Pith. sign in

REVIEW 4 major objections 4 minor 52 references

Hierarchical Implicit Neural Emulators

T0 review · 4 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 Coarse future states keep turbulent emulation stable for 200,000 steps.

desk verdict A genuinely new conditioning mechanism for neural emulators, with solid ablations and minimal compute overhead, but the headline stability number needs code and stronger baselines before I'd trust it. read the letter →

arxiv 2506.04528 v1 pith:5FYHVG3E submitted 2025-06-05 cs.LG

Ruoxi Jiang , Xiao Zhang , Karan Jakhar , Peter Y. Lu , Pedram Hassanzadeh , Michael Maire , Rebecca Willett This is my paper
classification cs.LG
keywords neuralemulatorautoregressivepredictionlong-termstabilityturbulentflowNavier-Stokeshierarchicalrepresentationimplicittime-steppingmultiscaledynamics
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

The paper tries to establish that an autoregressive neural emulator can stay stable over rollouts ten times longer than its training data, provided each next-step prediction is conditioned on its own coarse-grained predictions of several future states. The central idea is to use a hierarchy of downsampled future snapshots—one step ahead, two steps ahead, and so on—as an implicit refinement signal, the way implicit time-stepping methods stabilize numerical integration without iterative root solving. On 2D Navier-Stokes turbulence at Reynolds number $10^4$, the three-level model stays physically stable for 93% of $2\times10^5$-step rollouts, versus 38% for the best baseline, and cuts mean squared error by more than half at steps 25 to 50. If these results generalize, cheap long-horizon emulation of chaotic, multiscale systems becomes a more realistic tool for climate and fluid applications.

What carries the argument

The central object is the augmented state $(u_n, z^{(1)}_{n+1}, \ldots, z^{(L-1)}_{n+L-1})$, where $u_n$ is the current full-resolution field and each $z^{(l)}$ is a fixed spatial downsampling of a future state at increasing compression (rates $r_1=8$, $r_2=32$ for $L=3$). The emulator is trained to map this augmented state to the next physical state and the next shifted ladder of latent predictions, so that one forward pass both refines the immediate prediction and prepares the coarse future context for the following step. The hierarchy does the work: low-resolution future latents are cheap to predict and inject, they supervise the model toward long-term structure, and during rollout they provide a stabilizing feedback channel that baselines lacking future context do not have.

What would settle it

A direct test is to record, over a $2\times10^5$-step rollout, the error $\|\hat{z}^{(l)}_{n+l} - T^{(l)}(u_{n+l})\|$ between each predicted latent and the true coarse-grained state; the claimed mechanism is falsified if these latent errors grow large and lose correlation with the true future before or without the energy-based stability threshold being crossed.

Watch

Extended reading notes

Core claim

The paper's claim is that feeding an encoder-decoder emulator a ladder of spatially downsampled future latents $z^{(l)}_{n+l}$ as additional inputs turns a standard autoregressive model into a 'hierarchical implicit' emulator that refines each next-state prediction against anticipated future context. During training those latents are computed from the true trajectory; during rollout the model feeds back its own predicted latents, and the paper reports that this closed loop is what keeps long runs from collapsing. The headline evidence is on 2D turbulence at Reynolds number $10^4$: the $L=3$ configuration sustains a 93% stability rate over $2\times10^5$ steps, while the best baseline reaches only 38%, and it reduces MSE by over 50% at prediction steps 25–50 relative to the one-step baseline. The authors interpret this as the neural analogue of implicit time-stepping, with the iterative refinement executed in parallel through auxiliary low-resolution output heads rather than through an iterative solver.

Load-bearing premise

The load-bearing premise is that the model's self-predicted coarse-grained future states, fed back at inference, remain informative over 200,000 autoregressive steps; the paper trains on true future states and reports stability outcomes, but does not directly measure how those predicted latents drift during the rollout.

Editorial extensions

If this is right

  • On $Re=10^4$ 2D turbulence, the $L=3$ emulator achieves a 93% stability rate over $2\times10^5$ autoregressive steps (10 times the training length), measured by conserved kinetic energy staying within 5 standard deviations of the true mean; the best baseline, 2-Step History, reaches only 38%.
  • At prediction steps 25, 50, and 75, the $L=3$ model's rollout MSE is about 0.034, 0.140, and 0.324, compared with 0.080, 0.312, and 0.619 for the one-step baseline—a reduction of roughly 55% at step 50.
  • The same $L=3$ configuration preserves zonal-mean jet structure and the energy spectrum in $2\times10^5$-step rollouts, while baseline methods drift to averaged or exploding states; $L=2$ shows that even a two-level hierarchy helps and can recover after temporary energy excursions.
  • The added cost is small: $L=3$ uses 36.67M parameters and 0.031s per forward pass versus 34.88M and 0.030s for the one-step baseline.
  • On additional flows—$Re=5\times10^3$ without jets at 256×256 and $Re=10^4$ with jets at 512×512—$L=3$ keeps lower rollout MSE than the $L=1$ baseline at steps 25–75, suggesting the mechanism is not specific to one dataset.

Reading between the lines

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

  • Because the stability gain comes from feeding coarse future context back, a natural next test is whether the same hierarchy transfers to regimes with stronger nonlinearity or different aspect ratios, where the coarse variables may lose predictive memory more quickly.
  • The training-inference gap could be probed by adding a consistency loss that penalizes divergence between predicted and true latents during rollout, or by fine-tuning with teacher forcing; if stability improves further, the residual gap is a bottleneck rather than a feature.
  • A quantitative link to implicit methods could be made by fixing $L$ and increasing the emulation timestep $\Delta t$; if the hierarchy's stability margin grows with $L$ in the same way implicit schemes allow larger steps, the analogy would become a testable design principle.
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

4 major / 4 minor

Summary. The paper proposes a hierarchical implicit neural emulator for autoregressive prediction of turbulent flows. The model augments the current state with a hierarchy of coarse-grained future latent states z^{(l)}_{n+l}, trained with the true future latents as conditioning inputs (Eq. 10) and using the model's own previous predictions during inference. The authors report that on 2D Navier-Stokes at Re=10^4, the L=3 model reduces short-term MSE by over 50% at steps 25-50 relative to the L=1 baseline and achieves a 93% stability rate over 2e5 rollout steps, versus 38% for the best baseline, while adding minimal computational overhead (Tables 4-6, Fig. 3). The method is also tested on no-jet and 512x512 configurations (Table 3, Fig. 6).

Significance. If the empirical gains hold, the paper offers a simple and computationally cheap architectural modification that improves both short-term accuracy and long-term robustness of neural PDE emulators, a recognized bottleneck in the field. The empirical package has genuine strengths: MSE results in Tables 3-5 are averaged over 100 trials with standard deviations, the stability criterion is defined transparently in Appendix A.2, the downsampling ratio is ablated in Table 2, and Table 6 quantifies the small overhead. However, the causal attribution of the stability gain to the hierarchical future-conditioning mechanism is not fully established, and key state-of-the-art baselines are missing. The central mechanism needs a direct diagnostic before the claims can be accepted.

major comments (4)
  1. [Section 4.2 (Eq. 10), Section 5.1 (Fig. 3b)] The central causal claim is not directly verified. Eq. (10) trains f_theta on true future latents z^(1)_{n+1}, z^(2)_{n+2}, while inference feeds back the network's own one-step-old predictions. The paper reports only state-level MSE and energy-based stability over 2e5 steps. If the predicted latents drift, the model could in effect fall back to an L=1 baseline while still passing the energy criterion. Please report a rollout diagnostic for latent informativeness, for example the MSE or correlation of predicted latents z-hat^(l)_{n+l} against T^(l)(u_{n+l}) along the autoregressive rollout, or an inference-time ablation that zeroes or perturbs the latent inputs. Without such evidence, the 93% stability rate cannot be causally attributed to the hierarchical future-conditioning mechanism.
  2. [Section 5.1 (Fig. 3b), Appendix A.2] The long-term stability metric is a binary threshold on total energy only. Energy is a one-dimensional aggregate statistic; a trajectory can satisfy the 5-standard-deviation energy bound while losing phase structure or small-scale fidelity. The zonal-mean and energy-spectrum diagnostics in Fig. 4 are informative but are reported for visual comparison and only for the spectrum error; no stability-rate analog based on these statistics is given over the 100 long trials. Please quantify spectrum and zonal-mean errors over all long rollouts, or define stability via multiple statistics, so that the 93% number reflects preservation of physical structure rather than a single aggregate.
  3. [Section 5 (Tables 4-5), Related Work] The paper cites PDE-Refiner [37] and Thermalizer [45] as related methods for improving rollout stability but does not compare against either. Since the contribution is framed as a significant advance in long-term stability, the absence of these strong baselines leaves open the possibility that existing stabilization techniques achieve comparable stability with different trade-offs. Please add these comparisons, or justify their exclusion with concrete reasons such as unavailable code or incompatible setups.
  4. [Appendix A.3, Section 5.1] The experimental details are insufficient to reproduce the headline 93% stability number. There is no code or data release, no random seed specification, no training/validation split description, and no statement of how the first L-1 rollout steps are initialized for the L=2 and L=3 models (what values are used for z^(1)_1 and z^(2)_2 before any predictions exist). This initialization affects the comparison at early rollout steps and may influence long-term stability. Please specify the rollout initialization, report the data split and training hyperparameters, and release code and data or provide a clear reproducibility statement.
minor comments (4)
  1. [Figure 1 caption] The caption says 'up to 10^5 emulation steps' while the main text and Figure 3 use 2e5 steps; please harmonize the notation.
  2. [Section 5.2 and Table 2] The text states r1 is ablated over [2, 4, 8, 32], but Table 2 reports r1 = 16 as well; please correct the inconsistent list.
  3. [Appendix A.3] The architecture paragraph contains an unresolved citation 'following the design in []'; a reference should be inserted.
  4. [Equation (12)] Please specify which distance d is used in the main experiments; the text mentions l1 or l2, while Appendix A.3 later refers to an l1 + l2 loss for the FNO comparison, leaving the main loss unclear.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the method is a supervised neural emulator evaluated on held-out rollouts, and no prediction reduces to a fitted input or self-citation chain.

full rationale

The paper's central claims are empirical: a UNet-based emulator is trained with Eq. (10) and loss Eq. (12), where true future latents z^(l)_{n+l} are used as conditioning inputs during training, and at inference the model feeds back its own predicted latents. This is a standard teacher-forcing then closed-loop setup, not a circular reduction: the reported short-term MSE (Fig. 3a, Table 4) and long-term stability rates (Fig. 3b) are computed on held-out initial conditions against a ground-truth trajectory or its invariant statistics. The downsampling ratios r1, r2 are selected by ablation on the same dataset (Table 2), which is hyperparameter tuning rather than fitting the headline result itself; the choice does not force any particular stability rate. The 5-standard-deviation energy threshold in Section A.2 is derived from the true dynamics' energy distribution, not from model outputs, so it is a metric definition rather than a self-fulfilling prediction. Self-citations in the paper ([2], [25], [34], [48]) appear as related work and as the data-generation tool (py2d); none is load-bearing for the claimed stability improvement, and no uniqueness theorem is imported from the authors' prior work. The implicit-solver analogy is explicitly motivational ('Drawing inspiration,' 'analogy with the implicit time-stepping methods') and is not used to derive a result that is equivalent to its input. The training-inference gap on latent feedback is a validity concern about whether the predicted latents remain informative, but it is not a circularity: the long-run stability claim is an empirical outcome that could in principle have failed. No equation in the paper reduces to another by construction, and no fitted parameter is renamed as a prediction. Therefore no circular step can be exhibited, and the honest finding is no significant circularity.

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

The method introduces no new physical entities. The latent future states are downsampled versions of the true state (Eq. 11), and the extra structure is entirely within the neural network.

free parameters (3)
  • downsampling rate r1 (L=2 and L=3) = 8 (L=3, 256x256 grid); 32 for L=2
    Chosen via ablation in Table 2; the value is selected by rollout MSE on the same experimental setup it is then evaluated on.
  • downsampling rate r2 (L=3) = 32
    Sets the second hierarchy level; no separate ablation shown, likely chosen together with r1.
  • stability threshold (5 standard deviations of energy) = 5
    Defines the binary stability rate that produces the 93% headline number; derived from the observation that training data lie within 4 standard deviations, but still a hand-chosen criterion.
assumptions (3)
  • domain assumption Energy is a sufficient statistic for long-term physical stability
    Section 5.1 defines stability purely by energy within 5 standard deviations; other statistics (zonal mean, spectrum) are shown separately but the rate itself is energy-only.
  • domain assumption The turbulent flow is statistically stationary and ergodic
    Long-term rollouts are compared to a fixed reference energy/spectrum under the assumption that the statistics do not depend on initial condition (Section A.2).
  • ad hoc to paper Implicit time-stepping stability theory carries over to learned updates with future latent conditioning
    The motivation in Section 4.1 is an analogy, not a proof; the architecture does not actually solve an implicit equation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Implicit Neural Emulators." pith.science (2026). https://pith.science/paper/5FYHVG3E

@misc{pith2026250604528,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Implicit Neural Emulators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FYHVG3E}},
  note         = {Machine review of arXiv:2506.04528}
}
read the original abstract

Neural PDE solvers offer a powerful tool for modeling complex dynamical systems, but often struggle with error accumulation over long time horizons and maintaining stability and physical consistency. We introduce a multiscale implicit neural emulator that enhances long-term prediction accuracy by conditioning on a hierarchy of lower-dimensional future state representations. Drawing inspiration from the stability properties of numerical implicit time-stepping methods, our approach leverages predictions several steps ahead in time at increasing compression rates for next-timestep refinements. By actively adjusting the temporal downsampling ratios, our design enables the model to capture dynamics across multiple granularities and enforce long-range temporal coherence. Experiments on turbulent fluid dynamics show that our method achieves high short-term accuracy and produces long-term stable forecasts, significantly outperforming autoregressive baselines while adding minimal computational overhead.

Figures

Figures reproduced from arXiv: 2506.04528 by the authors.

Figure 1
Figure 1. In a chaotic, turbulent system, our emulator achieves accurate short-to-mid-term rollout [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Diagram of our hierarchical implicit neural emulator. Our model approximates the dynamics u using an implicit formulation with hierarchical latent representations. For each trajectory un, we augment its state with multiscale latent features z (l) n , obtained by downsampling un. The model predicts the future state uˆn+1 and long-horizon latent representations zˆ (1) n+2, zˆ (2) n+3, enabling it to capture long-term … view at source ↗
Figure 3
Figure 3. Short-term accuracy vs. long-term robustness. Left: MSE trend over a 100-step autoregressive rollout. Solid lines indicate the average performance, and shaded regions represent one standard deviation. Right: We evaluate the stability rate over 10 times the training sequence length across 100 trials with various initial conditions. Stability is defined by maintaining physical statistics (conserved kinetic energy). Ou… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Further investigation of long-term stability of 2 × 105 steps (10× the training length) rollout. (a): Time-averaged zonal mean of vorticity comparing ground truth (dashed black lines) with emulator runs. Only our systems with L = 2 and L = 3 accurately follow the groun…
Figure 5
Figure 5. Figure 5: Ablation studies on the design of the hierarchy. We evaluate our system across different hierarchy levels L and visualize the results using: (a) PCA autocorrelation and (c) energy spectrum averaging over 200 rollout steps, with corresponding deviations from the ground …
Figure 6
Figure 6. Figure 6: Visualization of rollout estimation across multiple datasets. We apply our approach to three different flows: (1) Re = 104 , 256 × 256 resolution featuring zonal jets, (2) Re = 5 × 103 , 256 × 256 resolution without zonal jets, and (3) Re = 104 , 512 × 512 resolution w…
Figure 7
Figure 7. Figure 7: Energy of true dynamics along extended timescale. We compute the energy evolution of the true dynamics over 10 times the training dataset length (2 × 105 time steps). We show that energy along the sampled true dynamics remains within the ±5 standard deviation of its me…
Figure 8
Figure 8. Figure 8: We apply our approach to flow dataset of [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Visualization of prediction and residual to ground truth across methods and prediction steps. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Visualization of prediction and residual to ground truth across methods and prediction steps. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Visualization of prediction and residual to ground truth across methods and prediction steps. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Energy v.s. rollout time steps of the predicted dynamics. The dashed horizontal line shows the maximum and minimum values computed using 5 standard deviations from the mean of the true dynamics. Our L = 3 method is able to maintain energy along the long-term predictio…
Figure 13
Figure 13. Figure 13: Long-Term Rollout Visualization. Long-term dynamics corresponding to Trial 1 in [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Visualization of long-term rollout. We visualize the long-term dynamics corresponds to trial 2 in [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]
Figure 15
Figure 15. Figure 15: Energy v.s. rollout time steps of the predicted dynamics. The dashed horizontal line shows the maximum and minimum values computed using 5 standard deviations from the mean of the true dynamics. Our L = 3 method is able to maintain energy along the long-term predictio…
Figure 16
Figure 16. Figure 16: Visualization of long-term rollout. We visualize the long-term dynamics corresponds to trial 3 in [PITH_FULL_IMAGE:figures/full_fig_p025_16.png]
Figure 17
Figure 17. Figure 17: Visualization of long-term rollout. We visualize the long-term dynamics corresponds to trial 4 in [PITH_FULL_IMAGE:figures/full_fig_p026_17.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 30 canonical work pages

  1. [37]

    Pde- refiner: Achieving accurate long rollouts with neural pde solvers.Advances in Neural Information Processing Systems, 36:67398–67433, 2023

    Phillip Lippe, Bas Veeling, Paris Perdikaris, Richard Turner, and Johannes Brandstetter. Pde- refiner: Achieving accurate long rollouts with neural pde solvers.Advances in Neural Information Processing Systems, 36:67398–67433, 2023

    work page 2023

  2. [45]

    Thermalizer: Stable autoregressive neural emulation of spatiotemporal chaos.arXiv preprint arXiv:2503.18731, 2025

    Chris Pedersen, Laure Zanna, and Joan Bruna. Thermalizer: Stable autoregressive neural emulation of spatiotemporal chaos.arXiv preprint arXiv:2503.18731, 2025

    arXiv 2025

  3. [1]

    Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators.arXiv preprint arXiv:2202.11214, 2022

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopad- hyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators.arXiv preprint arXiv:2202.11214, 2022

    arXiv 2022

  4. [2]

    Lu, and Rebecca Willett

    Ruoxi Jiang, Peter Y. Lu, and Rebecca Willett. Embed and emulate: Contrastive representations for simulation-based inference.arXiv preprint arXiv:2409.18402, 2024

    arXiv 2024

  5. [3]

    Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    2023

  6. [4]

    Neural general circulation models for weather and climate.Nature, 632(8027):1060–1066, 2024

    Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Klöwer, James Lottes, Stephan Rasp, Peter Düben, et al. Neural general circulation models for weather and climate.Nature, 632(8027):1060–1066, 2024

    2024

  7. [5]

    Ace: A fast, skillful learned global atmospheric model for climate prediction.arXiv preprint arXiv:2310.02074, 2023

    Oliver Watt-Meyer, Gideon Dresdner, Jeremy McGibbon, Spencer K Clark, Brian Henn, James Dun- can, Noah D Brenowitz, Karthik Kashinath, Michael S Pritchard, Boris Bonev, et al. Ace: A fast, skillful learned global atmospheric model for climate prediction.arXiv preprint arXiv:2310.02074, 2023

    arXiv

  8. [6]

    Owen, Mordechai Kornbluth, and Boris Kozinsky

    Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics.Nature Communications, 14(1):579, Feb 2023. ISSN 2041-1723. doi:10.1038/s41467-023- 36329-y

    doi:10.1038/s41467-023- 2023

Show all 52 references
  1. [7]

    An overview about neural networks potentials in molecular dynamics simulation.International Journal of Quantum Chemistry, 124(11):e27389, 2024

    RaidelMartin-Barrios, EdiselNavas-Conyedo, XuyiZhang, YunweiChen, andJorgeGulín-González. An overview about neural networks potentials in molecular dynamics simulation.International Journal of Quantum Chemistry, 124(11):e27389, 2024

    2024

  2. [8]

    Neural network and nearest neighbor algorithms for enhancing sampling of molecular dynamics.Journal of chemical theory and computation, 13(6):2489–2500, 2017

    Raimondas Galvelis and Yuji Sugita. Neural network and nearest neighbor algorithms for enhancing sampling of molecular dynamics.Journal of chemical theory and computation, 13(6):2489–2500, 2017

    2017

  3. [9]

    Field-level neural network emulator for cosmological n-body simulations.The Astrophysical Journal, 952(2):145, 2023

    Drew Jamieson, Yin Li, Renan Alves de Oliveira, Francisco Villaescusa-Navarro, Shirley Ho, and David N Spergel. Field-level neural network emulator for cosmological n-body simulations.The Astrophysical Journal, 952(2):145, 2023

    2023

  4. [10]

    Simple lessons from complex learning: what a neural network model learns about cosmic structure formation.PNAS nexus, 2(4):pgac250, 2023

    DrewJamieson, YinLi, SiyuHe, FranciscoVillaescusa-Navarro, ShirleyHo, RenanAlvesdeOliveira, and David N Spergel. Simple lessons from complex learning: what a neural network model learns about cosmic structure formation.PNAS nexus, 2(4):pgac250, 2023

    2023

  5. [11]

    Learning to predict the cosmological structure formation.Proceedings of the National Academy of Sciences, 116(28):13825–13832, 2019

    Siyu He, Yin Li, Yu Feng, Shirley Ho, Siamak Ravanbakhsh, Wei Chen, and Barnabás Póczos. Learning to predict the cosmological structure formation.Proceedings of the National Academy of Sciences, 116(28):13825–13832, 2019

    2019

  6. [12]

    When are dynamical systems learned from time series data statistically accurate?arXiv preprint arXiv:2411.06311, 2024

    Jeongjin Park, Nicole Yang, and Nisha Chandramoorthy. When are dynamical systems learned from time series data statistically accurate?arXiv preprint arXiv:2411.06311, 2024. 10

    2024 arXiv

  7. [13]

    On some limitations of current machine learning weather prediction models

    Massimo Bonavita. On some limitations of current machine learning weather prediction models. Geophysical Research Letters, 51(12):e2023GL107377, 2024

    2024

  8. [14]

    Qiang Sun, and Pedram Hassanzadeh

    Ashesh Chattopadhyay, Y. Qiang Sun, and Pedram Hassanzadeh. Challenges of learning multi- scale dynamics with ai weather models: Implications for stability and one solution, 2024. URL https://arxiv.org/abs/2304.07029

    2024 arXiv

  9. [15]

    Machine learning for climate physics and simulations.Annual Review of Condensed Matter Physics, 16, 2024

    Ching-Yao Lai, Pedram Hassanzadeh, Aditi Sheshadri, Maike Sonnewald, Raffaele Ferrari, and Venkatramani Balaji. Machine learning for climate physics and simulations.Annual Review of Condensed Matter Physics, 16, 2024

    2024

  10. [16]

    A comparison of data-driven approaches to build low-dimensional ocean models.Journal of Advances in Modeling Earth Systems, 13(9):e2021MS002537, 2021

    Niraj Agarwal, Dmitri Kondrashov, Peter Dueben, Eugene Ryzhov, and Pavel Berloff. A comparison of data-driven approaches to build low-dimensional ocean models.Journal of Advances in Modeling Earth Systems, 13(9):e2021MS002537, 2021

    2021

  11. [17]

    Gupta and Johannes Brandstetter

    Jayesh K. Gupta and Johannes Brandstetter. Towards multi-spatiotemporal-scale generalized pde modeling, 2022. URLhttps://arxiv.org/abs/2209.15616

    2022 arXiv

  12. [18]

    Message passing neural pde solvers

    Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers. arXiv preprint arXiv:2202.03376, 2022

    2022 arXiv

  13. [19]

    Pdebench: An extensive benchmark for scientific machine learning,

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Dan MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning,

  14. [20]

    Dyffusion: A dynamics-informed diffusion model for spatiotemporal forecasting.Advances in neural information processing systems, 36:45259–45287, 2023

    Salva Rühling Cachay, Bo Zhao, Hailey Joren, and Rose Yu. Dyffusion: A dynamics-informed diffusion model for spatiotemporal forecasting.Advances in neural information processing systems, 36:45259–45287, 2023

    2023

  15. [21]

    Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  16. [22]

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

    Tapas Tripura and Souvik Chakraborty. Wavelet neural operator: a neural operator for parametric partial differential equations.arXiv preprint arXiv:2205.02191, 2022

    2022 arXiv

  17. [23]

    Wavelet diffusion neural operator.arXiv preprint arXiv:2412.04833, 2024

    Peiyan Hu, Rui Wang, Xiang Zheng, Tao Zhang, Haodong Feng, Ruiqi Feng, Long Wei, Yue Wang, Zhi-Ming Ma, and Tailin Wu. Wavelet diffusion neural operator.arXiv preprint arXiv:2412.04833, 2024

    2024 arXiv

  18. [24]

    Learning chaotic dynamics in dissipative systems

    Zongyi Li, Miguel Liu-Schiaffini, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Learning chaotic dynamics in dissipative systems. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho,...

    2022

  19. [25]

    Lu, Elena Orlova, and Rebecca Willett

    Ruoxi Jiang, Peter Y. Lu, Elena Orlova, and Rebecca Willett. Training neural operators to preserve invariant measures of chaotic attractors.Advances in Neural Information Processing Systems, 36, 2024

    2024

  20. [26]

    Unconditional convergence of the euler semi-implicit scheme for the three-dimensional incompressible mhd equations.IMA Journal of Numerical Analysis, 35(2):767–801, 2015

    Yinnian He. Unconditional convergence of the euler semi-implicit scheme for the three-dimensional incompressible mhd equations.IMA Journal of Numerical Analysis, 35(2):767–801, 2015

    2015

  21. [27]

    U-Net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. InMICCAI, 2015

    2015

  22. [28]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017

    2017

  23. [29]

    Multiscale vision transformers

    Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. InIEEE/CVF International Conference on Computer Vision, 2021. 11

    2021

  24. [30]

    Progressive multigrid eigensolvers for multiscale spectral segmenta- tion

    Michael Maire and Stella X Yu. Progressive multigrid eigensolvers for multiscale spectral segmenta- tion. InProceedings of the IEEE International Conference on Computer Vision, pages 2184–2191, 2013

    2013

  25. [31]

    Multigrid neural architectures

    Tsung-Wei Ke, Michael Maire, and Stella X Yu. Multigrid neural architectures. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6665–6673, 2017

    2017

  26. [32]

    Tri Huynh, Michael Maire, and Matthew R. Walter. Multigrid neural memory. InProceedings of the 37th International Conference on Machine Learning, 2020

    2020

  27. [33]

    NVAE: A deep hierarchical variational autoencoder.Advances in neural information processing systems, 33:19667–19679, 2020

    Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder.Advances in neural information processing systems, 33:19667–19679, 2020

    2020

  28. [34]

    Nested diffusion models using hierarchical latent priors.arXiv preprint arXiv:2412.05984, 2024

    Xiao Zhang, Ruoxi Jiang, Rebecca Willett, and Michael Maire. Nested diffusion models using hierarchical latent priors.arXiv preprint arXiv:2412.05984, 2024

    2024 arXiv

  29. [35]

    Gencast: Diffusion- based ensemble forecasting for medium-range weather.arXiv preprint arXiv:2312.15796, 2023

    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Gencast: Diffusion- based ensemble forecasting for medium-range weather.arXiv preprint arXiv:2312.15796, 2023

    2023 arXiv

  30. [36]

    Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators.arXiv preprint arXiv:1910.03193, 2019

    Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators.arXiv preprint arXiv:1910.03193, 2019

    1910 arXiv

  31. [38]

    Multipdenet: Pde-embedded learning with multi-time-stepping for accelerated flow simulation.arXiv preprint arXiv:2501.15987, 2025

    Qi Wang, Yuan Mi, Haoyun Wang, Yi Zhang, Ruizhi Chengze, Hongsheng Liu, Ji-Rong Wen, and Hao Sun. Multipdenet: Pde-embedded learning with multi-time-stepping for accelerated flow simulation.arXiv preprint arXiv:2501.15987, 2025

    2025 arXiv

  32. [39]

    Implicit-explicit methods for time-dependent partial differential equations.SIAM Journal on Numerical Analysis, 32(3):797–823, 1995

    Uri M Ascher, Steven J Ruuth, and Brian TR Wetton. Implicit-explicit methods for time-dependent partial differential equations.SIAM Journal on Numerical Analysis, 32(3):797–823, 1995

    1995

  33. [40]

    Strong stability-preserving high-order time discretization methods.SIAM review, 43(1):89–112, 2001

    Sigal Gottlieb, Chi-Wang Shu, and Eitan Tadmor. Strong stability-preserving high-order time discretization methods.SIAM review, 43(1):89–112, 2001

    2001

  34. [41]

    On the stability of implicit-explicit linear multistep methods.Applied Numerical Mathematics, 25(2-3):193–205, 1997

    Jason Frank, Willem Hundsdorfer, and Jan G Verwer. On the stability of implicit-explicit linear multistep methods.Applied Numerical Mathematics, 25(2-3):193–205, 1997

    1997

  35. [42]

    Courier Corporation, 2012

    Eugene Isaacson and Herbert Bishop Keller.Analysis of numerical methods. Courier Corporation, 2012

    2012

  36. [43]

    Learning closed-form equations for subgrid-scale closures from high-fidelity data: Promises and challenges.Journal of Advances in Modeling Earth Systems, 16(7):e2023MS003874, 2024

    Karan Jakhar, Yifei Guan, Rambod Mojgani, Ashesh Chattopadhyay, and Pedram Hassanzadeh. Learning closed-form equations for subgrid-scale closures from high-fidelity data: Promises and challenges.Journal of Advances in Modeling Earth Systems, 16(7):e2023MS003874, 2024

    2024

  37. [44]

    Yifei Guan, Ashesh Chattopadhyay, Adam Subel, and Pedram Hassanzadeh. Stable a posteriori les of 2d turbulence using convolutional neural networks: Backscattering analysis and generalization to higher re via transfer learning.Journal of Computational Physics, 458:111090, 2022

    2022

  38. [46]

    Subgrid modelling for two- dimensional turbulence using neural networks.Journal of Fluid Mechanics, 858:122–144, 2019

    Romit Maulik, Omer San, Adil Rasheed, and Prakash Vedula. Subgrid modelling for two- dimensional turbulence using neural networks.Journal of Fluid Mechanics, 858:122–144, 2019

    2019

  39. [47]

    Cambridge University Press, 2017

    Geoffrey K Vallis.Atmospheric and oceanic fluid dynamics. Cambridge University Press, 2017

    2017

  40. [48]

    py2d: High-performance 2D Navier-Stokes solver in python (version 0.1), 2024

    Karan Jakhar, Rambod Mojgani, Moein Darman, Yifei Guan, and Pedram Hassanzadeh. py2d: High-performance 2D Navier-Stokes solver in python (version 0.1), 2024. URLhttps://github. com/envfluids/py2d. 12

    2024

  41. [49]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021

    2021

  42. [50]

    Turbulent flows.Measurement Science and Technology, 12(11):2020–2021, 2001

    Stephen B Pope. Turbulent flows.Measurement Science and Technology, 12(11):2020–2021, 2001

    2020

  43. [51]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede liu, Kaushik Bhat- tacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. InInternational Conference on Learning Representations, 2021. URL ht...

    2021

  44. [2024]

    URLhttps://arxiv.org/abs/2210.07182

    arXiv

Pith tools

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