Research map / Architectures and efficiency / Attention and transformers
Scalable Transformer for PDE Surrogate Modeling
Zijie Li, Dule Shu, Amir Barati Farimani
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 linear attention with factorized attention (ablation) | FactFormer · 2D Kolmogorov flow 128x128 | relative L2 error: 0.5853 -> 0.1529 | Improved |
| replace linear attention with factorized attention (ablation) | FactFormer · 2D Kolmogorov flow 128x128 | fwd+bwd time (s): 0.3013 -> 0.0954 | Improved |
| replace linear attention with factorized attention (ablation) | FactFormer · 2D Kolmogorov flow 128x128 | peak memory (MB): 12029 -> 5217 | Improved |
| use latent marching with pushforward instead of autoregressive training (ablation) | FactFormer · 2D Kolmogorov flow 256x256 | omega avg error: 0.8835 -> 0.1734 | Improved |
| use latent marching with pushforward instead of autoregressive training (ablation) | FactFormer · 3D smoke buoyancy | u final error: 0.7881 -> 0.4482 | Improved |
| use latent marching instead of autoregressive training (ablation) | FactFormer · 2D Kolmogorov inference time | inference time (s): 3.14 -> 1.38 | Improved |
| use latent marching instead of autoregressive training (ablation) | F-FNO2D · 2D Kolmogorov flow 256x256 | omega avg error: 0.1486 -> 0.2453 | Hurt |
| use FactFormer instead of Dil-ResNet | LM models · 3D isotropic turbulence | u avg error: 0.1250 -> 0.1670 | Hurt |
| use FactFormer instead of Dil-ResNet | LM models · 3D smoke buoyancy | d final error: 0.1062 -> 0.0941 | Improved |
| use FactFormer instead of Dil-ResNet | LM models · 3D smoke buoyancy | inference time (s): 6.94 -> 2.62 | Improved |
| use fewer attention heads (ablation) | FactFormer · 2D Kolmogorov flow 128x128 | prediction error | Hurt |
| increase kernel dimension (ablation) | FactFormer · 2D Kolmogorov flow 128x128 | prediction error | Improved |
Closest papers by meaning
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations (ICML 2025 · similarity 0.86)
- Transolver: A Fast Transformer Solver for PDEs on General Geometries (ICML 2024 · similarity 0.84)
- Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains (NeurIPS 2025 · similarity 0.84)
- Inducing Point Operator Transformer: A Flexible and Scalable Architecture for Solving PDEs (AAAI 2024 · similarity 0.84)
- Vectorized Conditional Neural Fields: A Framework for Solving Time-dependent Parametric Partial Differential Equations (ICML 2024 · similarity 0.84)
- Toward Efficient Kernel-Based Solvers for Nonlinear PDEs (ICML 2025 · similarity 0.84)
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.
