Graph Feedforward Network (GFN) - a novel neural network layer for resolution-invariant machine learning

Why GFNs? • Documentation • Installation • Quickstart • Citing

GFN is a generalisation of feedforward networks for graphical data.

[!IMPORTANT] The code reproducing the results in the GFN-ROM paper has now been moved to Oisin-M/GFN-ROM.

Why GFNs?

Many applications rely upon graphical data, which standard machine learning methods such as feedforward networks and convolutions cannot handle. GFNs present a novel approach of tackling this problem by extending existing machine learning approaches for use on graphical data. GFNs have very close links with neural operators and graph neural networks.

Key advantages of GFNs:

• Resolution invariance
• Equivalence to feedforward networks for single fidelity data (no deterioration in performance)
• Provable guarantees on performance for super- and sub-resolution
• Both fixed and adapative multifidelity training possible

Installation

gfn is readily available on PyPI.

pip install gfn-layer

Note: the package name on PyPI is gfn-layer, gfn refers to a different package.

Quickstart

See the user guide to get started!

Citing

If this work is useful to you, please cite

[1] Morrison, O. M., Pichi, F. and Hesthaven, J. S. (2024) ‘GFN: A graph feedforward network for resolution-invariant reduced operator learning in multifidelity applications’. Available at: arXiv and Computer Methods in Applied Mechanics and Engineering

@article{Morrison2024,
  title = {{GFN}: {A} graph feedforward network for resolution-invariant reduced operator learning in multifidelity applications},
  author = {Morrison, Oisín M. and Pichi, Federico and Hesthaven, Jan S.},
  journal = {Computer Methods in Applied Mechanics and Engineering},
  volume = {432},
  pages = {117458},
  year = {2024},
  doi = {10.1016/j.cma.2024.117458},
}