Pith. sign in

REVIEW 3 major objections 1 minor 48 references

Data-driven optimized high-order WENO schemes with low-dissipation and low-dispersion

T0 review · 3 major / 1 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Data-driven WENO5-JS/Z-NN schemes, trained via an approximate-dispersion-relation loss, match the exact spectrum of the PDE over a wider wavenumber range than classical WENO5-JS/Z while preserving shock capture.

desk verdict Plausible WENO idea in the abstract, but the supplied full text is a different paper, so the work is unverifiable as submitted. read the letter →

arxiv 2508.13190 v1 pith:4UDZ3PLR submitted 2025-08-15 math.NA cs.NA

Jinrui Zhou , Yiqi Gu , Song Jiang , Hua Shen , Liwei Xu , Guanyu Zhou This is my paper
classification math.NAcs.NA MSC 65M0665M1265M7068T07
keywords WENOspectralresolutiondispersion-dissipationneuralnetworkdata-drivenPDEshockcapturingapproximatedispersionrelationhigh-ordermethods
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

This paper proposes two data-driven WENO schemes, WENO5-JS-NN and WENO5-Z-NN, that insert a neural-network correction term into the classical weight functions. The network is trained to minimize reconstruction error on a large set of trigonometric stencils, guided by an approximate-dispersion-relation (ADR) bound that connects spectral error to reconstruction error. The result, if it holds, is a scheme with the exact spectrum of the underlying PDE more accurately matched over a wider range of wavenumbers than conventional WENO5-JS/Z, while still capturing shocks because the loss includes TVD and anti-dissipation penalties. A sympathetic reader would care because this offers a practical, trainable route toward low-dissipation, low-dispersion shock-capturing schemes without abandoning the WENO framework.

What carries the argument

The key object is the weight function with a neural-network compensation term, trained through a loss whose main term is the reconstruction error on trigonometric test functions with different wavenumbers, justified by the ADR bound that spectral error is bounded by reconstruction errors over such functions. The TVD and anti-dissipation penalties in the loss anchor the scheme's shock-capturing and stability behavior.

What would settle it

Train the network on wavenumbers up to some cutoff, then evaluate the actual dispersion relation (for example, the real and imaginary parts of the modified wavenumber) for wavenumbers well above that cutoff. If the scheme's dispersion error there is worse than WENO5-JS, or if a standard shock-turbulence test like the Shu-Osher problem develops visible oscillations or loses positivity, the central claim of broad-spectrum improvement with maintained shock capture would be in doubt.

Watch

Extended reading notes

Core claim

The central claim is that the spectral properties of WENO schemes can be improved by learning the nonlinear weights. The authors show that the approximate dispersion relation (ADR) of the scheme bounds its spectral error by the reconstruction errors of a family of trigonometric functions over many stencils. They therefore train a neural network to output a compensation term for the WENO5-JS and WENO5-Z weights, with a loss that combines these reconstruction errors, a total-variation-diminishing constraint, and an anti-dissipation penalty. According to the ADR analysis, the resulting WENO5-JS/Z-NN schemes match the exact dispersion relation more closely over a broader wavenumber band, while r

Load-bearing premise

The load-bearing premise is that minimizing reconstruction error on the sampled set of stencils and wavenumbers, with the TVD constraint applied only as a soft penalty, yields weights whose spectral accuracy transfers to the PDE solutions the scheme actually encounters, and that the ADR bound connecting spectral error to those reconstruction errors is valid as stated.

Editorial extensions

If this is right

  • WENO5-JS/Z-NN should resolve small-scale features like turbulence and acoustic waves on coarser grids than standard WENO5-JS/Z without adding explicit dissipation.
  • The same ADR-based loss can be exported to other WENO orders or other nonlinear shock-capturing families.
  • The trained weights can be frozen and used as drop-in replacements in existing WENO-based CFD codes, since the network is only an auxiliary term in the weight function.

Reading between the lines

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

  • If the ADR bound is tight, the paper effectively converts a notoriously hard problem—designing low-dissipation, low-dispersion, shock-capturing schemes—into a supervised regression problem, a different kind of approach for the field.
  • Because the TVD constraint is a penalty rather than a hard guarantee, there is a risk of spurious oscillations in extreme shock cases; a direct test would be to check discrete total variation against baseline WENO5-JS on a strong shock-adiabatic interaction.
  • The learned weights are likely to be problem-specific; a testable extension is to train on a mixture of wavenumber bands and compare resolution on problems with very different dominant wavenumbers.
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 / 1 minor

Summary. The paper, as described by its title and abstract, proposes data-driven optimized WENO schemes (WENO5-JS/Z-NN) in which a neural-network compensation term is added to the WENO5-JS/Z weight functions. The stated goal is to improve spectral properties—specifically the approximate dispersion relation (ADR)—so that the schemes resolve high-frequency waves and small-scale features better than classical WENO5-JS/Z, while still capturing discontinuities. The abstract claims that the spectral error can be bounded by reconstruction errors on trigonometric functions, and that the neural network is trained to minimize these errors under TVD and anti-dissipation penalties. The supplied full text, however, does not contain the WENO paper at all. It appears to be an unrelated NLP manuscript on homeless-bias detection, with sections on corpus construction, LLM evaluation, and related work. None of the WENO equations, proofs, training details, or numerical experiments are present.

Significance. If the claims in the abstract were demonstrated, the contribution would be potentially interesting to the CFD community: a data-driven WENO variant with a provable link between reconstruction error and spectral accuracy, combined with constraints intended to preserve shock capturing, could offer a practical way to improve resolution of fine-scale flow features. The proposed ADR-based bound is a plausible and useful theoretical bridge, and the idea of using a neural network to adjust WENO weights is consistent with current trends in learned numerical methods. However, because the manuscript body is entirely unrelated to the abstract, there is no verifiable technical content on which to base an assessment of novelty, correctness, or usefulness. The significance is therefore conditional and, as submitted, unassessable.

major comments (3)
  1. [Full text (Sections 2 and 3)] The body of the manuscript is a completely different paper on detecting bias against people experiencing homelessness, with content on corpus annotation and LLM evaluation. There is no WENO formulation, no approximate dispersion relation derivation, no neural network architecture, no loss function, and no numerical experiments. The central claims of the title and abstract are entirely unsupported by the provided text. This is a load-bearing issue that prevents any evaluation of the claimed results.
  2. [Abstract (unlabeled equation/bound)] The key bridge, that 'the spectral error of the schemes can be bounded by the reconstructed errors of a series of trigonometric functions with different wavenumbers,' is asserted without proof or derivation. No such bound appears in the body, and no reference is given. This bound is the theoretical justification for the training objective, so its absence is a fundamental gap.
  3. [Abstract (TVD constraint)] The abstract states that the TVD constraint and anti-dissipation penalization are incorporated into the loss function. As a soft penalty, this does not guarantee the non-oscillatory property of WENO. The manuscript provides no definition of these penalties, no evidence that the neural-network compensation preserves the WENO weighting requirements (nonnegativity, sum-to-one, smoothness-based adaptation), and no shock-capture test. The claim that the schemes 'maintain the ability to capture discontinuities' is therefore unverified.
minor comments (1)
  1. [Abstract] The notation 'WENO5-JS/Z-NN' is introduced without a definition of the NN compensation term; if the intended full text were available, this would need to be clearly specified in the method section.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the ADR evaluation is aligned with the training objective but not forced by construction, and the paper includes independent shock-capturing checks.

full rationale

The paper's derivation chain is: (i) ADR analysis bounds spectral error by reconstruction error of trigonometric functions; (ii) a neural-network compensation term is trained to minimize that reconstruction error over stencils; (iii) ADR is then reported as evidence of improved spectral accuracy. This creates a metric-alignment concern: if the stated bound were tight and the training covered the evaluated wavenumbers, the ADR improvement would be an expected consequence of the training objective rather than an independent discovery. However, the abstract only states a bound, not an equality, and the training is over a finite set of stencils and wavenumbers; the ADR curves for untrained wavenumbers and the shock-capturing tests (TVD penalty, discontinuity benchmarks) provide independent content. The abstract also asserts the ADR bound without proof (Abstract, sentence beginning 'By analyzing the approximate dispersion relation')—this is a support gap, not circularity. No self-citation chain or fitted-parameter-renamed-as-prediction is present. Therefore the central claim does not reduce to its inputs; score 2 reflects the mild closeness between the training loss and the headline spectral evaluation, while remaining within the 'no significant circularity' range.

Assumptions & free parameters 2 free parameters · 2 assumptions · 1 invented entities

All adjustable quantities are fitted: NN weights via the reconstruction-error objective, and loss-balancing coefficients by hand. The two axioms are the domain-specific bridges (spectral error to reconstruction error; soft TVD penalty to genuine non-oscillatory behavior). The only entity introduced is the NN-parameterized scheme itself, which has no independent falsifiable handle outside the paper. Because the body text supplied is a different manuscript, these entries reflect the abstract only.

free parameters (2)
  • Neural network weights of the compensation term = not reported in abstract
    The correction added to the WENO5-JS/Z weight function is parameterized by a neural network; the network weights are learned by minimizing reconstruction error across training stencils, i.e., fitted values.
  • Loss-function penalty coefficients (TVD constraint weight, anti-dissipation penalty weight) = not reported in abstract
    These terms are incorporated into the loss function; their balance determines shock-capturing versus spectral performance and must be chosen by hand or by tuning.
assumptions (2)
  • domain assumption The spectral error of a WENO scheme is bounded by its reconstruction errors on a series of trigonometric functions of different wavenumbers (ADR analysis).
    This is the theoretical bridge (abstract) that justifies using reconstruction-error minimization as a proxy for spectral accuracy; the proof is not visible in the supplied text.
  • domain assumption Weights produced by the trained network, constrained only through a soft TVD penalty in the loss, still satisfy the non-oscillatory requirement on discontinuities.
    The abstract claims shock capture is maintained, but the TVD constraint is a loss term, not a hard guarantee; whether it holds for the actual trained weights is an empirical claim.
invented entities (1)
  • WENO5-JS/Z-NN schemes (NN compensation term in the WENO5-JS/Z weights)
    purpose: Improve spectral resolution (dispersion and dissipation) of fifth-order WENO while keeping shock capture.
    The new scheme is an object designed within the paper; there is no falsifiable handle outside the paper. Its only evidence is the in-paper ADR comparison and the unreported discontinuity tests.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data-driven optimized high-order WENO schemes with low-dissipation and low-dispersion." pith.science (2026). https://pith.science/paper/4UDZ3PLR

@misc{pith2026250813190,
  author       = {Pith},
  title        = {Pith review of: Data-driven optimized high-order WENO schemes with low-dissipation and low-dispersion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UDZ3PLR}},
  note         = {Machine review of arXiv:2508.13190}
}
read the original abstract

Classical high-order weighted essentially non-oscillatory (WENO) schemes are designed to achieve optimal convergence order for smooth solutions and to maintain non-oscillatory behaviors for discontinuities. However, their spectral properties are not optimal, which limits the ability to capture high-frequency waves and small-scale features. In this paper, we propose a data-driven optimized method to improve the spectral properties of the WENO schemes. By analyzing the approximate dispersion relation (ADR), the spectral error of the schemes can be bounded by the reconstructed errors of a series of trigonometric functions with different wavenumbers. Therefore, we propose the new schemes WENO5-JS/Z-NN that introduce a compensation term parameterized by a neural network to the weight function of the WENO5-JS/Z schemes. The neural network is trained such that the generated weights can minimize the reconstructed errors over a large number of spatial stencils, and furthermore, improve the spectral accuracy. Meanwhile, the Total Variation Diminishing (TVD) constraint and anti-dissipation penalization are incorporated into the loss function to enhance the shock-capturing capability and preserve stability in simulating high-frequency waves. Compared to WENO5-JS/Z, our schemes maintain the ability to capture discontinuities while providing higher resolution for fine-scale flow features. The ADR indicates that the new schemes can match the exact spectrum more accurately over a broader range of wavenumbers.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 47 canonical work pages

  1. [1]

    Suresh, H

    A. Suresh, H. T. Huynh, Accurate monotonicity-preserving schemes with Runge--Kutta time stepping, J. Comput. Phys. 136 (1) (1997) 83--99

    work page 1997

  2. [2]

    Z. He, Y. Zhang, F. Gao, X. Li, B. Tian, An improved accurate monotonicity-preserving scheme for the Euler equations, Comput. Fluids 140 (2016) 1--10

    work page 2016

  3. [3]

    Harten, High resolution schemes for hyperbolic conservation laws, J

    A. Harten, High resolution schemes for hyperbolic conservation laws, J. Comput. Phys. 135 (2) (1997) 260--278

    work page 1997

  4. [4]

    Zhang, C

    D. Zhang, C. Jiang, D. Liang, L. Cheng, A review on TVD schemes and a refined flux-limiter for steady-state calculations, J. Comput. Phys. 302 (2015) 114--154

    work page 2015

  5. [5]

    Harten, B

    A. Harten, B. Engquist, S. Osher, S. R. Chakravarthy, Uniformly high order accurate essentially non-oscillatory schemes, III , J. Comput. Phys. 71 (2) (1987) 231--303

    work page 1987

  6. [6]

    C.-W. Shu, S. Osher, Efficient implementation of essentially non-oscillatory shock-capturing schemes, J. Comput. Phys. 77 (2) (1988) 439--471

    work page 1988

  7. [7]

    C.-W. Shu, S. Osher, Efficient implementation of essentially non-oscillatory shock-capturing schemes, II , J. Comput. Phys. 83 (1) (1989) 32--78

    work page 1989

  8. [8]

    X.-D. Liu, S. Osher, T. Chan, Weighted essentially non-oscillatory schemes, J. Comput. Phys. 115 (1) (1994) 200--212

    work page 1994

Show all 48 references
  1. [9]

    Jiang, C.-W

    G.-S. Jiang, C.-W. Shu, Efficient implementation of weighted ENO schemes, J. Comput. Phys. 126 (1) (1996) 202--228

    1996

  2. [10]

    A. K. Henrick, T. D. Aslam, J. M. Powers, Mapped weighted essentially non-oscillatory schemes: achieving optimal order near critical points, J. Comput. Phys. 207 (2) (2005) 542--567

    2005

  3. [11]

    Borges, M

    R. Borges, M. Carmona, B. Costa, W. S. Don, An improved weighted essentially non-oscillatory scheme for hyperbolic conservation laws, J. Comput. Phys. 227 (6) (2008) 3191--3211

    2008

  4. [12]

    D. S. Balsara, C.-W. Shu, Monotonicity preserving weighted essentially non-oscillatory schemes with increasingly high order of accuracy, J. Comput. Phys. 160 (2) (2000) 405--452

    2000

  5. [13]

    X. Deng, H. Zhang, Developing high-order weighted compact nonlinear schemes, J. Comput. Phys. 165 (1) (2000) 22--44

    2000

  6. [14]

    Nonomura, K

    T. Nonomura, K. Fujii, Robust explicit formulation of weighted compact nonlinear scheme, Comput. Fluids 85 (2013) 8--18

    2013

  7. [15]

    D. Levy, G. Puppo, G. Russo, Central WENO schemes for hyperbolic systemsof conservation laws, ESAIM: Math. Model. Numer. Anal. 33 (3) (1999) 547--571

    1999

  8. [16]

    D. Levy, G. Puppo, G. Russo, A third order central WENO scheme for 2D conservation laws, Appl. Numer. Math. 33 (1-4) (2000) 415--421

    2000

  9. [17]

    M. P. Mart \' n, E. M. Taylor, M. Wu, V. G. Weirs, A bandwidth-optimized WENO scheme for the effective direct numerical simulation of compressible turbulence, J. Comput. Phys. 220 (1) (2006) 270--289

    2006

  10. [18]

    Sun, Y.-X

    Z.-S. Sun, Y.-X. Ren, C. Larricq, S.-y. Zhang, Y.-c. Yang, A class of finite difference schemes with low dispersion and controllable dissipation for DNS of compressible turbulence, J. Comput. Phys. 230 (12) (2011) 4616--4635

    2011

  11. [19]

    Z.-s. Sun, L. Luo, Y.-x. Ren, S.-y. Zhang, A sixth order hybrid finite difference scheme based on the minimized dispersion and controllable dissipation technique, J. Comput. Phys. 270 (2014) 238--254

    2014

  12. [20]

    X. Hu, V. Tritschler, S. Pirozzoli, N. Adams, Dispersion-dissipation condition for finite difference schemes, arXiv:1204.5088 (2012)

    2012 arXiv

  13. [21]

    Y. Li, C. Chen, Y.-X. Ren, A class of high-order finite difference schemes with minimized dispersion and adaptive dissipation for solving compressible flows, J. Comput. Phys. 448 (2022) 110770

    2022

  14. [22]

    Acker, R

    F. Acker, R. d. R. Borges, B. Costa, An improved WENO-Z scheme, J. Comput. Phys. 313 (2016) 726--753

    2016

  15. [23]

    Luo, S.-p

    X. Luo, S.-p. Wu, Improvement of the WENO-Z+ scheme, Comput. Fluids 218 (2021) 104855

    2021

  16. [24]

    Luo, S.-p

    X. Luo, S.-p. Wu, An improved WENO-Z+ scheme for solving hyperbolic conservation laws, J. Comput. Phys. 445 (2021) 110608

    2021

  17. [25]

    L. Fu, X. Y. Hu, N. A. Adams, A family of high-order targeted ENO schemes for compressible-fluid simulations, J. Comput. Phys. 305 (2016) 333--359

    2016

  18. [26]

    Liang, L

    T. Liang, L. Fu, A new type of non-polynomial based TENO scheme for hyperbolic conservation laws, J. Comput. Phys. 497 (2024) 112618

    2024

  19. [27]

    F. M. de Lara, E. Ferrer, Accelerating high order discontinuous Galerkin solvers using neural networks: 1D Burgers' equation , Comput. Fluids 235 (2022) 105274

    2022

  20. [28]

    F. M. de Lara, E. Ferrer, Accelerating high order discontinuous Galerkin solvers using neural networks: 3D compressible Navier-Stokes equations , J. Comput. Phys. 489 (2023) 112253

    2023

  21. [29]

    Z. Xue, Y. Xia, C. Li, X. Yuan, A simplified multilayer perceptron detector for the hybrid WENO scheme, Comput. Fluids 244 (2022) 105584

    2022

  22. [30]

    Bar-Sinai, S

    Y. Bar-Sinai, S. Hoyer, J. Hickey, M. P. Brenner, Learning data-driven discretizations for partial differential equations, Proc. Natl. Acad. Sci. U.S.A. 116 (31) (2019) 15344--15349

    2019

  23. [31]

    Zhuang, D

    J. Zhuang, D. Kochkov, Y. Bar-Sinai, M. P. Brenner, S. Hoyer, Learned discretizations for passive scalar advection in a two-dimensional turbulent flow, Phys. Rev. Fluids 6 (6) (2021) 064605

    2021

  24. [32]

    Kochkov, J

    D. Kochkov, J. A. Smith, A. Alieva, Q. Wang, M. P. Brenner, S. Hoyer, Machine learning--accelerated computational fluid dynamics, Proc. Natl. Acad. Sci. U.S.A. 118 (21) (2021) e2101784118

    2021

  25. [33]

    Stevens, T

    B. Stevens, T. Colonius, Enhancement of shock-capturing methods via machine learning, Theor. Comput. Fluid Dyn. 34 (4) (2020) 483--496

    2020

  26. [34]

    D. A. Bezgin, S. J. Schmidt, N. A. Adams, WENO3-NN : A maximum-order three-point data-driven weighted essentially non-oscillatory scheme, J. Comput. Phys. 452 (2022) 110920

    2022

  27. [35]

    Nogueira, J

    X. Nogueira, J. Fern \'a ndez-Fidalgo, L. Ramos, I. Couceiro, L. Ram \' rez, Machine learning-based WENO5 scheme, Comput. Math. Appl. 168 (2024) 84--99

    2024

  28. [36]

    S. Fan, J. Qin, Y. Dong, Y. Jiang, X. Deng, WCNS3-MR-NN : A machine learning-based shock-capturing scheme with accuracy-preserving and high-resolution properties, J. Comput. Phys. 532 (2025) 113973

    2025

  29. [37]

    Y. Feng, F. S. Schranner, J. Winter, N. A. Adams, A deep reinforcement learning framework for dynamic optimization of numerical schemes for compressible flow simulations, J. Comput. Phys. 493 (2023) 112436

    2023

  30. [38]

    B. Liu, H. He, Q. Cheng, J. Wan, Y. Liu, A high-resolution finite difference scheme with neural network optimized dispersion and localized adaptive dissipation, Phys. Fluids 37 (3) (2025)

    2025

  31. [39]

    C.-B. Zhou, Q. Wang, Y.-X. Ren, Machine learning optimization of compact finite volume methods on unstructured grids, J. Comput. Phys. 500 (2024) 112746

    2024

  32. [40]

    H. Wang, Y. Cao, Z. Huang, Y. Liu, P. Hu, et al., Recent advances on machine learning for computational fluid dynamics: A survey, arXiv:2408.12171 (2024)

    2024 arXiv

  33. [41]

    Calin, Deep learning architectures, Springer, 2020

    O. Calin, Deep learning architectures, Springer, 2020

    2020

  34. [42]

    Pirozzoli, On the spectral properties of shock-capturing schemes, J

    S. Pirozzoli, On the spectral properties of shock-capturing schemes, J. Comput. Phys. 219 (2) (2006) 489--497

    2006

  35. [43]

    V. A. Titarev, E. F. Toro, Finite-volume WENO schemes for three-dimensional conservation laws, J. Comput. Phys. 201 (1) (2004) 238--260

    2004

  36. [44]

    G. M. Arshed, K. A. Hoffmann, Minimizing errors from linear and nonlinear weights of WENO scheme for broadband applications with shock waves, J. Comput. Phys. 246 (2013) 58--77

    2013

  37. [45]

    P. D. Lax, Weak solutions of nonlinear hyperbolic equations and their numerical computation, Commun. Pure Appl. Math. 7 (1) (1954) 159--193

    1954

  38. [46]

    Woodward, P

    P. Woodward, P. Colella, The numerical simulation of two-dimensional fluid flow with strong shocks, J. Comput. Phys. 54 (1) (1984) 115--173

    1984

  39. [47]

    P. D. Lax, X.-D. Liu, Solution of two-dimensional Riemann problems of gas dynamics by positive schemes, SIAM J. Sci. Comput. 19 (2) (1998) 319--340

    1998

  40. [48]

    Shi, Y.-T

    J. Shi, Y.-T. Zhang, C.-W. Shu, Resolution of high order WENO schemes for complicated flow structures, J. Comput. Phys. 186 (2) (2003) 690--696

    2003

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.