SciML / SciML/NeuralOperators.jl
Implement Transolver (Physics-Attention Transformer Solver)
Nobody has claimed this yet.
- Dominant language
- Julia
- Stars
- 41
- Forks
- 15
- Avg merge
- 13h 14m
- Merged PRs (30d)
- 12
Description
Summary
Implement Transolver, which introduces Physics-Attention that decomposes discretized domains into learnable physical-state slices.
Reference
- Wu et al., "Transolver: A Fast Transformer Solver for PDEs on General Geometries," ICML 2024. arXiv:2402.02366
Description
Transolver introduces "Physics-Attention" that groups mesh points into learnable slices based on similar physical states (rather than spatial proximity). Attention is applied to these physics-aware tokens, which better captures intricate physical correlations than raw mesh-point attention. Works on general geometries without requiring regular grids.
Key innovation: Instead of attending over all mesh points (quadratic cost), it learns to partition the domain into a small number of physically meaningful slices and operates on those.
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 ICML 2024 paper and its arXiv version to understand Transolver and Physics-Attention. The issue names no files, tests, or entry points; done means implementing the described physics-aware slicing and attention solver in NeuralOperators.jl.
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
- Needs clarification
- Newbie friendliness
- 25/100