Pith. sign in

REVIEW 3 major objections 4 minor 3 cited by

Recurrent Neural Operators: Stable Long-Term PDE Prediction

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Recurrently trained neural operators, which optimize on their own rollouts during training, reduce worst-case long-horizon forecast error growth from exponential to linear in time.

desk verdict Empirically useful, theoretically unproven: the paper shows recurrent training stabilizes MgNO but the headline exponential-to-linear bound assumes the stability it claims to establish. read the letter →

arxiv 2505.20721 v1 pith:INRTZETA submitted 2025-05-27 cs.LG cs.NAmath.NA

Zaijun Ye , Chen-Song Zhang , Wansheng Wang This is my paper · ORCID
classification cs.LGcs.NAmath.NA
keywords recurrenttrainingneuraloperatorsteacherforcingerroraccumulationautoregressiverollouttime-dependentPDEsMultigridOperatorlong-termprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Time-dependent PDE surrogates are usually trained one step at a time on ground-truth states, a scheme known as teacher forcing; at inference the model must instead feed on its own predictions, so errors compound and long-horizon forecasts diverge. The paper proposes Recurrent Neural Operators (RNOs), which during training recursively apply the learned operator to its own outputs over a temporal window, making the training dynamics match inference dynamics. The central claim, stated as Theorem 1, is that this recurrent training changes the worst-case error growth in autoregressive prediction from exponential in the time horizon $T$ to linear in $T$. If the claim is right, neural operators can be trained to produce stable forecasts far beyond their observation window without post-hoc correction, and the experiments show the Multigrid Neural Operator benefiting most.

What carries the argument

The load-bearing object is the error recursion $e_{n+1} \le e_n + \Delta t\,\|G_\theta(\hat u_n)-D(u_n)\| + O(\Delta t^2)$ that connects one-step approximation error to rollout error. Teacher forcing bounds the discrepancy via Lipschitz continuity of $G_\theta$, producing the geometric factor $(1+C\Delta t)^n \le e^{CT}$; recurrent training is assumed to control the discrepancy directly on the model's own trajectory, so summing the recursion yields $e_n \le e_0 + n(\epsilon\Delta t+O(\Delta t^2)) \le e_0 + T(\epsilon+O(\Delta t))$. The argument also rests on the universal approximation lemma and on boundedness of the multigrid operator layers, which justify the per-step accuracy $\epsilon$.

What would settle it

Run a recurrently trained operator on a benchmark whose predicted trajectory leaves the region where the operator is accurate, and measure the per-step growth rate $\ln(\|\hat u_n-u_n\|/\|\hat u_0-u_0\|)/n$ over several hundred rollout steps; the paper's claim predicts that this rate decays toward zero, whereas teacher forcing predicts it approaches a positive constant $C$. If the measured rate does not decay, the theorem's key premise is violated.

Watch

Extended reading notes

Core claim

The paper's central discovery is Theorem 1, a pair of error bounds for the autoregressive rollout of a neural operator $G_\theta$ used as a one-step Euler update. Under teacher forcing the worst-case error obeys $\max_{0\le n\le N}\|\hat u_n-u_n\| \le e^{CT}\|\hat u_0-u_0\| + (e^{CT}-1)(\epsilon+O(\Delta t))/C$, while under recurrent training it obeys $\|\hat u_0-u_0\| + T(\epsilon+O(\Delta t))$, provided the training objective keeps the discrepancy $\|G_\theta(\hat u_n)-D(u_n)\|$ bounded by $\epsilon$ along the model's own predicted trajectory. The qualitative contrast is the exponential factor $e^{CT}$ versus the linear factor $T$. Empirically, recurrently trained Multigrid Neural Operators reduce the mean relative $L^2$ error at step 50 by roughly 9 percent to 87 percent relative to their teacher-forced counterparts on heat, Allen-Cahn, Cahn-Hilliard, and Navier-Stokes benchmarks, and remain stable in rollouts up to step 500 on Allen-Cahn, while the same recurrent treatment of the Fourier Neural Operator gives inconsistent results.

Load-bearing premise

The linear error bound rests on the premise that recurrent training actually keeps the operator's per-step error small on its own predicted states, a stability property the paper assumes in the proof rather than derives from the training objective.

Editorial extensions

If this is right

  • Aligning training with autoregressive inference is what stabilizes long-horizon neural operator forecasts: optimizing on the model's own rollouts, rather than on ground-truth inputs, is the mechanism that suppresses compounding errors.
  • If the theorem's premise holds, worst-case rollout error grows at most linearly with the forecast horizon, making extrapolation beyond the training window a practical target for neural operators.
  • The benefit is architecture-dependent: in the reported experiments the Multigrid Neural Operator improves substantially under recurrent training while the Fourier Neural Operator does not, so the recipe is most effective when the operator layers are bounded and stable.
  • The best recurrent configuration matches or beats a post-hoc correction baseline with fewer parameters, but at the cost of higher per-epoch compute and memory, so the trade-off is compute during training versus robustness during inference.
  • Longer observation windows during recurrent training improve far-horizon extrapolation but weaken accuracy near the initial condition, revealing a practical trade-off in choosing how many steps to unroll.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the theorem is best read as a reduction: the hard problem becomes keeping the operator's discrepancy small on its own predicted trajectory, so training objectives that directly penalize rollout divergence, such as spectral regularization or input perturbation during unrolling, are natural next steps.
  • Beyond the paper, the mechanism is about the training/inference mismatch rather than about PDE-specific structure, so the same recurrent principle should transfer to other autoregressive surrogate models for spatiotemporal data whenever the underlying map is Lipschitz and training data come from a finite observation window.
  • Beyond the paper, the architecture-dependence suggests a testable hypothesis: operators with contractive or smoothing layers will benefit more from recurrent training than operators with long-range spectral mixing, because the recurrent objective gives high-frequency instabilities room to amplify.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes Recurrent Neural Operators (RNOs), a training strategy that replaces teacher forcing with an autoregressive rollout during training for neural operator models of time-dependent PDEs. The authors claim a theoretical result (Theorem 1) that recurrent training reduces worst-case error growth from exponential (teacher forcing) to linear in the time horizon, and they support this with experiments on heat conduction, Allen–Cahn, Cahn–Hilliard, and Navier–Stokes benchmarks using FNO and MgNO backbones. The empirical section reports that recurrently trained MgNO (r-MgNO) achieves lower long-rollout errors than teacher-forced MgNO, FNO variants, and a PDE-Refiner baseline, with ablations on observation length, time step, and data size, plus training cost measurements.

Significance. The empirical message is practically valuable: aligning training with autoregressive inference is a simple and effective way to improve long-horizon neural operator forecasts, and the experiments support this for MgNO in particular. The paper provides reproducible details, including an anonymous code link, data generation specifications, and hyperparameters, which strengthens the empirical contribution. However, the theoretical centerpiece—an exponential-to-linear improvement in worst-case error growth—is not established by the proof as written. The main result is conditional on an assumption that essentially restates the stability property it claims to derive, and this assumption is not a consequence of the training objective or the universal approximation lemma. If the theorem is reframed as a conditional statement or replaced by a rigorously derived bound, the paper's empirical findings could still stand, but the current theoretical claim is overstated in the abstract and Section 3.3.

major comments (3)
  1. [Appendix B.2, Theorem 1(ii)] The proof of the linear error bound assumes exactly the property the theorem claims to establish. The text states: 'assuming the training objective effectively minimizes the discrepancy term ||G_theta(hat u_n, f) - D(u_n, f)|| over the rollout' and 'the proof provided assumes this leads to the bound ||G_theta(hat u_n, f) - D(u_n, f)|| <= epsilon.' This discrepancy is the residual of the learned dynamics along the model's own trajectory, not the state error minimized by the RNO objective (7). The objective only controls sum_n ||hat u_n - u_n||; a small state error does not imply a small residual, since the residual can be as large as (e_{n+1} - e_n)/Delta t + O(Delta t). Thus the exponential-to-linear contrast in Theorem 1 is a conditional statement, not a proven consequence of the RNO training procedure.
  2. [Section 3.3, after Theorem 1] The main text itself hedges the recurrent bound with 'under the assumption that the training successfully enforces stability.' This assumption is not derived from Eq. (7) or from Lemma 1, and it is not stated as a hypothesis in the theorem. As written, the theorem overstates what has been shown. The theorem statement should either include this stability assumption explicitly (making the contribution a conditional guarantee) or the proof must derive the residual bound from the training objective, which would require substantially more argument.
  3. [Lemma 1 and its application in B.2] The universal approximation property in Lemma 1 guarantees approximation of D on a compact set C in the state space, typically the ground-truth states. Recurrent training does not automatically extend this guarantee to states on a model-generated rollout, which may drift outside that compact set. The proof of part (ii) applies the epsilon-approximation bound at the predicted states hat u_n without justifying that these states remain in the approximation regime. This is a second, distinct gap that would need to be addressed if the theorem is to be made rigorous.
minor comments (4)
  1. [Section 4.2, Table 1] The claimed reduction range '9.1% to 87.2%' for r-MgNO versus tf-MgNO at n=50 does not match the values in the table (e.g., Cahn–Hilliard improves from 3.4e-3 to 2.9e-3, about 14.7%, which is below 9.1%; Navier–Stokes improves about 36%). Please verify the range or clarify the calculation.
  2. [Section 4.3, Figure 2 caption] The caption refers to 'cyclical training,' but the paper's method is called 'recurrent training.' This appears to be a typo; please use consistent terminology.
  3. [Eq. (7) and Eq. (6)] In Eq. (6), the rollout starts at hat u_0 = u_0, but Eq. (7) sums errors from n=1 to N. This is fine, but the notation could be clarified to distinguish the initial condition from the predicted states; also, the sum in Eq. (5) runs from 0 to N-1 with u_{n+1} as target, while Eq. (7) runs to N, which may confuse readers about the training horizon.
  4. [Section 3.2, Eq. (7)] The objective in Eq. (7) sums the state errors but does not include the final state explicitly; if the goal is to minimize rollout error over the full window, this is appropriate, but the relationship to the residual term used in B.2 should be discussed explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1(ii) is conditional on assuming the very stability property it claims to establish: Appendix B.2 assumes ||G_theta(hat u_n)-D(u_n)|| <= epsilon along the model rollout without deriving it from objective (7).

  1. other [Section 3.3 / Appendix B.2, Theorem 1(ii), Eq. (9)]
    "In the recurrent training setting (assuming the training objective effectively minimizes the discrepancy term ∥Gθ(ˆun, f)− D(un, f)∥over the rollout), we bound this term directly using the approximation error ϵ. ... The proof provided assumes this leads to the bound: ∥Gθ(ˆuRN O n , f)− D(un, f)∥ ≤ϵ."

    The RNO objective (7), min_θ Σ_{n=1}^N ||û_n − u_n||, is a trajectory-state loss; it contains no term in ||Gθ(û_n,f) − D(u_n,f)||. From the recurrence (6), that residual equals ((û_{n+1} − û_n)/Δt) − ((u_{n+1} − u_n)/Δt) + O(Δt), so small per-step state errors imply only a residual bound of order (||e_{n+1}|| + ||e_n||)/Δt, not ≤ ε. Appendix B.2 nevertheless simply assumes the residual is ≤ ε along the model's own rollout, which is exactly the stability property Theorem 1(ii) purports to derive from recurrent training. The theorem's post-statement hedge, 'under the assumption that the training successfully enforces stability,' confirms that this is an unproven premise rather than a consequence of (7).

full rationale

Most of the paper is not circular. The empirical comparison (Table 1, Figures 2–7) is self-contained: r-MgNO versus tf-MgNO, tf-FNO, and PDE-Refiner is a direct benchmark with reported errors, and these results do not depend on Theorem 1. The MgNO universal approximation lemma is imported from [4], whose authors are not the present authors, so that citation is independent support rather than a self-citation chain. The teacher-forcing bound, Theorem 1(i), is a standard Lipschitz-plus-approximation argument and is not circular. The one load-bearing circular step is confined to Theorem 1(ii) and Appendix B.2: the proof assumes the exact rollout-residual bound that recurrent training is supposed to enforce. Since the paper itself labels this as an assumption, the headline theoretical contribution — that recurrent training reduces worst-case error growth from exponential to linear — is conditional on the target property and does not follow from objective (7). This warrants a score of 6 rather than a higher score because the evaluated method and empirical claims have independent content; it is not a lower score because the paper's central theoretical claim reduces to an assumed premise.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on the usual machine-learning ingredients: architecture choices, data generation, and hyperparameters. The theoretical exponential-to-linear comparison rests on an additional unproven premise, namely that recurrent training makes the operator accurate on its own predicted trajectory, which is essentially the conclusion of the theorem. That premise is the main burden of the paper's theory.

free parameters (4)
  • Rollout training horizon N = 10 steps in main benchmarks, 5 to 40 in ablation
    The RNO objective (7) sums errors over N steps; both error bounds and observed performance depend on N, and Fig. 3(a) shows that longer observation lengths improve extrapolation.
  • Time step Delta t = 0.2 in main benchmarks, 0.05 to 0.4 in ablation
    Discretization in Eq. (4); the ablation in Fig. 3(b) shows error decreases from 0.4 to 0.1 but increases at 0.05, so the claimed linear-in-Delta-t trend is not universal.
  • Data size and GP length scale = 1000 training and 200 test samples; length scale ell unspecified
    Initial conditions are sampled from a periodic Gaussian process, Eq. (15); Table 2 shows strong sensitivity to data size, and ell controls the smoothness and difficulty of the benchmark fields.
  • Training hyperparameters = lr=1e-3, weight decay=1e-5, batch=32, epochs=500, OneCycle scheduler
    These are chosen to make training work and are not derived; no sensitivity analysis is reported for them, and gradient clipping is applied only for certain initializations.
assumptions (5)
  • domain assumption Universal approximation of MgNO (Lemma 1, based on Theorem 3.1 of [4])
    Used in Appendix B to assert that an epsilon-accurate neural operator approximation exists for the dynamics operator on a compact set; no explicit compact set is identified for the PDE benchmarks.
  • domain assumption Lipschitz continuity of G_theta with constant C
    Used in the teacher-forcing bound in Theorem 1(i); activations and bounded linear operators may imply it, but C is never estimated or shown to be finite for the trained models.
  • ad hoc to paper Recurrent training enforces ||G_theta(hat u_n) - D(u_n)|| <= epsilon on the predicted rollout
    Appendix B.2 explicitly assumes this bound; it is the property the theorem's linear-growth result needs and is not derived from objective (7) or from the universal approximation lemma.
  • standard math Forward Euler local truncation error O(Delta t^2)
    Used to write u^{n+1} = u^n + Delta t D(u^n, f) + O(Delta t^2); standard for smooth solutions.
  • domain assumption Benchmark reference solutions are accurate after numerical generation and downsampling
    Appendix D.1 states that data is computed with traditional numerical methods and downsampled, but no solver tolerances, grid convergence checks, or verification results are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recurrent Neural Operators: Stable Long-Term PDE Prediction." pith.science (2026). https://pith.science/paper/INRTZETA

@misc{pith2026250520721,
  author       = {Pith},
  title        = {Pith review of: Recurrent Neural Operators: Stable Long-Term PDE Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INRTZETA}},
  note         = {Machine review of arXiv:2505.20721}
}
read the original abstract

Neural operators have emerged as powerful tools for learning solution operators of partial differential equations. However, in time-dependent problems, standard training strategies such as teacher forcing introduce a mismatch between training and inference, leading to compounding errors in long-term autoregressive predictions. To address this issue, we propose Recurrent Neural Operators (RNOs)-a novel framework that integrates recurrent training into neural operator architectures. Instead of conditioning each training step on ground-truth inputs, RNOs recursively apply the operator to their own predictions over a temporal window, effectively simulating inference-time dynamics during training. This alignment mitigates exposure bias and enhances robustness to error accumulation. Theoretically, we show that recurrent training can reduce the worst-case exponential error growth typical of teacher forcing to linear growth. Empirically, we demonstrate that recurrently trained Multigrid Neural Operators significantly outperform their teacher-forced counterparts in long-term accuracy and stability on standard benchmarks. Our results underscore the importance of aligning training with inference dynamics for robust temporal generalization in neural operator learning.

Figures

Figures reproduced from arXiv: 2505.20721 by the authors.

Figure 1
Figure 1. Recurrent Neural Operators. (Left) RNO aligns training with autoregressive inference, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Comparison of long-term rollout stability on the Allen–Cahn equation. Both panels show [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Ablation study on the Allen-Cahn equation using the r-MgNO model. (a) Effect of varying [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of predicted solutions for the transient heat conduction problem. The base [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Visualization of predicted solutions for the Allen-Cahn problem. The base model is MgNO. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Visualization of predicted solutions for the Cahn-Hilliard problem. The base model is [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Visualization of predicted solutions for the Navier-Stokes problem. The base model is [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. HERO: History-Enriched Rollout Training for Long-Horizon Autoregressive Neural Operators

    cs.LG 2026-07 conditional novelty 6.0 of 10

    HERO adds a margin-based 'beat your old self' objective to rollout training, dynamically selecting the worst reference trajectory from lagged/self/perturbed rollouts and reweighting the standard ground-truth gradient,...

  2. No Free Lunch in Flow Surrogates under Time-Varying Boundary Conditions: A Two-Regime Study

    math.NA 2026-07 conditional novelty 6.0 of 10

    No single flow-surrogate architecture transfers from a boundary-driven Stokes film to a self-sustained Kármán wake; time treatment decides the winner and pointwise RMSE ranks the wrong models.

  3. Scale-Aware Learning of Chaotic Dynamics on Unstructured Meshes via Binned Spectral Losses

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Adding graph-Laplacian band-power matching to the training loss improves long-horizon autoregressive forecasting of chaotic flows on unstructured meshes.

Reference graph

Works this paper leans on

33 extracted references · 29 canonical work pages · cited by 3 Pith papers

  1. [1]

    LNO: Laplace Neural Operator for Solving Differential Equations.Nature Machine Intelligence, 6(6):631–640, June 2024

    Qianying Cao, Somdatta Goswami, and George Em Karniadakis. LNO: Laplace Neural Operator for Solving Differential Equations.Nature Machine Intelligence, 6(6):631–640, June 2024

    work page 2024

  2. [2]

    Weinan E, Jiequn Han, and Arnulf Jentzen. Deep Learning-Based Numerical Methods for High- Dimensional Parabolic Partial Differential Equations and Backward Stochastic Differential Equations.Communications in Mathematics and Statistics, 5(4):349–380, December 2017

    work page 2017

  3. [3]

    The Deep Ritz Method: A Deep Learning-Based Numerical Algorithm for Solving Variational Problems.Communications in Mathematics and Statistics, 6(1):1–12, February 2018

    Weinan E and Bing Yu. The Deep Ritz Method: A Deep Learning-Based Numerical Algorithm for Solving Variational Problems.Communications in Mathematics and Statistics, 6(1):1–12, February 2018

    work page 2018

  4. [4]

    MgNO: Efficient Parameterization of Linear Operators via Multigrid

    Juncai He, Xinliang Liu, and Jinchao Xu. MgNO: Efficient Parameterization of Linear Operators via Multigrid. InThe Twelfth International Conference on Learning Representations, May 2024

    work page 2024

  5. [5]

    State-space models are accurate and efficient neural operators for dynamical systems

    Zheyuan Hu, Nazanin Ahmadi Daryakenari, Qianli Shen, Kenji Kawaguchi, and George Em Karniadakis. State-space models are accurate and efficient neural operators for dynamical systems. arXiv, January 2025. arXiv:2409.03231

    arXiv 2025

  6. [6]

    Deep backward schemes for high-dimensional nonlinear PDEs.Mathematics of Computation, 89(324):1547–1579, January 2020

    Côme Huré, Huyên Pham, and Xavier Warin. Deep backward schemes for high-dimensional nonlinear PDEs.Mathematics of Computation, 89(324):1547–1579, January 2020

    work page 2020

  7. [7]

    DPM: A Novel Training Method for Physics-Informed Neural Networks in Extrapolation

    Jungeun Kim, Kookjin Lee, Dongeun Lee, Sheo Yon Jhin, and Noseong Park. DPM: A Novel Training Method for Physics-Informed Neural Networks in Extrapolation. InProceedings of the AAAI Conference on Artificial Intelligence, May 2021

    work page 2021

  8. [8]

    Approximating Numerical Fluxes Using Fourier Neural Operators for Hyperbolic Conservation Laws.Communications in Computational Physics, 37(2):420–456, January 2025

    Taeyoung Kim and Myungjoo Kang. Approximating Numerical Fluxes Using Fourier Neural Operators for Hyperbolic Conservation Laws.Communications in Computational Physics, 37(2):420–456, January 2025

    work page 2025

Show all 33 references
  1. [9]

    On Universal Approximation and Error Bounds for Fourier Neural Operators.Journal of Machine Learning Research, 22(290):1–76, January 2021

    Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. On Universal Approximation and Error Bounds for Fourier Neural Operators.Journal of Machine Learning Research, 22(290):1–76, January 2021

    2021

  2. [10]

    Neural Operator: Learning Maps Between Function Spaces With Applica- tions to PDEs.Journal of Machine Learning Research, 24(89):1–97, January 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, and Andrew Stuart. Neural Operator: Learning Maps Between Function Spaces With Applica- tions to PDEs.Journal of Machine Learning Research, 24(89):1–97, January 2023

    2023

  3. [11]

    Fourier Neural Operator with Learned Deformations for PDEs on General Geometries.The Journal of Machine Learning Research, 24(1):1532–4435, January 2023

    Zongyi Li, Daniel Zhengyu Huang, and Burigede Liu. Fourier Neural Operator with Learned Deformations for PDEs on General Geometries.The Journal of Machine Learning Research, 24(1):1532–4435, January 2023

    2023

  4. [12]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier Neural Operator for Parametric Partial Differential Equations. InInternational Conference on Learning Representations, May 2021

    2021

  5. [13]

    Geometry-Informed Neural Operator for Large-Scale 3D PDEs

    Zongyi Li, Nikola Borislavov Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Prakash Otta, Mohammad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, and Anima Anandkumar. Geometry-Informed Neural Operator for Large-Scale 3D PDEs. In Proceedings o...

    2023

  6. [14]

    Veeling, Paris Perdikaris, Richard E

    Phillip Lippe, Bastiaan S. Veeling, Paris Perdikaris, Richard E. Turner, and Johannes Brandstet- ter. PDE-Refiner: Achieving Accurate Long Rollouts with Neural PDE Solvers. InProceedings of the 37th International Conference on Neural Information Processing Systems, December 2023

    2023

  7. [15]

    Mitigating spectral bias for the multiscale operator learning.Journal of Computational Physics, 506:112944, June 2024

    Xinliang Liu, Bo Xu, Shuhao Cao, and Lei Zhang. Mitigating spectral bias for the multiscale operator learning.Journal of Computational Physics, 506:112944, June 2024

    2024

  8. [16]

    A MgNO Method for Multiphase Flow in Porous Media

    Xinliang Liu, Xia Yang, Chen-Song Zhang, Lian Zhang, and Li Zhao. A MgNO Method for Multiphase Flow in Porous Media. InProceedings of the 16th Annual Meeting Conference on Porous Media, April 2025

    2025

  9. [17]

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

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218–229, March 2021. 10

    2021

  10. [18]

    Towards Stability of Autoregressive Neural Operators.Transactions on machine learning research, October 2023

    Michael McCabe, Peter Harrington, Shashank Subramanian, and Jed Brown. Towards Stability of Autoregressive Neural Operators.Transactions on machine learning research, October 2023

    2023

  11. [19]

    Neural Operator Learning for Long-Time Integration in Dynamical Systems with Recurrent Neural Networks

    Katarzyna Michałowska, Somdatta Goswami, George Em Karniadakis, and Signe Riemer- Sørensen. Neural Operator Learning for Long-Time Integration in Dynamical Systems with Recurrent Neural Networks. In2024 International Joint Conference on Neural Networks, June 2024

    2024

  12. [20]

    FourCastNet: A Global Data- driven High-resolution Weather Model using Adaptive Fourier Neural Operators

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, Pedram Hassanzadeh, Karthik Kashinath, and Animashree Anandkumar. FourCastNet: A Global Data- driven High...

    2023

  13. [21]

    Forward–Backward Stochastic Neural Networks: Deep Learning of High- Dimensional Partial Differential Equations

    Maziar Raissi. Forward–Backward Stochastic Neural Networks: Deep Learning of High- Dimensional Partial Differential Equations. InPeter Carr Gedenkschrift, Research Advances in Mathematical Finance, pages 637–655. December 2023

    2023

  14. [22]

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707, February 2019

    2019

  15. [23]

    Finite Operator Learning: Bridging Neural Operators and Numerical Methods for Efficient Parametric Solution and Optimization of PDEs

    Shahed Rezaei, Reza Najian Asl, Kianoosh Taghikhani, Ahmad Moeineddin, Michael Kaliske, and Markus Apel. Finite Operator Learning: Bridging Neural Operators and Numerical Methods for Efficient Parametric Solution and Optimization of PDEs. arXiv, October 2024. arXiv:2407.04157

    2024

  16. [24]

    DGM: A deep learning algorithm for solving partial differential equations.Journal of Computational Physics, 375:1339–1364, December 2018

    Justin Sirignano and Konstantinos Spiliopoulos. DGM: A deep learning algorithm for solving partial differential equations.Journal of Computational Physics, 375:1339–1364, December 2018

    2018

  17. [25]

    Factorized Fourier Neural Operators

    Alasdair Tran, Alexander Mathews, Lexing Xie, and Cheng Soon Ong. Factorized Fourier Neural Operators. InThe Eleventh International Conference on Learning Representations, May 2023

    2023

  18. [26]

    COAST: Intelligent Time-Adaptive Neural Operators

    Zhikai Wu, Shiyang Zhang, Sizhuang He, Sifan Wang, Min Zhu, Anran Jiao, Lu Lu, and David van Dijk. COAST: Intelligent Time-Adaptive Neural Operators. arXiv, February 2025. arXiv:2502.08574

    2025

  19. [27]

    Amortized Fourier Neural Operators

    Zipeng Xiao, Siqi Kou, Zhongkai Hao, Bokai Lin, and Zhijie Deng. Amortized Fourier Neural Operators. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, December 2024

    2024

  20. [28]

    Transfer Learning Enhanced DeepONet for Long- Time Prediction of Evolution Equations

    Wuzhe Xu, Yulong Lu, and Li Wang. Transfer Learning Enhanced DeepONet for Long- Time Prediction of Evolution Equations. InProceedings of the AAAI Conference on Artificial Intelligence, June 2023

    2023

  21. [29]

    FBSJNN: A Theoretically Interpretable and Efficiently Deep Learning method for Solving Partial Integro-Differential Equations

    Zaijun Ye and Wansheng Wang. FBSJNN: A Theoretically Interpretable and Efficiently Deep Learning method for Solving Partial Integro-Differential Equations. arXiv, December 2024. arXiv:2412.11010

    2024 arXiv

  22. [30]

    Reliable extrapolation of deep neural operators informed by physics or sparse observations.Computer Methods in Applied Mechanics and Engineering, 412:116064, July 2023

    Min Zhu, Handi Zhang, Anran Jiao, George Em Karniadakis, and Lu Lu. Reliable extrapolation of deep neural operators informed by physics or sparse observations.Computer Methods in Applied Mechanics and Engineering, 412:116064, July 2023. 11 A Formal Representation of Neural Ope...

    2023

  23. [31]

    Downward Path:I (r1) κ1 → Rr2 r1 → · · · → RrJ rJ−1 ,

  24. [32]

    Coarse Solve:I (rJ ) κJ ,

  25. [33]

    Upward Path:P rJ−1 rJ → · · · → Pr1 r2 → I(r1) κ2J−1 , with resolution hierarchyr j wherer 1 > r2 >· · ·> rJ denotes progressively coarser grids. 12 B Proof of Theorem 1 We begin by noting that the residual correction iterations, restriction operations, prolongation opera- tio...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.