Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Iterative Refinement Neural Operators are Learned Fixed-Point Solvers: A Principled Approach to Spectral Bias Mitigation
Xiaotian Liu, Shuyuan Shang, Xiaopeng Wang, Pu Ren, Yaoqing Yang
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 iterative refinement (IRNO) to a pre-trained FNO | FNO · TR-2D | VRMSE: 0.2394 -> 0.1309 | Improved |
| add iterative refinement (IRNO) to TFNO | TFNO · Active Matter | VRMSE: 0.1981 -> 0.0387 | Improved |
| add IRNO on top of a strong WDSR base | WDSR · ERA5 16x super-resolution | ACC: 0.9091 -> 0.9104 | Improved |
| use progressive spectral loss (lambda 1->2) instead of fixed lambda (ablation) | IRNO (TFNO base) · Active Matter | VRMSE: 0.051-0.070 -> 0.039 | Improved |
| add deep supervision over every refinement step (ablation) | IRNO (1-D synthetic) · 1-D synthetic Jacobian monotonicity | fraction with m > 0: 3.9% -> 100% | Improved |
| combine IRNO with HFS frequency-aware architecture | HFS · Active Matter | VRMSE: 0.0631 -> 0.0486 | Improved |
| use IRNO trained on a weaker base operator (TFNO) to refine a stronger one (FNO) | FNO · TR-2D | VRMSE improvement (%): 58.53% vs 45.32% for same-operator IRNO | Improved |
| step size alpha 0.3-0.5 instead of 0.2 (ablation) | IRNO (TFNO) · TR-2D | stability beyond training horizon | Hurt |
| step size alpha 0.6 (ablation) | IRNO (TFNO) · TR-2D | VRMSE: 0.12 -> 0.24 | Hurt |
| swap refinement backbone (ResNet, ConvNext, FNO) or normalization (BatchNorm, LayerNorm, GroupNorm) (ablation) | IRNO · Active Matter / TR-2D | VRMSE reduction: > 71% VRMSE reduction for all backbones | No clear effect |
| apply IRNO to RIGNO on irregular meshes with autoregressive rollout | RIGNO · CE-Gauss | relative L2 error (%): 16.617 -> 13.080 at t = 7 | Improved |
| use iterative refinement instead of a capacity-matched monolithic residual model | FNO · ERA5 | ACC at 1100 GFLOPs: 0.79 -> 0.84 | Improved |
Closest papers by meaning
- Spectral-Refiner: Accurate Fine-Tuning of Spatiotemporal Fourier Neural Operator for Turbulent Flows (ICLR 2025 · similarity 0.88)
- Guaranteed Approximation Bounds for Mixed-Precision Neural Operators (ICLR 2024 · similarity 0.85)
- Harnessing the Power of Neural Operators with Automatically Encoded Conservation Laws (ICML 2024 · similarity 0.85)
- Hierarchical Neural Operator Transformer with Learnable Frequency-aware Loss Prior for Arbitrary-scale Super-resolution (ICML 2024 · similarity 0.85)
- RIGNO: A Graph-based Framework For Robust And Accurate Operator Learning For PDEs On Arbitrary Domains (NeurIPS 2025 · similarity 0.85)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.84)
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.
