146 | ✅ Active (Google) | ~100x speed vs Meep. Limited to 2.5D. Only simple dielectrics supported |
Assessment: Meep = general-purpose reference, fdtdz = speed-focused (2.5D), flaport = educational. Lacking a general-purpose 3D GPU FDTD.
2.2 Integration Candidates (High Priority)
| Solver | Language | GPU | AD | License | ⭐ | Status | Key Strengths |
|---|---|---|---|---|---|---|---|
| FDTDX | Python/JAX | JAX (multi-GPU) | ✅ | MIT | ~203 | ✅ Very active | Optimal for large-scale 3D inverse design. Multi-GPU. Memory-efficient gradients using Maxwell time-reversal. JOSS published |
| Khronos.jl | Julia | CUDA/ROCm/Metal/OneAPI | ✅ | MIT | ~66 | ✅ Active (Meta) | Multi-vendor GPU (NVIDIA+AMD+Apple+Intel). Pure Julia. Differentiable |
| ceviche | Python/autograd | ❌ | ✅ | MIT | ~390 | ⚠️ Low activity | Pioneer of differentiable EM. 2D FDFD+FDTD. Stanford Fan Group |
FDTDX details:
- JAX-based fully differentiable 3D FDTD
- Multi-GPU scaling → billions of grid cells simulation possible
- Memory-efficient gradients leveraging time-reversibility of Maxwell's equations
- Published in JOSS (Journal of Open Source Software) (2025)
- MIT license + active development → favorable for long-term integration
Khronos.jl details:
- Developed by Meta Research
- Only solver with multi-vendor GPU support (CUDA, ROCm, Metal, OneAPI)
- Vendor-independent GPU code based on KernelAbstractions.jl
- DFT monitors + convergence-based automatic termination
- Julia ecosystem dependency is a drawback
2.3 Special Purpose / Reference
| Solver | Language | GPU | License | ⭐ | Key Features |
|---|---|---|---|---|---|
| openEMS | C++/MATLAB/Python | ❌(OpenMP/MPI) | GPL-3.0 | ~628 | EC-FDTD, cylindrical coordinates, RF/antenna/microwave specialized |
| gprMax | Python/Cython | CUDA | GPL-3.0 | ~788 | GPR (Ground Penetrating Radar) specialized, CUDA 30x speedup. Soil models |
| EMOPT | Python/C | ❌(MPI) | BSD-3 | ~110 | FDFD (2D/3D) + CW-FDTD, shape optimization (boundary smoothing), adjoint method |
| fdtd3d | C++ | CUDA/MPI | GPL-3.0 | ~150 | MPI+OpenMP+CUDA, cross-architecture (x64/ARM/RISC-V/Wasm) |
| GSvit | C/C++ | CUDA | GPL-2.0 | N/A | Nanoscale optics (SNOM, roughness), GPU accelerated |
| Luminescent.jl | Julia | CUDA | MIT | ~60 | Differentiable FDTD (Zygote.jl), semiconductor photonics+acoustics+RF |
| Angora | C++ | ❌ | GPL | Small | Biomedical scattering specialized |
2.4 Commercial / Non-Open-Source (Reference)
| Solver | Notes |
|---|---|
| Tidy3D (Flexcompute) | Python client is LGPL but computation engine is commercial cloud. Cannot run locally. Very fast |
| Lumerical (Ansys) | Fully commercial. Industry standard GUI. CUDA GPU solver. Academic discounts available |
3. Technology Trend Analysis
3.1 Differentiable EM Simulation (Differentiable EM)
The biggest trend of 2024-2026. Built-in AD (automatic differentiation) is becoming essential for inverse design and topology optimization.
| Generation | Representative Solvers | AD Framework | Performance |
|---|---|---|---|
| 1st generation (2019) | ceviche, grcwa | autograd (CPU) | Slow, 2D |
| 2nd generation (2021-23) | torcwa, flaport/fdtd | PyTorch | GPU accelerated, limited 3D |
| 3rd generation (2024-26) | fmmax, FDTDX, fdtdz, meent | JAX | JIT+multi-GPU, large-scale 3D |
Conclusion: The JAX ecosystem is emerging as the mainstream for EM solvers. PyTorch-based solvers remain viable, but JAX's JIT compilation + vmap + pmap are better suited for EM simulation.
3.2 GPU Acceleration Status
| Approach | Solver Examples | Speedup |
|---|---|---|
| PyTorch CUDA | torcwa, flaport | 5-20x |
| JAX JIT+CUDA | fmmax, FDTDX, fdtdz, meent | 10-100x |
| Custom CUDA kernels | fdtdz | ~100x (vs Meep) |
| Julia CUDA.jl | Khronos.jl, RCWA.jl | 5-10x |
| Multi-GPU | FDTDX, Khronos.jl | Linear scaling |
3.3 License Distribution
| License | Solver Count | Commercial Use |
|---|---|---|
| MIT | 10 (meent, fmmax, fdtdz, FDTDX, ceviche, rcwa4d, EMpy, flaport, Khronos.jl, Luminescent.jl) | ✅ Free |
| BSD-3 | 2 (rcwa_tf, EMOPT) | ✅ Free |
| GPL/LGPL | 12+ (torcwa, grcwa, S4, nannos, RCWA.jl, Meep, openEMS, gprMax, torchrdit, etc.) | ⚠️ Restricted |
| AGPL | 1 (inkstone) | ❌ Very restricted |
4. COMPASS Integration Recommendations
Tier 1: Strongly Recommended (MIT License, Active Development, High Performance)
| Solver | Type | Rationale |
|---|---|---|
| fmmax | RCWA | MIT, Meta-backed, best convergence (vector FMM), JAX batching. Potential synergy with meent's JAX backend |
| FDTDX | FDTD | MIT, multi-GPU 3D, fully differentiable, JOSS published. Complements fdtdz's 2.5D limitation |
Tier 2: Recommended for Review (Special Strengths)
| Solver | Type | Rationale |
|---|---|---|
| torchrdit | RCWA (R-DIT) | Major speedup by eliminating eigenvalue decomposition. GPL is an obstacle |
| S4 (phoebe-p fork) | RCWA | C++ performance reference implementation. Useful for verification. GPL |
| ceviche | FDTD/FDFD | 2D differentiable EM. For rapid prototyping. MIT |
Tier 3: Long-Term Watch
| Solver | Type | Rationale |
|---|---|---|
| Khronos.jl | FDTD | Multi-vendor GPU is attractive but Julia dependency |
| inkstone | RCWA | Unique tensor permittivity support but AGPL |
| Luminescent.jl | FDTD | Julia differentiable FDTD. Still early stage |
Existing Integrated Solver Assessment
| Solver | Retention Recommendation | Notes |
|---|---|---|
| meent ✅ | Actively retain | MIT, active, 3 backends, best flexibility |
| torcwa ⚠️ | Retain but monitor | LGPL, development stalled. Potential replacement by meent |
| grcwa ❌ | Consider deprecation | GPL, discontinued since 2020. CPU only. Inferior |
| Meep ✅ | Actively retain | General-purpose reference. Lack of GPU support is regrettable |
| flaport ⚠️ | Retain but monitor | MIT, low activity. Only valuable for education/prototyping |
| fdtdz ✅ | Retain | MIT, Google-backed, extreme speed. Acknowledge 2.5D limitation |
5. Complete Solver Summary Tables
RCWA (20 solvers)
| # | Solver | Language | GPU | AD | License | ⭐ | Status |
|---|---|---|---|---|---|---|---|
| 1 | meent | Py (NumPy/JAX/PyTorch) | ✅ | ✅ | MIT | 112 | ✅ Active |
| 2 | fmmax | Py/JAX | ✅ | ✅ | MIT | 137 | ✅ Active |
| 3 | torcwa | Py/PyTorch | ✅ | ✅ | LGPL | 171 | ⚠️ |
| 4 | S4 | C++/Lua/Py | ❌ | ❌ | GPL-2.0 | 166 | ⚠️ Fork |
| 5 | EMpy | Py | ❌ | ❌ | MIT | 219 | ✅ |
| 6 | rcwa (edmundsj) | Py | ❌ | ❌ | MIT | 134 | ⚠️ |
| 7 | grcwa | Py/autograd | ❌ | ✅ | GPL | 94 | ❌ Discontinued |
| 8 | inkstone | Py/NumPy | ❌ | ❌ | AGPL | 63 | ⚠️ |
| 9 | rcwa_tf | Py/TF | ✅ | ✅ | BSD-3 | 51 | ⚠️ |
| 10 | RCWA.jl | Julia | ✅ | ❌ | GPL-3.0 | 46 | ✅ |
| 11 | rcwa4d | Py | ❌ | ❌ | MIT | 40 | ⚠️ |
| 12 | MESH | C++ | ❌ | ❌ | GPL-3.0 | 33 | ⚠️ |
| 13 | EMUstack | Fortran/Py | ❌ | ❌ | GPL-3.0 | 28 | ⚠️ |
| 14 | nannos | Py | ✅ | ✅ | GPL-3.0 | 20 | ⚠️ |
| 15 | torchrdit | Py/PyTorch | ✅ | ✅ | GPL-3.0 | 11 | ✅ |
| 16 | RETICOLO | MATLAB | ❌ | ❌ | Freeware | N/A | ✅ |
| 17 | PPML | MATLAB | ❌ | ❌ | Free | N/A | ⚠️ |
| 18-20 | 3 educational solvers | Py | ❌ | ❌ | Various | <10 | ⚠️ |
FDTD (17 solvers)
| # | Solver | Language | GPU | AD | License | ⭐ | Status |
|---|---|---|---|---|---|---|---|
| 1 | Meep | C++/Py | ❌ | ✅ adj | GPL-2.0+ | 1,500 | ✅ Active |
| 2 | gprMax | Py/Cython | CUDA | ❌ | GPL-3.0 | 788 | ✅ |
| 3 | flaport/fdtd | Py/PyTorch | ✅ | ✅ | MIT | 650 | ⚠️ |
| 4 | openEMS | C++ | ❌ | ❌ | GPL-3.0 | 628 | ✅ |
| 5 | ceviche | Py | ❌ | ✅ | MIT | 390 | ⚠️ |
| 6 | FDTDX | Py/JAX | ✅ Multi | ✅ | MIT | 203 | ✅ Very active |
| 7 | Tidy3D | Py | ☁️ Cloud | ✅ | LGPL/Commercial | 164 | ✅ (Non-open-source) |
| 8 | fdtd3d | C++ | CUDA/MPI | ❌ | GPL-3.0 | 150 | ✅ |
| 9 | fdtdz | C++/CUDA/JAX | ✅ | ✅ | MIT | 146 | ✅ |
| 10 | EMOPT | Py/C | ❌ | ✅ adj | BSD-3 | 110 | ⚠️ |
| 11 | PhotonTorch | Py/PyTorch | ✅ | ✅ | MIT | 81 | ⚠️ (Circuit sim) |
| 12 | Khronos.jl | Julia | ✅ Multi-vendor | ✅ | MIT | 66 | ✅ |
| 13 | Luminescent.jl | Julia | ✅ | ✅ | MIT | 60 | ✅ |
| 14 | GSvit | C/C++ | CUDA | ❌ | GPL-2.0 | N/A | ✅ |
| 15 | Angora | C++ | ❌ | ❌ | GPL | Small | ⚠️ |
| 16 | MaxwellFDTD.jl | Julia | ❌ | ❌ | N/A | Small | ⚠️ |
| 17 | REMS | Rust | ❌ | ❌ | N/A | Tiny | ❌ (1D PoC) |
RCWA vs FDTD Solver Comparison
Compare simulated quantum efficiency (QE) curves from RCWA and FDTD solvers. Adjust pixel pitch and solver parameters to see how results and performance change.
RCWA (Fourier order = 9)
FDTD (grid = 20 nm)
RCWA
Time estimate:137 ms
Memory:6 MB
Periodic structures:Yes
Arbitrary geometry:Limited
FDTD
Time estimate:188 ms
Memory:3 MB
Periodic structures:Yes
Arbitrary geometry:Yes
Agreement
Max |Delta QE|:2.2%
Avg |Delta QE|:0.9%
Status:Good agreement
Note: As of 2026, no Rust-based RCWA solver exists. For FDTD, only a 1D PoC (REMS) exists.