Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Test-time Generalization for Physics through Neural Operator Splitting
Louis Serrano, Jiequn Han, Edouard Oyallon, Shirley Ho, Rudy Morel
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 test-time beam search over operator compositions with operator splitting | DISCO · Unseen PDE combination: Advection+Diffusion | NRMSE: 0.170 -> 0.015 | Improved |
| add test-time beam search over operator compositions | DISCO · Unseen PDE combination: Euler+Diffusion | NRMSE: 0.572 -> 0.066 | Improved |
| use beam search instead of uniform sampling to pick operators | operator composition search · Unseen PDE combination: All Three | NRMSE: 0.075 -> 0.036 | Improved |
| use beam search instead of uniform sampling | operator composition search · Unseen PDE combination: Nonlinear Adv.+Diffusion | NRMSE: 0.068 -> 0.056 | Improved |
| add test-time beam search | DISCO · Parameter extrapolation: Adv.-Diff. D | NRMSE: 0.159 -> 0.002 | Improved |
| add test-time beam search | DISCO · Parameter extrapolation: Combined gamma | NRMSE: 1.007 -> 0.022 | Improved |
| add test-time beam search | DISCO · Parameter extrapolation: Adv.-Diff. c | NRMSE: 0.768 -> 0.052 | Improved |
| increase number of uniform-search trials (ablation) | uniform search · OOD advection-diffusion | fitting and prediction NRMSE | Improved |
| increase dictionary size from 64 to 256 under Mixed 0.5 training (ablation) | beam search · Mixed-physics training (Mixed 0.5) | NRMSE improvement over direct: 0% -> 41% | Improved |
| increase dictionary size from 64 to 256 under Mixed 1.0 training (ablation) | beam search · Mixed-physics training (Mixed 1.0) | NRMSE improvement over direct: 15% -> 32% | Improved |
| increase unseen nonlinear perturbation eps from 0 to 1.00 (ablation) | beam search · Adv-Diff with Burgers perturbation | Beam NRMSE: 0.055 -> 0.797 | Hurt |
| improve accuracy of individual learned operators via longer pretraining (epoch 50 to 500) (ablation) | Strang splitting of heat and dispersion operators · 1D heat-dispersion | Split Rollout NRMSE: 1.22e-01 -> 8.85e-03 | Improved |
Closest papers by meaning
- Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning (NeurIPS 2024 · similarity 0.88)
- DISCO: learning to DISCover an evolution Operator for multi-physics-agnostic prediction (ICML 2025 · similarity 0.88)
- HyPINO: Multi-Physics Neural Operators via HyperPINNs and the Method of Manufactured Solutions (NeurIPS 2025 · similarity 0.87)
- RIGNO: A Graph-based Framework For Robust And Accurate Operator Learning For PDEs On Arbitrary Domains (NeurIPS 2025 · similarity 0.87)
- Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators (NeurIPS 2024 · similarity 0.87)
- Boosting Generalization in Parametric PDE Neural Solvers through Adaptive Conditioning (NeurIPS 2024 · similarity 0.86)
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.
