Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries
Hang Zhou, Haixu Wu, Haonan Shangguan, Yuezhou Ma, Huikun Weng, Jianmin Wang, Mingsheng Long
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 |
|---|---|---|---|
| use Transolver-3 instead of AB-UPT | Neural PDE solver · NASA-CRM / AhmedML / DrivAerML | relative L2 error: DrivAerML u 5.93 -> 4.14; NASA-CRM ps 9.77 -> 8.71 | Improved |
| use Transolver-3 instead of AB-UPT on AhmedML volume pressure | Neural PDE solver · AhmedML | volume pressure pv error: 2.07 -> 2.16 | Hurt |
| use Transolver-3 on Aircraft dataset with shocks | Transolver++ / AB-UPT · Aircraft | surface pressure error: 6.40 -> 5.80 (Transolver++); 6.93 -> 5.80 (AB-UPT) | Improved |
| reorder Physics-Attention with faster slice and deslice (move projections into slice domain) (ablation) | Physics-Attention · Single-GPU capacity | cells per GPU / latency: about 10% more cells than Transolver++; latency about 60% lower; GFLOPs 20% lower | Improved |
| add geometry slice tiling (ablation) | Transolver-3 · Single-GPU capacity | max cells: boosts capacity by roughly 90% (1.9x Transolver++, 2.9M cells) | Improved |
| reduce tile size from 800k to 100k to 5k (ablation) | Geometry slice tiling · 800k input training | memory / latency: memory 22.16 -> 12.28 -> 11.04 GB; latency 404 -> 429 -> 1307 ms | Mixed |
| increase training subset size n from 1k to 100k (ablation) | Geometry amortized training · DrivAerML | relative L2 error: ps 4.84 -> 3.71; pv 7.94 -> 5.72 | Improved |
| train on small random subset (n=1000) (ablation) | Geometry amortized training · DrivAerML | relative L2 error | No clear effect |
| use more cells for physical state cache up to full mesh (ablation) | Physical state caching · DrivAerML | relative L2 error | Improved |
| increase evaluation surface cell density (ablation) | Full mesh decoding · DrivAerML | drag and lift coefficient error | Improved |
| use decoupled inference compared with AB-UPT | Transolver-3 · DrivAerML | inference memory / latency: memory 11.2 -> 6.0 GB; latency 220 -> 260 s | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Transolver++ — Transolver: A Fast Transformer Solver for PDEs on General Geometries (ICML 2024) · better on 5 of 5 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 0.94)
- Transolver: A Fast Transformer Solver for PDEs on General Geometries (ICML 2024 · similarity 0.93)
- Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains (NeurIPS 2025 · similarity 0.87)
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations (ICML 2025 · similarity 0.86)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.85)
- Efficient Generative Transformer Operators for Million-Point PDEs (NeurIPS 2026 · similarity 0.85)
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.
