Research map / Architectures and efficiency / Attention and transformers
MSPT: Efficient Large-Scale Physical Modeling via Parallelized Multi-Scale Attention
Pedro M. P. Curvo, Jan-Willem van de Meent, Maksim Zhdanov
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Transolver global slices with MSPT patch attention plus pooled supernodes | MSPT · Navier-Stokes | relative L2 (x1e-2): 9.00 -> 6.32 | Improved |
| MSPT versus Transolver on Elasticity | MSPT · Elasticity | relative L2 (x1e-2): 0.64 -> 0.48 | Improved |
| MSPT versus Transolver on Darcy | MSPT · Darcy | relative L2 (x1e-2): 0.57 -> 0.63 | Hurt |
| MSPT versus Erwin on Elasticity | MSPT · Elasticity | relative L2 (x1e-2): 0.34 -> 0.48 | Hurt |
| MSPT versus Transolver on ShapeNet-Car volume | MSPT · ShapeNet-Car volume | relative L2 (x1e-2): 2.07 -> 1.89 | Improved |
| MSPT versus Transolver on AhmedML volume | MSPT · AhmedML volume | relative L2 (x1e-2): 2.05 -> 2.04 | No clear effect |
| MSPT versus Transolver on AhmedML surface | MSPT · AhmedML surface | relative L2 (x1e-2): 3.45 -> 3.22 | Improved |
| vary number of patches K from 32 to 1024 (ablation) | MSPT · ShapeNet-Car test loss | test loss (x1e-2): 6.08 (K=32), 6.83 (K=128), 5.99 (K=1024) | Mixed |
| use max pooling instead of mean pooling for supernodes (ablation) | MSPT · ShapeNet-Car test loss | test loss | Hurt |
| use learned linear projection instead of mean pooling (ablation) | MSPT · ShapeNet-Car test loss | test loss | No clear effect |
| increase supernodes per patch Q above 1 (ablation) | MSPT · ShapeNet-Car test loss | test loss | Improved |
| MSPT instead of Transolver for memory at 500k points | MSPT · 500k points efficiency | peak memory (GB): 42.8 -> 26.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Erwin — Erwin: A Tree-based Hierarchical Transformer for Large-scale Physical Systems (ICML 2025) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries (ICML 2025 · similarity 0.88)
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations (ICML 2025 · similarity 0.88)
- GeoTransolver: Learning Physics on Irregular Domains using Multi-scale Geometry Aware Physics Attention Transformer (NeurIPS 2026 · similarity 0.87)
- Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators (NeurIPS 2024 · similarity 0.87)
- Transolver: A Fast Transformer Solver for PDEs on General Geometries (ICML 2024 · similarity 0.86)
- Erwin: A Tree-based Hierarchical Transformer for Large-scale Physical Systems (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
