FNETs

This repo contains the code, data and weights used in the paper: High-Fidelity Prediction of Perturbed Optical Fields using Fourier Feature Networks.

Citation

All python code and model weights are free to used under a standard MIT licence.

Please cite our associated paper (available on ArXiV):

@article{jandrell2025fnet,
  title        = {High-Fidelity Prediction of Perturbed Optical Fields using Fourier Feature Networks},
  author       = {Jandrell, Joshua R. and Cox, Mitchell A.},
  year         = {2025},
  month        = sep,
  journal      = {arXiv preprint},
  volume       = {arXiv:2508.19751},
  doi          = {10.48550/arXiv.2508.19751},
  url          = {https://arxiv.org/abs/2508.19751}
}

Important

Licencing and Citation does not cover the training data (.h5) files. These files contain reformatted data originally collected by M. W. Matthès, Y. Bromberg, J. de Rosny and S. M. Popoff for their paper Learning and Avoiding Disorder in Multimode Fibers.

Original data can be retrieved from this repo.

⚡Quick Demo

For a brief demonstration of how to load model weights and plots of how well model predictions match data see plot_fits.ipynb

Pre-Trained Model Weights

Weights for pretrained pytorch models can be found in the Data/Weights directory.

You can use these files to load full pytorch models without including or importing any additional code:

import torch
model = torch.load(<path-to-weights> weights_only=False)

Tip

All models predict real and imaginary complex components as two concatenated tensors of floating point values. The custom oil/complex module can be used to convert these into a single tensor of torch.complex values.

from oil.complex import ComplexConverter
re_im_cc = ComplexConverter(expand_method='concat',representation='real-imag')
raw_pred = model(x)
complex_pred = re_im_cc.float_to_complex(raw_pred)

Additional Source Code

Source code for a Fourier feature pytorch module, neural networks models and other useful custom pytorch code can be found in the custom oil module. Feel free to use this code as you wish! We hope you find it helpful (please cite our paper if you do).

The oil module contains some scripts we wrote to make pytorch run a bit smoother and be a bit simpler to use. The version included in this repository is incomplete and is provided (without warranty) under the this repo’s general MIT license.