Research map / Label-efficient and robust learning / Recognition and long-tail
GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training
Haixu Wu, Minghao Guo, Zongyi Li, Zhiyang Dou, Mingsheng Long, Kaiming He, Wojciech Matusik
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 |
|---|---|---|---|
| add dynamics-lifted GeoPT pre-training (vs training from scratch) | Transolver · DTCHull | relative L2 / data requirement: 60% data requirement reduction | Improved |
| add dynamics-lifted GeoPT pre-training | Transolver · NASA-CRM | relative L2: moderate improvement | Improved |
| add dynamics-lifted GeoPT pre-training | Transolver · DrivAerML | data saving / convergence speed: 40% data saving, 2x faster converging | Improved |
| add dynamics-lifted GeoPT pre-training | Transolver · AirCraft, NASA-CRM, DTCHull, Car-Crash | data saving: 20-60% fewer physics samples; up to 2x faster convergence | Improved |
| pre-train with geometry-only supervision (vector distance) instead of training from scratch (ablation) | Transolver · DrivAerML | relative L2 | Hurt |
| condition on Hunyuan3D geometry representation (geometry-only conditioning) (ablation) | Transolver · industrial benchmarks | relative L2 | No clear effect |
| pre-train backbone with GeoPT instead of using GeoPT features as conditioning (ablation) | Transolver · industrial benchmarks | relative L2 | Improved |
| increase model depth from 8 to 32 layers with GeoPT pre-training (ablation) | Transolver · limited-data industrial simulation | relative L2 | Improved |
| reduce the number of unique pre-training geometries (ablation) | GeoPT · downstream physics tasks | relative L2 | Hurt |
| reduce sampled dynamic trajectories per geometry to 6% (ablation) | GeoPT · DrivAerML (fixed inflow) | relative L2: 6% comparable to 100 samples | No clear effect |
| increase dynamic trajectories per geometry (ablation) | GeoPT · AirCraft (varied speed, AoA, sideslip) | relative L2 | Improved |
| apply GeoPT to radiosity simulation | Transolver · Radiosity (Cornell box) | MAE: 9.7×10−2 -> 9.0×10−2 | Improved |
Closest papers by meaning
- GeoTransolver: Learning Physics on Irregular Domains using Multi-scale Geometry Aware Physics Attention Transformer (NeurIPS 2026 · similarity 0.88)
- Multiple Physics Pretraining for Spatiotemporal Surrogate Models (NeurIPS 2024 · similarity 0.86)
- Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators (NeurIPS 2024 · similarity 0.86)
- Space and time continuous physics simulation from partial observations (ICLR 2024 · similarity 0.85)
- Neural Modes: Self-supervised Learning of Nonlinear Modal Subspaces (CVPR 2024 · similarity 0.85)
- A meshfree exterior calculus for generalizable and data-efficient learning of physics from point clouds (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.
