Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Data-Driven Adaptive Gradient Recovery for Unstructured Finite Volume Computations

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A geometry-aware neural network that corrects the gradient reconstruction inside a second-order finite-volume scheme makes coarse-mesh solutions of the 2D Euler equations 20–60% more accurate than the conventional solver, at no per-mesh…

desk verdict A plausible extension of learned gradient corrections to unstructured meshes, but the 20–60% gains are shown only on meshes close to the training family and the computational speedup claim is unexplained. read the letter →

arxiv 2507.16571 v1 pith:MBYA6G34 submitted 2025-07-22 math.NA cs.AIcs.NAmath.AP

classification math.NAcs.AIcs.NAmath.AP MSC 65M0835L6568T07
keywords machinelearningfinitevolumeunstructuredmeshhyperbolicconservationlawsgradientreconstructionDeepONetEulerequationsshockcapturing
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 argues that a small, geometry-aware neural network can replace the gradient-reconstruction step of a second-order finite-volume solver on unstructured triangular meshes, making coarse-mesh solutions of the 2D Euler equations 20–60% more accurate than the conventional scheme at no extra cost per mesh. The network is trained once on fine-grid reference solutions projected onto a coarse periodic mesh, then used as a local, physics-constrained correction operator that adjusts the weights in Green-Gauss or least-squares gradient formulas. The authors show the corrected scheme preserves symmetry and shock structure on standard Riemann problems and on the forward-facing step, and that its mesh-convergence rate exceeds the baseline solver's. If the result holds beyond the tested meshes, it would mean high-fidelity hyperbolic flow simulations can be run on coarser grids than classical accuracy requirements dictate, with the neural network acting as a subgrid correction.

What carries the argument

The key object is a modified DeepONet used as a local gradient-correction operator. It takes as input the neighbor-cell solution differences $u_{N_i} - u_i$ and three inter-cell angles $\theta_{jik}, \theta_{kil}, \theta_{lij}$ describing local geometry, and outputs per-neighbor weights $\alpha_j$ that plug into the corrected Green-Gauss gradient (Eq. 19) or least-squares gradient (Eq. 20). The angle inputs make the operator rotation-equivariant; the difference inputs enforce exactness on constants; the closed-form correction formulas hard-constrain consistency, leaving the network only the freedom to adjust the stencil weights. Physics is enforced softly through entropy-inequality and TVD penalizations plus $\ell^1$ weight regularization in the loss.

What would settle it

Take the published trained operator (no retraining) and apply it, on the same 2D Euler equation, to two meshes outside its training family: one refined by bisection instead of 4-triangle subdivision, and one with strongly anisotropic cells (aspect ratio above, say, 5). If the L1 error against the fine reference solution is no better, or worse, than the classical second-order solver's, the claimed mesh-transferability and 20–60% gain fail exactly where the paper says they might.

Watch

Extended reading notes

Core claim

The central discovery is that a learned, local correction of the gradient used in a MUSCL finite-volume scheme yields both higher accuracy and lower per-mesh cost than the traditional second-order solver. The correction is written as per-neighbor coefficients $\alpha_j$ added to the Green-Gauss and least-squares formulas, so the network's output is a small set of weights rather than a boundary flux or a full solution. Because the formulas still sum the same neighbor data, the scheme remains conservative, and the network is hard-constrained to preserve first-order consistency and rotation equivariance. Trained on random periodic Riemann problems and smooth waves with entropy, total-variation, and weight regularizers, the operator generalizes to the literature Riemann suite and to a boundary-fitted step geometry, with reported gains of roughly 22–34% on individual Riemann cases, about 40% averaged over the suite, and just over 60% on the forward-facing step.

Load-bearing premise

The whole 20–60% gain rests on the assumption that one fixed set of network weights, trained on a single 2614-cell periodic triangular mesh and a narrow set of local geometric inputs, keeps producing good gradient corrections when the mesh, boundary conditions, and flow conditions change; the paper's own closing section concedes that significant mesh-quality variations reduce accuracy or require retraining.

Editorial extensions

If this is right

  • On meshes similar to the training family, the 20–60% accuracy gain is available at no per-mesh time penalty, so users can choose coarser grids for the same error tolerance.
  • The corrected least-squares gradient, not the Green-Gauss one, is the robust choice: the paper reports better precision, symmetry, and convergence for the LSQ variant.
  • Because the operator is local and boundary cells are left to the classical reconstruction, the method slots into existing finite-volume codes without changing the flux, limiter, or boundary-condition treatment.
  • Averaged over the 18 Riemann test cases, the gain reaches about 40% at long integration times, implying the correction does not merely reduce initial transient error.

Reading between the lines

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

  • The learned operator behaves like a mesh-dependent subgrid model; a natural testable extension is to include cell-distance or anisotropy measures among the inputs, which the authors themselves flag, and to train once on a family of meshes spanning several refinement strategies.
  • If the correction is truly local and conservative, the same trained operator should transfer to other hyperbolic systems with the same eigenstructure (e.g. shallow-water equations) without retraining, merely by changing which primitive variables are fed in; the paper does not test this.
  • The near-zero time penalty suggests most of the cost is the flux evaluation, so combining the learned gradient with a cheaper flux or a larger allowable time step could compound the gain; this is an extrapolation, not a claim in the paper.
  • The negative initial gain on the forward-facing step and the degradation at extreme shock strengths indicate a robustness question: a practical deployment would need a fail-safe that falls back to the classical gradient when the network's correction is detected to increase local variation.
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 a data-driven correction to gradient reconstruction in a second-order finite volume solver for the 2D Euler equations on unstructured triangular meshes. A DeepONet-type architecture takes local angles and neighbor state differences as inputs and outputs per-neighbor weights alpha_j that modify Green-Gauss or least-squares gradient estimates. Training minimizes an L2 distance to a fine-grid reference solution plus entropy, TVD, and L1 regularization penalties. The method is tested on 18 two-dimensional Riemann problems and the forward-facing step, reporting 20-60% accuracy gains, improved mesh convergence, and reduced per-mesh computation time relative to the traditional solver. The manuscript also includes a limitations section that concedes sensitivity to mesh quality, anisotropy, and refinement strategy.

Significance. If substantiated, the approach would be a valuable contribution to ML-augmented CFD: it respects the conservation form, uses a local and rotation-invariant operator, and is tested on a nontrivial geometry with several Riemann benchmarks. The authors also provide a physically motivated regularization strategy and a convergence study. However, the claims currently outrun the evidence. The central generalization claim is not supported for mesh families outside the training distribution, the advertised 'first-order constraint' is not actually enforced, and the accuracy evaluation is single-run and uses a reference produced by the same finite volume scheme. These issues are load-bearing for the abstract's headline statements, so the manuscript needs substantial revision before the results can be accepted as stated.

major comments (4)
  1. [§4.5, §5.2, §8] The central generalization claim is not supported by the experimental design. Training is performed on a single family of regular triangular meshes: a periodic [0,1]^2 mesh with about 2614 cells and its uniform 4:1 refinement, and the network inputs are limited to three local angles and neighbor state differences, with no cell-size or anisotropy information. The validation meshes (5150 and 13089 cells) are regular triangular meshes of the same general type on the square and step domains, and the paper provides no mesh statistics (angle distributions, aspect ratios) to show they lie outside the training distribution. Section 8 explicitly concedes that 'significant variations in mesh quality, anisotropy, or refinement strategies reduce accuracy or require retraining.' The abstract's statement of 'robust generalization to complex flow configurations or geometries' is therefore not established for anisotropic, adaptive, or strongly boundary-fitted meshes. Please either provide quantitative evidence on a genuinely different mesh family or revise the generalization claim to the mesh class actually tested.
  2. [Abstract, §4.2, Eq. (20)] The claim that the architecture 'ensures a first-order constraint on the learned operator' is unsupported and, as stated, false. Equations (19) and (20) introduce free per-neighbor weights alpha_j, but no constraint is imposed on alpha_j. For a linear function u, the LSQ base gradient is exact, but the correction term in Eq. (20), M^{-1} Σ_j ω_j alpha_j (∆x_j, ∆y_j)^T (∇u · ∆r_j), is not zero for generic alpha_j. Thus the learned operator does not generally reproduce linear functions exactly, and the scheme's consistency order is not guaranteed by construction. Either enforce discrete moment conditions on alpha_j (for example, Σ_j ω_j alpha_j ∆x_j^2 = 0 and analogous conditions) or remove the first-order-constraint claim from the abstract and introduction.
  3. [§4.3, §4.5, §5] The gain metric is defined relative to a fine-grid reference computed with the same finite volume scheme and projected onto the coarse mesh. Since the training loss in Eq. (23) is exactly the L2 distance to that same projected reference, the reported 20-60% gains partly quantify how well the ML model imitates the fine-grid version of its own reference solver, not how well it approximates the exact solution. This is especially relevant for the shock-dominated Riemann problems, where the fine-grid reference itself carries discretization error. The paper would be materially strengthened by adding at least one validation case with an independent reference (an analytic solution, an exact Riemann solver, or a different high-order scheme) and by reporting results over several training seeds with means and standard deviations, since the current numbers are single-run estimates.
  4. [§7, Fig. 12] The efficiency claim that the ML solver is faster per mesh than the traditional LSQ solver is counterintuitive, because the ML version adds a neural-network evaluation to the same LSQ machinery. Figure 12 is presented without implementation details: whether the baseline and ML solvers use identical TensorFlow kernels, whether timings include the NN forward pass and the LSQ solve separately, and how many time steps are averaged. Since the abstract states that the proposed algorithm 'is faster and more accurate than the traditional second-order finite volume solver,' this claim needs a fair, reproducible benchmark (same mesh, same number of time steps, same hardware, with and without the NN component) rather than a single curve obtained from the convergence study. If the speed advantage is implementation-dependent, that should be stated explicitly.
minor comments (5)
  1. [§5.2] The text says 'supersonic outflow on the left boundary of the domain'; this should presumably be the right boundary, since the inflow is specified on the left boundary.
  2. [Title] The title contains a typo: 'V olume' should be 'Volume'.
  3. [Eq. (24)] The entropy penalty in Eq. (24) is written as max(0, ...)^2 without parentheses; clarify whether the maximum is applied before squaring and whether the term represents a one-step discrete entropy production residual.
  4. [§4.2] The sentence 'The flatten and un-flatten layers allows us to establish a link...' has a subject-verb agreement error; also, 'flatten' should likely be 'flattening' or 'flatten operations.'
  5. [Appendix A, Figure 13] The caption reads 'Figure .13' and should be 'Figure 13'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned corrections are trained against a fine-grid reference and then evaluated on held-out Riemann cases and a different geometry, so the claimed gains are not forced by construction.

full rationale

The derivation chain is self-contained. The only quantity that could look like a circular target is the reference solution u_ref in Eq. (23), which is generated by the same MUSCL finite-volume scheme on a fine grid and projected back to the coarse grid; the training loss is the relative L2 error against this reference. However, the reported gains are not defined by the training loss: the paper evaluates the learned correction on held-out configurations, namely the 18 Lax-Liu Riemann problems (including cases 6 and 11) and the Woodward-Colella forward-facing step with 13089 cells and slip/supersonic boundary conditions, using an L1 error against the same fine-grid reference. Those cases are not used to train or select the 1332 parameters, so a gain on them is not forced by construction; a model that merely memorized training initial conditions would not produce the reported gains. The repeated citations to de Romémont et al. (2024) are for dataset construction, loss form, and non-rolling training, but the unstructured-mesh architecture, angle inputs, gradient corrections (Eqs. 19-20), and the validation protocol are implemented and tested in this paper; no uniqueness theorem or external ansatz is imported as a load-bearing premise. The Section 8 limitations about mesh quality, anisotropy, and retraining are an honest generalization caveat, not evidence that the reported benchmarks are circular.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method relies on 1332 fitted network parameters plus three hand-chosen loss weights and a hand-set time stepping constant. The main unproven premises are the adequacy of the fine-grid FV solution as ground truth, the completeness of the angle-based local representation, and cross-mesh transfer of the learned operator. No invented physical entities.

free parameters (6)
  • Neural network weights (1332 trainable parameters) = Trained on 16,000 solution frames
    The correction coefficients alpha_j in Eqs. (19)-(20) are outputs of this network; they are the central fitted object of the method.
  • Entropy penalty weight lambda_ent = 1e5
    Hand-chosen in Sec. 4.6; part of loss Eq. (23), far outside the stated 1-10 percent guideline for regularizer weights.
  • TVD penalty weight lambda_TVD = 1e-6
    Hand-chosen in Sec. 4.6; same inconsistency with the stated guideline.
  • L1 regularization weight lambda_reg = 1e-4
    Hand-chosen in Sec. 4.6; controls weight magnitudes.
  • CFL-like constant Co = 0.03 (training), 0.01 (inference)
    Defines the time step in Eq. (14); chosen for stability and to keep time discretization error small, and it influences the loss landscape and generalization.
  • Initial condition amplitude max = 6 (scaled Euler variables range [-6,6])
    Sets the magnitude of training data in Appendix A; chosen to cover the range of values met in practice.
assumptions (5)
  • domain assumption The fine-grid solution produced by the traditional second-order FV scheme is an adequate proxy for the exact Euler solution.
    Used as the supervised target u_ref in Eq. (23). If the fine-grid scheme smears shocks or is dissipative, the learned corrections inherit those errors, so the reported gains are relative to this reference, not to the exact solution.
  • domain assumption Three angles and neighbor solution differences fully determine the optimal local gradient correction on any triangular mesh.
    The network inputs in Sec. 4.2 are restricted to u_Ni - u_i and angles theta_jik, theta_kil, theta_lij; no distances or cell shape measures are used, yet the method is claimed to generalize across meshes.
  • domain assumption The trained local operator transfers to meshes and geometries different from the training mesh.
    Training uses one mesh family (2614 to 10456 cells, uniform refinement, periodic); tests use 5150 and 13089 cells, boundary-fitted, with inflow/outflow/wall conditions. Section 8 concedes this transfer fails for large mesh quality or refinement changes.
  • domain assumption The Rusanov flux and Venkatakrishnan limiter keep the ML-modified scheme stable.
    Algorithm 1 changes only the gradient reconstruction; stability and conservation rest on the unchanged flux and limiter of the baseline FV solver.
  • ad hoc to paper Entropy and TVD soft constraints are sufficient to select physically relevant weak solutions.
    The penalties in Eqs. (24) and the TVD term are added with hand-chosen weights and are described by the authors as heuristic and problem-specific (Sec. 8).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data-Driven Adaptive Gradient Recovery for Unstructured Finite Volume Computations." pith.science (2026). https://pith.science/paper/MBYA6G34

@misc{pith2026250716571,
  author       = {Pith},
  title        = {Pith review of: Data-Driven Adaptive Gradient Recovery for Unstructured Finite Volume Computations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBYA6G34}},
  note         = {Machine review of arXiv:2507.16571}
}
read the original abstract

We present a novel data-driven approach for enhancing gradient reconstruction in unstructured finite volume methods for hyperbolic conservation laws, specifically for the 2D Euler equations. Our approach extends previous structured-grid methodologies to unstructured meshes through a modified DeepONet architecture that incorporates local geometry in the neural network. The architecture employs local mesh topology to ensure rotation invariance, while also ensuring first-order constraint on the learned operator. The training methodology incorporates physics-informed regularization through entropy penalization, total variation diminishing penalization, and parameter regularization to ensure physically consistent solutions, particularly in shock-dominated regions. The model is trained on high-fidelity datasets solutions derived from sine waves and randomized piecewise constant initial conditions with periodic boundary conditions, enabling robust generalization to complex flow configurations or geometries. Validation test cases from the literature, including challenging geometry configuration, demonstrates substantial improvements in accuracy compared to traditional second-order finite volume schemes. The method achieves gains of 20-60% in solution accuracy while enhancing computational efficiency. A convergence study has been conveyed and reveal improved mesh convergence rates compared to the conventional solver. The proposed algorithm is faster and more accurate than the traditional second-order finite volume solver, enabling high-fidelity simulations on coarser grids while preserving the stability and conservation properties essential for hyperbolic conservation laws. This work is a part of a new generation of solvers that are built by combining Machine-Learning (ML) tools with traditional numerical schemes, all while ensuring physical constraint on the results.

Figures

Figures reproduced from arXiv: 2507.16571 by the authors.

Figure 1
Figure 1. Unstructured mesh in 2D The red bullets denote the center of the cells. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Ghost cell approach at boundaries, the ghost cell is represented in dashed line [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. ML architecture inspired from the DeepONet architecture Lu et al. [2019]. Namely, [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Division of the domain Ω for the Riemman test cases All Riemann problems are posed such that the solutions to the four one-dimensional Riemann problems between quadrants exhibit precisely one propagating disturbance (i.e., a shock, rarefaction, or contact waves). All r…
Figure 5
Figure 5. Figure 5: Results for the 2D Riemann problem case 6 [Lax and Liu, 1998]. Density color map is overlayed by 30 density contours. The computations were performed [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Results for the 2D Riemann problem case 11 [Lax and Liu, 1998]. Density color map is overlayed by 30 density contours. The computations were performed [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: All gains across iterations for the 18 test cases present in Lax and Liu [1998] with the LSQ based gradient (Eq. (20). Each di [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: All gains across iterations for the 18 test cases present in Lax and Liu [1998] with the Green-Gauss based gradient (Eq. (19). Each di [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Results for the 2D forward facing stepWoodward and Colella [1984] at [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Slice at y = 0.5 of density of the solution for the forward facing step represented [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Mesh convergence with all Riemann initial conditions present in Lax and Liu [1998] computed for [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Computation time vs error on a log-log scale. The points used are the same as on Figure 11. [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Guarantees by Construction for Learned Finite Volume Schemes on Steady Supersonic Flow

    physics.flu-dyn 2026-07 conditional novelty 6.0 of 10

    A cost-matched evaluation of a guaranteed-admissible learned finite-volume scheme for 2D Euler finds that the unlearned skeleton with guarantee machinery is the most accurate equal-mesh scheme and never loses at equal...

  2. JAX-FVM: A differentiable, entropy-stable finite volume solver on unstructured meshes for compressible flows

    math.NA 2026-07 conditional novelty 6.0 of 10

    JAX-FVM is a differentiable, entropy-stable finite volume solver for 2D compressible Euler/Navier-Stokes equations on unstructured meshes, built in JAX.

Reference graph

Works this paper leans on

59 extracted references · 52 canonical work pages · cited by 2 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  2. [2]

    , author Barham, P

    author Abadi, M. , author Barham, P. , author Chen, J. , author Chen, Z. , author Davis, A. , author Dean, J. , author Devin, M. , author Ghemawat, S. , author Irving, G. , author Isard, M. , et al., year 2016 . title \ TensorFlow \ : a system for \ Large-Scale \ machine learning , in: booktitle 12th USENIX symposium on operating systems design and implem...

  3. [3]

    , author Modi, A

    author Agrawal, A. , author Modi, A. , author Passos, A. , author Lavoie, A. , author Agarwal, A. , author Shankar, A. , author Ganichev, I. , author Levenberg, J. , author Hong, M. , author Monga, R. , et al., year 2019 . title Tensorflow eager: A multi-stage, python-embedded dsl for machine learning . journal Proceedings of Machine Learning and Systems ...

  4. [4]

    , author Hoyer, S

    author Bar-Sinai, Y. , author Hoyer, S. , author Hickey, J. , author Brenner, M.P. , year 2019 . title Learning data-driven discretizations for partial differential equations . journal Proceedings of the National Academy of Sciences volume 116 , pages 15344--15349

  5. [5]

    , author Pearlmutter, B.A

    author Baydin, A.G. , author Pearlmutter, B.A. , author Radul, A.A. , author Siskind, J.M. , year 2018 . title Automatic differentiation in machine learning: a survey . journal Journal of Marchine Learning Research volume 18 , pages 1--43

  6. [6]

    , author Holmes, P

    author Berkooz, G. , author Holmes, P. , author Lumley, J.L. , year 1993 . title The proper orthogonal decomposition in the analysis of turbulent flows . journal Annual review of fluid mechanics volume 25 , pages 539--575

  7. [7]

    , author Schmidt, S.J

    author Bezgin, D.A. , author Schmidt, S.J. , author Adams, N.A. , year 2021 . title A data-driven physics-informed finite-volume scheme for nonclassical undercompressive shocks . journal Journal of Computational Physics volume 437 , pages 110324

  8. [8]

    , author Frostig, R

    author Bradbury, J. , author Frostig, R. , author Hawkins, P. , author Johnson, M.J. , author Leary, C. , author Maclaurin, D. , author Necula, G. , author Paszke, A. , author VanderPlas, J. , author Wanderman-Milne, S. , et al., year 2018 . title Jax: composable transformations of python+ numpy programs

Show all 59 references
  1. [9]

    , author Hesthaven, J.S

    author Bruno, O.P. , author Hesthaven, J.S. , author Leibovici, D.V. , year 2022 . title Fc-based shock-dynamics solver with neural-network localized artificial-viscosity assignment . journal Journal of Computational Physics: X volume 15 , pages 100110

  2. [10]

    , author Damodaran, M

    author Bui-Thanh, T. , author Damodaran, M. , author Willcox, K. , year 2003 . title Proper orthogonal decomposition extensions for parametric applications in compressible aerodynamics , in: booktitle 21st AIAA applied aerodynamics conference , p. pages 4213

  3. [11]

    , author Zou, Q

    author Cen, J. , author Zou, Q. , year 2024 . title Deep finite volume method for partial differential equations . journal Journal of Computational Physics volume 517 , pages 113307

  4. [12]

    , author Liang, C

    author Chen, X. , author Liang, C. , author Huang, D. , author Real, E. , author Wang, K. , author Pham, H. , author Dong, X. , author Luong, T. , author Hsieh, C.J. , author Lu, Y. , et al., year 2023 . title Symbolic discovery of optimization algorithms . journal Advances in...

  5. [13]

    , author Meneghetti, L

    author Coscia, D. , author Meneghetti, L. , author Demo, N. , author Stabile, G. , author Rozza, G. , year 2023 . title A continuous convolutional trainable filter for modelling unstructured data . journal Computational Mechanics volume 72 , pages 253--265

  6. [14]

    , year 1989

    author Cybenko, G. , year 1989 . title Approximation by superpositions of a sigmoidal function . journal Mathematics of control, signals and systems volume 2 , pages 303--314

  7. [15]

    , year 2005

    author Dafermos, C.M. , year 2005 . title Hyperbolic conservation laws in continuum physics . volume volume 3 . publisher Springer

  8. [16]

    , author Pain, C

    author Fang, F. , author Pain, C. , author Navon, I. , author Gorman, G. , author Piggott, M. , author Allison, P. , author Farrell, P. , author Goddard, A. , year 2009 . title A pod reduced order unstructured mesh ocean modelling method for moderate reynolds number flows . jo...

  9. [17]

    , author Schoenholz, S.S

    author Gilmer, J. , author Schoenholz, S.S. , author Riley, P.F. , author Vinyals, O. , author Dahl, G.E. , year 2017 . title Neural message passing for quantum chemistry , in: booktitle International conference on machine learning , organization PMLR . pp. pages 1263--1272

  10. [18]

    , author Raviart, P.A

    author Godlewski, E. , author Raviart, P.A. , year 2013 . title Numerical approximation of hyperbolic systems of conservation laws . volume volume 118 . publisher Springer Science & Business Media

  11. [19]

    , author Ying, Z

    author Hamilton, W. , author Ying, Z. , author Leskovec, J. , year 2017 . title Inductive representation learning on large graphs . journal Advances in neural information processing systems volume 30

  12. [20]

    , author Houston, P

    author Hartmann, R. , author Houston, P. , year 2002 . title Adaptive discontinuous galerkin finite element methods for the compressible euler equations . journal Journal of Computational Physics volume 183 , pages 508--532

  13. [21]

    , author Stinchcombe, M

    author Hornik, K. , author Stinchcombe, M. , author White, H. , year 1989 . title Multilayer feedforward networks are universal approximators . journal Neural networks volume 2 , pages 359--366

  14. [22]

    , year 2018

    author Innes, M. , year 2018 . title Flux: Elegant machine learning with julia . journal Journal of Open Source Software volume 3 , pages 602

  15. [23]

    , author Arafat, N.A

    author Jessica, L.S.E. , author Arafat, N.A. , author Lim, W.X. , author Chan, W.L. , author Kong, A.W.K. , year 2023 . title Finite volume features, global geometry representations, and residual training for deep learning-based cfd simulation . journal arXiv preprint arXiv:2311.14464

  16. [24]

    , author Zhang, Z

    author Kharazmi, E. , author Zhang, Z. , author Karniadakis, G.E. , year 2019 . title Variational physics-informed neural networks for solving partial differential equations . journal arXiv preprint arXiv:1912.00873

  17. [25]

    , author Kim, J

    author Kim, T. , author Kim, J. , author Tae, Y. , author Park, C. , author Choi, J.H. , author Choo, J. , year 2021 . title Reversible instance normalization for accurate time-series forecasting against distribution shift . journal ICLR 2022

  18. [26]

    , author Smith, J.A

    author Kochkov, D. , author Smith, J.A. , author Alieva, A. , author Wang, Q. , author Brenner, M.P. , author Hoyer, S. , year 2021 . title Machine learning--accelerated computational fluid dynamics . journal Proceedings of the National Academy of Sciences volume 118 , pages e...

  19. [27]

    , year 1973

    author Lax, P.D. , year 1973 . title Hyperbolic systems of conservation laws and the mathematical theory of shock waves . publisher SIAM

  20. [28]

    , year 2005

    author Lax, P.D. , year 2005 . title Weak solutions of nonlinear hyperbolic equations and their numerical computation , in: booktitle Selected Papers Volume I . publisher Springer , pp. pages 198--232

  21. [29]

    , author Liu, X.D

    author Lax, P.D. , author Liu, X.D. , year 1998 . title Solution of two-dimensional riemann problems of gas dynamics by positive schemes . journal SIAM Journal on Scientific Computing volume 19 , pages 319--340

  22. [30]

    , year 1992

    author LeVeque, R.J. , year 1992 . title Numerical methods for conservation laws . volume volume 214 . publisher Springer

  23. [31]

    , author Zou, S

    author Li, T. , author Zou, S. , author Chang, X. , author Zhang, L. , author Deng, X. , year 2023 . title Finite volume graph network (fvgn): Predicting unsteady incompressible fluid dynamics with finite volume informed neural network . journal arXiv preprint arXiv:2309.10050

  24. [32]

    , author Arora, S

    author Li, Z. , author Arora, S. , year 2019 . title An exponential learning rate schedule for deep learning . journal arXiv preprint arXiv:1910.07454

  25. [33]

    , author Wang, M

    author Liu, B. , author Wang, M. , author Foroosh, H. , author Tappen, M. , author Pensky, M. , year 2015 . title Sparse convolutional neural networks , in: booktitle Proceedings of the IEEE conference on computer vision and pattern recognition , pp. pages 806--814

  26. [34]

    , author Jin, P

    author Lu, L. , author Jin, P. , author Karniadakis, G.E. , year 2019 . title Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators . journal arXiv preprint arXiv:1910.03193

  27. [35]

    , author Ray, D

    author Magiera, J. , author Ray, D. , author Hesthaven, J.S. , author Rohde, C. , year 2020 . title Constraint-aware neural networks for riemann problems . journal Journal of Computational Physics volume 409 , pages 109345

  28. [36]

    , author Salmoiraghi, F

    author Manzoni, A. , author Salmoiraghi, F. , author Heltai, L. , year 2015 . title Reduced basis isogeometric methods (rb-iga) for the real-time simulation of potential flows about parametrized naca airfoils . journal Computer Methods in Applied Mechanics and Engineering volu...

  29. [37]

    , author M \"u ller, N

    author Morand, V. , author M \"u ller, N. , author Weightman, R. , author Piccoli, B. , author Keimer, A. , author Bayen, A.M. , year 2024 . title Deep learning of first-order nonlinear hyperbolic conservation law solvers . journal Journal of Computational Physics volume 511 ,...

  30. [38]

    , author McKerns, M

    author Nguyen-Fotiadis, N. , author McKerns, M. , author Sornborger, A. , year 2022 . title Machine learning changes the rules for flux limiters . journal Physics of Fluids volume 34

  31. [39]

    , author Gross, S

    author Paszke, A. , author Gross, S. , author Chintala, S. , author Chanan, G. , author Yang, E. , author DeVito, Z. , author Lin, Z. , author Desmaison, A. , author Antiga, L. , author Lerer, A. , year 2017 . title Automatic differentiation in pytorch

  32. [40]

    , author Manickam, I

    author Patel, R.G. , author Manickam, I. , author Trask, N.A. , author Wood, M.A. , author Lee, M. , author Tomas, I. , author Cyr, E.C. , year 2022 . title Thermodynamically consistent physics-informed neural networks for hyperbolic systems . journal Journal of Computational ...

  33. [41]

    , author Karniadakis, G.E

    author Raissi, M. , author Karniadakis, G.E. , year 2018 . title Hidden physics models: Machine learning of nonlinear partial differential equations . journal Journal of Computational Physics volume 357 , pages 125--141

  34. [42]

    , author Perdikaris, P

    author Raissi, M. , author Perdikaris, P. , author Karniadakis, G.E. , year 2019 . title Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations . journal Journal of Computational p...

  35. [43]

    , author Hill, C

    author Ranade, R. , author Hill, C. , author Pathak, J. , year 2021 . title Discretizationnet: A machine-learning based solver for navier--stokes equations using finite volume discretization . journal Computer Methods in Applied Mechanics and Engineering volume 378 , pages 113722

  36. [44]

    , year 1986

    author Roe, P.L. , year 1986 . title Characteristic-based schemes for the euler equations . journal Annual review of fluid mechanics volume 18 , pages 337--365

  37. [45]

    , author Renac, F

    author de Romémont, G. , author Renac, F. , author Nunez, J. , author Chinesta, F. , year 2024 . title A data-driven learned discretization approach in finite volume schemes for hyperbolic conservation laws and varying boundary conditions . journal arXiv preprint arXiv:2412.07541

  38. [46]

    , year 1961

    author Rusanov, V.V. , year 1961 . title The calculation of the interaction of non-stationary shock waves with barriers . journal Z. Vycisl. Mat. i Mat. Fiz volume 1 , pages 267--279

  39. [47]

    , author Godwin, J

    author Sanchez-Gonzalez, A. , author Godwin, J. , author Pfaff, T. , author Ying, R. , author Leskovec, J. , author Battaglia, P. , year 2020 . title Learning to simulate complex physics with graph networks , in: booktitle International conference on machine learning , organiz...

  40. [48]

    , author Keim, J

    author Schwarz, A. , author Keim, J. , author Chiocchetti, S. , author Beck, A. , year 2023 . title A reinforcement learning based slope limiter for second-order finite volume schemes . journal PAMM volume 23 , pages e202200207

  41. [49]

    , year 2008

    author Goncalv \`e s da Silva, E. , year 2008 . title R \'e solution num \'e rique des \'e quations d'Euler 1D . https://cel.archives-ouvertes.fr/cel-00556980. note lecture

  42. [50]

    , author Colonius, T

    author Stevens, B. , author Colonius, T. , year 2020 a. title Enhancement of shock-capturing methods via machine learning . journal Theoretical and Computational Fluid Dynamics volume 34 , pages 483--496

  43. [51]

    , author Colonius, T

    author Stevens, B. , author Colonius, T. , year 2020 b. title Finitenet: A fully convolutional lstm network architecture for time-dependent partial differential equations . journal arXiv preprint arXiv:2002.03014

  44. [52]

    , author Billett, S

    author Toro, E.F. , author Billett, S. , year 2000 . title Centred tvd schemes for hyperbolic conservation laws . journal IMA Journal of Numerical Analysis volume 20 , pages 47--79

  45. [53]

    , year 1979

    author Van Leer, B. , year 1979 . title Towards the ultimate conservative difference scheme. v. a second-order sequel to godunov's method . journal Journal of computational Physics volume 32 , pages 101--136

  46. [54]

    , author Cucurull, G

    author Velickovic, P. , author Cucurull, G. , author Casanova, A. , author Romero, A. , author Lio, P. , author Bengio, Y. , et al., year 2017 . title Graph attention networks . journal stat volume 1050 , pages 10--48550

  47. [55]

    , year 1995

    author Venkatakrishnan, V. , year 1995 . title Convergence to steady state solutions of the euler equations on unstructured grids with limiters . journal Journal of computational physics volume 118 , pages 120--130

  48. [56]

    , author Colella, P

    author Woodward, P. , author Colella, P. , year 1984 . title The numerical simulation of two-dimensional fluid flow with strong shocks . journal Journal of computational physics volume 54 , pages 115--173

  49. [57]

    , et al., year 2018

    author Yu, B. , et al., year 2018 . title The deep ritz method: a deep learning-based numerical algorithm for solving variational problems . journal Communications in Mathematics and Statistics volume 6 , pages 1--12

  50. [58]

    , author Wang, Q

    author Zhou, C.B. , author Wang, Q. , author Ren, Y.X. , year 2024 . title Machine learning optimization of compact finite volume methods on unstructured grids . journal Journal of Computational Physics volume 500 , pages 112746

  51. [59]

    , author Kochkov, D

    author Zhuang, J. , author Kochkov, D. , author Bar-Sinai, Y. , author Brenner, M.P. , author Hoyer, S. , year 2021 . title Learned discretizations for passive scalar advection in a two-dimensional turbulent flow . journal Physical Review Fluids volume 6 , pages 064605

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.