Pith. sign in

REVIEW 3 major objections 6 minor 54 references

A "Neural" Riemann solver for Relativistic Hydrodynamics

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Small neural networks can replace the root-finding in an exact relativistic Riemann solver, matching its accuracy at about one-fourteenth of the cost.

desk verdict Clever hybrid ML/exact Riemann solver, but the headline speed claim is undermined by the paper's own report that its exact-solver baseline fails and falls back to HLLE. read the letter →

arxiv 2505.18914 v1 pith:CEAO5S4Z submitted 2025-05-25 gr-qc

Carlo Musolino This is my paper · ORCID
classification gr-qc
keywords relativistichydrodynamicsRiemannproblemneuralnetworksmixtureofexpertsshock-capturingschemesroot-findingaccelerationGodunovmethods
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 claims that the expensive iterative root-finding inside an exact relativistic Riemann solver can be replaced by a handful of small feed-forward neural networks without giving up accuracy. One network per wave pattern predicts the pressure across the contact discontinuity, and two further networks reconstruct states when the cell interface lies inside a rarefaction fan; the analytic wave structure and flux formulas are kept intact. Trained on exact solutions sampled from a synthetic distribution and embedded in a Godunov-type scheme, the "neural" solver matches the exact solver's L1 errors on one-dimensional shock-tube benchmarks while running about 14 times faster, and beats the approximate HLLE and HLLC solvers in most first-order runs. If the result holds, exact-quality relativistic Riemann fluxes become cheap enough for large-scale simulations.

What carries the argument

The mechanism is a physics-informed mixture-of-experts: three pattern-specific pressure networks (double shock, double rarefaction, shock-rarefaction) plus two direction-specific rarefaction state networks, selected by the analytic wave-pattern criterion of the exact solver. Each network learns an implicitly defined solution map x* such that f(x*, theta)=0, trained on low-discrepancy quasi-random samples of exact solutions with a loss that blends the residual of the physical constraint with a robust loss against the exact root; the output is constrained by a sigmoid and affine rescaling to the physical range. Because only a single scalar (the contact pressure, or the rarefaction sound speed) is predicted, the rest of the Riemann fan is recovered analytically, so network error is confined to the one quantity that the rest of the flux depends on nonlinearly.

What would settle it

Re-run the five-network solver on the Problem 2 blastwave at 800 zones for the same set of training seeds used in the paper; if a non-negligible fraction of seeds produce L1 density errors an order of magnitude above HLLC, or the run fails outright, the distribution-coverage assumption is falsified. For the efficiency claim, time the 800-zone first-order Problem 4 run and check whether the neural solver is actually about 14 times faster than the exact solver; if the speedup is not there, the central practical claim collapses.

Watch

Extended reading notes

Core claim

The central discovery is that the exact relativistic Riemann solution decomposes into smooth, single-valued maps that small networks can learn: the contact pressure as a function of the two initial states for each of the three wave patterns (double shock, double rarefaction, shock-rarefaction), and the rarefaction-fan state at the interface. Wave-pattern classification and flux selection remain analytic, so the networks only replace root-finding; all other primitive variables are recovered from the predicted pressure using the exact wave relations. In the paper's 800-zone first-order benchmark the neural solver reaches the same L1 density error as the exact solver (marginally lower, in fact) at roughly one-fourteenth of the runtime, and in first-order Godunov tests it consistently outperforms HLLE and HLLC except in the supersonic blastwave Problem 3, where solver differences wash out. High-order WENO results are comparable to, but not consistently better than, HLLC, and strongly shocked problems show small post-shock oscillations that converge away with resolution.

Load-bearing premise

The load-bearing premise is that the synthetic training distribution (densities from $10^{-2}$ to $10^{2}$, pressures from $10^{-7}$ to $10^{3}$.5, velocities from -0.99 to 0.99, with pattern-specific pressure-ratio bounds) covers the states that show up in a real simulation; the paper itself shows this can break down when a problem sits near the boundary, as Problem 2's pressure jump of order $10^{7}$ does for some training seeds.

Editorial extensions

If this is right

  • If the accuracy claim holds, exact-quality relativistic Riemann fluxes could be used in production hydrodynamics codes for roughly 1/14 of the exact solver's cost per interface.
  • The neural solver's runtime is non-iterative and seed-stable, unlike exact root-finding, so per-interface cost is predictable for large parallel simulations.
  • In first-order Godunov schemes, the trained solver sharpens contact discontinuities and rarefaction structure compared with HLLE and HLLC on the tested shock tubes.
  • In second-order WENO runs, all solvers converge and the neural solver's errors are comparable to HLLC, with small non-divergent post-shock oscillations on strongly shocked problems.
  • Residual-based ensembling of several small networks removes the failure outliers seen for some seeds in Problem 2 while changing the runtime only modestly.

Reading between the lines

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

  • An unstated consequence is that the same hybrid design could accelerate any Riemann solver whose bottleneck is a scalar root-finding step, not just relativistic hydrodynamics, as long as a reliable wave-pattern classifier exists.
  • The paper's fully differentiable implementation opens a path to end-to-end training where the network is tuned against the downstream simulation error rather than against offline exact roots, but that direction is not explored here.
  • The distribution-boundary failure suggests a practical test: before deploying a trained solver in a given simulation campaign, one should check the campaign's extreme initial states against the training box; adaptive or curriculum sampling could extend coverage without enlarging the networks.
  • Ensemble selection by the physical residual is cheap here, but in more complex settings where the residual is expensive, the paper's proposal would need a learned confidence score; that is an extension, not something demonstrated.
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 / 6 minor

Summary. The paper presents a hybrid data-driven Riemann solver for one-dimensional special-relativistic hydrodynamics. The exact solver's expensive root-finding steps are replaced by five compact feedforward networks: three pattern-specific “pressure predictor” networks (double shock, double rarefaction, shock–rarefaction) that predict the contact pressure p*, and two “rarefaction fan” networks that reconstruct the state at the interface when it lies inside a rarefaction. The remaining steps of the Riemann fan are computed analytically. The networks are trained on synthetic Sobol-sampled data using a loss that combines the residual of the nonlinear equation with the distance to the exact root. The solver is embedded in a first- and second-order finite-volume scheme and compared with HLLE, HLLC, and an exact solver on four canonical one-dimensional shock-tube problems. The central claim, stated in the abstract and Section 6.5, is that the neural solver achieves accuracies comparable to the exact algorithm at a fraction of the cost, with a reported speedup of about 14x on one 800-zone first-order test while matching its L1 error.

Significance. If the central claims are correct, the paper makes a useful contribution: it shows a modular, interpretable way to replace the iterative parts of an exact relativistic Riemann solver with small learned components, and it provides careful ablations on dataset size, model size, and random seeds, together with an ensembling strategy that reduces variance. The authors are unusually transparent about failure modes: Appendix B.2 explicitly reports outlier seeds for Problem 2, and Appendix C studies residual-based ensembling. However, the main quantitative claim rests on an exact-solver baseline that the paper itself reports as occasionally failing and falling back to HLLE, and the training labels and reference solutions are generated with that same solver. Because the accuracy comparison and the speed comparison are both anchored to this baseline, the headline result is not yet established. The paper's strength is the architecture and the extensive ablation study; its weakness is the validity of the exact-solver reference used for comparison.

major comments (3)
  1. [§6.5, footnote 4; §3.3 and Appendix A] The central speed/accuracy claim rests on an exact Riemann solver that the paper itself reports as unreliable. Footnote 4 states that the neural solver's L1 error is marginally lower than the exact solver's because of “occasional failures in the root-finding step (in which case the solver reverts to HLLE)”, without identifying the cause. Since the same implementation generates the training labels in Appendix A and the reference solutions used throughout Section 6, an undiagnosed root-finding flaw can contaminate the labels, the reference solutions, and the accuracy comparison simultaneously. The 14x speedup and the “comparable accuracy” statement in Section 6.5 therefore need to be re-established against a validated, independently implemented exact Riemann solver, with the failure and fallback rate reported explicitly.
  2. [Appendix B.2, Fig. 13; §6.2] The representativeness of the synthetic Sobol training distribution is a load-bearing assumption, and the paper's own ablation shows that it can fail near the boundary. For Problem 2, whose pressure jump is of order 10^7, several training seeds produce L1 errors more than an order of magnitude above HLLC or outright simulation failure (Appendix B.2, Fig. 13). The main text hides this from the central accuracy claim by excluding a single outlier from Problem 2 in Fig. 8. The authors should either extend the training distribution to cover such states, make the residual-based fallback or ensembling of Section 7 and Appendix C a mandatory component of the solver, or state the restricted validity domain of the claimed robustness.
  3. [§6.4, Figs. 6–9] The claim that the neural solver achieves accuracy comparable to the exact algorithm is not supported at high order for strong shocks. In Section 6.4 the neural solution develops post-shock oscillations even with first-order reconstruction, and Fig. 9 shows that HLLC outperforms the neural solver in Problem 4 with second-order reconstruction. Since the abstract and Section 8 present general accuracy claims without these qualifications, the conclusions should be rephrased to specify reconstruction order and problem regime, and to quantify the amplitude and boundedness of the oscillations that were actually verified.
minor comments (6)
  1. [§6.3] The sentence beginning “he second-order solution computed with the neural Riemann solver” is missing the leading “T” and should read “The second-order solution”.
  2. [Acknowledgements] The phrase “We am also sincerely grateful” is ungrammatical and should be “I am” or “We are”.
  3. [Appendix B.3 and Appendix C] The text says networks are trained on “1017 samples” and “10 17 samples”; this should be 2^17 samples, consistent with Appendix B.2 and Fig. 13.
  4. [References] Reference [1] contains the typo “hydrpdynamics” and should read “hydrodynamics”.
  5. [Appendix B.2 and Fig. 13 caption] The text describes the marker for the main-text dataset size as a vertical dashed line, while the Fig. 13 caption calls it a vertical solid line; the two descriptions should be made consistent.
  6. [§9] For reproducibility, the trained model weights and datasets should be archived alongside the code rather than “available from the author upon reasonable request”.

Circularity Check

1 steps flagged · score 2.0 of 10

Partial fit-to-teacher dependency in the exact-accuracy claim, but held-out benchmarks keep the central result non-circular.

  1. fitted input called prediction [Sec. 5 (Eq. 22) and Sec. 6.5 (Fig. 10)]
    "Each network is trained by minimizing a loss function that blends two objectives: Lθ(x) = H(f(x,θ),0) + H(log(x), log(x∗)) ... x∗ is the exact root obtained offline ... Our neural solver achieves accuracy comparable to the exact solver while being approximately 14 times faster."

    The baseline for the headline accuracy claim is the same exact solver that produced the offline labels x* in Eq. (22). The second loss term directly regresses the network output onto the exact solver's root, so 'comparable to the exact algorithm' is partly a restatement of the training objective rather than an independent validation. This is only partial circularity: the L1 errors come from held-out shock-tube evolutions and are also compared against independent HLLE and HLLC solvers, so the central practical claim is not forced by construction.

full rationale

The paper transparently trains all learned components on exact-solver labels (Eq. 22; Appendix A), so the statement that the neural solver achieves accuracy comparable to the exact algorithm is partly a regression-fidelity statement to the teacher. This is a surrogate-modeling caveat rather than a hidden derivation. The evaluation uses four canonical shock tubes not used as training samples, compares against independent approximate solvers (HLLE and HLLC), and includes seed and dataset ablations, so the reported downstream L1 errors and the 14x speedup are not statistically forced by the training objective. The evidence also contains acknowledged limitations that are correctness/robustness concerns rather than circularity: footnote 4 reports occasional failures in the exact solver's root-finding step with fallback to HLLE, and Appendix B.2 shows Problem 2 outliers caused by pressure jumps near the training-distribution boundary. No load-bearing self-citation or imported uniqueness claim is present. Overall, the central claim retains independent content; the score of 2 reflects only the partial dependency of the exact-accuracy comparison on the training labels.

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

The central claim rests on the correctness of the exact solver used for labels, the chosen training distribution, and hand-set hyperparameters. No new physical entities are introduced; the learned networks and ensembling are computational components, not postulates.

free parameters (6)
  • Neural network weights and biases (all five networks) = not enumerated, trained on 2^17 Sobol samples per network
    The core fitted parameters; the accuracy claim depends directly on their trained values.
  • Learnable output scaling beta (Eq. 21) = learned per network
    Scales the sigmoid output and is part of the trained model.
  • Input rescaling bounds for training data = densities in [1e-2, 1e2], pressures in [1e-7, 1e3.5], velocities in [-0.99, 0.99]
    Chosen by hand; they define the training distribution and therefore the generalization envelope.
  • Pressure ratio sampling ranges per pattern = log10(pL/pR): [0.1, 1.5] for g2S, [0.1, 1.0] for g2R, [0.1, 7.0] for gSR
    Hand-chosen bounds restrict each solver's input domain; Problem 2's pressure jump lies near the gSR boundary.
  • Network and dataset hyperparameters = 2 hidden layers, 64/32 neurons, 2^17 samples, 100 epochs, learning rate 1e-2, seed 42
    Chosen by the author; sensitivity to these choices is partly explored in Appendix B.
  • Fallback threshold eps_c = 1e-10
    Hand-fixed threshold below which the solver falls back to the HLLE flux; affects near-continuous interfaces.
assumptions (5)
  • domain assumption Ideal gas (gamma-law) equation of state h = 1 + gamma p / ((gamma-1) rho), with gamma = 5/3 in all tests
    The entire derivation and data generation assume this EOS; extensions to tabulated or other EOSs are deferred.
  • domain assumption The exact Riemann solver of [4-6] provides correct ground truth for training labels and reference solutions
    Supervised training and error evaluation both rely on this external algorithm's correctness.
  • domain assumption The wave pattern can be classified solely by the invariant relative velocity v_LR, following Rezzolla and Zanotti [6]
    Used in step 1 of the exact solution and in routing to the correct gp network.
  • ad hoc to paper The synthetic training distribution (Appendix A) is representative of the states encountered in the test problems and in future production use
    Load-bearing for generalization; Appendix B.2 shows it fails for a few seeds on Problem 2's pressure jump of order 1e7.
  • standard math Universal Approximation Theorem and Implicit Function Theorem justify learnability of the root maps
    Invoked in Section 4 to justify network approximation of the implicitly defined root x*(theta).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A "Neural" Riemann solver for Relativistic Hydrodynamics." pith.science (2026). https://pith.science/paper/CEAO5S4Z

@misc{pith2026250518914,
  author       = {Pith},
  title        = {Pith review of: A "Neural" Riemann solver for Relativistic Hydrodynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CEAO5S4Z}},
  note         = {Machine review of arXiv:2505.18914}
}
read the original abstract

In this paper, we present an approach to solving the Riemann problem in one-dimensional relativistic hydrodynamics, where the most computationally expensive steps of the exact solver are replaced by compact, highly specialized neural networks. The resulting "neural" Riemann solver is integrated into a high-resolution shock-capturing scheme and tested on a range of canonical problems, demonstrating both robustness and efficiency. By constraining the learned components to the root-finding of single-valued functions, the method retains physical interpretability while significantly accelerating the computation. The solver is shown to achieve accuracies comparable to the exact algorithm at a fraction of the cost, suggesting that this approach may offer a viable path toward more efficient Riemann solvers for use in large-scale numerical relativity simulations of astrophysical systems.

Figures

Figures reproduced from arXiv: 2505.18914 by the authors.

Figure 1
Figure 1. Schematic architecture of the neural Riemann solver. Neural networks are [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Density profile at t = 0.4 for the shocktube Problem 1, obtained with the first order scheme on a grid consisting of 100 points. The exact solution is shown as a dotted line and the numerical results obtained with the HLLE and neural Riemann solvers are shown with dashed lines. While the methods perform similarly at the rar￾efaction waves, the neural solver resolves the contact discontinuity more accurately than HLL… view at source ↗
Figure 3
Figure 3. Profiles of density (solid line), pressure (dash-dotted line) and velocity [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Same as Fig [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Same as Fig [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: compares the solutions obtained with first-order schemes using the HLLE and neural Riemann solvers. As in previous cases, the neural method exhibits a sharper resolution of the contact discontinuity than its traditional counterpart. However, the zoomed inset reveals th…
Figure 7
Figure 7. Figure 7: Same as Fig [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: L1 error norms for the density across the four shock tube problems solved using the first-order method, plotted as a function of the number of grid points. The neural solver consistently outperforms HLLE and HLLC across all problems, with the exception of Problem 3, wh…
Figure 9
Figure 9. Figure 9: Same as Fig [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Comparison of L1 density error (blue, left bars), Riemann solver runtime (green, center bars), and total timestep runtime (orange, right bars) for four differ￾ent Riemann solvers, applied to Problem 4 on a uniform grid with 800 zones using a first-order scheme. All te…
Figure 11
Figure 11. Figure 11: Training and test loss trajectories (root in blue and residual in orange) for [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Training and test loss trajectories for each learned Riemann solver com [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: L1 density error across problems for networks with two hidden layers of 64 neurons each. The rarefaction networks have half as many neurons as the pressure predictors and are trained on a dataset 1/10 the size. Seed-to-seed variability is shown as a shaded region, whi…
Figure 14
Figure 14. Figure 14: L1 density error across problems for networks with two hidden layers of varying widths, rarefaction solvers always comprising half as many neurons as corresponding gp MLPs. All networks are trained on 1017 samples. Seed-to-seed variability is shown as a shaded region,…
Figure 15
Figure 15. Figure 15: Comparison of L1 density error across problems for single seed networks and equivalent ensembles. The ensembles are composed of 2, 4 and 8 networks of 64 neurons per layer, and are compared to appropriately scaled single-seed networks of 128, 256 and 512 neurons per h…
Figure 16
Figure 16. Figure 16: Riemann solver and total RHS computation runtime per call of single-seed [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 28 canonical work pages

  1. [1]

    S. K. Godunov, A difference method for numerical calculations of discontinuous solutions of the equations of hydrpdynamics, Mat. Sb. 47, 271 (1959), In Russian

    work page 1959

  2. [2]

    J. M. Martí and E. Müller, Extension of the Piecewise Parabolic Method to One- Dimensional Relativistic Hydrodynamics, Journal of Computational Physics123, 1 (1996), doi:10.1006/jcph.1996.0001

    arXiv 1996

  3. [3]

    Colella and P

    P . Colella and P . R. Woodward, The piecewise parabolic method (ppm) for gas- dynamical simulations, Journal of Computational Physics 54(1), 174 (1984), doi:DOI: 10.1016/0021-9991(84)90143-8

    doi:10.1016/0021-9991(84)90143-8 1984

  4. [4]

    J. M. Martí and E. Müller, The analytical solution of the riemann problem in relativistic hydrodynamics, J. Fluid Mech. 258, 317 (1994)

    work page 1994

  5. [5]

    J. A. Pons, J. M. Martí and E. Müller,The exact solution of the Riemann problem with non- zero tangential velocities in relativistic hydrodynamics, J. Fluid Mech. 422, 125 (2000)

    work page 2000

  6. [6]

    Rezzolla and O

    L. Rezzolla and O. Zanotti, An improved exact riemann solver for relativistic hydrodynam- ics, Journ. of Fluid Mech. 449, 395 (2001)

    work page 2001

  7. [7]

    Harten, J

    A. Harten, J. M. Hyman, P . D. Lax and B. Keyfitz,On finite-difference approximations and entropy conditions for shocks, Communications on Pure and Applied Mathematics 29(3), 297 (1976), doi:10.1002 /cpa.3160290305

    work page 1976

  8. [8]

    Einfeldt, On Godunov-type methods for gas dynamics , SIAM J

    B. Einfeldt, On Godunov-type methods for gas dynamics , SIAM J. Numer. Anal. 25, 294 (1988)

    work page 1988

Show all 54 references
  1. [9]

    Schneider, U

    V . Schneider, U. Katscher, D. Rischke, B. Waldhauser, J. Maruhn and C.-D. Munz,New al- gorithms for ultra-relativistic numerical hydrodynamics, Journal of Computational Physics 105(1), 92 (1993), doi:https: //doi.org/10.1006/jcph.1993.1056

    1993

  2. [10]

    Duncan, L

    C. Duncan, L. Sczaniecki and G. Shastri, Three dimensional relativistic hydrodynamics I. the code, tests, and application to off-center collisions Preprint

  3. [11]

    E. F . Toro, M. Spruce and W . Speares,Restoration of the contact surface in the HLL-Riemann solver, Shock Waves 4, 25 (1994), doi:10.1007 /BF01414629

    1994

  4. [12]

    Mignone and G

    A. Mignone and G. Bodo,An HLLC Riemann solver for relativistic flows - I. Hydrodynamics, Mon. Not. R. Astron. Soc. 364, 126 (2005), doi:10.1111 /j.1365-2966.2005.09546.x, arXiv:astro-ph/0506414

    2005

  5. [13]

    Kiuchi, L

    K. Kiuchi, L. E. Held, Y. Sekiguchi and M. Shibata, Implementation of advanced rie- mann solvers in a neutrino-radiation magnetohydrodynamics code in numerical relativity and its application to a binary neutron star merger , Physical Review D 106(12) (2022), doi:10.1103/physre...

    2022 doi

  6. [14]

    Rackauckas, Y

    C. Rackauckas, Y. Ma, J. Martensen, C. Warner, K. Zubov, R. Supekar, D. Skinner, A. Ra- madhan and A. Edelman, Universal differential equations for scientific machine learning (2021), 2001.04385

    2021 arXiv

  7. [15]

    L. Lu, P . Jin, G. Pang, Z. Zhang and G. E. Karniadakis, Learning nonlinear operators via deeponet based on the universal approximation theorem of operators , Nature Machine Intelligence 3(3), 218–229 (2021), doi:10.1038 /s42256-021-00302-5. 32 SciPost Physics Submission

    2021

  8. [16]

    Raissi, P

    M. Raissi, P . Perdikaris and G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , Journal of Computational Physics 378, 686 (2019), doi:10.1016/j.jcp.2018.10.045

    2019 doi

  9. [17]

    Udrescu and M

    S.-M. Udrescu and M. Tegmark, Ai feynman: a physics-inspired method for symbolic re- gression (2020), 1905.11481

    2020 arXiv

  10. [18]

    Udrescu, A

    S.-M. Udrescu, A. Tan, J. Feng, O. Neto, T . Wu and M. Tegmark,Ai feynman 2.0: Pareto- optimal symbolic regression exploiting graph modularity (2020), 2006.10782

    2020 arXiv

  11. [19]

    Kacmaz, R

    S. Kacmaz, R. Haas and E. A. Huerta, Machine learning-driven conservative-to-primitive conversion in hybrid piecewise polytropic and tabulated equations of state (2025), 2412. 07836

    2025

  12. [20]

    Mudimadugula, F

    R. Mudimadugula, F . Schianchi, A. Neuweiler, T . Wouters, H. Gieg and T . Dietrich,Em- ploying deep-learning techniques for the conservative-to-primitive recovery in binary neu- tron star simulations (2025), 2503.08289

    2025 arXiv

  13. [21]

    Galeazzi, W

    F . Galeazzi, W . Kastaun, L. Rezzolla and J. A. Font, Implementation of a simplified ap- proach to radiative transfer in general relativity , Phys. Rev. D 88(6), 064009 (2013), doi:10.1103/PhysRevD.88.064009, 1306.4953

    2013 arXiv

  14. [22]

    Kastaun, J

    W . Kastaun, J. V . Kalinani and R. Ciolfi,Robust recovery of primitive variables in relativistic ideal magnetohydrodynamics, Physical Review D 103(2), 023018 (2021)

    2021

  15. [23]

    J. C.-H. Wang and J.-P . Hickey , Fluxnet: A physics-informed learning-based riemann solver for transcritical flows with non-ideal thermodynamics , Com- puter Methods in Applied Mechanics and Engineering 411, 116070 (2023), doi:https://doi.org/10.1016/j.cma.2023.116070

    2023

  16. [24]

    D. G. Patsatzis, M. di Bernardo, L. Russo and C. Siettos, Gorinns: Godunov-riemann informed neural networks for learning hyperbolic conservation laws (2024), 2410.22193

    2024 arXiv

  17. [25]

    Ruggeri, I

    M. Ruggeri, I. Roy , T . Mueterthies, Michael J. and ˜Gruenwald and C. Scalo, Neural- network-based riemann solver for real fluids and high explo sives; application to compu- tational fluid dynamics, Physics of Fluids 34(11), 116121 (2022)

    2022

  18. [26]

    Magiera, D

    J. Magiera, D. Ray , J. S. Hesthaven and C. Rohde, Constraint-aware neural net- works for riemann problems , Journal of Computational Physics 409, 109345 (2020), doi:https://doi.org/10.1016/j.jcp.2020.109345

    2020

  19. [27]

    Machine learning approaches for the solution of the riemann problem in fluid dynam- ics: a case study , Communications on Applied Mathematics and Computation (2024), doi:https://doi.org/10.1007/s42967-023-00334-1

    2024 doi

  20. [28]

    Peyvan, V

    A. Peyvan, V . Oommen, A. D. Jagtap and G. E. K. adakis, Riemannonets: Interpretable neural operators for riemann problems , Computer Methods in Applied Mechanics and Engineering 426, 116996 (2024), doi:https: //doi.org/10.1016/j.cma.2024.116996

    2024

  21. [29]

    Shazeer, A

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton and J. Dean, Out- rageously large neural networks: The sparsely-gated mixture-of-experts layer (2017), 1701.06538

    2017 arXiv

  22. [30]

    J. M. Martí and E. Müller, Numerical hydrodynamics in special relativity , Living Rev. Relativ.2, 3 (1999), astro-ph/9906333. 33 SciPost Physics Submission

    1999 arXiv

  23. [31]

    Banyuls, J

    F . Banyuls, J. A. Font, J. M. Ibáñez, J. M. Martí and J. A. Miralles,Numerical 3+1 general- relativistic hydrodynamics: A local characteristic approach, Astrophys. J. 476, 221 (1997), doi:10.1086/303604

    1997 doi

  24. [32]

    Moesta, B

    P . Moesta, B. Mundim, J. Faber, S. Noble, T . Bode, R. Haas, F . Loeffler, C. Ott, C. Reisswig and E. Schnetter, General relativistic magneto-hydrodynamics with the Einstein Toolkit, In APS Meeting Abstracts, p. 10001 (2013)

    2013

  25. [33]

    Radice and L

    D. Radice and L. Rezzolla, THC: a new high-order finite-difference high-resolution shock- capturing code for special-relativistic hydrodynamics, Astron. Astrophys. 547, A26 (2012), doi:10.1051/0004-6361/201219735, 1206.6502

    2012 arXiv

  26. [34]

    Radice, L

    D. Radice, L. Rezzolla and F . Galeazzi, High-Order Numerical-Relativity Simulations of Binary Neutron Stars, In N. V . Pogorelov, E. Audit and G. P . Zank, eds.,Numerical Modeling of Space Plasma Flows ASTRONUM-2014 , vol. 498 of Astronomical Society of the Pacific Conference...

    2015 arXiv

  27. [35]

    E. R. Most, L. J. Papenfort and L. Rezzolla,Beyond second-order convergence in simulations of magnetized binary neutron stars with realistic microphysics, Mon. Not. R. Astron. Soc. 490(3), 3588 (2019), doi:10.1093 /mnras/stz2809, 1907.10328

    2019 arXiv

  28. [36]

    Rezzolla and O

    L. Rezzolla and O. Zanotti, Relativistic Hydrodynamics , Ox- ford University Press, ISBN 978-0-19-174650-5, 978-0-19-852890-6, doi:10.1093/acprof:oso/9780198528906.001.0001 (2013)

    2013

  29. [37]

    Hornik, Approximation capabilities of multilayer feedforward networks , Neural Net- works 4, 251 (1991)

    K. Hornik, Approximation capabilities of multilayer feedforward networks , Neural Net- works 4, 251 (1991)

    1991

  30. [38]

    P . J. Huber, Robust Estimation of a Location Parameter , The Annals of Mathematical Statistics 35(1), 73 (1964), doi:10.1214 /aoms/1177703732

    1964

  31. [39]

    Girshick, Fast r-cnn (2015), 1504.08083

    R. Girshick, Fast r-cnn (2015), 1504.08083

    2015 arXiv

  32. [40]

    D. P . Kingma and J. Ba, Adam: A method for stochastic optimization , arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  33. [41]

    Bengio, Practical recommendations for gradient-based training of deep architectures (2012), 1206.5533

    Y. Bengio, Practical recommendations for gradient-based training of deep architectures (2012), 1206.5533

    2012 arXiv

  34. [42]

    Pascanu, T

    R. Pascanu, T . Mikolov and Y. Bengio, On the difficulty of training recurrent neural net- works, In International Conference on Machine Learning, pp. 1310–1318 (2013)

    2013

  35. [43]

    T . Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman and C. Finn, Gradient surgery for multi-task learning, In Proceedings of the 34th International Conference on Neural Infor- mation Processing Systems, NIPS ’20. Curran Associates Inc., Red Hook, NY, USA, ISBN 9781713829546 (2020)

    2020

  36. [44]

    X.-D. Liu, S. Osher and T . Chan,Weighted Essentially Non-oscillatory Schemes, Journal of Computational Physics 115, 200 (1994), doi:10.1006 /jcph.1994.1187

    1994

  37. [45]

    Jiang and C.-W

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

    1996

  38. [46]

    C. W . Shu, High Order ENO and WENO Schemes for Computational Fluid Dynamics , In T . J. Barth and H. Deconinck, eds., High-Order Methods for Computational Physics , pp. 439–582. Springer (1999). 34 SciPost Physics Submission

    1999

  39. [47]

    J. M. Hyman,The method of lines solution of partial differential equations, Tech. Rep. COO- 3077-139, ERDA Mathematics and Computing Laboratory , Courant Institute of Mathe- matical Sciences, New York University (1976)

    1976

  40. [48]

    Ansel, E

    J. Ansel, E. Yang, H. He, N. Gimelshein, A. Jain, M. Voznesensky , B. Bao, P . Bell, D. Berard, E. Burovski, G. Chauhan, A. Chourdia et al., Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation , In Proceedings of the 29th ACM...

    2024

  41. [49]

    R. T . Q. Chen,torchdiffeq (2018)

    2018

  42. [50]

    J. M. Martí and E. Müller, Numerical hydrodynamics in special relativity , Living Rev. Relativ.6, 7; http://www.livingreviews.org/lrr (2003)

    2003

  43. [51]

    J. M. Stone, P . D. Mullen, D. Fielding, P . Grete, M. Guo, P . Kempski, E. R. Most, C. J. White and G. N. Wong, Athenak: A performance-portable version of the athena ++ amr framework (2024), 2409.16053

    2024 arXiv

  44. [52]

    J. V . Kalinani, L. Ji, L. Ennoggi, F . G. L. Armengol, L. T . Sanches, B.-J. Tsao, S. R. Brandt, M. Campanelli, R. Ciolfi, B. Giacomazzo, R. Haas, E. Schnetter et al., Asterx: a new open-source gpu-accelerated grmhd code for dynamical spacetimes (2024), 2406.11669

    2024 arXiv

  45. [53]

    Shankar, P

    S. Shankar, P . Mösta, S. R. Brandt, R. Haas, E. Schnetter and Y. de Graaf,Gram-x: a new gpu-accelerated dynamical spacetime grmhd code for exascale computing with the einstein toolkit, Classical and Quantum Gravity 40(20), 205009 (2023), doi:10.1088 /1361- 6382/acf2d9

    2023

  46. [54]

    T . G. Dietterich, Ensemble methods in machine learning , In International workshop on multiple classifier systems, pp. 1–15. Springer (2000). 35

    2000

Pith tools

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