pith. sign in

arxiv: 2605.24041 · v2 · pith:M6VKBBNVnew · submitted 2026-05-21 · 💻 cs.LG · cs.AI

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 This is my paper

Pith reviewed 2026-06-30 17:01 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords neural operatorsspectral biasiterative refinementfixed-point iterationcontraction mappingprogressive spectral lossresidual correctionturbulent flow
0
0 comments X

The pith

Adding a learned refinement module turns neural operators into fixed-point solvers that converge and cut high-frequency errors.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that a pre-trained neural operator can be improved by attaching a separate learned module whose output is fed back into itself repeatedly until it stops changing. This process decomposes any prediction into an initial coarse guess followed by successive small corrections, mirroring how classical solvers refine solutions step by step. The authors prove that, when the module satisfies local contraction conditions, the iteration reaches a single stable result rather than oscillating or diverging. Training uses a loss that gradually places heavier weight on high-frequency mismatches, directly addressing the tendency of neural operators to miss fine details. A reader would care because this yields measurable drops in overall error and especially large gains on the small-scale features that matter for accurate physical simulations.

Core claim

The Iterative Refinement Neural Operator augments any pre-trained neural operator with a learned refinement module that is applied iteratively through fixed-point iteration. The prediction is formed as a coarse initialization plus successive residual corrections. Under local assumptions the induced operator is a contraction mapping and therefore converges to a unique fixed point. A progressive spectral loss is introduced that increases the penalty on high-frequency components at each refinement step during training.

What carries the argument

The learned refinement module applied repeatedly via fixed-point iteration, together with the progressive spectral loss that raises the weight on high-frequency residuals over steps.

If this is right

  • Error drops consistently on multiple physical systems, reaching a 56.05 percent reduction on turbulent flow.
  • On Active Matter the normalized error in high frequencies falls to 1.48-2.04 percent of the base operator's error.
  • The same error ratios hold when the iteration count exceeds the number of steps seen during training.
  • Low- and mid-frequency errors also decrease, to 27.72-36.10 percent and 5.07-6.68 percent of the base error respectively.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The fixed-point view may let practitioners combine the learned module with classical numerical correctors inside a single solver loop.
  • The same refinement construction could be tried on other single-pass models that exhibit frequency-dependent accuracy loss.
  • Checking whether the contraction property holds for random initial guesses rather than the network's own coarse output would test robustness outside the paper's local assumptions.

Load-bearing premise

The local conditions that make the refinement operator a contraction mapping are satisfied for the trained networks and target systems.

What would settle it

A test case in which repeated application of the refinement module either diverges or leaves high-frequency error unchanged or larger than the base operator on a held-out physical system.

Figures

Figures reproduced from arXiv: 2605.24041 by Pu Ren, Shuyuan Shang, Xiaopeng Wang, Xiaotian Liu, Yaoqing Yang.

Figure 1
Figure 1. Figure 1: Iterative Refinement on ERA5 16× Super-Resolution. The base opera￾tor (FNO) captures large-scale atmospheric structure but under-resolves fine-scale details. Successive refinement steps (k = 1 → 12) from IRNO progressively reduce high-frequency error while preserving global coherence. Horizontal dashed line indicates the training cutoff at K = 6; iterations beyond this point demonstrate stable extrapolatio… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Iterative Refinement Neural Operator (IRNO). (a) Standard neural operators approximate the solution in a monolithic, single forward evalua￾tion, often losing fine-scale details. (b) IRNO reformulates inference as a dynamic process. A Base Operator provides a coarse initialization, which is then iteratively corrected by a shared-weight refinement operator Φθ. At each step k, the network conc… view at source ↗
Figure 3
Figure 3. Figure 3: Empirical validation of bias–error floor relationship. Scatter plots of mini￾mum error mink ∥ek∥ as a function of bias magnitude ∥b∥ = ∥Φθ(x, y)∥ over k = 24 refinement steps. (a) Active Matter: Pearson r = 0.933 (p ≪ 10−10). (b) TR-2D: Pearson r = 0.949 (p ≪ 10−10). Both systems exhibit a strong linear dependence between the asymptotic error floor and the bias, consistent with Corollary 3.3. 3.2 Main Resu… view at source ↗
Figure 4
Figure 4. Figure 4: Convergence behavior across physical systems. VRMSE for TR-2D and Active Matter, and RFNE and ACC for ERA5, plotted as a function of refinement step k. Vertical dashed lines indicate the training cutoff (FNO at K = 6 and TFNO/WDSR at K = 4). Error metrics decrease or reach a stable plateau for k > K, while ACC increases and stabilizes. 4.2 Global Convergence Behavior Across Physical Regimes We evaluate whe… view at source ↗
Figure 5
Figure 5. Figure 5: Spectral Error Evolution under iterative refinement (Active Matter, FNO). (a) Median normalized spectral error ratios E˜(k) (ω) across the test set, with shaded interquartile ranges (25–75%). IRNO exhibits consistent attenuation of mid-to-high fre￾quency error with increasing refinement steps, with stable behavior near the Nyquist limit ω = 128, indicated by a vertical dashed line. (b) Spectral Mean Square… view at source ↗
Figure 6
Figure 6. Figure 6: Frequency-band error ratios across refinement steps (Active Matter, FNO). Normalized error ratios R (B) k for each tensor field, computed separately over (a) low-, (b) mid-, and (c) high-frequency bands as a function of refinement step k. The vertical dashed line marks the training cutoff at K = 6. Lower values indicate stronger error reduction relative to the base operator. denote the spectral error of sa… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation: Step size α sensitivity (TR-2D, TFNO). Convergence behavior for α ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6}, trained with K = 4 steps (dashed line). Small step sizes converge slowly; moderate step sizes (α ∈ [0.2, 0.4]) achieve optimal balance; α = 0.6 diverges rapidly beyond the training horizon, violating the contraction condition q = ∥I−αA(x)∥op < 1 from Theorem 3.1. depicted in the theory section 3. W… view at source ↗
Figure 8
Figure 8. Figure 8: Cost–performance trade-offs under iterative refinement (ERA5, FNO). ACC and RFNE plotted against computational cost (FLOPs) and memory usage during inference. Each marker corresponds to a refinement step of the iterative U-Net, with step indices annotated. Baselines (FNO, SRCNN, and 15× U-Net) are shown for reference. IRNO achieves improved ACC or lower error at comparable or lower computational cost. See … view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of the estimated strong-monotonicity constant [PITH_FULL_IMAGE:figures/full_fig_p030_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative results on ERA5 for Kinetic Energy (top) and Temperature Field [PITH_FULL_IMAGE:figures/full_fig_p041_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Temporal evolution of ACC on ERA5 under 8 [PITH_FULL_IMAGE:figures/full_fig_p042_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative results on ERA5 for Temperature Field under 16 [PITH_FULL_IMAGE:figures/full_fig_p043_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Temporal evolution of ACC on ERA5 under 16 [PITH_FULL_IMAGE:figures/full_fig_p044_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison between Ground Truth, Base FNO, and single-step refine [PITH_FULL_IMAGE:figures/full_fig_p045_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative results on TR-2D with FNO as base operator. The model is trained [PITH_FULL_IMAGE:figures/full_fig_p046_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative results on TR-2D with TFNO as base operator. The model is trained [PITH_FULL_IMAGE:figures/full_fig_p046_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative results on Active Matter with different base operators. We visualize [PITH_FULL_IMAGE:figures/full_fig_p047_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: VRMSE vs. refinement step k up to 8× the training horizon (K = 6, dashed vertical line) for α = 0.05 (stable) and α = 0.20 (diverges) with FNO-based IRNO on Active Matter. 47 [PITH_FULL_IMAGE:figures/full_fig_p047_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Spectral error energy vs. radial frequency for HFS-ResUNet and IRNO on Active [PITH_FULL_IMAGE:figures/full_fig_p048_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Spectral error vs. radial frequency for F-Adapter (left) and per-frequency im [PITH_FULL_IMAGE:figures/full_fig_p048_20.png] view at source ↗
read the original abstract

Neural operators serve as fast, data-driven surrogates for scientific modeling but typically rely on a monolithic, single-pass inference procedure that struggles to resolve high-frequency details, a limitation known as spectral bias. We introduce the Iterative Refinement Neural Operator (IRNO), which augments pre-trained operators with a learned refinement module iteratively applied via fixed-point iteration. IRNO decomposes the prediction into a coarse initialization followed by successive residual corrections, paralleling classical numerical solvers. Under local assumptions, we establish contraction of the induced operator, ensuring convergence to a unique fixed point. To explicitly target high-frequency errors, we propose a progressive spectral loss that adaptively increases penalty on high-frequency components over refinement steps during training. Across physical systems, IRNO consistently lowers error, with up to 56.05% improvement on turbulent flow. On Active Matter, spectral analysis reveals that, relative to base operator, the normalized error ratios decrease to 27.72-36.10% in low-, 5.07-6.68% in mid-, and 1.48-2.04% in high-frequencies, remaining stable beyond the trained iteration count. Code is available at https://github.com/xiaotianliu-dartmouth/Iterative_Refinement_Neural_Operator

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces the Iterative Refinement Neural Operator (IRNO) that augments a pre-trained neural operator with a learned refinement module applied iteratively as a fixed-point iteration. It decomposes the solution into a coarse initialization plus successive residual corrections. Under unspecified local assumptions, the authors prove that the induced operator is contractive and therefore converges to a unique fixed point. A progressive spectral loss is introduced that increases the penalty on high-frequency components across refinement steps. Experiments on turbulent flow and Active Matter report error reductions (up to 56.05 % on turbulent flow) and show that normalized high-frequency error ratios drop to 1.48–2.04 % while remaining stable past the training horizon. Reproducible code is released.

Significance. If the contraction result can be made operational, the work supplies a principled, architecture-agnostic route to spectral-bias mitigation that parallels classical iterative solvers. The explicit framing as learned fixed-point iteration, the progressive loss design, and the public code release are concrete strengths that would allow the community to test the method on additional PDEs.

major comments (2)
  1. [§3] §3 (Contraction Mapping): The local assumptions required for the refinement operator to be contractive (Lipschitz constant of the learned residual map <1, domain restrictions, regularity of the base operator) are invoked but never stated explicitly or bounded; without these the convergence theorem is not falsifiable and the advertised stability beyond the training horizon cannot be verified.
  2. [§5] §5 (Experiments on turbulent flow and Active Matter): No post-hoc numerical diagnostic is reported that checks whether the contraction condition actually holds on the test trajectories; the claim that error ratios remain stable at iteration counts exceeding those seen in training therefore rests on an unverified premise.
minor comments (1)
  1. [§4] Notation for the progressive spectral loss (Eq. (X)) should include an explicit schedule for the frequency-weighting parameter so that the training procedure is fully reproducible from the text alone.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive report and the positive assessment of the work's significance. We address each major comment below and commit to revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [§3] §3 (Contraction Mapping): The local assumptions required for the refinement operator to be contractive (Lipschitz constant of the learned residual map <1, domain restrictions, regularity of the base operator) are invoked but never stated explicitly or bounded; without these the convergence theorem is not falsifiable and the advertised stability beyond the training horizon cannot be verified.

    Authors: We agree that the local assumptions underlying the contraction result were not stated with sufficient explicitness or bounds in the original manuscript. In the revised version we will add a dedicated subsection that precisely enumerates the assumptions (Lipschitz constant of the residual map strictly less than one on a suitable ball, domain restrictions, and regularity of the base operator), together with any available quantitative bounds derived from the training procedure. This will make the theorem directly falsifiable and will clarify the conditions under which the observed stability beyond the training horizon is expected to hold. revision: yes

  2. Referee: [§5] §5 (Experiments on turbulent flow and Active Matter): No post-hoc numerical diagnostic is reported that checks whether the contraction condition actually holds on the test trajectories; the claim that error ratios remain stable at iteration counts exceeding those seen in training therefore rests on an unverified premise.

    Authors: We acknowledge that an explicit numerical check of the contraction condition on held-out test trajectories is currently absent. In the revised manuscript we will insert a new diagnostic subsection that reports an empirical estimate of the Lipschitz constant of the learned refinement operator evaluated on the test sets for both turbulent flow and active-matter problems. This will directly verify whether the contraction premise holds out of sample and thereby substantiate the stability claims. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper introduces IRNO by augmenting a base neural operator with a learned refinement module and applies the classical contraction mapping theorem under local assumptions to prove convergence to a unique fixed point. The progressive spectral loss is presented as an explicit training design choice that increases high-frequency penalties over iterations. No equation or claim reduces a reported prediction or fixed-point result to a fitted parameter by construction, and no load-bearing step depends on a self-citation chain. The derivation remains self-contained as a methodological proposal whose central guarantees rest on standard analysis rather than tautological re-labeling of inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central contribution is the IRNO method and loss; convergence rests on a domain assumption about local contraction properties rather than new invented entities or fitted constants.

axioms (1)
  • domain assumption Local assumptions ensure the induced refinement operator is a contraction mapping
    Invoked in the abstract to establish convergence to a unique fixed point.

pith-pipeline@v0.9.1-grok · 5775 in / 1247 out tokens · 58694 ms · 2026-06-30T17:01:45.015215+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 13 canonical work pages · 1 internal anchor

  1. [1]

    doi: 10.1088/0004-637x/765/1/39

    ISSN 1538-4357. doi: 10.1088/0004-637x/765/1/39. URL http://dx.doi.org/10.1088/0004-637X/765/1/39. Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. Deep equilibrium models. InAdvances in Neural Information Processing Systems (NeurIPS),

    work page doi:10.1088/0004-637x/765/1/39

  2. [2]

    Yuchen Fan, Jiahui Yu, and Thomas S Huang

    doi: 10.52202/079017-0201. Yuchen Fan, Jiahui Yu, and Thomas S Huang. Wide-activated deep residual networks based restoration for bpg-compressed images. InProceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition Workshops, pages 2621–2624,

    work page doi:10.52202/079017-0201

  3. [3]

    URLhttps://doi

    doi: 10.1088/1361-6420/aa9a90. URLhttps://doi. org/10.1088/1361-6420/aa9a90. Sheikh Md Shakeel Hassan, Arthur Feeney, Akash Dhruv, Jihoon Kim, Youngjoon Suh, Jaiyoung Ryu, Yoonjin Won, and Aparna Chandramowlishwaran. Bubbleml: A multi- phase multiphysics dataset and benchmarks for machine learning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, ...

    work page doi:10.1088/1361-6420/aa9a90

  4. [4]

    Deep Residual Learning for Image Recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. InProceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR ’16, pages 770–778. IEEE, June

    2016

  5. [5]

    Deep Residual Learning for Image Recognition,

    doi: 10.1109/CVPR.2016.90. URLhttp://ieeexplore.ieee.org/document/7780459. Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr´ as Hor´ anyi, Joaqu´ ın Mu˜ noz- Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis.Quarterly journal of the royal meteorological society, 146(730):1999–2049,

    work page doi:10.1109/cvpr.2016.90 2016

  6. [6]

    doi: https://doi.org/10.1016/ j.neunet.2025.108027

    ISSN 0893-6080. doi: https://doi.org/10.1016/ j.neunet.2025.108027. URLhttps://www.sciencedirect.com/science/article/pii/ S0893608025009074. Jean Kossaifi, Nikola Kovachki, Zongyi Li, David Pitt, Miguel Liu-Schiaffini, Valentin Du- ruisseaux, Robert Joseph George, Boris Bonev, Kamyar Azizzadenesheli, Julius Berner, and Anima Anandkumar. A library for lear...

    work page arXiv 2025

  7. [7]

    Fourier Neural Operator for Parametric Partial Differential Equations

    URLhttps://openreview.net/forum?id=c8P9NQVtmnO. See also arXiv:2010.08895 (2020). Xinliang Liu, Bo Xu, Shuhao Cao, and Lei Zhang. Mitigating spectral bias for the multiscale operator learning.Journal of Computational Physics, 506:112944,

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  8. [8]

    doi: https://doi.org/10.1016/j.jcp.2024.112944

    ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2024.112944. URLhttps://www.sciencedirect.com/ science/article/pii/S0021999124001931. Lu Lu, Pengzhan Jin, Guofei Pang, Zongren Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of opera- tors.Nature Machine Intelligence, 3(3):218–229,

    work page doi:10.1016/j.jcp.2024.112944 2024

  9. [9]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

    doi: 10.1038/s42256-021-00302-5. Xihaier Luo, Xiaoning Qian, and Byung-Jun Yoon. Hierarchical neural operator transformer with learnable frequency-aware loss prior for arbitrary-scale super-resolution. InForty-first International Conference on Machine Learning,

    work page doi:10.1038/s42256-021-00302-5

  10. [10]

    doi: 10.1098/rspa.2024.0819

    ISSN 1364-5021. doi: 10.1098/rspa.2024.0819. URL https://doi.org/10.1098/rspa.2024.0819. Shaoxiang Qin, Fuyuan Lyu, Wenhui Peng, Dingyang Geng, Ju Wang, Xing Tang, Sylvie Leroyer, Naiping Gao, Xue Liu, and Liangzhu Leon Wang. Toward a better understanding of fourier neural operators from a spectral perspective,

    work page doi:10.1098/rspa.2024.0819 2024

  11. [11]

    Toward a better understanding of fourier neural operators from a spectral perspective.arXiv preprint arXiv:2404.07200,

    URLhttps://arxiv.org/ abs/2404.07200. Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred A. Ham- precht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. InInternational Conference on Machine Learning (ICML), volume 97 ofProceedings of Machine Learning Research, pages 5301–5310,

    work page arXiv

  12. [12]

    URLhttps://doi.org/10.1214/ 20-EJS1735

    doi: 10.1214/20-EJS1735. URLhttps://doi.org/10.1214/ 20-EJS1735. Pu Ren, N. Benjamin Erichson, Junyi Guo, Shashank Subramanian, Omer San, Zarija Lukic, and Michael W. Mahoney. Superbench: A super-resolution benchmark dataset for scien- tific machine learning.Journal of Data-centric Machine Learning Research,

    work page doi:10.1214/20-ejs1735

  13. [13]

    Dataset Certification, Reproducibility Certification

    URL https://openreview.net/forum?id=OJ6zUcWldW. Dataset Certification, Reproducibility Certification. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi, editors,Medical Image Computing and Computer-Assisted Inter...

    2015

  14. [14]

    Frank Werner and Bernd Hofmann

    URLhttps: //arxiv.org/abs/2512.05132. Frank Werner and Bernd Hofmann. Convergence analysis of (statistical) inverse problems under conditional stability estimates.Inverse Problems, 36(1):015004, dec

    work page arXiv

  15. [15]

    URLhttps://doi.org/10.1088/1361-6420/ab4cd7

    doi: 10.1088/1361-6420/ab4cd7. URLhttps://doi.org/10.1088/1361-6420/ab4cd7. 20 R. Wienands and W. Joppich.Practical Fourier Analysis for Multigrid Methods. Numerical Insights. Taylor & Francis,

    work page doi:10.1088/1361-6420/ab4cd7

  16. [16]

    URLhttps://books.google

    ISBN 9781584884927. URLhttps://books.google. com/books?id=IOSux5GxacsC. Zhikai Wu, Shiyang Zhang, Sizhuang He, Sifan Wang, Min Zhu, Anran Jiao, Lu Lu, and David van Dijk. COAST: Intelligent time-adaptive neural operators. In2nd AI for Math Workshop @ ICML 2025,

    2025

  17. [17]

    net/forum?id=6QJZDAIhfk

    URLhttps://openreview. net/forum?id=6QJZDAIhfk. 21 A Limitations and Future Work The proposed Iterative Refinement Neural Operator (IRNO) demonstrates stable conver- gence, reduced spectral error, and improved cost–performance trade-offs across multiple physical systems. Several open questions remain that point toward productive directions for future rese...

    2019

  18. [18]

    More recent approaches have sought to mitigate spectral bias through frequency-specific boosting and scaling mechanisms

    was proposed to enable nested feature computation that better captures multiscale solution spaces, by employing a scale-adaptive 22 interaction range and self-attention mechanisms over a hierarchy of levels. More recent approaches have sought to mitigate spectral bias through frequency-specific boosting and scaling mechanisms. SpecBoost was introduced as ...

    2024

  19. [19]

    In most cases, these models are trained asdirect solversthat approximate the solution operator through a single forward evaluation

    parameterize mappings in spectral or basis- function spaces, consistent with classical spectral/pseudo-spectral methods. In most cases, these models are trained asdirect solversthat approximate the solution operator through a single forward evaluation. A closer conceptual link to our framework arises in implicit layers, Deep Equilibrium Mod- els (DEQs)Bai...

    2019

  20. [20]

    Case 2.c >0.The function (1−q)r−cr 2 is maximized atr ∗ = 1−q 2c , yielding maximum value (1−q)2 4c

    Case 1.c= 0.The right-hand side is (1−q)r, giving the condition ∥b∥ ≤ (1−q)r α . Case 2.c >0.The function (1−q)r−cr 2 is maximized atr ∗ = 1−q 2c , yielding maximum value (1−q)2 4c . Therefore, a validr∈(0, δ] exists whenever ∥b∥ ≤ (1−q) 2 4αc , in which case the two positive roots of (1−q)r−cr 2 =α∥b∥are r± = (1−q)± p (1−q) 2 −4αc∥b∥ 2c , andB r(y) is fo...

    2020

  21. [21]

    The test sets for the corresponding interpolation and ex- trapolation tasks are from 2009 and 2014–2015, respectively

    The validation sets for interpolation and extrapolation (look-back) evaluation are drawn from 2012 and 2007, respectively. The test sets for the corresponding interpolation and ex- trapolation tasks are from 2009 and 2014–2015, respectively. All methods are trained and evaluated on the same splits to ensure fair comparison. For visualizations in the main ...

    2012

  22. [22]

    The refinement operator is optimized using AdamW with an initial learning rate of 3×10 −4 and weight decay 1×10 −5

    27.6 27.7 6.9 FNO + SRCNN× ×6.6 FNO + UNet (×15)× ×37.3 FNO + FNO× ×9.5 Refinement Operator Training.Algorithm 1 provides the complete training procedure for IRNO, detailing the progressive refinement steps, combined spatial-spectral losses, and fixed-point regularization. The refinement operator is optimized using AdamW with an initial learning rate of 3...

    2026

  23. [23]

    Two complementary strategies can protect against divergence

    Withα= 0.05, refinement remains stable throughout without divergence. Two complementary strategies can protect against divergence. First, step-size scheduling reducesαaskincreases. Second, adaptive stopping halts iteration when∥Φ θ(x, hk)∥falls below a user-defined threshold tied to the bias level α∥b∥/(1−q) from Corollary 3.3. HFS-ResUNet and IRNO Comple...

    1946

  24. [24]

    IRNO.Table 20 compares F-Adapter [Zhang et al., 2026] and IRNO (FNO base) on Active Matter

    BatchNorm (ours) 0.1102 53.52% 0.1042 56.05% LayerNorm 0.1038 56.61% 0.0988 58.70% GroupNorm 0.1031 57.46% 0.1006 58.50% F-Adapter vs. IRNO.Table 20 compares F-Adapter [Zhang et al., 2026] and IRNO (FNO base) on Active Matter. F-Adapter targets minimal-parameter adaptation, achieving 2.31% VRMSE reduction with gains concentrated in mid- and high-frequency...

    2026