Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Neural Entropy-stable conservative flux form neural networks for learning hyperbolic conservation laws

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a neural network can learn unknown hyperbolic conservation laws with built-in conservation and entropy dissipation by learning flux and entropy together from data.

desk verdict Worth a serious look, but the advertised entropy-stability guarantee is not actually proved as written: the flux in (3.4) is an arithmetic mean, not the path-integral entropy-conservative flux that Theorem 2.6 requires. read the letter →

arxiv 2507.01795 v1 pith:TG77IPOK submitted 2025-07-02 math.NA cs.LGcs.NAmath-phmath.MP

Lizuo Liu , Lu Zhang , Anne Gelb This is my paper
classification math.NAcs.LGcs.NAmath-phmath.MP MSC 65M0868T0765M2265M3265D25
keywords hyperbolicconservationlawsentropystabilityneuralnetworksconservativefluxformlearneddata-drivenPDEmodelingshockcapturinginputconvex
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 seeks to establish that a neural network can learn unknown hyperbolic conservation laws directly from solution trajectories, with conservation and entropy dissipation built into the architecture rather than imposed by a chosen numerical scheme. The proposed NESCFN jointly learns a numerical flux, a local wave speed, and a convex entropy function, and uses them in a Rusanov-type entropy-stable flux update. If the claim holds, data-driven surrogates for shock-forming systems can remain stable and conservative over long time horizons, capture shock speeds, and tolerate noisy training data, even when the training window contains no shocks. The stability guarantee is internal: it holds with respect to the learned entropy, and physical entropy compatibility is not separately verified.

What carries the argument

The central object is the learned entropy-stable numerical flux in (3.5), a Rusanov-type flux whose base is the average $F^{\mu,*}_{j+1/2}=\frac12(F^\mu(u^+_{j+1/2})+F^\mu(u^-_{j+1/2}))$, whose dissipation coefficient is a learned wave speed that enforces the CFL condition, and whose dissipation term uses $(\eta_\theta'')^{-1}[\eta_\theta'(u)]_{j+1/2}$. The second load-bearing component is the input convex neural network $\eta_\theta$, a feedforward network with non-negative weights and convex non-decreasing activations whose structure guarantees the learned entropy is convex and whose Hessian appears directly in the dissipation; the two-stage loss then penalizes deviation from the symmetry condition of Theorem 2.3 to keep the learned system hyperbolic.

What would settle it

For 1D Burgers' equation with a quadratic learned entropy, compute both the arithmetic-average flux $F^{\mu,*}_{j+1/2}=\frac12(F^\mu(u^+_{j+1/2})+F^\mu(u^-_{j+1/2}))$ and the exact integral entropy-conservative flux $\int_0^1 F^\mu(v_j+\xi(v_{j+1}-v_j))\,d\xi$, and track the discrete entropy remainder (4.10) on a fine grid; if the remainder turns positive with the average but stays non-positive with the integral flux, the claimed guarantee fails exactly at the base-flux assumption.

Watch

Extended reading notes

Core claim

The central claim is that the semidiscrete scheme formed by the learned flux (3.5) is conservative by construction and entropy-stable for the learned conservation law $\partial_t u + \partial_x F^\mu(u)=0$ with respect to the learned entropy $\eta_\theta$. The learned flux combines an average of neural flux evaluations at reconstructed interface states with a dissipation term in which the inverse learned entropy Hessian acts on the jump of learned entropy variables, with a CFL-aware learned wave speed. Thanks to the convexity of the input convex neural network and the symmetry penalty in training, the learned system is designed to be hyperbolic. The numerical experiments claim that this construction predicts shock formation and propagation speed without future-time oracle data, maintains conservation and non-positive discrete entropy remainder over long horizons, and degrades gracefully under noise.

Load-bearing premise

The argument assumes that the average of the learned flux at the two sides of a cell edge serves as the entropy-conservative flux required by Theorem 2.6; for genuinely nonlinear fluxes such as Burgers this average differs from the path-integral entropy-conservative flux, and the added dissipation may not cover that difference.

Editorial extensions

If this is right

  • Training on smooth data that predates shock formation still yields correct shock location and propagation speed, so the entropy-stable structure supplies extrapolation that unconstrained regression would not.
  • The same architecture handles scalar, system, and two-dimensional cases without redesigning the discretization, since no base scheme needs to be chosen in advance.
  • Conservation and entropy dissipation are maintained over horizons longer than the training window and under the tested noise levels, making the surrogate suitable for long-term forecasting.
  • The guarantee is relative to the learned entropy rather than a physical entropy; for real systems, shock admissibility is only as good as the learned entropy's approximation of a true entropy.

Reading between the lines

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

  • Because the arithmetic average in (3.4) is not generally the path-integral entropy-conservative flux assumed by Theorem 2.6, the sharpest testable question is whether the dissipation term still dominates the entropy production of the difference for generic learned fluxes.
  • A practical extension would validate the learned entropy against a known physical entropy, for example the thermodynamic entropy in Euler's equations, since the method alone does not tie its learned entropy to physical admissibility.
  • Replacing the arithmetic average with a quadrature of the path integral would restore the theorem's hypothesis and let the error analysis separate discretization error from structural violation.
  • A learned Roe-type diffusion matrix, using the eigendecomposition of the learned flux Jacobian, is a natural next step that could reduce the extra diffusion the paper observes in the 2D and high-noise experiments.
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 / 4 minor

Summary. The paper proposes NESCFN, a conservative flux-form neural network for learning hyperbolic conservation laws and an associated entropy function directly from solution trajectories. The architecture parameterizes a flux F_mu, an input-convex neural entropy eta_theta, and a wave-speed surrogate rho_w. The semi-discrete update (3.8) uses the flux (3.5), which combines an arithmetic average of neural fluxes (3.4) with a Rusanov-type dissipation term constructed from eta''_theta. Training minimizes a recurrent prediction loss together with soft penalties intended to enforce the symmetry condition of Theorem 2.3. Experiments cover 1D and 2D Burgers, shallow water, and Euler equations, including noisy data and out-of-distribution initial conditions. The paper claims that Theorem 2.6 guarantees entropy stability of (3.5) for the learned system (3.7) with respect to the learned entropy eta_theta, and that hyperbolicity is obtained from Theorem 2.3 via the soft symmetry penalties.

Significance. If the structural guarantees were valid, the paper would provide a valuable fully data-driven route to entropy-stable, conservative surrogates for unknown hyperbolic systems, with strong empirical support in long-time shock prediction, noise robustness, and generalization to unseen initial data. The numerical study is extensive and includes conservation and entropy remainder diagnostics. However, the central theoretical claim is not established: Theorem 2.6 requires an entropy-conservative path-integral flux, while Eq. (3.4) uses an arithmetic mean; hyperbolicity is only softly penalized; and the entropy metric (4.10) is internal to the learned entropy and does not check the physical entropy inequality. The contribution is therefore conditional, and the manuscript needs substantial revision of either the architecture or the claims.

major comments (3)
  1. [Section 3.1, Eqs. (3.4) and (3.5)] The claim that the flux (3.5) is entropy-stable for the learned conservation law (3.7) with respect to eta_theta does not follow from Theorem 2.6. Theorem 2.6 requires the base flux f*_{j+1/2} to be the entropy-conservative flux of (2.8)/(2.12), i.e., a path integral in entropy variables. Equation (3.4) instead defines F_mu,*_{j+1/2} as the arithmetic average of the neural flux at reconstructed states. For a generic nonlinear flux these differ: for the Burgers flux f(u)=u^2/2 with eta_theta(u)=u^2/2, the path integral is (u_L^2+u_L u_R+u_R^2)/6, while (3.4) gives (u_L^2+u_R^2)/4, a difference of (u_R-u_L)^2/12. The paper provides no estimate showing that the Rusanov-type dissipation in (3.5) dominates this extra entropy production, and it also does not analyze the additional mismatch that Theorem 2.6 is stated for cell averages while (3.5) uses reconstructed interface states. The stated guarantee therefore needs either a proof or a substantial qualification.
  2. [Section 3.1 and Section 4.2, Eqs. (4.6)-(4.7)] Hyperbolicity of the learned system is not enforced. The sentence following Eq. (3.7) says that if the training procedure enforces the required symmetric structure in Theorem 2.3, then the learned conservation law is guaranteed to be hyperbolic, but the implemented losses (4.6) and (4.7) are only soft L1/L2 penalty terms, and the second training stage fine-tunes only the final layer of eta_theta. No exact structural constraint is imposed, so the identity eta''_theta(F_mu)' = (eta''_theta(F_mu)')^T holds only approximately, if at all. Since Theorem 2.3 requires exact symmetry, the claim that the learned system is guaranteed hyperbolic is unsupported.
  3. [Section 4.3, Eq. (4.10)] The discrete entropy remainder J(u) in (4.10) is not a test of the entropy inequality for the learned entropy pair. For a genuine entropy pair (eta_theta, G_theta) one must have G_theta' = eta_theta'^T F_mu', and the discrete entropy inequality involves numerical entropy fluxes at cell interfaces. Equation (4.10) instead substitutes the pointwise product eta_theta'(hat u_j)^T F_mu(hat u_j) and does not include interface flux differences or the numerical entropy flux from Theorem 2.6. Moreover, the dissipation term in (3.5) is constructed from eta''_theta, so non-positivity of J is largely a consequence of the architecture and the fitted entropy rather than an independent validation against the physical entropy. The abstract's claim that the method 'ensures entropy dissipation' is therefore stronger than what the numerical metric demonstrates.
minor comments (4)
  1. [Eq. (3.5)] The second argument of the max in (3.5) is rho_w(u^-_{j-1/2}), which appears to be a typo for rho_w(u^-_{j+1/2}); using the wrong interface in the wave-speed estimate makes the formula as written inconsistent.
  2. [Eq. (4.6)] The symmetry penalty term in (4.6) is notationally garbled: the expression [(F_mu)']^T[(hat u_j)eta''_theta]^T(hat u_j) does not parse as a matrix product; it should presumably be something like ||eta''_theta(F_mu)' - (eta''_theta(F_mu)')^T||_F^2.
  3. [Section 4.4.1] The text says fully connected networks are used for the learned flux function in (3.3) and the maximum wave speed in (3.7), but (3.7) is the learned conservation law; the wave-speed surrogate is rho_w and appears in (3.5).
  4. [References and typos] References [21] and [22] are the same Kurganov-Tadmor paper and should be consolidated; there are also minor typos such as 'breifly' in Section 4.4 and 'pertubed' in Section 5.2.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the learned entropy-stability claim rests on an unverified precondition (arithmetic-mean base flux), which is a correctness gap rather than a circular reduction; empirical validation is externally benchmarked against PyClaw.

full rationale

The paper's derivation chain does not exhibit circularity. The learned flux F_mu and learned entropy eta_theta are fitted to solution trajectories generated by an external solver (PyClaw), and the reported long-time predictions and shock speeds are compared against unseen reference solutions, so the empirical claims are not equivalent to the training targets. The central theoretical assertion that the numerical flux (3.5) is entropy-stable relies on Theorem 2.6, but the base flux is approximated in (3.4) by an arithmetic mean rather than the path-integral entropy-conservative flux (2.8) required by that theorem; for a nonlinear flux such as Burgers these differ. This is a real applicability gap and a correctness risk, but it is not circularity because the conclusion is not assumed or fitted. Likewise, the discrete entropy remainder (4.10) is evaluated using the learned entropy and learned flux, making the metric self-referential, but J<=0 is not imposed in the loss functions (4.5)-(4.7), and the paper explicitly claims stability only 'with respect to the learned entropy eta_theta.' Self-citations, such as [26], provide architectural inspiration rather than load-bearing justification for the main guarantee, which is instead based on Tadmor's classical entropy-stable framework. Overall, the paper's claimed derivation does not reduce to its own inputs by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central claim rests on learned network parameters for the flux, entropy, and wave speed, plus hand-chosen regularization constants and loss weights. The key unstated assumptions are that the averaged neural flux satisfies the entropy-conservation requirement of Theorem 2.6 and that soft penalties enforce the symmetry condition needed for hyperbolicity. The learned entropy is a new entity with no independent validation against a physical entropy.

free parameters (5)
  • F_mu network parameters = trained on solution trajectories
    Learned flux approximation; fitted to training data via the recurrent loss in (4.5).
  • eta_theta network parameters = trained with data plus structural penalties
    Learned entropy; constrained by ICNN convexity and the penalty terms in (4.6)-(4.7).
  • rho_w network parameters = trained
    Learned maximum wave speed surrogate used in the dissipation coefficient in (3.5).
  • Hessian regularization constants C1, C2 = C1=10, C2=0.3
    Hand-chosen constants for the exponentially decaying Hessian perturbation in (3.6).
  • Loss weights lambda1, lambda2 = e.g., 1e-3/1e-2 (Burgers), 1e-5/1e-3 (Euler)
    Weights balancing data fidelity and structural penalties in (4.6)-(4.7); chosen per experiment.
assumptions (5)
  • domain assumption The system generating the data is a hyperbolic conservation law of the form (2.1) with unknown flux.
    Stated in the introduction and Section 4.1; the whole recovery problem is predicated on this assumption.
  • domain assumption Training trajectories produced by PyClaw with the HLLE solver faithfully represent solutions of the target PDE at the chosen discretization.
    All experiments use synthetic data from numerical solvers; the learned model inherits any approximation error from this data.
  • ad hoc to paper The arithmetic average in (3.4) is taken to be an entropy-conservative flux as required by Theorem 2.6.
    The entropy-stability proof requires the base flux to be the path-integral flux (2.8); Eq. (3.4) uses an endpoint average, which generally differs for nonlinear fluxes.
  • ad hoc to paper The soft penalty terms in (4.6)-(4.7) approximately enforce the symmetry condition of Theorem 2.3, hence hyperbolicity of the learned system.
    The paper states hyperbolicity holds 'if the training procedure is designed to enforce' the symmetric structure; the penalties are not exact constraints.
  • standard math Standard theorems on entropy pairs and entropy-stable schemes (Theorems 2.3-2.6) from [14, 34, 35] are valid and applicable.
    Used to justify the flux construction and the entropy-stability claim; not proved in the paper.
invented entities (2)
  • Learned entropy function eta_theta(u)
    purpose: Provides the entropy variable and entropy Hessian used in the dissipation term of the numerical flux, making the learned scheme entropy-stable by construction.
    No comparison is made with the physical entropy of the underlying systems (e.g., u^2/2 for Burgers or the physical entropy for Euler); the only validation is the self-referential inequality J(u) <= 0 in (4.10).
  • Learned maximum wave speed surrogate rho_w(u)
    purpose: Provides the scalar dissipation coefficient in the Rusanov-type flux (3.5); its tanh construction is designed to keep the value within a CFL-type bound.
    The paper does not compare the learned wave speed to the true spectral radius of the flux Jacobian; it is fitted to data and constrained by construction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Entropy-stable conservative flux form neural networks for learning hyperbolic conservation laws." pith.science (2026). https://pith.science/paper/TG77IPOK

@misc{pith2026250701795,
  author       = {Pith},
  title        = {Pith review of: Neural Entropy-stable conservative flux form neural networks for learning hyperbolic conservation laws},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TG77IPOK}},
  note         = {Machine review of arXiv:2507.01795}
}
read the original abstract

We propose a neural entropy-stable conservative flux form neural network (NESCFN) for learning hyperbolic conservation laws and their associated entropy functions directly from solution trajectories, without requiring any predefined numerical discretization. While recent neural network architectures have successfully integrated classical numerical principles into learned models, most rely on prior knowledge of the governing equations or assume a fixed discretization. Our approach removes this dependency by embedding entropy-stable design principles into the learning process itself, enabling the discovery of physically consistent dynamics in a fully data-driven setting. By jointly learning both the numerical flux function and a corresponding entropy, the proposed method ensures conservation and entropy dissipation, critical for long-term stability and fidelity in the system of hyperbolic conservation laws. Numerical results demonstrate that the method achieves stability and conservation over extended time horizons and accurately captures shock propagation speeds, even without oracle access to future-time solution profiles in the training data.

Figures

Figures reproduced from arXiv: 2507.01795 by the authors.

Figure 5.1
Figure 5.1. Comparison of the reference solution (black solid line) to 1D Burgers’ equation with [PITH_FULL_IMAGE:figures/full_fig_p021_5_1.png] view at source ↗
Figure 5.2
Figure 5.2. Discrete conserved quantity remainder C(u), defined in (4.8), for the 1D Burgers’ equation over the time interval t ∈ [0, 3] with noise levels ξ = 0, .25, .5, 1 in (5.16). Figures 5.2 and 5.3 further illustrate the robustness of the NESCFN method to noise [PITH_FULL_IMAGE:figures/full_fig_p021_5_2.png] view at source ↗
Figure 5.3
Figure 5.3. Discrete entropy remainder J (u), defined in (4.10), for the 1D Burgers’ equation over the time interval t ∈ [0, 3], evaluated across all testing initial conditions specified in (5.3). From (left) to (right) the plots correspond to training data with respective noise levels ξ = 0, .25, .5, and 1 in (5.16). 5.2.2. Shallow water equation. We now examine the impact of noise in the training data for the shallow water eq… view at source ↗
Figures from the paper (9 more)
Figure 5.4
Figure 5.4. Figure 5.4: Comparison of the reference solution (black solid line) of height [PITH_FULL_IMAGE:figures/full_fig_p022_5_4.png]
Figure 5.5
Figure 5.5. Figure 5.5: Discrete conserved quantity remainder (4.8) of height C(h) (left) and the momentum C(hu) (right) for the shallow water equations with noise levels ξ = 0, .25, .5, 1 in (5.17) [PITH_FULL_IMAGE:figures/full_fig_p023_5_5.png]
Figure 5.6
Figure 5.6. Figure 5.6: Discrete entropy remainder J ([h, hu] ⊤), defined in (4.10), for the shallow water equations, evaluated across all testing initial conditions specified in (5.6). From (left) to (right), the plots correspond to training data with noise levels ξ = 0, .25, .5, and 1, re…
Figure 5.7
Figure 5.7. Figure 5.7: Comparison of the reference solution (black solid line) of density [PITH_FULL_IMAGE:figures/full_fig_p023_5_7.png]
Figure 5.8
Figure 5.8. Figure 5.8: Discrete conserved quantity remainder C(ρ)(left), C(E)(right), defined in (4.8), for Euler’s equation with noise levels ξ = 0, .25, .5, 1 in (5.18) [PITH_FULL_IMAGE:figures/full_fig_p024_5_8.png]
Figure 5.9
Figure 5.9. Figure 5.9: Discrete entropy remainder J ([ρ, ρu, E] ⊤), defined in (4.10), for Euler’s equation, evaluated across all testing initial conditions specified in (5.9)–(5.10). From (left) to (right), the plots correspond to training data with noise levels ξ = 0, .25, .5, and 1, res…
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 5.10
Figure 5.10. Figure 5.10: Comparison of the reference solution of u in 2D Burgers’ equation (5.11) with the NESCFN model predictions for noise coefficient ξ = 1: (left) reference solution at t = .8, (middle-left) predictions at t = .8, (middle-right) reference solution at t = 1.6, (right) pr…
Figure 5.11
Figure 5.11. Figure 5.11: Discrete entropy remainder J (u), defined in (4.10), for the 2D Burgers’ equation, evaluated across all testing initial conditions specified in (5.13)–(5.15). From (left) to (right), the plots correspond to training data with noise levels ξ = 0, .25, .5, and 1, resp…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  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. FluxNet: Learning Capacity-Constrained Local Transport Operators for Conservative and Bounded PDE Surrogates

    cond-mat.mtrl-sci 2026-02 conditional novelty 6.0 of 10

    FluxNet is a family of neural PDE time-steppers whose learned transport plans make discrete conservation exact and one-sided bound preservation structural, with empirical dual-bound control via a consistency-regulariz...

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · cited by 2 Pith papers

  1. [26]

    L. Liu, T. Li, A. Gelb, and Y. Lee. Entropy stable conservative flux form neural networks. ENTROPY-STABLE LEARNED HYPERBOLIC CONSER V ATION LA WS 27 arXiv preprint arXiv:2411.01746 , 2024

    work page Pith review arXiv 2024

  2. [1]

    B. Amos, L. Xu, and J. Z. Kolter. Input convex neural networks. In International conference on machine learning , pages 146–155. PMLR, 2017. 26 LIZUO LIU, LU ZHANG AND ANNE GELB

    work page 2017

  3. [2]

    T. J. Barth. Numerical Methods for Gasdynamic Systems on Unstructured Meshes , pages 195–285. Springer Berlin Heidelberg, Berlin, Heidelberg, 1999

    work page 1999

  4. [3]

    G. R. Bigg, M. R. Wadley, D. P. Stevens, and J. A. Johnson. Modelling the dynamics and thermodynamics of icebergs. Cold Regions Science and Technology, 26(2):113–135, 1997

    work page 1997

  5. [4]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, and Q. Zhang. JAX: composable trans- formations of Python+NumPy programs, 2018

    work page 2018

  6. [5]

    Z. Chen, V. Churchill, K. Wu, and D. Xiu. Deep neural network modeling of unknown partial differential equations in nodal space. Journal of Computational Physics , 449:110782, 2022

    work page 2022

  7. [6]

    Z. Chen, A. Gelb, and Y. Lee. Learning the dynamics for unknown hyperbolic conservation laws using deep neural networks. SIAM Journal on Scientific Computing , 46(2):A825– A850, 2024

    work page 2024

  8. [7]

    Churchill and D

    V. Churchill and D. Xiu. Flow map learning for unknown dynamical systems: Overview, imple- mentation, and benchmarks. Journal of Machine Learning for Modeling and Computing , 4(2):173–201, 2023

    work page 2023

Show all 38 references
  1. [8]

    Clawpack software, 2020

    Clawpack Development Team. Clawpack software, 2020. Version 5.7.1

    2020

  2. [9]

    Dulny, A

    A. Dulny, A. Hotho, and A. Krause. NeuralPDE: Modelling dynamical systems from data. In R. Bergmann, L. Malburg, S. C. Rodermund, and I. J. Timm, editors, KI 2022: Advances in Artificial Intelligence , pages 75–89, Cham, 2022. Springer International Publishing

    2022

  3. [10]

    U. S. Fjordholm, S. Mishra, and E. Tadmor. Arbitrarily high-order accurate entropy stable essentially nonoscillatory schemes for systems of conservation laws. SIAM Journal on Numerical Analysis, 50(2):544–573, 2012

    2012

  4. [11]

    Forgione and D

    M. Forgione and D. Piga. dynoNet: A neural network architecture for learning dynamical systems. International Journal of Adaptive Control and Signal Processing , 35(4):612–626, 2021

    2021

  5. [12]

    Gajamannage, D

    K. Gajamannage, D. I. Jayathilake, Y. Park, and E. M. Bollt. Recurrent neural networks for dynamical systems: Applications to ordinary differential equations, collective motion, and hydrological modeling. Chaos: An Interdisciplinary Journal of Nonlinear Science , 33(1):013109, 01 2023

    2023

  6. [13]

    Girard, S

    L. Girard, S. Bouillon, J. Weiss, D. Amitrano, T. Fichefet, and V. Legat. A new modeling framework for sea-ice mechanics based on elasto-brittle rheology. Annals of Glaciology , 52(57):123–132, 2011

    2011

  7. [14]

    Godlewski and P.-A

    E. Godlewski and P.-A. Raviart. Numerical approximation of hyperbolic systems of conservation laws, volume 118. Springer Science & Business Media, 2013

    2013

  8. [15]

    Gottlieb and C.-W

    S. Gottlieb and C.-W. Shu. Total variation diminishing Runge-Kutta schemes. Math. Comput., 67(221):73–85, jan 1998

    1998

  9. [16]

    J. R. Holton and G. J. Hakim. An Introduction to Dynamic Meteorology . Academic Press, Oxford, 5th edition, 2012

    2012

  10. [17]

    Ismail and P

    F. Ismail and P. L. Roe. Affordable, entropy-consistent Euler flux functions ii: Entropy pro- duction at shocks. Journal of Computational Physics , 228(15):5410–5436, 2009

    2009

  11. [18]

    Kast and J

    M. Kast and J. S. Hesthaven. Positional embeddings for solving PDEs with evolutional deep neural networks. Journal of Computational Physics , 508:112986, 2024

    2024

  12. [19]

    D. I. Ketcheson, K. T. Mandli, A. J. Ahmadia, A. Alghamdi, M. Quezada de Luna, M. Parsani, M. G. Knepley, and M. Emmett. PyClaw: Accessible, Extensible, Scalable Tools for Wave Propagation Problems. SIAM Journal on Scientific Computing , 34(4):C210–C231, Nov. 2012

    2012

  13. [20]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Y. Bengio and Y. LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015

    2015

  14. [21]

    Kurganov and E

    A. Kurganov and E. Tadmor. New high-resolution central schemes for nonlinear conservation laws and convection–diffusion equations. Journal of computational physics, 160(1):241–282, 2000

    2000

  15. [22]

    Kurganov and E

    A. Kurganov and E. Tadmor. New high-resolution central schemes for nonlinear conservation laws and convection–diffusion equations. Journal of Computational Physics , 160(1):241– 282, 2000

    2000

  16. [23]

    P. G. LeFloch, J. M. Mercier, and C. Rohde. Fully discrete, entropy conservative schemes of arbitraryorder. SIAM Journal on Numerical Analysis , 40(5):1968–1992, 2002

    1968

  17. [24]

    R. J. LeVeque. Finite Volume Methods for Hyperbolic Problems . Cambridge Texts in Applied Mathematics. Cambridge University Press, 2002

    2002

  18. [25]

    Liu and W

    L. Liu and W. Cai. DeepPropNet – a recursive deep propagator neural network for learning evolution PDE operators, 2022

    2022

  19. [27]

    L. Liu, K. Nath, and W. Cai. A causality-deeponet for causal responses of linear dynamical systems. Communications in Computational Physics , 35(5):1194–1228, 2024

    2024

  20. [28]

    Z. Long, Y. Lu, X. Ma, and B. Dong. PDE-Net: Learning PDEs from data, 2018

    2018

  21. [29]

    D. G. Patsatzis, M. di Bernardo, L. Russo, and C. Siettos. Gorinns: Godunov-riemann in- formed neural networks for learning hyperbolic conservation laws. Journal of Computa- tional Physics , 534:114002, 2025

    2025

  22. [30]

    R. S. Pritchard. An Elastic-Plastic Constitutive Law for Sea Ice. Journal of Applied Mechanics, 42(2):379–384, 06 1975

    1975

  23. [31]

    P. Roe. Approximate riemann solvers, parameter vectors, and difference schemes. Journal of Computational Physics , 43(2):357–372, 1981

    1981

  24. [32]

    Shu and S

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

    1988

  25. [33]

    S. W. Suh, J. F. MacArt, L. N. Olson, and J. B. Freund. A TVD neural network closure and application to turbulent combustion. Journal of Computational Physics , 523:113638, 2025

    2025

  26. [34]

    E. Tadmor. The numerical viscosity of entropy stable schemes for systems of conservation laws. i. Mathematics of Computation , 49:91–103, 1987

    1987

  27. [35]

    E. Tadmor. Entropy stable schemes. In Handbook of Numerical Analysis , volume 17, pages 467–493. Elsevier, 2016

    2016

  28. [36]

    Y. Tong, S. Xiong, X. He, S. Yang, Z. Wang, R. Tao, R. Liu, and B. Zhu. Roenet: Predict- ing discontinuity of hyperbolic systems from continuous data. International Journal for Numerical Methods in Engineering , 125(6):e7406, 2024

    2024

  29. [37]

    C. B. Vreugdenhil. Numerical Methods for Shallow-Water Flow . Springer, Dordrecht, 1994

    1994

  30. [38]

    Y. Yin, M. Kirchmeyer, J.-Y. Franceschi, A. Rakotomamonjy, and P. Gallinari. Continuous PDE dynamics forecasting with implicit neural representations. In International Conference on Learning Representations (ICLR), 2023

    2023

Pith tools

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