SciML / SciML/NeuralOperators.jl
Implement Adaptive Fourier Neural Operator (AFNO)
Nobody has claimed this yet.
- Dominant language
- Julia
- Stars
- 41
- Forks
- 15
- Avg merge
- 13h 14m
- Merged PRs (30d)
- 12
Description
Summary
Implement AFNO, which bridges vision transformers and neural operators via Fourier-domain token mixing.
Reference
- Guibas et al., "Adaptive Fourier Neural Operators: Efficient Token Mixers for Transformers," ICLR 2022. arXiv:2111.13587
Description
AFNO uses block-diagonal structure in Fourier space with adaptive weight sharing and soft-thresholding sparsification for token mixing. This achieves quasi-linear complexity and provides an efficient alternative to standard self-attention. It is the backbone of FourCastNet for global weather forecasting.
Key components:
- Fourier-domain token mixing (FFT → block-diagonal multiply → soft-threshold → IFFT)
- Block-diagonal weight structure with adaptive sharing
- Sparsification via soft-thresholding in Fourier space
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reading the linked Guibas et al. AFNO paper and comparing its FFT, block-diagonal mixing, soft-thresholding, and IFFT steps with the repository's existing neural-operator implementations. Done means AFNO is implemented with adaptive weight sharing and Fourier-space sparsification as described in the issue, with behavior validated against the reference design.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- julia
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100