Research map / Architectures and efficiency / Mixture of experts
NESTOR: A Nested MOE-based Neural Operator for Large-Scale PDE Pre-Training
Dengdi Sun, Xiaoya Zhou, Xiao Wang, Hao Si, Wanli Lyu, Jin Tang, Bin Luo
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 token-level Sub-MoE (ablation) | NESTOR · PDEBench 6 sub-tasks | Avg L2RE: 0.0197 -> 0.0173 | Improved |
| add load balancing loss (ablation) | NESTOR · PDEBench 6 sub-tasks | Avg L2RE: 0.0178 -> 0.0173 | Mixed |
| fuse AFNO and FlashAttention via MoE instead of simple addition (ablation) | NESTOR · PDEBench 6 sub-tasks | Avg L2RE: 0.0196 -> 0.0173 | Improved |
| increase non-shared experts from 6 to 12 (ablation) | NESTOR · FT-500 average | Avg L2RE: 0.0213 -> 0.0230 | Hurt |
| increase non-shared experts from 2 to 6 (ablation) | NESTOR · FT-500 average | Avg L2RE: 0.0222 -> 0.0213 | Improved |
| increase non-shared experts from 6 to 12 at FT-200 (ablation) | NESTOR · FT-200 average | Avg L2RE: 0.0246 -> 0.0832 | Hurt |
| increase pre-training datasets from 3 to 12 (ablation) | NESTOR · FT-500 average | Avg L2RE: 0.0234 -> 0.0208 | Improved |
| fine-tune 500 epochs instead of 200 (ablation) | NESTOR · PDEBench CNS eta=1 avg | L2RE: 0.0161 -> 0.0114 | Improved |
| NESTOR instead of DPOT-T at pre-training | PDE pre-training · PDEBench CNS eta=0.1 avg | L2RE: 0.0176 -> 0.0161 | Improved |
| NESTOR instead of DPOT-T at pre-training | PDE pre-training · PDEArena NS | L2RE: 0.125 -> 0.132 | Hurt |
| fine-tune from pre-trained weights instead of training from scratch | 2D turbulence 512x512 · 2D high-resolution turbulence | L2RE: 47.3% improvement | Improved |
Closest papers by meaning
- Mixture-of-Experts Operator Transformer for Large-Scale PDE Pre-Training (NeurIPS 2025 · similarity 0.91)
- DPOT: Auto-Regressive Denoising Operator Transformer for Large-Scale PDE Pre-Training (ICML 2024 · similarity 0.87)
- Pretraining Codomain Attention Neural Operators for Solving Multiphysics PDEs (NeurIPS 2024 · similarity 0.86)
- RIGNO: A Graph-based Framework For Robust And Accurate Operator Learning For PDEs On Arbitrary Domains (NeurIPS 2025 · similarity 0.86)
- Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning (NeurIPS 2024 · similarity 0.86)
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · 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.
