difftmm 0.1.2

Differentiable transfer matrix method (TMM) in PyTorch for multi-layer thin film optics — supports anisotropic media, GPU batching, and gradient-based inverse design.

DiffTMM: Differentiable Transfer Matrix Method

A PyTorch-based differentiable thin film solver for multi-layer optical coatings. Supports both isotropic and anisotropic materials with full autograd for inverse design.

Advantages over NumPy TMM

Feature	NumPy TMM (sbyrnes321/tmm)	DiffTMM
Differentiable	No	Yes (PyTorch autograd)
GPU acceleration	No (CPU only)	Yes (CUDA)
Batch processing	No (sequential)	Yes (vectorized)
Anisotropic materials	No (isotropic only)	Yes (4x4 transfer matrix)
Speed (batch=16)	1x baseline	~200x (isotropic 2x2)

Installation

git clone https://github.com/singer-yang/DiffTMM.git
cd DiffTMM
pip install torch numpy matplotlib scipy

Quick Start

Forward Simulation (1_forward_simu.ipynb)

Initialize a film stack with known refractive indices and thicknesses, then compute Fresnel coefficients at arbitrary wavelengths and angles.

import torch
from difftmm import IsotropicFilmSolver

# Define film stack: Glass | Ta2O5 | SiO2 | Ta2O5 | Glass
solver = IsotropicFilmSolver(
    n_in=1.5,                         # incident medium
    n_out=1.5,                        # exit medium
    n_layers_list=[2.10, 1.46, 2.10], # interior layer indices
    d_layers=[0.080, 0.120, 0.080],   # thicknesses in um
    device=torch.device("cuda"),
)

# Compute Fresnel coefficients: ts, tp, rs, rp
angles = torch.linspace(0, 1.2, 100, device=solver.device)
ts, tp, rs, rp = solver.simulate(theta=angles, wvln=[0.45, 0.55, 0.65])
# Output shape: (n_mirrors, n_wvlns, n_angles)

Inverse Design via Differentiable Optimization (2_inverse_design.ipynb)

Given target Fresnel coefficients, recover unknown film thicknesses using gradient-based optimization.

import torch
from difftmm import create_jones_matrix_isotropic

# Film stack with unknown thicknesses
n_list = torch.tensor([2.10, 1.46, 2.10, 1.46, 2.10], device="cuda")
d_param = torch.nn.Parameter(torch.randn(5, device="cuda") * 0.5)

def param_to_thickness(p):
    return torch.sigmoid(p) * 0.19 + 0.01  # map to [0.01, 0.20] um

# Optimization loop
optimizer = torch.optim.Adam([d_param], lr=0.02)
for step in range(3000):
    optimizer.zero_grad()
    d = param_to_thickness(d_param)
    pred = forward_tmm(n_list, d, n_in=1.0, n_out=1.52, inp=inp)
    loss = ((pred - target).real ** 2 + (pred - target).imag ** 2).mean()
    loss.backward()
    optimizer.step()

Result: Layer thicknesses recovered from random initialization:

Layer     GT (nm)   Recovered (nm)    Error (nm)
  1        60.00           60.00          0.00
  2       130.00          130.00          0.00
  3        85.00           85.00          0.00
  4       110.00          110.00          0.00
  5        70.00           70.00          0.00

Two Solvers

• difftmm.IsotropicFilmSolver — Fast 2x2 transfer matrix method for isotropic materials (~200x faster than NumPy TMM)
• difftmm.FilmSolver (also AnisotropicFilmSolver) — General 4x4 transfer matrix method for both isotropic and anisotropic materials

Both solvers share the same API:

solver = Solver(
    n_in=1.0,                  # incident medium refractive index
    n_out=1.52,                # exit medium refractive index
    n_layers_list=[2.1, 1.46], # interior layer refractive indices
    d_layers=[0.08, 0.12],     # thicknesses in um (optional, random if None)
    device=torch.device("cuda"),
)
ts, tp, rs, rp = solver.simulate(theta, wvln)

Accuracy Validation

Validated against the reference NumPy TMM library (sbyrnes321/tmm) on surface plasmon resonance (SPR) calculations:

The anisotropic 4x4 solver is validated for energy conservation, isotropic limit accuracy, cross-polarization coupling, and reciprocity:

Performance Benchmarks

Speed (batch=16, NVIDIA A100)

Layers	TMM NumPy (s)	Anisotropic 4x4 (s)	Isotropic 2x2 (s)	Speedup (4x4)	Speedup (2x2)
3	0.389	0.043	0.002	9.1x	250.6x
11	0.782	0.135	0.004	5.8x	192.3x
25	1.463	0.291	0.008	5.0x	194.7x
39	2.145	0.455	0.011	4.7x	197.2x

GPU Memory (batch=16, forward + backward)

The isotropic 2x2 solver uses ~5x less GPU memory than the anisotropic 4x4 solver. NumPy TMM is CPU-only (0 GPU memory).

Repository Structure

├── difftmm/                        # Importable package
│   ├── __init__.py                 #   Public API
│   ├── film_solver_isotropic.py    #   2x2 isotropic solver (fast)
│   └── film_solver_anisotropic.py  #   4x4 anisotropic solver (general)
├── 1_forward_simu.ipynb            # Example: forward simulation
├── 2_inverse_design.ipynb          # Example: differentiable inverse design
├── tmm_numpy/                      # Reference NumPy TMM library (isotropic only)
│   ├── tmm_core.py                 #   Steven Byrnes' TMM implementation
│   └── manual.pdf                  #   Physics reference
├── benchmarks/                     # Accuracy and performance benchmarks
│   ├── 1_compare_angle_response_*.py
│   ├── 2_compare_speed.py
│   └── 3_compare_memory.py
├── pyproject.toml                  # Packaging metadata
├── CITATION.cff                    # Citation metadata
└── README.md

Physics

• 2x2 transfer matrix method: Standard formulation for isotropic multi-layer films
• 4x4 transfer matrix method: General formulation for anisotropic (birefringent) media
• Snell's law, Fresnel equations, evanescent wave handling beyond critical angle
• Bidirectional propagation (forward and reverse through the film stack)
• Complete polarization handling via Jones calculus

References

• S. J. Byrnes, "Multilayer optical calculations," arXiv:1603.02720
• Steven Byrnes' TMM library: github.com/sbyrnes321/tmm
• Yang, X., Liu, Z., Nie, Z., Fan, Q., Shi, Z., Bonar, J., & Heidrich, W. (2026). "End-to-end differentiable design of geometric waveguide displays." arXiv preprint arXiv:2601.04370

License

DiffTMM is licensed under the Apache License 2.0.

The bundled NumPy TMM reference library (tmm_numpy/) is by Steven Byrnes and is licensed under the MIT License.