mmf-tm-simulated / README.md
callucallum's picture
Added DOI to 'Code, weights and splits' in Related section
f309ad6
|
Raw History Blame Contribute Delete
3.9 kB
metadata
license: cc-by-4.0
viewer: false
pretty_name: Simulated 1m multimode-fibre transmission matrices
tags:
  - optics
  - multimode-fibre
  - transmission-matrix
  - simulation

Simulated 1 m multimode-fibre transmission matrices

DOI

Citing: please use the BibTeX below, not the auto-generated Hugging Face entry (which names the organisation as author).

5,000 simulated transmission matrices (TMs) of a 1 m step-index multimode fibre with random bends, generated with the transfer matrix method (TMM), plus the matrices that convert them to the pixel basis. They were used to pretrain the autoencoder in the code repository, where five 1m TMs are multiplied to form one simulated 5m "composite" TM.

Simulated fibre

Fibre step-index, core diameter 50µm, NA 0.22, refractive index 1.457
Length 1m
Wavelength 632.8nm
Guided modes kept 386 (scalar LP modes, one polarisation, one azimuthal orientation)
Bends random bend radii and directions, saved alongside each TM
Filter only TMs with a 2-norm condition number below 250 were kept

Contents

mmf-tm-simulated/
├── TransferMatrixMethod/
│   ├── TM0000.pkl ... TM0999.pkl            TMs made by the CPU script
│   ├── TM1000-GPU.pkl ... TM4999-GPU.pkl    TMs made by the GPU script
│   ├── dist_pim_to_real.pkl                 distal-facet matrix: propagation-invariant modes (PIMs) -> pixel basis
│   ├── prox_real_to_pim.pkl                 proximal-facet matrix: pixel basis -> PIMs
│   ├── fibre_metadata.pkl                   parameters of the simulated fibre
│   └── orthonormal_random_inputs_371x371_145.npy   the 145 random orthonormal input patterns (371 x 371)
└── README.md

Each TM*.pkl is about 1.2MB and holds the TM in the 386-mode basis under the key tm_mode. All .pkl and .npy files are stored with Git LFS.

From a mode-basis TM to a pixel-basis TM

With u_in the input patterns (flattened to a column per pattern):

TM_pixel = dist_pim_to_real @ tm_mode @ prox_real_to_pim @ u_in

For a composite 5m fibre, tm_mode is replaced by the product of five 1m TMs in propagation order. Measurement noise is not stored in the files; the code repository adds it during training. See training/ and src/ there for the exact procedure.

Loading

import pickle

with open("TransferMatrixMethod/TM0000.pkl", "rb") as f:
    tm = pickle.load(f)["tm_mode"]      # 386 x 386, complex

pickle files can run code when loaded: only load files you downloaded from this repository.

Related

Acknowledgements and licence

The TMM simulation method and the original MATLAB fibre simulation code were developed by David B. Phillips and Robert J. Kilpatrick.

Licence: CC BY 4.0.

Contact: Callum J. Wilson, cw2011@cam.ac.uk

Citation

If you use this dataset, please cite it.

@dataset{wilson2026mmftmsimulated,
  author    = {Wilson, Callum J. and Gordon, George S.D.},
  title     = {Simulated 1m multimode-fibre transmission matrices},
  year      = {2026},
  publisher = {Hugging Face},
  version   = {v1.0.0},
  doi       = {https://doi.org/10.57967/hf/11677},
  url       = {https://huggingface.co/datasets/OPTIMlab/mmf-tm-simulated}
}