Research map / Probabilistic and scientific modeling / Differential equations and PDEs
DeltaPhi: Physical States Residual Learning for Neural Operators in Data-Limited PDE Solving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeltaPhi, which reformulates neural-operator PDE solving as learning residuals between a query and a retrieved similar training sample, giving implicit data augmentation for data-limited settings.
Problem
Neural operators for PDE solving need large amounts of high-quality data, which is expensive to collect and often limited in resolution and biased in distribution, hurting generalization.
Approach
Instead of mapping input to solution directly, the model takes an input and a similar auxiliary training sample (input and solution) and predicts the solution residual relative to the auxiliary solution; random sampling among top-K similar training samples during training yields implicit data augmentation, justified by the stability (Lipschitz continuity) of PDE solution operators.
Key results
- On irregular domains, NORM (DeltaPhi) reduces relative error vs NORM direct by up to 50% (Heat Transfer, 2.70e-4 to 1.35e-4) and 41% on Pipe Turbulence.
- On regular Darcy Flow with 100 training samples, FFNO error drops from 5.22e-2 to 2.93e-2 (43.76% relative gain) with DeltaPhi.
- On Navier-Stokes gains are modest (e.g., FNO 4.86%, Resnet 1.87%), with the authors attributing this to chaotic low-viscosity dynamics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DeltaPhi residual learning to NORM | NORM · Irregular Darcy Flow, Pipe Turbulence, Heat Transfer, Composite, and Blood Flow | relative error: 1.05e-2 -> 9.75e-3; 1.01e-2 -> 5.96e-3; 2.70e-4 -> 1.35e-4; 9.99e-3 -> 9.18e-3; 4.82e-2 -> 4.29e-2 | Helped |
| Apply DeltaPhi residual learning to FNO | FNO · Darcy Flow and Navier-Stokes | relative L2 error: 3.70e-2 -> 3.31e-2; 2.24e-1 -> 2.13e-1 | Helped |
| Apply DeltaPhi to FNO for zero-shot resolution generalization | FNO · Darcy Flow, train size 100 and 900; train resolution 85 × 85, test resolution 421 × 421 | relative error: 7.29e-2 -> 6.91e-2; 6.73e-2 -> 4.91e-2 | Helped |
| Replace similarity-based retrieval with random auxiliary sampling (ablation) | FNO (DeltaPhi) · Darcy Flow, train sizes 300, 500, and 700 | relative error: 1.41e-2 -> 1.62e-2; 1.02e-2 -> 1.29e-2; 8.37e-3 -> 1.10e-2 | Hurt |
| Use small retrieval range K=5 rather than K=20 (ablation) | FNO (DeltaPhi) · Darcy Flow test Split2 | relative error: 5.00e-2 -> 5.01e-2 | Hurt |
| Use retrieval range K=20 rather than direct learning (ablation) | FNO (DeltaPhi) · Darcy Flow, train size 100 | relative error: 3.57e-2 -> 3.31e-2 | Helped |
| Use partial rather than full auxiliary input (ablation) | FNO (DeltaPhi) · Navier-Stokes, train size 100 | relative error: 2.14e-1 -> 2.13e-1 | Helped |
| Add auxiliary solution u_k_i(x) to customized input (ablation) | FNO (DeltaPhi) · Navier-Stokes, train size 100 | relative error: 2.34e-1 -> 2.13e-1; 2.37e-1 -> 2.13e-1 | Helped |
| Add similarity score to customized input (ablation) | FNO (DeltaPhi) · Navier-Stokes, train size 100 | relative error: 2.13e-1 -> 2.13e-1; 2.37e-1 -> 2.34e-1 | Helped |
| Replace cosine similarity with Euclidean distance (ablation) | FNO (DeltaPhi) · Darcy Flow | relative error: 3.31e-2 -> 3.31e-2 | No clear effect |
| Replace cosine similarity with Manhattan distance (ablation) | FNO (DeltaPhi) · Darcy Flow | relative error: 3.31e-2 -> 3.28e-2 | Helped |
| Replace fixed-range retrieval with importance sampling (ablation) | DeltaPhi · Heat Transfer | relative error: 1.35e-4 -> 2.02e-4 | Hurt |
Surprising findings
- Learned residual operators approximately respect the symmetry of the residual mapping (symmetry loss drops from 2.01 to 3.22e-2) without explicit constraints.
- Importance sampling over all training samples slightly outperforms fixed top-K retrieval on Darcy Flow (3.22e-2 vs 3.31e-2).
Limitations
- Gains are smaller on chaotic low-viscosity Navier-Stokes, where cosine similarity of initial states may not capture trajectory similarity.
- Relies on a single cosine similarity metric for retrieval, which may not be optimal for all physical systems.
- Evaluated only on relatively small benchmark datasets; the authors note more realistic, costly-data applications remain to be studied.
Experimental setup
Datasets: Darcy Flow (421x421) Navier-Stokes (viscosity 1e-5, 64x64) Irregular Darcy Pipe Turbulence Heat Transfer Composite Blood Flow
Benchmarks: Darcy Flow Navier-Stokes Irregular Darcy Pipe Turbulence Heat Transfer Composite Blood Flow
Metrics: relative L2 error
Baselines: FNO FFNO CFNO GNOT Galerkin MiOnet ResNet GraphSAGE DeepONet POD-DeepONet NORM
Builds on: Fourier Neural Operator DeepONet ResMem NORM
Closest papers by meaning
- Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning (NeurIPS 2024 · similarity 0.87)
- Physics-Informed Deep Inverse Operator Networks for Solving PDE Inverse Problems (ICLR 2025 · similarity 0.86)
- RIGNO: A Graph-based Framework For Robust And Accurate Operator Learning For PDEs On Arbitrary Domains (NeurIPS 2025 · similarity 0.86)
- Active Learning for Neural PDE Solvers (ICLR 2025 · similarity 0.86)
- PhysGuard: Fisher-Guided Gradient Projection for Sim-to-Real Neural PDE Surrogates (NeurIPS 2026 · similarity 0.86)
- Adaptive Coordinate Transforms for Neural Operators (NeurIPS 2026 · 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.