Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Locally Subspace-Informed Neural Operators for Efficient Multiscale PDE Solving
Alexander Rudikov, Vladimir Fanaskov, Sergei Stepanov, Buzheng Shan, Ekaterina Muravleva, Yalchin Efendiev, Ivan Oseledets
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 subspace alignment loss instead of RBFL2 loss (ablation) | GMsFEM-NO · Richards, simple RHS (100x100, Nbf=8) | relative L2 error: 3.46% -> 1.87% | Improved |
| use subspace alignment loss instead of RBFL2 loss (ablation) | GMsFEM-NO · Richards, complex RHS (100x100, Nbf=4) | relative L2 error: 9.78% -> 6.11% | Improved |
| add projection regularization to SAL (ablation) | GMsFEM-NO · Richards, simple RHS (250x250) | relative L2 error: 1.82% -> 1.72% | Improved |
| add projection regularization to SAL (ablation) | GMsFEM-NO · small grids (100x100) | relative L2 error: 1.06% vs 1.06% | No clear effect |
| predict GMsFEM basis with neural operator instead of local eigenproblems | GMsFEM · Basis generation, 100x100x100 | time (s): 10547.2 -> 1.33 | Improved |
| predict GMsFEM basis subspace instead of full solution with standalone NO | F-FNO / GNOT / Transolver++ · Richards, complex RHS (250x250) | relative L2 error: 4.45% -> 1.60% | Improved |
| use GMsFEM-NO instead of F-FNO when training data is reduced to 200 samples (ablation) | GMsFEM-NO · Richards, simple RHS (250x250) | relative L2 error: 6.52% -> 2.07% | Improved |
| test standalone NO on out-of-distribution forcing term | standalone NO · Richards, complex RHS (100x100) | relative L2 error: 6503% | Hurt |
| train on 100 grid and test on 500 grid instead of training on 250 grid (ablation) | GMsFEM-NO · Richards, simple RHS (500x500 test) | relative L2 error: 2.25% -> 4.70% | Hurt |
| use 15 random test vectors instead of 10 (ablation) | GMsFEM-NO SAL-PR · Diffusion, simple RHS (100x100) | relative L2 error: 1.06% -> 1.31% | Hurt |
| use Nbf=4 instead of 8 basis functions (ablation) | GMsFEM-NO · Richards, complex RHS (100x100) | relative L2 error: 2.99% -> 6.11% | Hurt |
| replace GMsFEM eigenbasis with NO-predicted basis | GMsFEM · 3D 100x100x100 diffusion simple RHS | relative L2 error: 1.62% -> 1.68% | 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
- UGrid: An Efficient-And-Rigorous Neural Multigrid Solver for Linear PDEs (ICML 2024 · similarity 0.89)
- BENO: Boundary-embedded Neural Operators for Elliptic PDEs (ICLR 2024 · similarity 0.87)
- Holistic Physics Solver: Learning PDEs in a Unified Spectral-Physical Space (ICML 2025 · similarity 0.87)
- MC^2: Monte Carlo Correction for Fast Elliptic PDE Solving (NeurIPS 2026 · similarity 0.87)
- Neural Spectral Methods: Self-supervised learning in the spectral domain (ICLR 2024 · similarity 0.87)
- Neural Green’s Functions (NeurIPS 2025 · similarity 0.87)
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.
