Pith. sign in

REVIEW 3 major objections 5 minor 65 references

PMNO: A novel physics guided multi-step neural operator predictor for partial differential equations

T0 review · 3 major / 5 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 A neural operator that consumes several past solution states and is trained against an implicit backward-differentiation-formula residual can extrapolate PDE solutions far beyond its training window, with valid Gray-Scott predictions…

desk verdict The multi-step + BDF idea is sound and the gains over data-only baselines are real, but the BDF mechanism itself is never tested in isolation — the PI baselines are already PMNO variants. read the letter →

arxiv 2506.01598 v1 pith:YFPMKQIP submitted 2025-06-02 cs.LG physics.comp-ph

Jin Song , Kenji Kawaguchi , Zhenya Yan This is my paper
classification cs.LGphysics.comp-ph MSC 68T0765L0665M06
keywords neuraloperatorspartialdifferentialequationslong-horizonpredictionbackwarddifferentiationformulaphysics-informedlearningmulti-stepmethodsFourieroperatorDeepONet
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

PMNO—physics guided multi-step neural operator—is an attempt to fix two weaknesses of neural operator surrogates for PDEs: short valid prediction horizons and the need for large training datasets. The forward predictor consumes $k$ past solution states and outputs the next state as a learned weighted combination of those states and neural-operator updates, while the backpropagation loss forces each rollout stage to satisfy the residual of an implicit $k$-step backward differentiation formula for the governing differential operator. The paper reports that this combination keeps errors below a 10% relative-RMSE threshold far past the training window across five physical systems, with the 3D Gray-Scott model reaching valid prediction time 82. It also reports stable training with only the first $k$ states labeled, causal weighting that avoids staged training, and resolution-invariant extrapolation inherited from the FNO or DeepONet backbone. The claim, taken sympathetically, is that borrowing the temporal structure of classical linear multi-step integrators—not adding network capacity—is what buys the longer horizon.

What carries the argument

The load-bearing mechanism is the pairing of a learned multi-step predictor with an implicit backward differentiation formula (BDF) residual as the training loss. BDF is a family of implicit linear multi-step schemes for $u_t = \mathcal{N}(u)$ that uses derivative information at the current step; the $k$-step BDF is zero-stable for $k \le 6$. The forward ansatz is $P[u_i,\ldots,u_{i+k-1}] = \sum_{j=0}^{k-1}[\lambda_j u_{i+j} + \Delta t\,\delta_j G(u_{i+j})]$, which mirrors the explicit linear multi-step formula, while the backpropagation loss enforces $\sum_{j=0}^{k} a_j u_{i+j} = \Delta t\, b_k \mathcal{N}[u_{i+k}]$ with $k=5$ by default. A double-ended queue rolls the window forward, and the causal weights $\omega_i = \exp(-\epsilon \sum_{j<i} L_j)$ prevent the optimizer from minimizing later residuals before earlier ones are small. This is what lets the model learn the dynamics from the first $k$ labeled states and then extend them.

What would settle it

Compute the BDF-5 residual $\|\sum_{j=0}^{5} a_j u_{i+j} - \Delta t\, b_5 \mathcal{N}[u_{i+5}]\|$ on the reference Gray-Scott solution at the training resolution ($32^3$, $\Delta t=0.4$); if this residual is already large where the model still predicts well, the reported extrapolation gain cannot come from the physics guidance, and a data-only single-step FNO should be retested on the same data budget.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the temporal structure of the predictor, not only the backbone architecture, determines how far a learned PDE surrogate can be trusted. PMNO replaces the single-state operator input with a window of $k$ past states and defines the next state as $$P[u_i,\ldots,u_{i+k-1}] = \sum_{j=0}^{k-1}\bigl[\lambda_j u_{i+j} + \$\Delta$ t\,\delta_j G(u_{i+j})\bigr],$$ with learnable coefficients $\lambda_j,\delta_j$ and a neural operator $G$ (FNO or DeepONet). Training unrolls this predictor and, at each stage, minimizes the BDF residual $$\Bigl\|\sum_{j=0}^{k} a_j u_{i+j} - \$\Delta$ t\, b_k \mathcal{N}[u_{i+k}]\Bigr\|$_2^{2}$,$$ where $\mathcal{N}$ is the PDE operator computed by automatic differentiation or finite differences and the $a_j,b_k$ are fixed BDF coefficients. The paper reports that this physics-residual training, plus multi-step history and causal weighting, keeps the relative RMSE below the 0.1 validity threshold for much longer horizons than data-only baselines, with valid prediction time 82 on the 3D Gray-Scott system and 24 on spiral-wave reaction-diffusion, while retaining extrapolation on finer spatial grids.

Load-bearing premise

The load-bearing premise is that the BDF time step and the spatial finite-difference evaluation of the PDE operator on the training grid are accurate enough that the residual really describes the true physics; if those discretizations are coarse, the physics guidance would teach the network the discretized dynamics rather than the true one.

Editorial extensions

If this is right

  • A model trained once on a coarse grid can be rolled out on a finer grid without retraining, because the input states are simply re-sampled on the target grid and the backbone is FNO or DeepONet.
  • Only the first $k$ states need to be labeled; the rest of the trajectory is constrained by physics residuals, so the same accuracy should be attainable with fewer training trajectories than single-step data-driven operators.
  • Increasing the step order $k$ up to the BDF stability bound ($k \le 6$) extends the valid prediction horizon; on the 2D advection test, the paper finds monotone improvement from $k=1$ to $k=6$.
  • The causal weighting scheme lets the model train end-to-end over a long recurrent rollout without multi-stage pre-training or fine-tuning, at negligible extra computational cost.
  • For a new PDE application, the same training loop can be used with a problem-appropriate backbone, such as FNO for periodic domains and DeepONet for irregular or mesh-free domains.

Reading between the lines

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

  • Ablation the paper does not run: hold the multi-step input fixed and swap the implicit BDF residual for an explicit Euler residual at the same $k$; if the horizon gain disappears, the implicit scheme—not the extra history—is the active ingredient.
  • Replace BDF with another implicit scheme such as Crank-Nicolson or Adams-Moulton; comparable horizons would identify implicit stability as the mechanism, while a clear drop would single out BDF's stiff-decay property.
  • The resolution-invariance claim is tested only at moderate refinement, for example $32^3$ to $64^3$; evaluating broadband initial data at $256^3$ after training at $32^3$ would stress whether spectral aliasing limits the claimed invariance.
  • An $\epsilon$ sweep for the causal weight schedule would show how sensitive the end-to-end training claim is to the choice of schedule steepness, since the paper fixes one schedule per experiment.
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 / 5 minor

Summary. The paper introduces PMNO, a recurrent neural operator predictor that (i) takes multiple past solution states as input with learned weights, (ii) trains with a physics residual written in Backward Differentiation Formula form, and (iii) uses a causal weighting schedule. Experiments on five PDE systems report valid-prediction times, resolution-invariant extrapolation, and sensitivity studies for the multi-step order k and the recurrent length L. The central claim is that the BDF-guided backpropagation scheme, rather than more data or a larger network, is responsible for the improved long-horizon extrapolation.

Significance. If the attribution claim is established, the paper makes a useful contribution: it connects classical linear multistep theory to recurrent neural operator training, and it demonstrates consistent improvements over purely data-driven recurrent baselines across advection, reaction-diffusion, irregular-domain heat, and complex-valued wave problems. Strengths of the manuscript include the release of code and data, the use of fixed BDF coefficients rather than fitted parameters, hard-constraint boundary treatment, and explicit sensitivity analyses for k and L. However, the central scientific claim is currently not isolated by the experimental design: the table rows labeled PIdeepONet and PIFNO already include the multi-step architecture and causal training, so the reported gains cannot be attributed specifically to BDF. The contribution is therefore significant but conditional on an ablation that the manuscript does not contain.

major comments (3)
  1. [Section 4 and Table 1] The central attribution of the extrapolation gains to BDF is not isolated. Section 4 states that 'we employ the multi-step neural operator structure for the general operator-based prediction methods in the following examples, highlighting the key role of implicit physics guidance in backpropagation.' Thus the rows PIdeepONet-MLP, PIdeepONet-CNN, and PIFNO are not the standard physics-informed operator baselines of Refs. [40,41]; they are PMNO instantiations with multi-step inputs, BDF-form residual, causal weighting, and learned λ,δ weights. No experiment replaces the BDF residual in Eq. (6) with a continuous PDE residual or an explicit Euler residual while holding the multi-step architecture, causal weighting, training data, and optimization budget fixed. Consequently, the comparison in Table 1 (e.g., PIFNO Tv=82 vs FNO Tv=76 for Gray-Scott) does not establish that BDF, rather than multi-step input or causal training, is the cause of the improvement. This missing control is load-bearing for the paper's main claim, and the manuscript should be revised to include such an ablation.
  2. [Section 5.2 and Table 3] The reported performance metrics lack error bars, repeated random seeds, or standard deviations. All valid-time values in Table 1 and all error curves in Figures 2-12 appear to be single-run results. Since Tv is defined as a crossing time of a threshold, it is sensitive to stochastic training, and a single run does not support the 'state-of-the-art' claim. In addition, the causal weighting steepness ε in Eq. (9) is a user-set hyperparameter and is a claimed contributor to training efficiency, but its values are not reported in Table 3 and its sensitivity is not studied; Section 5.2 varies only k and L. The authors should report ε for each experiment and include at least a small seed ensemble.
  3. [Section 3.1 and Eq. (6)] The naming of the backpropagation loss as an 'implicit time-stepping scheme based on BDF' is imprecise. Eq. (4), the forward predictor, is explicitly constructed with bk=0, while Eq. (6) is a PDE residual written in BDF form; the learned weights λj,δj are not tied to the BDF coefficients aj,bk. This is not itself an error, because Eq. (6) can act as a soft physical constraint on the recurrent outputs, but the text should clarify that the method does not implement BDF as the forward integrator. Without that clarification, the reader may over-interpret the 'implicit BDF' phrase in the abstract and Section 1.
minor comments (5)
  1. [Section 3.1, Eq. (3)] The definitions following Eq. (3) contain a typo: 'αj = −aj/ak and αj = ∆tbj/ak' should read 'αj = −aj/ak and βj = ∆tbj/ak'.
  2. [Appendix C, Table 4] The branch-net structure notation '[Conv2d, Tanh]+[1002, 100]' is ambiguous; it likely means a layer of width 100 followed by a layer of 100 outputs, but the notation should be made unambiguous (e.g., 100^2 or 100,2).
  3. [Section 4] There is a typo in the text: 'ConLSTM' should be 'ConvLSTM' in the sentence describing the comparison method.
  4. [Section 3.2, Eq. (16)] The definition of the star-shaped domain is missing parentheses around the rational expression; as printed, the bound '0 ≤ r ≤ R 1+0.5 sin(5θ)/1+0.5|sin(2.5θ)|' is ambiguous.
  5. [Section 3.3, Eq. (9)] The causal weighting matrix uses 0 entries on the diagonal for all i, so ω_i does not depend on the current residual L_i; the text should state explicitly that this is intentional and that the weights are computed from previously accumulated residuals.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the physics residual uses fixed BDF coefficients and the extrapolation targets are not fitted inputs.

full rationale

The derivation chain is self-contained. The forward predictor in Eq. (4) is a learned residual multi-step operator, and the backpropagation loss in Eq. (6) enforces the fixed-coefficient BDF relation sum a_j u_{i+j} - Delta t b_k N[u_{i+k}] = 0. The BDF coefficients in Appendix B are standard numerical-analysis constants (Gear/Dahlquist), not parameters fitted to any target. The only labeled data are the initial k steps; future states are recurrent model outputs, and the reported valid times are evaluated against independent reference solutions, so the prediction is not fitted to the measured quantity by construction. Using the same PDE in the physics loss and in evaluation is the standard physics-informed setup and does not make the output equal to its input. No load-bearing self-citation chain is present: the cited PIdeepONet/PIFNO baselines [40,41], causal training [53], and operator approximation theorems [28-32] are external to this paper's authors and are not used to forbid alternatives. The paper's main attribution weakness is an omitted control isolating the BDF residual from the multi-step architecture and causal weighting; all physics-guided methods in Table 1 are PMNO-style instantiations, as Section 4 states, so the specific source of the extrapolation gain is underdetermined. That is a missing ablation concerning attribution, not an equivalence by construction, and therefore does not raise the circularity score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. Its reliance on domain knowledge is through the PDE operator and the BDF discretization, both of which are standard inputs. The free parameters are the learned network weights and the hand-set causal weighting parameter.

free parameters (3)
  • Forward weights {λj, δj}_{j=0}^{k-1} = Learned during training (values not reported)
    These weights combine identity and neural operator terms in Eq. (4); they are optimized with Adam and are central to the predictor.
  • Causal weighting steepness ε = Not reported
    Controls the exponential decay of the causal loss weights in Eq. (9); set by hand, no value given in the paper.
  • Internal parameters of neural operator G = Learned
    Parameters of FNO or DeepONet used as the operator backbone; standard network parameters.
assumptions (4)
  • domain assumption The governing PDE operator N and its parameters are exactly known and match the data generation process.
    Used in the BDF residual loss, Eq. (6). If the PDE is misspecified, the physics guidance would be wrong.
  • domain assumption The k-step BDF scheme with k=5 is stable and accurate for the chosen time steps and PDEs.
    The loss treats the BDF residual as a valid training target; stability and accuracy for specific Δt values (e.g., Δt=0.4 for Gray-Scott) are not verified in the paper.
  • ad hoc to paper Spatial derivatives in N can be computed accurately enough by automatic differentiation or finite differences at the training resolution.
    The paper does not analyze the truncation error of finite differences on coarse grids (e.g., 32^3 for 3D problems), yet these derivatives are central to the physics loss.
  • standard math Neural operators G satisfy the universal approximation property, allowing the residual error to be controlled at each stage.
    Invoked in Section 3.1 to justify that the operator can approximate the dynamics; this is a standard result from the cited neural operator literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PMNO: A novel physics guided multi-step neural operator predictor for partial differential equations." pith.science (2026). https://pith.science/paper/YFPMKQIP

@misc{pith2026250601598,
  author       = {Pith},
  title        = {Pith review of: PMNO: A novel physics guided multi-step neural operator predictor for partial differential equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YFPMKQIP}},
  note         = {Machine review of arXiv:2506.01598}
}
read the original abstract

Neural operators, which aim to approximate mappings between infinite-dimensional function spaces, have been widely applied in the simulation and prediction of physical systems. However, the limited representational capacity of network architectures, combined with their heavy reliance on large-scale data, often hinder effective training and result in poor extrapolation performance. In this paper, inspired by traditional numerical methods, we propose a novel physics guided multi-step neural operator (PMNO) architecture to address these challenges in long-horizon prediction of complex physical systems. Distinct from general operator learning methods, the PMNO framework replaces the single-step input with multi-step historical data in the forward pass and introduces an implicit time-stepping scheme based on the Backward Differentiation Formula (BDF) during backpropagation. This design not only strengthens the model's extrapolation capacity but also facilitates more efficient and stable training with fewer data samples, especially for long-term predictions. Meanwhile, a causal training strategy is employed to circumvent the need for multi-stage training and to ensure efficient end-to-end optimization. The neural operator architecture possesses resolution-invariant properties, enabling the trained model to perform fast extrapolation on arbitrary spatial resolutions. We demonstrate the superior predictive performance of PMNO predictor across a diverse range of physical systems, including 2D linear system, modeling over irregular domain, complex-valued wave dynamics, and reaction-diffusion processes. Depending on the specific problem setting, various neural operator architectures, including FNO, DeepONet, and their variants, can be seamlessly integrated into the PMNO framework.

Figures

Figures reproduced from arXiv: 2506.01598 by the authors.

Figure 1
Figure 1. The architectural overview of the PMNO. (a) Forward process and backpropagation. (b) Structure of the multi-step neural operator predictor. effective implicit methods, we primarily adopt the Backward Differentiation Formula (BDF) [10] due to its favorable stability properties for training our neural operator model. Besides, it involves only the information of current-step function N [x, ti+k ; ui+k ] on the right-ha… view at source ↗
Figure 2
Figure 2. 2D advection equation: (a) Error propagation curves, where the gray region on the left represents the training domain, while the yellow one on the right indicates the extrapolation domain. (b) The reference solution at t = 3. (c) The predictions and corresponding errors (error = u − u ∗ ) from each predictor at t = 3. (v1(t), v2(t)) and decreases rapidly away from this point. Hence, the solution can be assumed to va… view at source ↗
Figure 3
Figure 3. λ-ω RD equations: Error propagation curves, where the gray region on the left represents the training domain, while the yellow one on the right indicates the extrapolation domain [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: λ-ω RD equations: The reference solutions and the predictions from each predictor at different times t = 2.5, t = 5 and t = 15. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: 2D heat equation with an irregular domain: (a) Error propagation curves, where the gray region on the left represents the training domain, while the yellow one on the right indicates the extrapolation domain. (b) The reference solution at t = 3.5. (c) The predictions a…
Figure 6
Figure 6. Figure 6: 3D complex-valued wave dynamics: (a) Error propagation curves, where the gray region on the left represents the training domain, while the yellow one on the right indicates the extrapolation domain. (b) The corresponding loss curves. sampling points, is able to accurat…
Figure 7
Figure 7. Figure 7: 3D complex-valued wave dynamics: Isosurfaces of |U| at values 0.2, 0.85, 1.5 for predictions and references from each predictor at different times t = 5.25, t = 8.25, t = 10.5 and t = 14.25. isosurfaces of the corresponding error against reference solution from each pr…
Figure 8
Figure 8. Figure 8: 3D complex-valued wave dynamics: Isosurfaces of |U − U∗ | at values 0.01, 0.03, 0.05 for corresponding error against reference solution from each predictor at different times t = 5.25, t = 8.25, t = 10.5 and t = 14.25. behavior, the relative error consistently stays un…
Figure 9
Figure 9. Figure 9: 3D Gray–Scott model: (a) Error propagation curves, where the gray region on the left represents the training domain, while the yellow one on the right indicates the extrapolation domain. (b) The corresponding loss curves. 4.5 3D Gray–Scott model In the last example, we…
Figure 10
Figure 10. Figure 10: 3D Gray–Scott model: Isosurfaces of u at values 0.3, 0.4, 0.5 and v at values 0.1, 0.15, 0.2 for each predictor and the corresponding error at values 0.002, 0.016, 0.03 against reference solution at t = 40. 5 Algorithm analysis 5.1 Computational cost Firstly, we take …
Figure 11
Figure 11. Figure 11: Impact of k on the performance of the PIFNO predictor for 2D advection equation. (a) Error propaga￾tion curves, where the gray region on the left represents the training domain, while the yellow one on the right indicates the extrapolation domain. (b) The correspondin…
Figure 12
Figure 12. Figure 12: Impact of L on the performance of the PIdeepONet-CNN predictor for λ-ω RD equations. (a) Error propagation curves at k = 5. (b) The corresponding loss curves. begin decreasing significantly. These observations highlight the need to carefully balance the training durat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 59 canonical work pages

  1. [1]

    Infeld, Nonlinear Waves, Solitons and Chaos, 2nd ed., Cambridge Univ

    E. Infeld, Nonlinear Waves, Solitons and Chaos, 2nd ed., Cambridge Univ . Press, Cambridge, 2000

    work page 2000

  2. [2]

    R. C. Hilborn, Chaos and Nonlinear Dynamics: An Introduc tion for Scientists and Engineers, 2nd ed., Ox- ford Univ . Press, Oxford, 2000

    work page 2000

  3. [3]

    S. H. Strogatz, Nonlinear Dynamics and Chaos with Applic ations to Physics, Biology , Chemistry and Engi- neering, 3rd ed., CRC Press, New York, 2024

    work page 2024

  4. [4]

    Hasegawa, Y

    A. Hasegawa, Y . Kodama, Solitons in Optical Communicati ons, Oxford Univ . Press, Oxford, 1995

    work page 1995

  5. [5]

    Pitaevskii, S

    L. Pitaevskii, S. Stringari, Bose–Einstein Condensati on, Oxford Univ . Press, Oxford, 2003

    work page 2003

  6. [6]

    Y . V . Kartashov , G. E. Astrakharchik, B. A. Malomed, L. Torner, Frontiers in multidimensional self-trapping of nonlinear fields and matter, Nat. Rev . Phys. 1 (2019) 185–197. 24 T able 6:Summary of FNO architectures and hyperparameters for diffe rent physical systems, where K is integral kernel operator parameterized by Fourier transforms with d ifferen...

    work page 2019

  7. [7]

    A. R. Osborne, Nonlinear Ocean Waves, Academic Press, Ne w York, 2009

    work page 2009

  8. [8]

    M. Ghil, R. Vautard, Interdecadal oscillations and the w arming trend in global temperature time series, Nature 350 (1991) 324-327

    work page 1991

Show all 65 references
  1. [9]

    Yan, V ector financial rogue waves, Phys

    Z. Yan, V ector financial rogue waves, Phys. Lett. A 375 (20 11) 4274-4279

  2. [10]

    C. W. Gear, The numerical integration of ordinary diffe rential equations, Math. Comput. 21 (1967) 146-156

    1967

  3. [11]

    Yang, Nonlinear Waves in Integrable and Nonintegrab le Systems, SIAM, 2010

    J. Yang, Nonlinear Waves in Integrable and Nonintegrab le Systems, SIAM, 2010

    2010

  4. [12]

    Stoer, R

    J. Stoer, R. Bulirsch, Introduction to Numerical Analy sis, Springer, 2013

    2013

  5. [13]

    J. Shen, T. Tang, L. L. Wang, Spectral Methods: Algorith ms, Analysis and Applications, Springer, 2011

    2011

  6. [14]

    R. J. LeV eque, Finite difference methods for ordinary a nd partial differential equations: steady-state and time-dependent problems, SIAM, 2007

    2007

  7. [15]

    O. C. Zienkiewicz, R. L. Taylor, J. Z. Zhu, The finite elem ent method: its basis and fundamentals, Elsevier, 2005

    2005

  8. [16]

    Raissi, P

    M. Raissi, P . Perdikaris, G. E. Karniadakis, Machine le arning of linear diferential equations using Gaussian processes, J. Comput. Phys. 348 (2017) 683-693

    2017

  9. [17]

    Raissi, P

    M. Raissi, P . Perdikaris, G. E. Karniadakis, Physics-i nformed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial diferential equations, J. Comput. Phys. 378 (2019) 686-707

    2019

  10. [18]

    J. Han, A. Jentzen, W. E, Solving high-dimensional part ial differential equations using deep learning, Proc. Natl. Acad. Sci. 115 (2018) 8505-8510

    2018

  11. [19]

    Bar-Sinai, S

    Y . Bar-Sinai, S. Hoyer, J. Hickey , M. P . Brenner, Learni ng data-driven discretizations for partial diferential equations, Proc. Natl Acad. Sci. 116 (2019) 15344-15349

    2019

  12. [20]

    Sanchez-Gonzalez, J

    A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Le skovec, P . Battaglia, Learning to simulate complex physics with graph networks, in International Conference on Machine Learning, PMLR, 2020, pp. 8459-8468

    2020

  13. [21]

    Pathak, B

    J. Pathak, B. Hunt, M. Girvan, Z. Lu, E. Ott, Model-free p rediction of large spatiotemporally chaotic systems from data: a reservoir computing approach, Phys. Rev . Lett. 120 (2018) 024102

    2018

  14. [22]

    X. Shi, Z. Chen, H. Wang, D. Y . Yeung, W. K. Wong, W. C. Woo, Convolutional LSTM network: A machine learning approach for precipitation nowcasting, Advances in neural information processing systems, 28 (2015). 25

    2015

  15. [23]

    Z. Long, Y . Lu, X. Ma, B. Dong, Pde-net: Learning pdes fro m data, in International conference on machine learning , PMLR, 2018, pp. 3208-3216

    2018

  16. [24]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jone s, A. N. Gomez, L. U. Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing s ystems, 30 (2017)

    2017

  17. [25]

    Geneva, N

    N. Geneva, N. Zabaras, Transformers for modeling physi cal systems, Neural Netw . 146 (2022) 272-289

    2022

  18. [26]

    LeCun, Y

    Y . LeCun, Y . Bengio, G. Hinton, Deep learning, Nature 52 1 (2015) 436-444

    2015

  19. [27]

    Karpatne, G

    A. Karpatne, G. Atluri, J. H. Faghmous, et al. Theory-gu ided data science: a new paradigm for scientific discovery from data, IEEE Trans. Knowl. Data Eng. 29 (2017) 2 318-2331

    2017

  20. [28]

    L. Lu, P . Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learn ing nonlinear operators via DeepONet based on the universal approximation theorem of operators, Nat. Mac h. Intell. 3 (2021) 218-229

    2021

  21. [29]

    T. Chen, H. Chen, Universal approximation to nonlinear operators by neural networks with arbitrary acti- vation functions and its application to dynamical systems, IEEE Trans. Neural Networks 6 (1995) 911-917

    1995

  22. [30]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhatt acharya, A. Stuart, A. Anandkumar, Neural operator: Learning maps between function spaces with applications to pdes, J. Mach. Learn. Res. 24 (2023) 1-97

    2023

  23. [31]

    Kovachki, S

    N. Kovachki, S. Lanthaler, S. Mishra, On universal appr oximation and error bounds for fourier neural oper- ators, J. Mach. Learn. Res. 22 (2021) 1-76

    2021

  24. [32]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhatt acharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric partial differential equations, 2 020, arXiv preprint arXiv:2010.08895

    2010 arXiv

  25. [33]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhatt acharya, A. Stuart, A. Anandkumar, Neural operator: Graph kernel network for partial differential equations, 2 020, arXiv preprint arXiv:2003.03485

    2003 arXiv

  26. [34]

    Gupta, X

    G. Gupta, X. Xiao, P . Bogdan, Multiwavelet-based opera tor learning for differential equations, Advances in neural information processing systems, 34 (2021) 24048-24 062

    2021

  27. [35]

    Tripura, S

    T. Tripura, S. Chakraborty , Wavelet neural operator: a neural operator for parametric partial differential equations, 2022, arXiv preprint arXiv:2205.02191

    2022 arXiv

  28. [36]

    Michalowska, S

    K. Michalowska, S. Goswami, G. E. Karniadakis, S. Rieme r-Sorensen, Neural operator learning for long-time integration in dynamical systems with recurrent neural net works, in 2024 International Joint Conference on Neural Networks (IJCNN) (2024) 1-8

    2024

  29. [37]

    Karumuri, L

    S. Karumuri, L. Graham-Brady , S. Goswami, Physics-inf ormed latent neural operator for real-time predic- tions of complex physical systems, 2025, preprint arXiv:25 01.08428

    2025

  30. [38]

    Q. Cao, S. Goswami, T. Tripura, S. Chakraborty , and G. E.Karniadakis, Deep neural operators can predict the real-time response of floating offshore structures under ir regular waves, Comput. Struct. 291 (2024) 107228

    2024

  31. [39]

    Y . Wang, Z. Li, Z. Yuan, W. Peng, T. Liu, J. Wang, Predicti on of turbulent channel flow using Fourier neural operator-based machine-learning strategy , Phys. Rev . Fluids 9 (2024) 084604

    2024

  32. [40]

    Z. Li, H. Zheng, N. Kovachki, D. Jin, H. Chen, B. Liu, A. An andkumar, Physics-informed neural operator for learning partial differential equations, ACM/JMS J. Da ta Sci. 1 (2024) 1-27

    2024

  33. [41]

    S. Wang, H. Wang, P . Perdikaris, Learning the solution o perator of parametric partial differential equations with physics-informed DeepONets, Sci. Adv . 7 (2021) eabi8605

    2021

  34. [42]

    P . R. Vlachas, W. Byeon, Z. Y . Wan, T. P . Sapsis, P . Koumou tsakos, Data-driven forecasting of high- dimensional chaotic systems with long short-term memory ne tworks, Proc. R. Soc. A 474 (2018) 20170844

    2018

  35. [43]

    Chang, L

    B. Chang, L. Meng, E. Haber, F. Tung, D. Begert, Multi-le vel residual networks from dynamical systems view , 2017, arXiv preprint arXiv:1710.10348

    2017 arXiv

  36. [44]

    Zhang, Z

    X. Zhang, Z. Li, C. Loy , D. Lin, Polynet: A pursuit of stru ctural diversity in very deep networks, in Proceed- ings of the IEEE Conference on Computer Vision and Pattern Re cognition, 2017, pp. 718-726

    2017

  37. [45]

    Larsson, M

    G. Larsson, M. Maire, G. Shakhnarovich, Fractalnet: Ul tra-deep neural networks without residuals, 2016, arXiv preprint arXiv:1605.07648

    2016 arXiv

  38. [46]

    Y . Lu, A. Zhong, Q. Li, B. Dong, Beyond finite layer neural networks: Bridging deep architectures and numerical differential equations, in International confe rence on machine learning, PMLR, 2018, pp. 3276- 3285

    2018

  39. [47]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning fo r image recognition, in Proceedings of the IEEE conference on computer vision and pattern recognition, 201 6, pp. 770-778

  40. [48]

    G. G. Dahlquist, A special stability problem for linear multistep methods, BIT Numer. Math. 3 (1963) 27-43

    1963

  41. [49]

    L. Lu, R. Pestourie, W. Yao, Z. Wang, F. V erdugo, S. G. Joh nson, Physics-informed neural networks with hard constraints for inverse design, SIAM J. Sci. Comput. 43 (2021) B1105-B1132

    2021

  42. [50]

    Y . Du, T. A. Zaki, Evolutional deep neural network, Phys . Rev . E 104 (2021) 045303

    2021

  43. [51]

    S. Dong, N. Ni, A method for representing periodic funct ions and enforcing exactly periodic boundary conditions with deep neural networks, J. Comput. Phys. 435 ( 2021) 110242

    2021

  44. [52]

    C. Rao, P . Ren, Q. Wang, O. Buyukozturk, H. Sun, Y . Liu, En coding physics to learn reaction–diffusion processes, Nat. Mach. Intell. 5 (2023) 765-779. 26

    2023

  45. [53]

    S. Wang, S. Sankaran, P . Perdikaris, Respecting causal ity for training physics-informed neural networks, Comput. Methods Appl. Mech. Engrg. 421 (2024) 116813

    2024

  46. [54]

    D. P . Kingma, J. Ba, Adam: a method for stochastic optimi zation, 2014, preprint, arXiv:1412.6980

    2014 arXiv

  47. [55]

    L. Lu, X. Meng, S. Cai, Z. Mao, S. Goswami, Z. Zhang, G. E. K arniadakis, A comprehensive and fair compar- ison of two neural operators (with practical extensions) ba sed on fair data, Comput. Methods Appl. Mech. Engrg. 393 (2022) 114778

    2022

  48. [56]

    Y . Mei, Y . Zhang, X. Zhu, R. Gou, J. Gao, Fully Convolutional Network enhanced DeepONet-based surrogate of predicting the travel-time fields, IEEE Trans. Geosci. Re mote Sens. (2024)

    2024

  49. [57]

    L. D. Landau, E. M. Lifshitz, Fluid Mechanics: V olume 6, Elsevier, 1987

    1987

  50. [58]

    E. E. Holmes, M. A. Lewis, J. E. Banks, R. R. V eit, Partial diferential equations in ecology: spatial interactions and population dynamics, Ecology 75 (1994) 17-29

    1994

  51. [59]

    V ervloet, F

    D. V ervloet, F. Kapteijn, J. Nijenhuis, J. R. van Ommen, Fischer–Tropsch reaction–difusion in a cobalt catalyst particle: aspects of activity and selectivity for a variable chain growth probability , Catal. Sci. Technol. 2 (2012) 1221-1233

    2012

  52. [60]

    Halatek, E

    J. Halatek, E. Frey , Rethinking pattern formation in re action–difusion systems, Nat. Phys. 14 (2018) 507-514

    2018

  53. [61]

    Y . F. Sun, P . Parra-Rivas, C. Milian, Y . V . Kartashov , M. Ferraro, F. Mangini, R. Jauberteau, F. Talenti, S. Wabnitz, Robust three-dimensional high-order solitons an d breathers in driven dissipative systems: A kerr cavity realization, Phys. Rev . Lett. 131 (2023) 137201

    2023

  54. [62]

    Bruna, B

    J. Bruna, B. Peherstorfer, E. Vanden-Eijnden, Neural G alerkin schemes with active learning for high- dimensional evolution equations, J. Comput. Phys. 496 (202 4) 112588

  55. [63]

    Griewank, A

    A. Griewank, A. Walther, Evaluating derivatives: prin ciples and techniques of algorithmic differentiation, SIAM, 2008

    2008

  56. [64]

    Z. Shi, Z. Hu, M. Lin, K. Kawaguchi, Stochastic Taylor De rivative Estimator: Efficient amortization for arbitrary differential operators, 2024, arXiv preprint ar Xiv:2412.00088

    2024 arXiv

  57. [65]

    Glorot, Y

    X. Glorot, Y . Bengio, Understanding the difficulty of tr aining deep feedforward neural networks, J. Mach. Learn. Res. 9 (2010) 249-256. 27

    2010

Pith tools

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