Pith. sign in

REVIEW 4 major objections 5 minor 85 references

Hierarchical-embedding autoencoder with a predictor (HEAP) as efficient architecture for learning long-term evolution of complex multi-scale physical systems

T0 review · 4 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 The paper claims that representing a physical state as a hierarchy of scale-specific embeddings, advanced in sync by a Markovian predictor, yields multifold improvements in long-term statistical accuracy for Hasegawa-Wakatani turbulence…

desk verdict Plausible new architecture with a genuine empirical win, but the submission withholds the details needed to verify the mechanism; the hierarchy claim is promising, not proven. read the letter →

arxiv 2505.18857 v1 pith:WAESFBKW submitted 2025-05-24 cs.AI physics.plasm-ph

Alexander Khrabry , Edward Startsev , Andrew Powis , Igor Kaganovich This is my paper
classification cs.AIphysics.plasm-ph
keywords hierarchicalembeddingsmulti-scalephysicalsystemsplasmaturbulenceHasegawa-WakataniautoencodersurrogatemodelingResNetlong-termrollouts
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 argues that multi-scale physical systems can be predicted far more accurately and efficiently by encoding the state as a stack of embeddings, one per scale, and advancing all of them in sync, rather than compressing everything into a single embedding as conventional autoencoder-plus-ResNet surrogates do. It introduces HEAP, a hierarchical fully-convolutional autoencoder that produces these layered embeddings and a predictor that couples neighboring layers with convolutional operators. Tested on Hasegawa-Wakatani plasma turbulence, models with three embedding layers improve long-term prediction of statistical quantities—spatial and temporal spectra, autocorrelations, and vortex counts—by a factor of three or more over flat baselines, with only modestly higher training cost. The paper's central thesis is that the separation-of-scales principle, where structures interact only locally across scales, is sufficient to capture the cascade dynamics.

What carries the argument

The machinery is a hierarchical fully-convolutional autoencoder (HAE) paired with a layered ResNet-style predictor. The encoder applies strided convolutions at each level: one convolution diverts information into an embedding layer for that scale, while another passes information down to coarser scales, giving each embedding layer a small channel count on an ever-coarser grid (for 128x128 input, layers at 32x32, 16x16, and so on). The predictor advances all embedding layers synchronously: each internal step is a block combining convolutions within each layer and between adjacent layers, with skip connections and layer normalization. The defining assumption is the separation-of-scales locality: structures of similar scale interact when in contact, and different scales interact only when the smaller structure is part of the larger one, so distant small features do not need direct interaction modeling.

What would settle it

Train HEAP and a flat ResNet baseline on a 2D incompressible fluid system, where pressure couples distant regions instantaneously, using the same data volume and compute budget. If the multifold improvement in long-term spectral and autocorrelation accuracy shrinks to parity, the locality assumption is falsified for that class of systems; a more direct test would inject a small localized perturbation that is known to affect a distant small-scale structure within one time step and measure whether HEAP's predictor can reproduce that causal chain.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that separating the state representation by physical scale, not just by channel depth, is the key to accurate long-term forecasting of multi-scale turbulence. Hierarchical embeddings preserve spatial resolution at each level, and a Markovian predictor that advances all levels in sync, using convolutions both within and between levels, reproduces the long-time statistics of Hasegawa-Wakatani turbulence over 3000 rollout steps. The deeper claim is that this architecture outperforms flat ResNet-style surrogate models by a factor of three or more across spectral, autocorrelation, and vortex-count metrics, while using similar or at most twice the training compute; a single-embedding flat model performs worst.

Load-bearing premise

The load-bearing premise is that structures of different scales interact only locally: a smaller structure matters only when it is inside or touching a larger structure, so distant small features never need to communicate directly.

Editorial extensions

If this is right

  • If the claim holds, surrogate models for multi-scale turbulence can become both more accurate and more data-efficient by explicitly representing scale hierarchy, rather than by increasing model depth or using global attention.
  • Long-term rollout statistics such as spectra, autocorrelations, and vortex counts become practical training and selection criteria, since the architecture preserves them over thousands of steps.
  • The same design extends to 3D systems with 3D convolutions and to unstructured grids via graph convolutions, as the paper notes.
  • The number of embedding layers is a hyperparameter tied to the system's dominant scale range; for Hasegawa-Wakatani turbulence, three levels suffice, and deeper models provide no further improvement.
  • Because the predictor is Markovian and all physical fields are fed in, no recurrent memory is needed for systems with first-order-in-time dynamics.

Reading between the lines

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

  • The locality principle suggests HEAP should excel on systems with a clear scale cascade but may lose its advantage on systems with direct long-range coupling between small structures, such as incompressible flows where pressure propagates instantaneously; this is directly testable.
  • The plateau in performance beyond three embedding layers implies that only a limited range of active scales dominates Hasegawa-Wakatani turbulence; the embedding hierarchy could serve as a diagnostic for identifying the active scale range in other systems.
  • If correct, the approach implies that the bottleneck for turbulence surrogates is not predictor capacity but the scale decomposition of the latent representation; similar hierarchical latent structures could be added to transformer-based or Fourier-based operators.
  • Because exact trajectory matching is impossible for chaotic systems, the paper's statistical evaluation suggests a practical extension: calibrate the ensemble spread of independently trained predictors to provide forecast uncertainty.
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 / 5 minor

Summary. The paper proposes HEAP, a hierarchical fully-convolutional autoencoder with a predictor, for surrogate modeling of multi-scale physical systems. The architecture encodes a state into multiple embedding layers at different spatial resolutions, with the predictor advancing all embedding layers in synchrony under the assumption that scale interactions are local. The method is evaluated on Hasegawa-Wakatani turbulence generated by the independent BOUT++ solver, comparing hierarchical models H1-H5 with flat fully-convolutional models C1-C3 on long autoregressive rollouts. The authors report a multifold improvement in the accuracy of statistical characteristics such as spatial and temporal Fourier spectra, temporal autocorrelations, and the number of local extrema in the potential field. Implementation details, computational-complexity analysis, and ablation results are advertised but withheld in the appendices.

Significance. If the claimed improvement is indeed attributable to the hierarchical embedding structure, this would be a useful contribution to data-driven surrogate modeling of plasma turbulence and related multi-scale systems. The evaluation is physically meaningful: metrics are computed against held-out simulation statistics, and the benchmark data come from an independent solver. The authors are transparent about their seed-selection protocol, which is a strength. However, the central claim that the hierarchical architecture, rather than additional model capacity or favorable seed selection, causes the improvement is not established by the reported experiments. The missing controlled comparisons and withheld implementation details are load-bearing for the paper's main conclusion.

major comments (4)
  1. [§4.1, §4.3] The headline improvement is confounded with model capacity. H3 has three 8-channel embedding layers, giving 32×32×8 + 16×16×8 + 8×8×8 = 10,752 embedding units, whereas H1=C1 has 32×32×8 = 8,192 units and an 8-channel predictor input; H3's predictor operates on 24 input channels. The factor-of-3-or-more error reduction reported in §4.3 is therefore not shown to be caused by the hierarchical structure rather than by the larger embedding and predictor capacity. The C2/C3 baselines in §4.1 and Appendix C do not control for this: they keep total embedding units at 8,192 by increasing channel counts (32 and 128 respectively), so they neither match H3's total embedding units nor its channel distribution, and their results are reported only for the two best realizations. Please add parameter-matched or FLOP-matched flat baselines (for example, a single-layer embedding with the same total number of embedding units and predictor parameters as H3), report parameter counts and inference FLOPs, and report the corresponding roll-out errors.
  2. [§4.3] The evaluation selects realizations on the basis of the outcome metrics: sixteen seeds are trained, realizations are ranked by their worst-performing metric, and best/mean/std are reported for the top six. This is a best-case comparison, and the 'multifold improvement' may reflect favorable initialization for H3 rather than the architecture. The 10%-20% of non-converging realizations are excluded without reporting whether their frequency differs across models. Please report the full distribution over all sixteen seeds (or at least medians and interquartile ranges), state the non-convergence count per model, and justify that the top-six selection rule does not bias the comparison in favor of the hierarchical models.
  3. [Appendix B.5, B.1, B.3, D.1] The manuscript withholds the information needed to assess both the architecture and the central efficiency claim. Appendix B.5, which should contain the computational-complexity analysis supporting 'scales linearly with the system's size' in §3, is empty; Appendices B.1 and B.3 provide no layer details; and Appendix D.1, the ablation claimed to show that cross-layer interactions are crucial, contains only a caption instructing readers to contact the authors. No parameter counts or wall-clock/inference-cost measurements appear anywhere in the paper, and Fig. 7 shows hierarchical predictor training to be about twice as expensive as H1. Please provide the full implementation details, the complexity derivation, parameter counts, FLOPs, and the complete ablation results, including the cross-layer-interaction ablation.
  4. [§3, §4] The paper's motivating assumption in §3 is that structures of different scales interact only locally and that small-scale features far apart need not interact directly. This is a strong physical assumption, and the single Hasegawa-Wakatani test case (α=0.01, κ=0.5) does not probe its limits; the authors explicitly acknowledge in the Introduction that incompressible fluids have infinite-speed information propagation, for which the efficiency gain may disappear. Since the title and abstract claim a general architecture for complex multi-scale systems, the scope of validity should be substantiated or clearly narrowed. At minimum, please report the effective receptive field or test on a second system with nonlocal coupling, or revise the claims to be explicitly about systems satisfying the locality assumption.
minor comments (5)
  1. [§2] The text refers to 'Hasegawa-Wakatani (WH) turbulence'; the abbreviation should be HW.
  2. [Appendix C] The text says ten realizations were trained for the C1 model with N=8 and N=14, but then refers to a deep predictor 'N=12' as an exception; the figure caption also lists N=12. Please reconcile the values.
  3. [Fig. 4 caption] The caption reads 'c denotes of the number of embedding channels'; this should be 'c denotes the number of embedding channels.'
  4. [References] Reference [1] is listed as 'Under review. 2025.' with no title or author information; this placeholder should be replaced with a full citation or removed.
  5. [Appendix A] There is a typo: 'the field confinig the plasma' should be 'confining.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the HEAP advantage claim is an empirical benchmark against independent BOUT++ simulations, not a quantity fitted or defined into the model.

full rationale

The derivation is self-contained. The paper's central claim—that hierarchical embeddings improve long-term statistical prediction—is established by rolling out trained models for 3000 time steps and comparing output FFT spectra, autocorrelations, and extrema counts to BOUT++ reference data. The predictor is trained on one-step embedding transitions and the autoencoder on reconstruction; none of the evaluation metrics (spatial/temporal spectra, autocorrelation, extrema statistics) enters the training objective, so the reported 'predictions' are not fitted to the metrics used for judging them. H3 is not defined as 'the model that wins'; it is one of five hierarchical depths compared by measured error, and the reported improvement is an empirical result rather than a consequence of definition. There are no load-bearing self-citations or imported uniqueness theorems: reference [1] is an anonymous 'Under review' item and is not used to justify the architecture. Two non-circular transparency gaps should be flagged because the review instructions require it: Appendix B.5, cited for 'Computational complexity in this model scales linearly with the system's size,' is empty, and Appendices B.1, B.3, and D.1 explicitly withhold implementation details until after review. In addition, H3 has more embedding units than H1 and the paper's 'top 6 of 16 seeds' selection can inflate apparent gains. These concerns bear on soundness, reproducibility, and whether extra capacity rather than the hierarchical design drives the improvement, but they are not circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the local-interaction hypothesis, a Markovian state assumption, and the representativeness of a single HW turbulence case. None of these are derived or tested beyond the one benchmark.

free parameters (4)
  • Number of hierarchical embedding layers = 3 chosen as best of H1-H5
    The number of layers is tuned on rollout performance; deeper models H4/H5 do not improve, so layer count is a selected hyperparameter.
  • Embedding channels per layer = 8
    Fixed hand choice for all models; affects model capacity and the balance between layers.
  • Predictor internal steps N = 4
    Chosen default; tests with N=8 and N=14 are reported only for the baseline C1, not for HEAP.
  • Training hyperparameters = not stated
    Learning rate, batch size, epochs, normalization, and regularization details are withheld in Appendix B.
assumptions (6)
  • domain assumption Small-scale features far apart do not need to be modeled as interacting directly.
    Section 3, first paragraph: this is the physical premise behind using local convolutions on each hierarchical layer; it is load-bearing and untested outside the HW case.
  • domain assumption A single time-step state is sufficient to predict all future states.
    Section 3: 'Since the equations feature only first order derivatives in time, a single state of the system is sufficient to predict all future states.' Removes any need for recurrence.
  • domain assumption Statistical aggregates (spectra, autocorrelations, extrema counts) are sufficient validation of long-term evolution.
    Section 4 compares only time-averaged statistical characteristics, not pointwise trajectory accuracy, because HW turbulence is chaotic.
  • domain assumption Performance on HW turbulence transfers to other continuous multi-scale systems.
    Introduction states expected applicability to other systems; no experiment outside HW is provided.
  • domain assumption BOUT++ simulation is accurate ground truth.
    No grid convergence or solver accuracy study is reported; BOUT++ output is treated as ground truth.
  • domain assumption Convolutional kernels with shared weights are sufficient to capture cross-scale interactions.
    The predictor uses a combination of convolutional operators inside each internal step; if cross-scale interactions require nonlocal operators, the architecture's efficiency claim weakens.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical-embedding autoencoder with a predictor (HEAP) as efficient architecture for learning long-term evolution of complex multi-scale physical systems." pith.science (2026). https://pith.science/paper/WAESFBKW

@misc{pith2026250518857,
  author       = {Pith},
  title        = {Pith review of: Hierarchical-embedding autoencoder with a predictor (HEAP) as efficient architecture for learning long-term evolution of complex multi-scale physical systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WAESFBKW}},
  note         = {Machine review of arXiv:2505.18857}
}
read the original abstract

We propose a novel efficient architecture for learning long-term evolution in complex multi-scale physical systems which is based on the idea of separation of scales. Structures of various scales that dynamically emerge in the system interact with each other only locally. Structures of similar scale can interact directly when they are in contact and indirectly when they are parts of larger structures that interact directly. This enables modeling a multi-scale system in an efficient way, where interactions between small-scale features that are apart from each other do not need to be modeled. The hierarchical fully-convolutional autoencoder transforms the state of a physical system not just into a single embedding layer, as it is done conventionally, but into a series of embedding layers which encode structures of various scales preserving spatial information at a corresponding resolution level. Shallower layers embed smaller structures on a finer grid, while deeper layers embed larger structures on a coarser grid. The predictor advances all embedding layers in sync. Interactions between features of various scales are modeled using a combination of convolutional operators. We compare the performance of our model to variations of a conventional ResNet architecture in application to the Hasegawa-Wakatani turbulence. A multifold improvement in long-term prediction accuracy was observed for crucial statistical characteristics of this system.

Figures

Figures reproduced from arXiv: 2505.18857 by the authors.

Figure 1
Figure 1. Various architectures of convolutional autoencoders. Blue rectangles represent encoded [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Multi-scale nature of structures encoded by various encoder layers. Deeper layers encode [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Predictors to couple with a conventional fully-convolutional AE (a) and hierarchical AE (b) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Variants of the models tested; c denotes of the number of embedding channels. Each [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparing spatial spectra of recovered n and ϕ fields with various autoencoders (a). Nor￾malized overall deviation from the ground truth (b), and training time (c). The errors in the FFT spectra of recovered fields by all models are rather small. There is no need to im…
Figure 6
Figure 6. Figure 6: Comparing capabilities of the hierarchical models H1-H5 in long-sequence rollouts. Spatial [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparing number of local minima and maxima in the field of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Details of the model will be provided upon completion of the paper review. In the meantime, [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Details of the model will be provided upon completion of the paper review. In the meantime, [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Comparing long sequence prediction errors and predictor training time of fully convolu [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Details of the model will be made available once the paper review is complete. For further [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Long-term predictions of HW turbulence generated by the H3 model (best realization). [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Long-term predictions of HW turbulence generated by the H3 model (best realization). [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Comparing input and otput (recovered) fileds of [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 44 canonical work pages

  1. [2]

    Universal physics transformers: A framework for efficiently scaling neural operators.Advances in Neural Information Processing Systems, 37:25152–25194, 2024

    Benedikt Alkin, Andreas Fürst, Simon Schmid, Lukas Gruber, Markus Holzleitner, and Johannes Brandstetter. Universal physics transformers: A framework for efficiently scaling neural operators.Advances in Neural Information Processing Systems, 37:25152–25194, 2024

    2024

  2. [3]

    Relational inductive biases, deep learning, and graph networks.arXiv preprint arXiv:1806.01261, 2018

    Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks.arXiv preprint arXiv:1806.01261, 2018

    arXiv 2018

  3. [4]

    Deep neural networks for data-driven les closure models.Journal of Computational Physics, 398:108910, 2019

    Andrea Beck, David Flad, and Claus-Dieter Munz. Deep neural networks for data-driven les closure models.Journal of Computational Physics, 398:108910, 2019

    2019

  4. [5]

    Combining differentiable pde solvers and graph neural networks for fluid flow prediction

    Filipe De Avila Belbute-Peres, Thomas Economon, and Zico Kolter. Combining differentiable pde solvers and graph neural networks for fluid flow prediction. Ininternational conference on machine learning, pages 2402–2411. PMLR, 2020

    2020

  5. [6]

    Prediction of aerodynamic flow fields using convolutional neural networks.Computational Mechanics, 64:525–545, 2019

    Saakaar Bhatnagar, Yaser Afshar, Shaowu Pan, Karthik Duraisamy, and Shailendra Kaushik. Prediction of aerodynamic flow fields using convolutional neural networks.Computational Mechanics, 64:525–545, 2019

    2019

  6. [7]

    Worrall, and Max Welling

    Johannes Brandstetter, Daniel E. Worrall, and Max Welling. Message passing neural PDE solvers. InInternational Conference on Learning Representations, 2022. URL https:// openreview.net/forum?id=vSix3HPYKSU

    2022

  7. [8]

    Clifford neural layers for PDE modeling

    Johannes Brandstetter, Rianne van den Berg, Max Welling, and Jayesh K Gupta. Clifford neural layers for PDE modeling. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum?id=okwxL_c4x84

    work page 2023

  8. [9]

    Climformer-a spherical transformer model for long-term climate projections

    Salva Rühling Cachay, Peetak Mitra, Haruki Hirasawa, Sookyung Kim, Subhashis Hazarika, Dipti Hingmire, Phil Rasch, Hansi Singh, and Kalai Ramea. Climformer-a spherical transformer model for long-term climate projections. InProceedings of the Machine Learning and the Phys-ical Sciences Workshop, NeurIPS 2022, 2022

    work page 2022

Show all 85 references
  1. [10]

    Choose a transformer: Fourier or galerkin.Advances in neural information processing systems, 34:24924–24940, 2021

    Shuhao Cao. Choose a transformer: Fourier or galerkin.Advances in neural information processing systems, 34:24924–24940, 2021. 9

    2021

  2. [11]

    Deep spatial transformers for autoregressive data-driven forecasting of geophysical turbulence

    Ashesh Chattopadhyay, Mustafa Mustafa, Pedram Hassanzadeh, and Karthik Kashinath. Deep spatial transformers for autoregressive data-driven forecasting of geophysical turbulence. In Proceedings of the 10th international conference on climate informatics, pages 106–112, 2020

    2020

  3. [12]

    Deep learning method based on physics informed neural network with resnet block for solving fluid flow problems.Water, 13(4):423, 2021

    Chen Cheng and Guang-Tao Zhang. Deep learning method based on physics informed neural network with resnet block for solving fluid flow problems.Water, 13(4):423, 2021

    2021

  4. [13]

    Generative-machine-learning surrogate model of plasma turbulence.Physical Review E, 111(1):L013202, 2025

    B Clavier, D Zarzoso, D del Castillo-Negrete, and E Frénod. Generative-machine-learning surrogate model of plasma turbulence.Physical Review E, 111(1):L013202, 2025

    2025

  5. [14]

    Comparing different nonlinear dimen- sionality reduction techniques for data-driven unsteady fluid flow modeling.Physics of Fluids, 34(11), 2022

    Hunor Csala, Scott Dawson, and Amirhossein Arzani. Comparing different nonlinear dimen- sionality reduction techniques for data-driven unsteady fluid flow modeling.Physics of Fluids, 34(11), 2022

    2022

  6. [15]

    Magnet: A graph u-net architecture for mesh-based simulations.Engineering Applications of Artificial Intelligence, 133:108055, 2024

    Saurabh Deshpande, Stéphane PA Bordas, and Jakub Lengiewicz. Magnet: A graph u-net architecture for mesh-based simulations.Engineering Applications of Artificial Intelligence, 133:108055, 2024

    2024

  7. [16]

    Physics-informed neural networks as surrogate models of hydrodynamic simulators.Science of the Total Environment, 912:168814, 2024

    James Donnelly, Alireza Daneshkhah, and Soroush Abolfathi. Physics-informed neural networks as surrogate models of hydrodynamic simulators.Science of the Total Environment, 912:168814, 2024

    2024

  8. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:...

    2010 arXiv

  9. [18]

    Dudson, M.V

    B.D. Dudson, M.V . Umansky, X.Q. Xu, P.B. Snyder, and H.R. Wilson. Bout++: A framework for parallel plasma fluid simulations.Computer Physics Communications, 180(9):1467–1480, September 2009. ISSN 0010-4655. doi: 10.1016/j.cpc.2009.03.008. URL http://dx.doi. org/10.1016/j.cpc....

    2009 doi

  10. [19]

    Deep encoder–decoder hierarchical convolutional neural networks for conjugate heat transfer surro- gate modeling.Applied Energy, 372:123723, 2024

    Takiah Ebbs-Picken, David A Romero, Carlos M Da Silva, and Cristina H Amon. Deep encoder–decoder hierarchical convolutional neural networks for conjugate heat transfer surro- gate modeling.Applied Energy, 372:123723, 2024

    2024

  11. [20]

    Deep neural networks for nonlinear model order reduction of unsteady flows.Physics of Fluids, 32(10), 2020

    Hamidreza Eivazi, Hadi Veisi, Mohammad Hossein Naderi, and Vahid Esfahanian. Deep neural networks for nonlinear model order reduction of unsteady flows.Physics of Fluids, 32(10), 2020

    2020

  12. [21]

    Physics-informed neural networks for solving reynolds-averaged navier–stokes equations.Physics of Fluids, 34 (7), 2022

    Hamidreza Eivazi, Mojtaba Tahani, Philipp Schlatter, and Ricardo Vinuesa. Physics-informed neural networks for solving reynolds-averaged navier–stokes equations.Physics of Fluids, 34 (7), 2022

    2022

  13. [22]

    Multiscale meshgraphnets.arXiv preprint arXiv:2210.00612, 2022

    Meire Fortunato, Tobias Pfaff, Peter Wirnsberger, Alexander Pritzel, and Peter Battaglia. Multiscale meshgraphnets.arXiv preprint arXiv:2210.00612, 2022

    2022 arXiv

  14. [23]

    Scientific machine learning based reduced-order models for plasma turbulence simulations.Physics of Plasmas, 31(11), 2024

    Constantin Gahr, Ionu¸ t-Gabriel Farca¸ s, and Frank Jenko. Scientific machine learning based reduced-order models for plasma turbulence simulations.Physics of Plasmas, 31(11), 2024

    2024

  15. [24]

    Earthformer: Exploring space-time transformers for earth system forecasting.Advances in Neural Information Processing Systems, 35:25390–25403, 2022

    Zhihan Gao, Xingjian Shi, Hao Wang, Yi Zhu, Yuyang Bernie Wang, Mu Li, and Dit-Yan Yeung. Earthformer: Exploring space-time transformers for earth system forecasting.Advances in Neural Information Processing Systems, 35:25390–25403, 2022

    2022

  16. [25]

    Mesh-based gnn surrogates for time-independent pdes.Scientific reports, 14(1):3394, 2024

    Rini Jasmine Gladstone, Helia Rahmani, Vishvas Suryakumar, Hadi Meidani, Marta D’Elia, and Ahmad Zareei. Mesh-based gnn surrogates for time-independent pdes.Scientific reports, 14(1):3394, 2024

    2024

  17. [26]

    Physics-preserving ai-accelerated simulations of plasma turbulence.arXiv preprint arXiv:2309.16400, 2023

    Robin Greif, Frank Jenko, and Nils Thuerey. Physics-preserving ai-accelerated simulations of plasma turbulence.arXiv preprint arXiv:2309.16400, 2023

    2023 arXiv

  18. [27]

    Hood: Hierarchical graphs for generalized modelling of clothing dynamics

    Artur Grigorev, Michael J Black, and Otmar Hilliges. Hood: Hierarchical graphs for generalized modelling of clothing dynamics. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16965–16974, 2023. 10

    2023

  19. [28]

    A comparison of neural network architectures for data-driven reduced-order modeling.Computer Methods in Applied Mechanics and Engineering, 393:114764, 2022

    Anthony Gruber, Max Gunzburger, Lili Ju, and Zhu Wang. A comparison of neural network architectures for data-driven reduced-order modeling.Computer Methods in Applied Mechanics and Engineering, 393:114764, 2022

    2022

  20. [29]

    Towards multi-spatiotemporal-scale generalized pde modeling.arXiv preprint arXiv:2209.15616, 2022

    Jayesh K Gupta and Johannes Brandstetter. Towards multi-spatiotemporal-scale generalized pde modeling.arXiv preprint arXiv:2209.15616, 2022

    2022 arXiv

  21. [30]

    Gnot: A general neural operator transformer for operator learning

    Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. Gnot: A general neural operator transformer for operator learning. InInternational Conference on Machine Learning, pages 12556–12569. PMLR, 2023

    2023

  22. [31]

    Springer Science & Business Media, 2012

    Akira Hasegawa.Plasma instabilities and nonlinear effects, volume 8. Springer Science & Business Media, 2012

    2012

  23. [32]

    Pseudo-three-dimensional turbulence in magnetized nonuniform plasma.The physics of Fluids, 21(1):87–92, 1978

    Akira Hasegawa and Kunioki Mima. Pseudo-three-dimensional turbulence in magnetized nonuniform plasma.The physics of Fluids, 21(1):87–92, 1978

    1978

  24. [33]

    Plasma edge turbulence.Physical Review Letters, 50 (9):682, 1983

    Akira Hasegawa and Masahiro Wakatani. Plasma edge turbulence.Physical Review Letters, 50 (9):682, 1983

    1983

  25. [34]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  26. [35]

    Turbulence model reduction by deep learning

    Robin A Heinonen and Patrick H Diamond. Turbulence model reduction by deep learning. Physical Review E, 101(6):061201, 2020

    2020

  27. [36]

    Group equivariant fourier neural operators for partial differential equations.arXiv preprint arXiv:2306.05697, 2023

    Jacob Helwig, Xuan Zhang, Cong Fu, Jerry Kurtin, Stephan Wojtowytsch, and Shuiwang Ji. Group equivariant fourier neural operators for partial differential equations.arXiv preprint arXiv:2306.05697, 2023

    2023 arXiv

  28. [37]

    Reduced-order modeling of fluid flows with transformers.Physics of Fluids, 35(5), 2023

    AmirPouya Hemmasian and Amir Barati Farimani. Reduced-order modeling of fluid flows with transformers.Physics of Fluids, 35(5), 2023

    2023

  29. [38]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017

    2017

  30. [39]

    The one hundred layers tiramisu: Fully convolutional densenets for semantic segmentation

    Simon Jégou, Michal Drozdzal, David Vazquez, Adriana Romero, and Yoshua Bengio. The one hundred layers tiramisu: Fully convolutional densenets for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 11–19, 2017

    2017

  31. [40]

    Nsfnets (navier-stokes flow nets): Physics-informed neural networks for the incompressible navier-stokes equations.Journal of Computational Physics, 426:109951, 2021

    Xiaowei Jin, Shengze Cai, Hui Li, and George Em Karniadakis. Nsfnets (navier-stokes flow nets): Physics-informed neural networks for the incompressible navier-stokes equations.Journal of Computational Physics, 426:109951, 2021

    2021

  32. [41]

    Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    2021

  33. [42]

    Deep fluids: A generative network for parameterized fluid simulations

    Byungsoo Kim, Vinicius C Azevedo, Nils Thuerey, Theodore Kim, Markus Gross, and Barbara Solenthaler. Deep fluids: A generative network for parameterized fluid simulations. InComputer graphics forum, volume 38, pages 59–70. Wiley Online Library, 2019

    2019

  34. [43]

    Smith, Ayya Alieva, Qing Wang, Michael P

    Dmitrii Kochkov, Jamie A. Smith, Ayya Alieva, Qing Wang, Michael P. Brenner, and Stephan Hoyer. Machine learning–accelerated computational fluid dynamics.Proceedings of the National Academy of Sciences, 118(21):e2101784118, 2021. doi: 10.1073/pnas.2101784118. URLhttps://www.pn...

    2021 doi

  35. [44]

    Andrey Nikolaevich Kolmogorov. A refinement of previous hypotheses concerning the local structure of turbulence in a viscous incompressible fluid at high reynolds number.Journal of Fluid Mechanics, 13(1):82–85, 1962

    1962

  36. [45]

    Fundamental statistical descriptions of plasma turbulence in magnetic fields

    John A Krommes. Fundamental statistical descriptions of plasma turbulence in magnetic fields. Physics Reports, 360(1-4):1–352, 2002. 11

    2002

  37. [46]

    Harnessing equivariance: Modeling turbulence with graph neural networks.arXiv preprint arXiv:2504.07741, 2025

    Marius Kurz, Andrea Beck, and Benjamin Sanderse. Harnessing equivariance: Modeling turbulence with graph neural networks.arXiv preprint arXiv:2504.07741, 2025

    2025 arXiv

  38. [47]

    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

  39. [48]

    Identification of high order closure terms from fully kinetic simulations using machine learning.Physics of Plasmas, 29(3), 2022

    Brecht Laperre, Jorge Amaya, Sara Jamal, and Giovanni Lapenta. Identification of high order closure terms from fully kinetic simulations using machine learning.Physics of Plasmas, 29(3), 2022

    2022

  40. [49]

    Training convolutional neural networks to estimate turbulent sub-grid scale reaction rates

    Corentin J Lapeyre, Antony Misdariis, Nicolas Cazard, Denis Veynante, and Thierry Poinsot. Training convolutional neural networks to estimate turbulent sub-grid scale reaction rates. Combustion and Flame, 203:255–264, 2019

    2019

  41. [50]

    Transformer for partial differential equations’ operator learning.Transactions on Machine Learning Research, 2023

    Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning.Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URLhttps://openreview.net/forum?id=EPPqt3uERT

    2023

  42. [51]

    Scalable transformer for pde surrogate modeling

    Zijie Li, Dule Shu, and Amir Barati Farimani. Scalable transformer for pde surrogate modeling. Advances in Neural Information Processing Systems, 36:28010–28039, 2023

    2023

  43. [52]

    Fourier neural operator for parametric partial differen- tial 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 differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  44. [53]

    Current and emerging deep-learning methods for the simulation of fluid dynamics.Proceedings of the Royal Society A, 479(2275):20230058, 2023

    Mario Lino, Stathi Fotiadis, Anil A Bharath, and Chris D Cantwell. Current and emerging deep-learning methods for the simulation of fluid dynamics.Proceedings of the Royal Society A, 479(2275):20230058, 2023

    2023

  45. [54]

    Veeling, Paris Perdikaris, Richard E Turner, and Johannes Brandstetter

    Phillip Lippe, Bastiaan S. Veeling, Paris Perdikaris, Richard E Turner, and Johannes Brandstetter. PDE-refiner: Achieving accurate long rollouts with neural PDE solvers. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview. net/for...

    2023

  46. [55]

    Enhancing fourier neural operators with local spatial features.arXiv preprint arXiv:2503.17797, 2025

    Chaoyu Liu, Davide Murari, Chris Budd, Lihao Liu, and Carola-Bibiane Schönlieb. Enhancing fourier neural operators with local spatial features.arXiv preprint arXiv:2503.17797, 2025

    2025 arXiv

  47. [56]

    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, 2021

    2021

  48. [57]

    Stacked convolutional auto-encoders for hierarchical feature extraction

    Jonathan Masci, Ueli Meier, Dan Cire¸ san, and Jürgen Schmidhuber. Stacked convolutional auto-encoders for hierarchical feature extraction. InArtificial neural networks and machine learning–ICANN 2011: 21st international conference on artificial neural networks, espoo, Finland...

    2011

  49. [58]

    Multiple physics pretraining for spatiotemporal surrogate models.Advances in Neural Information Processing Systems, 37:119301–119335, 2024

    Michael McCabe, Bruno Régaldo-Saint Blancard, Liam Parker, Ruben Ohana, Miles Cranmer, Alberto Bietti, Michael Eickenberg, Siavash Golkar, Geraud Krawezik, Francois Lanusse, et al. Multiple physics pretraining for spatiotemporal surrogate models.Advances in Neural Information ...

    2024

  50. [59]

    Climax: A foundation model for weather and climate.arXiv preprint arXiv:2301.10343, 2023

    Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K Gupta, and Aditya Grover. Climax: A foundation model for weather and climate.arXiv preprint arXiv:2301.10343, 2023

    2023 arXiv

  51. [60]

    Cfdnet: a deep learning-based accelerator for fluid simulations

    Octavi Obiols-Sales, Abhinav Vishnu, Nicholas Malaya, and Aparna Chandramowliswharan. Cfdnet: a deep learning-based accelerator for fluid simulations. InProceedings of the 34th ACM International Conference on Supercomputing, ICS ’20, New York, NY , USA, 2020. Association for C...

    2020

  52. [61]

    Learning mesh- based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh- based simulation with graph networks. InInternational conference on learning representations, 2020. 12

    2020

  53. [62]

    Transform once: Efficient operator learning in frequency domain.Advances in Neural Information Processing Systems, 35:7947–7959, 2022

    Michael Poli, Stefano Massaroli, Federico Berto, Jinkyoo Park, Tri Dao, Christopher Ré, and Stefano Ermon. Transform once: Efficient operator learning in frequency domain.Advances in Neural Information Processing Systems, 35:7947–7959, 2022

    2022

  54. [63]

    U-no: U-shaped neural operators.arXiv preprint arXiv:2204.11127, 2022

    Md Ashiqur Rahman, Zachary E Ross, and Kamyar Azizzadenesheli. U-no: U-shaped neural operators.arXiv preprint arXiv:2204.11127, 2022

    2022 arXiv

  55. [64]

    Maziar Raissi, Paris Perdikaris, and George E 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, 2019

    2019

  56. [65]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part I...

    2015

  57. [66]

    Graph networks as learnable physics engines for inference and control

    Alvaro Sanchez-Gonzalez, Nicolas Heess, Jost Tobias Springenberg, Josh Merel, Martin Ried- miller, Raia Hadsell, and Peter Battaglia. Graph networks as learnable physics engines for inference and control. InInternational conference on machine learning, pages 4470–4479. PMLR, 2018

    2018

  58. [67]

    Learning to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. InInternational conference on machine learning, pages 8459–8468. PMLR, 2020

    2020

  59. [68]

    Stacked u-nets: a no-frills approach to natural image segmentation.arXiv preprint arXiv:1804.10343, 2018

    Sohil Shah, Pallabi Ghosh, Larry S Davis, and Tom Goldstein. Stacked u-nets: a no-frills approach to natural image segmentation.arXiv preprint arXiv:1804.10343, 2018

    2018 arXiv

  60. [69]

    Self-attention with relative position repre- sentations.arXiv preprint arXiv:1803.02155, 2018

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position repre- sentations.arXiv preprint arXiv:1803.02155, 2018

    2018 arXiv

  61. [70]

    Learned coarse models for efficient turbulence simulation

    Kim Stachenfeld, Drummond Buschman Fielding, Dmitrii Kochkov, Miles Cranmer, Tobias Pfaff, Jonathan Godwin, Can Cui, Shirley Ho, Peter Battaglia, and Alvaro Sanchez-Gonzalez. Learned coarse models for efficient turbulence simulation. InInternational Conference on Learn- ing Re...

    2022

  62. [71]

    Adam Subel, Ashesh Chattopadhyay, Yifei Guan, and Pedram Hassanzadeh. Data-driven subgrid-scale modeling of forced burgers turbulence using deep learning with generalization to higher reynolds numbers via transfer learning.Physics of Fluids, 33(3), 2021

    2021

  63. [72]

    Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022

    2022

  64. [73]

    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, 2023. URL https://openreview.net/forum?id=tmIiMPl4IPa

    2023

  65. [74]

    Solver-in-the- loop: Learning from differentiable physics to interact with iterative pde-solvers.Advances in neural information processing systems, 33:6111–6122, 2020

    Kiwon Um, Robert Brand, Yun Raymond Fei, Philipp Holl, and Nils Thuerey. Solver-in-the- loop: Learning from differentiable physics to interact with iterative pde-solvers.Advances in neural information processing systems, 33:6111–6122, 2020

    2020

  66. [75]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  67. [76]

    Extracting and composing robust features with denoising autoencoders

    Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. InProceedings of the 25th international conference on Machine learning, pages 1096–1103, 2008

    2008

  68. [77]

    Enhancing computational fluid dynamics with machine learning.Nature Computational Science, 2(6):358–366, 2022

    Ricardo Vinuesa and Steven L Brunton. Enhancing computational fluid dynamics with machine learning.Nature Computational Science, 2(6):358–366, 2022. 13

    2022

  69. [78]

    Recent advances on machine learning for computational fluid dynamics: A survey.arXiv preprint arXiv:2408.12171, 2024

    Haixin Wang, Yadi Cao, Zijie Huang, Yuxuan Liu, Peiyan Hu, Xiao Luo, Zezheng Song, Wanjia Zhao, Jilin Liu, Jinan Sun, et al. Recent advances on machine learning for computational fluid dynamics: A survey.arXiv preprint arXiv:2408.12171, 2024

    2024 arXiv

  70. [79]

    Towards physics- informed deep learning for turbulent flow prediction

    Rui Wang, Karthik Kashinath, Mustafa Mustafa, Adrian Albert, and Rose Yu. Towards physics- informed deep learning for turbulent flow prediction. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’20, page 1457–1466, New Yor...

    2020

  71. [80]

    Learning the solution operator of paramet- ric partial differential equations with physics-informed deeponets.Science advances, 7(40): eabi8605, 2021

    Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of paramet- ric partial differential equations with physics-informed deeponets.Science advances, 7(40): eabi8605, 2021

    2021

  72. [81]

    DCW industries La Canada, CA, 1998

    David C Wilcox et al.Turbulence modeling for CFD, volume 2. DCW industries La Canada, CA, 1998

    1998

  73. [82]

    Rethinking and improving relative position encoding for vision transformer

    Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 10033–10041, 2021

    2021

  74. [83]

    W-net: A deep model for fully unsupervised image segmentation

    Xide Xia and Brian Kulis. W-net: A deep model for fully unsupervised image segmentation. arXiv preprint arXiv:1711.08506, 2017

    2017 arXiv

  75. [84]

    Sinenet: Learning temporal dynamics in time-dependent partial differential equations

    Xuan Zhang, Jacob Helwig, Yuchao Lin, Yaochen Xie, Cong Fu, Stephan Wojtowytsch, and Shuiwang Ji. Sinenet: Learning temporal dynamics in time-dependent partial differential equations. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openrev...

    2024

  76. [85]

    Pinnsformer: A transformer-based framework for physics-informed neural networks.arXiv preprint arXiv:2307.11833, 2023

    Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics-informed neural networks.arXiv preprint arXiv:2307.11833, 2023

    2023 arXiv

  77. [86]

    Neuralpde: Automating physics-informed neural networks (pinns) with error approximations.arXiv preprint arXiv:2107.09443, 2021

    Kirill Zubov, Zoe McCarthy, Yingbo Ma, Francesco Calisto, Valerio Pagliarino, Simone Azeglio, Luca Bottero, Emmanuel Luján, Valentin Sulzer, Ashutosh Bharambe, et al. Neuralpde: Automating physics-informed neural networks (pinns) with error approximations.arXiv preprint arXiv:...

    2021 arXiv

Pith tools

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