Research map / Probabilistic and scientific modeling / Differential equations and PDEs
From Cheap Geometry to Expensive Physics: Elevating Neural Operators via Latent Shape Pretraining
Zhizhou Zhang, Youjia Wu, Kaixuan Zhang, Yanjia Wang
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 |
|---|---|---|---|
| feed stage-1 pretrained occupancy VAE latents to the operator instead of raw point clouds | GNOT · Electrostatics, mesh query | relative L2 (x10^-2): 4.2 -> 3.3 | Improved |
| feed stage-1 pretrained latents instead of raw point clouds | LNO · Stress, mesh query | relative L2 (x10^-2): 26.5 -> 13.6 | Improved |
| feed stage-1 pretrained latents instead of raw point clouds | Transolver · AirfRans (near), random query | relative L2 (x10^-2): 15.0 -> 10.8 | Improved |
| feed stage-1 pretrained latents instead of raw point clouds | GNOT · Inductor (3D), mesh query | relative L2 (x10^-2): 7.0 -> 7.1 | No clear effect |
| feed stage-1 pretrained latents instead of raw point clouds | LNO · AirfRans (near), mesh query | relative L2 (x10^-2): 27.4 -> 27.1 | No clear effect |
| use random queries instead of mesh queries | two-stage framework · all four datasets | size of improvement | Improved |
| pretrain VAE on more geometry data (VAE3 vs VAE1) (ablation) | GNOT with VAE latents · Stress, random query | relative L2 (x10^-2): 9.1 -> 8.3 | Improved |
| pretrain VAE only on the same geometries as stage 2 (VAE1) (ablation) | Transolver · Stress, mesh query | relative L2 (x10^-2): 11.5 -> 12.0 | Hurt |
| lower KL weight from 0.001 to 0.0001 (ablation) | GNOT with VAE latents · AirfRans (near), mesh query | relative L2 (x10^-2): 5.6 -> 5.3 | Improved |
| use deterministic autoencoder instead of VAE (ablation) | Transolver · AirfRans (near), mesh query | relative L2 (x10^-2): 12.7 -> 11.3 | Improved |
| use shortest-vector proxy task instead of occupancy (ablation) | GNOT · AirfRans, mesh query | relative L2 (x10^-2): 5.6 -> 4.8 | Improved |
| use SDF proxy task pretraining with mesh query (ablation) | GNOT · AirfRans, mesh query | relative L2 (x10^-2): 5.2 -> 5.3 | No clear effect |
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 4 of 4 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
- Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning (NeurIPS 2024 · similarity 0.92)
- Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains (NeurIPS 2025 · similarity 0.91)
- Adaptive Coordinate Transforms for Neural Operators (NeurIPS 2026 · similarity 0.90)
- RIGNO: A Graph-based Framework For Robust And Accurate Operator Learning For PDEs On Arbitrary Domains (NeurIPS 2025 · similarity 0.90)
- Reference Neural Operators: Learning the Smooth Dependence of Solutions of PDEs on Geometric Deformations (ICML 2024 · similarity 0.89)
- Latent Neural Operator for Solving Forward and Inverse PDE Problems (NeurIPS 2024 · similarity 0.89)
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.
