Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Learning the Optimal Hydrodynamic Closure

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that the optimal hydrodynamic closure for linear kinetic equations is the slow spectral closure, and that its wave-number-dependent transport coefficients can be learned from density-fluctuation data, reproducing…

desk verdict The exact Shakhov spectral closure is a real analytical result, but the learned extension is fitting rather than proof of optimality, and the claims need softening before they are publishable as stated. read the letter →

arxiv 2501.13938 v3 pith:3O32ON4U submitted 2025-01-16 physics.flu-dyn math-phmath.MP

Florian Kogelbauer , Candi Zheng , Ilya Karlin This is my paper · ORCID
classification physics.flu-dynmath-phmath.MP MSC 76P0582C40 PACS 47.45.-n51.10.+y
keywords rarefiedgasdynamicshydrodynamicclosureslowspectralgeneralizedtransportcoefficientsmachinelearningdensityfluctuationsShakhovmodelDSMC
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 sets out to establish that the moment closure problem for linear kinetic equations has a uniquely optimal answer: the slow spectral closure, the constitutive law obtained on the invariant manifold spanned by the five hydrodynamic eigenmodes. It then shows that this closure, expressed through wave-number-dependent generalized transport coefficients, can be learned from density-fluctuation spectra, bypassing explicit eigenvalue calculations. The learned transport laws match Shakhov-model and DSMC spectra for Knudsen numbers up to $O(10)$, where Navier-Stokes and R13 deviate strongly, and they continue to work beyond the critical wave number where the analytic closure ceases to exist. If correct, this yields a macroscopic, five-field fluid model that is reliable at high rarefaction without any smallness assumption on the Knudsen number.

What carries the argument

The load-bearing object is the $5\times 5$ generalized transport matrix $T(k)$ in Fourier space, whose entries $\tau_i(k)$ map the five hydrodynamic moments $h=(\hat{\rho},\hat{u}_\parallel,\hat{u}_{\perp,1},\hat{u}_{\perp,2},\hat{T})$ to their time derivatives. In classical hydrodynamics these entries would be constants such as viscosity and conductivity; here they are nonlinear functions of wave number. The slow spectral closure constructs these coefficients from the hydrodynamic eigenvalue branches and the spectral temperature, while the learned model parameterizes two of them as neural outputs $N_2(k)$ and $N_6(k)$, fixes the other four, and then solves the linear system for the fluctuation spectra. This matrix carries the argument because once it is known, the density, velocity, and temperature spectra, as well as the time evolution of sharp initial profiles, are determined by the linear balance laws.

What would settle it

Train the network on a random subset of wave numbers within the Knudsen range 0 to 10, then evaluate the predicted density and velocity spectra at held-out wave numbers; systematic deviation from fresh Shakhov or DSMC outputs at those held-out values would show that the reduced parameterization is not the true optimal closure.

Watch

Extended reading notes

Core claim

The central discovery is that the dynamically optimal linear hydrodynamics of a kinetic model is not a local constitutive law but a wave-number-dependent transport matrix $T(k)$ whose entries $\tau_i(k)$ are generalized transport coefficients; moreover, these coefficients can be identified from light-scattering-type fluctuation data rather than from the operator's eigenvectors. For the linear Shakhov model, the exact spectral closure computed from the slow eigenvalues agrees with the kinetic model up to the critical wave number, and the neural-network-learned extension of the same matrix reproduces the density, velocity, and temperature spectra for Knudsen numbers up to $O(10)$, while Navier-Stokes and R13 fail. The paper interprets this agreement as empirical validation that the slow spectral closure is dynamically optimal on the linear level.

Load-bearing premise

Everything rests on the assumption that the reduced parameterization of Appendix F, with only $N_2(k)$ and $N_6(k)$ learned and $\tau_1=-k$, $\tau_3=-k$, $\tau_4=0$, $\tau_5=-2k/3$ fixed, is rich enough to represent the true optimal closure for all wave numbers up to Knudsen number 10.

Editorial extensions

If this is right

  • A closed system with only the five hydrodynamic fields reproduces kinetic-level spectra at Knudsen numbers of order 10, so costly kinetic simulations could be replaced by this macroscopic model in that regime.
  • The density-trained transport coefficients also reproduce velocity fluctuation spectra, an out-of-sample check that the learned closure captures the dynamics rather than overfitting one observable.
  • The learned coefficients evolve sharp density and temperature profiles correctly in the transient regime, indicating applicability beyond steady fluctuation spectra.
  • Because no smallness assumption on the Knudsen number enters, the same structural learning scheme can be applied to other linear kinetic models whenever fluctuation data are available.

Reading between the lines

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

  • A natural next test, not performed in the paper, is to let all six transport coefficients vary instead of fixing four; showing that the learned solution converges to the same fixed forms would confirm the reduced parameterization, while showing a better fit would mean the optimal closure was not fully identified.
  • Since the learning target is a wave-number-dependent matrix rather than trajectories, the same pipeline should transfer to the full hard-sphere Boltzmann operator, where explicit spectral data are unavailable but DSMC spectra can be generated; the paper only gestures at this.
  • The linear optimality established here suggests a route to nonlinear closures by linearizing about local Maxwellians with state-dependent learned coefficients; this is an extension the paper does not claim.
  • The exact spectral closure's critical wave number could serve as a built-in guard: outside its domain the learned coefficients are unconstrained by theory except entropy dissipation, so checking dissipation balance for every learned matrix would be a cheap physical consistency test.
Share X Bluesky LinkedIn Reddit HN

Formalized claims in Lean

  1. Claim #1: The central discovery is that the dynamically optimal linear hydrodynamics of a kinetic model is not a local constitutive law but a wave-number-dependent transport matrix $T(k)$ whose entries $\tau_i(k)$ are generalized transport coefficients; moreover, these coefficients can be identified from light-scattering-type fluctuation data rather than from the operator's eigenvectors. For the linear Shak

Signed reviews

No signed human review yet.

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 to learn the generalized transport coefficients of the slow spectral closure for linearized rarefied-gas hydrodynamics from density fluctuation spectra. It derives the exact spectral closure for the Shakhov model in Appendix D, parameterizes the transport matrix in a reduced two-function neural-network form in Appendix F, and trains the network on Shakhov and DSMC fluctuation spectra for Knudsen numbers between 0 and 10. The trained model is compared with density and velocity fluctuation spectra and with time-dependent moment advection, and is reported to outperform the Navier–Stokes and R13 models throughout this range.

Significance. If the learning scheme genuinely recovered the optimal closure beyond the critical wave number, this would be an important result: it would provide a data-driven route from kinetic theory to stable macroscopic hydrodynamics over a wide rarefaction range. The explicit Shakhov spectral closure and the closed-form density spectrum in Eq. (B36) are valuable analytical ingredients, and the comparisons to Shakhov and DSMC reference data are appropriate benchmarks. However, the central optimality claim is not yet supported, because the reduced parameterization is not proven sufficient to represent the exact closure and because the learned model is fitted to the very spectra used for validation.

major comments (3)
  1. [Appendix F, Eq. (F2); Section IV] The learned model fixes tau1 = -k, tau3 = -k, tau4 = 0, tau5 = -2k/3 and learns only tau2 and tau6 through N2 and N6. The density spectrum in Eq. (B36) depends explicitly on tau1, tau3, tau4, and tau5, and Figure 6 shows that the exact Shakhov transport coefficients deviate from their leading-order small-wave-number forms as k grows. Since no proof is given that these fixed forms remain valid up to Kn = 10, the two learned functions can compensate for errors in the fixed coefficients during fitting, so the agreement in Figures 2 and 3 does not by itself establish that the learned model coincides with the optimal spectral closure. I recommend comparing the learned tau_i with the exact formulas (A15) below the critical wave number and performing an ablation or a full-parameter learning run to test the sufficiency of the two-function parameterization.
  2. [Appendix E, Eq. (E2); Section III; Figure 3] The claim in Section III that velocity fluctuations provide an out-of-sample test is contradicted by the training loss in Eq. (E2), which includes the term (rho u_parallel - rho u_parallel^spec)^2. The velocity spectra shown in Figure 3 are therefore part of the training objective, not an independent validation. Similarly, the density spectra in Figure 2 are of the same kind used in the loss. To support the generalization claim, the authors should train on density spectra only and validate on velocity and temperature spectra held out from the loss, or otherwise exclude the test observables from training.
  3. [Section III; Section IV; Abstract] The manuscript states that the learned model proves the optimality of the slow spectral closure and that any other closure is less accurate, but the learned model is not the slow spectral closure: it is a two-function surrogate fitted to kinetic data. The exact spectral closure is independently derived up to the critical wave number, but the extension beyond criticality is one of infinitely many possible closures, and the data-fitting procedure does not select it by an optimality criterion. The paper should either soften the optimality claim to describe a data-driven closure in the spectral-closure family or provide a quantitative optimality test, for example comparing the learned transport coefficients with the exact Shakhov coefficients below kcrit and testing against genuinely held-out data.
minor comments (4)
  1. [Throughout] There are several typographical errors, for example 'constitute laws' should be 'constitutive laws', 'obstinate' should be 'obtained', 'equilibirum' should be 'equilibrium', and 'extended extended hydrodynamicist' in the introduction is garbled.
  2. [Section III; Figure 2] The text says the learned hydrodynamics 'extends to arbitrarily large wave numbers', but the training and inference are restricted to Knudsen numbers from 0 to 10, and the paper reports slight deviations at Kn = 10; 'arbitrarily large' is stronger than what is demonstrated.
  3. [Appendix E; Section III] The manuscript does not state whether the trained neural-network parameters, generation scripts, or DSMC/Shakhov data sets are publicly available; for a machine-learning paper this information would materially aid reproducibility.
  4. [Section II; Appendix D] The statement that dynamic optimality of the spectral closure is a direct consequence of the time-scale separation could use a specific reference or proof in the present paper, since the claim is used to interpret all subsequent numerical comparisons.

Circularity Check

1 steps flagged · score 6.0 of 10

The learned model's density-spectrum agreement (Fig. 2) is the same quantity minimized by the training loss, so that headline validation is by construction; exact spectral closure and out-of-sample velocity/advection tests provide independent content, making the circularity partial.

  1. fitted input called prediction [Section III, Eq. (E2); Section IV, Fig. 2]
    "The training procedure minimizes the mean squared error (MSE) between predicted and reference spectra. Let ρ^2_spec(k,ω), ρu∥,spec(k,ω), and ρT_spec(k,ω) be the reference spectra and define the loss function L as ... The learned hydrodynamic closure (red curve in Figure 2) is also in close agreement with the both the Shakhov as well as the DSMC data."

    The network's learned outputs N2(k) and N6(k) enter the transport coefficients τ2 and τ6 via Eq. (F2), and Eq. (B36) makes the predicted density fluctuation spectrum an explicit function of those τ_i. The loss (E2) is minimized against exactly the Shakhov/DSMC density, ρu∥ and ρT spectra. Therefore the agreement of the red learned curves with the blue/purple reference curves in Fig. 2 is the training objective itself, not an independent check. Presenting this agreement as evidence that the learned closure is accurate and that the spectral closure is optimal is thus a fitted-input validation. The paper does label Fig.

full rationale

The exact spectral closure for k up to k_crit is a genuine first-principles derivation: the paper gives the Shakhov spectrum and the closed-form transport coefficients (A14)-(A15), and the agreement of the exact brown curves with the full Shakhov and DSMC spectra is not a fit. That part is self-contained and non-circular. The circularity is confined to the learned extension: the network is trained with a weighted MSE loss (E2) on density, ρu∥ and ρT spectra, and the same density spectra are then exhibited in Fig. 2 as a main result. Because the loss is minimized against those reference curves, the close match is statistically forced rather than independently predicted. Two features limit the severity: the paper explicitly calls Fig. 2 a comparison to training data, and it provides genuinely out-of-sample tests in the velocity fluctuation spectra (Fig. 3) and the sharp-density advection problem (Fig. 4). Those tests support generalization but do not convert the Fig. 2 density-spectrum agreement into an independent test. A further non-circular but important gap is that Eq. (F2) fixes τ1, τ3, τ4 and τ5 to their small-k Chapman-Enskog/Navier-Stokes forms while learning only τ2 and τ6; Figure 6's caption states that the exact Shakhov coefficients deviate from those leading-order approximations for larger wave numbers, and the density spectrum (B36) depends on all τ_i, so the learned N2 and N6 could compensate for fixed forms without the model actually being the exact spectral closure. This is a representation/identifiability limitation rather than a circular reduction, so it does not by itself raise the score. The optimality theorem is also cited to the authors' prior Ref. [31], but because the present paper derives the explicit Shakhov closure and compares against independent kinetic and DSMC benchmarks, that normal self-citation is not treated as a separate circular step.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on prior spectral closure theory, on Shakhov and DSMC data as ground truth, and on a reduced two-function parameterization of the transport coefficients. No new physical entities such as particles, forces, or dimensions are introduced.

free parameters (3)
  • N2(k), learned correction to tau2 = Neural network output, inferred from data
    The value of N2(k) is fit to the reference Shakhov and DSMC spectra through the loss function (E2).
  • N6(k), learned correction to tau6 = Neural network output, inferred from data
    The value of N6(k) is fit to the reference spectra, and it directly controls tau6 in Eq. (F2).
  • tau1, tau3, tau4, tau5 functional forms = tau1=-k, tau3=-k, tau4=0, tau5=-2k/3
    These transport coefficients are fixed by hand in Eq. (F2) rather than derived from the exact Shakhov closure or learned from data; the paper gives no justification that this reduced set spans the optimal closure beyond criticality.
assumptions (4)
  • domain assumption The five hydrodynamic eigenmodes of the linear kinetic operator form an attracting slow manifold and define a dynamically optimal linear closure.
    This is the central theorem of slow spectral closure from prior work by the same authors, cited as [31] and [39]; it is not reproved here and is used as the basis for the transport matrix structure.
  • domain assumption The transport matrix has the block structure of Eq. (7) with nonzero entries only among density, longitudinal velocity, and temperature.
    The paper assumes this spectral-closure structure for all k, including beyond the critical wave number where the analytical closure ceases to exist.
  • ad hoc to paper The simplified two-function parameterization in Eq. (F2) is sufficient to represent the optimal generalized transport coefficients.
    This restriction is introduced for numerical stability and training simplicity; the paper does not prove that the fixed tau1, tau3, tau4, and tau5 forms are dynamically optimal beyond the critical wave number.
  • domain assumption The linearized Shakhov model and the VHS DSMC simulation are accurate ground truth for rarefied gas fluctuation spectra.
    The paper uses Shakhov and DSMC data as the reference kinetic truth and does not compare against experimental Rayleigh-Brillouin scattering data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning the Optimal Hydrodynamic Closure." pith.science (2026). https://pith.science/paper/3O32ON4U

@misc{pith2026250113938,
  author       = {Pith},
  title        = {Pith review of: Learning the Optimal Hydrodynamic Closure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3O32ON4U}},
  note         = {Machine review of arXiv:2501.13938}
}
read the original abstract

We present the optimal hydrodynamic model for rarefied gas flows relative to a given kinetic model by combining the recent theory of slow spectral closure with machine learning techniques. We learn generalized transport coefficients from density fluctuation data for the Shakhov model as well as Monte Carlo Simulations and demonstrate that our approach decisively outperforms previously proposed constitutive laws for higher-order hydrodynamics. The novel hydrodynamic model is in close alignment with the underlying kinetic models, thus proving the optimality of the slow spectral closure. Our theory is independent on any smallness assumption of the Knudsen number and is formulated solely in terms of macroscopic observables.

Figures

Figures reproduced from arXiv: 2501.13938 by the authors.

Figure 1
Figure 1. FIG. 1. Schematics of the learning algorithm: The neural net [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Comparison of density fluctuation curves (horizontal axis temporal frequency [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Comparison of velocity fluctuation curves (horizontal axis temporal frequency [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Comparison of moment advection for density and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Schematics of the light-scattering experiment: An [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. The transport coefficients of the linear Shakhov model in dependence on wave number (0 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. The spatial Fourier transform of the initial den [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. On the Relation of Exact Hydrodynamics to the Chapman-Enskog Series

    math-ph 2025-06 conditional novelty 4.0 of 10

    For a one-dimensional BGK kinetic model, the Chapman-Enskog expansion is a local Taylor approximation to the exact spectral hydrodynamic mode and diverges for every nonzero wave number, while the spectral closure rema...

Reference graph

Works this paper leans on

62 extracted references · 55 canonical work pages · cited by 1 Pith paper

  1. [31]

    Cabr´ e, E

    X. Cabr´ e, E. Fontich, and R. De la Llave, The param- eterization method for invariant manifolds I: manifolds associated to non-resonant subspaces, Indiana University mathematics journal , 283 (2003)

    work page 2003

  2. [1]

    V. E. Zakharov, V. S. L’vov, and G. Falkovich, Kolmogorov spectra of turbulence I: Wave turbulence (Springer Science & Business Media, 2012)

    2012

  3. [2]

    The analytic closure serves as a further benchmark to ensure the accuracy of the learned closure up to the cirtical wave number

    is also in close agreement with the both the Shakhov as well as the DSMC data. The analytic closure serves as a further benchmark to ensure the accuracy of the learned closure up to the cirtical wave number. While the ana- lytical closure is limited to frequencies below the critical wave number, the learned hydrodynamics extends to ar- bitrarily large wav...

    work page

  4. [3]

    As a further illustration of the optimality of the spec- trally closed hydrodynamics, we compute the time evolu- tion of density and temperature directly

    Again, the the learned curves show excellent agree- ment with the Shakhov data, while the R13 and Navier– Stokes spectra show huge deviations for larger Knudsen numbers. As a further illustration of the optimality of the spec- trally closed hydrodynamics, we compute the time evolu- tion of density and temperature directly. Figure 4 shows the advection of ...

    work page

  5. [4]

    Karplus, Development of multiscale models for com- plex chemical systems: From h+ h 2 to biomolecules (no- bel lecture)., Angewandte Chemie International Edition 53(2014)

    M. Karplus, Development of multiscale models for com- plex chemical systems: From h+ h 2 to biomolecules (no- bel lecture)., Angewandte Chemie International Edition 53(2014)

    work page 2014

  6. [5]

    Q. Zeng, A. Yu, and G. Lu, Multiscale modeling and sim- ulation of polymer nanocomposites, Progress in polymer science33, 191 (2008)

    work page 2008

  7. [6]

    S. A. Schaaf and P. L. Chambr´ e,Flow of rarefied gases, Vol. 8 (Princeton University Press Princeton, 1961)

    work page 1961

  8. [7]

    Akhlaghi, E

    H. Akhlaghi, E. Roohi, and S. Stefanov, A comprehensive review on micro-and nano-scale gas flow effects: Slip- jump phenomena, Knudsen paradox, thermally-driven flows, and Knudsen pumps, Physics Reports997, 1 (2023)

    work page 2023

Show all 62 references
  1. [8]

    Bird,Molecular Gas Dynamics and the Direct Simu- lation of Gas Flows(Clarendon Press, 1994)

    G. Bird,Molecular Gas Dynamics and the Direct Simu- lation of Gas Flows(Clarendon Press, 1994)

    1994

  2. [9]

    Jenny, M

    P. Jenny, M. Torrilhon, and S. Heinz, A solution algo- rithm for the fluid dynamic equations based on a stochas- 9 tic model for molecular motion, Journal of computational physics229, 1077 (2010)

    2010

  3. [10]

    Shakhov, Generalization of the Krook kinetic relax- ation equation, Fluid dynamics3, 95 (1968)

    E. Shakhov, Generalization of the Krook kinetic relax- ation equation, Fluid dynamics3, 95 (1968)

    1968

  4. [11]

    Struchtrup and M

    H. Struchtrup and M. Torrilhon, Regularization of Grad’s 13 moment equations: Derivation and linear analysis, Physics of Fluids15, 2668 (2003)

    2003

  5. [12]

    Wu and X.-J

    L. Wu and X.-J. Gu, On the accuracy of macroscopic equations for linearized rarefied gas flows, Advances in Aerodynamics2, 1 (2020)

    2020

  6. [13]

    Zwanzig,Nonequilibrium statistical mechanics(Ox- ford university press, 2001)

    R. Zwanzig,Nonequilibrium statistical mechanics(Ox- ford university press, 2001)

    2001

  7. [14]

    Hilbert, Mathematical problems, Bulletin of American Mathematical Society37, 407 (2000)

    D. Hilbert, Mathematical problems, Bulletin of American Mathematical Society37, 407 (2000)

    2000

  8. [15]

    Saint-Raymond,Hydrodynamic limits of the Boltz- mann equation, 1971 (Springer Science & Business Me- dia, 2009)

    L. Saint-Raymond,Hydrodynamic limits of the Boltz- mann equation, 1971 (Springer Science & Business Me- dia, 2009)

    1971

  9. [16]

    Chapman and T

    S. Chapman and T. Cowling,The Mathematical Theory of Non-uniform Gases, Cambridge Mathematical Library (Cambridge University Press, Cambridge, 1990)

    1990

  10. [17]

    Grad, On the kinetic theory of rarefied gases, Commu- nications on pure and applied mathematics2, 331 (1949)

    H. Grad, On the kinetic theory of rarefied gases, Commu- nications on pure and applied mathematics2, 331 (1949)

    1949

  11. [18]

    Burnett, The distribution of velocities in a slightly non-uniform gas, Proceedings of the London Mathemat- ical Society2, 385 (1935)

    D. Burnett, The distribution of velocities in a slightly non-uniform gas, Proceedings of the London Mathemat- ical Society2, 385 (1935)

    1935

  12. [19]

    A. V. Bobylev, Instabilities in the Chapman–Enskog ex- pansion and hyperbolic Burnett equations, Journal of statistical physics124, 371 (2006)

    2006

  13. [20]

    A. V. Bobylev, The Chapman–Enskog and Grad methods for solving the Boltzmann equation, inAkademiia Nauk SSSR Doklady, Vol. 262 (1982) pp. 71–75

    1982

  14. [21]

    Slemrod, From Boltzmann to Euler: Hilbert’s 6th problem revisited, Computers & Mathematics with Ap- plications65, 1497 (2013), Grasping Complexity

    M. Slemrod, From Boltzmann to Euler: Hilbert’s 6th problem revisited, Computers & Mathematics with Ap- plications65, 1497 (2013), Grasping Complexity

    2013

  15. [22]

    Kogelbauer and I

    F. Kogelbauer and I. Karlin, On the relation of exact hydrodynamics to the Chapman–Enskog series, arXiv preprint arXiv:2506.17441 (2025)

    2025 arXiv

  16. [23]

    A. N. Gorban and I. V. Karlin,Invariant Manifolds for Physical and Chemical Kinetics, Lecture Notes in Physics, Vol. 660 (Springer Science & Business Media, Heidelberg, 2005)

    2005

  17. [24]

    Gorban and I

    A. Gorban and I. Karlin, Hilbert’s 6th problem: Ex- act and approximate hydrodynamic manifolds for kinetic equations, Bulletin of the American Mathematical Soci- ety51, 187 (2014)

    2014

  18. [25]

    E. N. Lorenz, The slow manifold–what is it?, Journal of the atmospheric sciences49, 2449 (1992)

    1992

  19. [26]

    MacKay, Slow manifolds, inEnergy localisation and Transfer(World Scientific, 2004) pp

    R. MacKay, Slow manifolds, inEnergy localisation and Transfer(World Scientific, 2004) pp. 149–192

    2004

  20. [27]

    M. R. Roussel and S. J. Fraser, Invariant manifold meth- ods for metabolic model reduction, Chaos: An Interdis- ciplinary Journal of Nonlinear Science11, 196 (2001)

    2001

  21. [28]

    Ax ˚ as, M

    J. Ax ˚ as, M. Cenedese, and G. Haller, Fast data-driven model reduction for nonlinear dynamical systems, Non- linear Dynamics111, 7941 (2023)

    2023

  22. [29]

    Fenichel, Geometric singular perturbation theory for ordinary differential equations, Journal of differential equations31, 53 (1979)

    N. Fenichel, Geometric singular perturbation theory for ordinary differential equations, Journal of differential equations31, 53 (1979)

    1979

  23. [30]

    de La Llave, Invariant manifolds associated to nonres- onant spectral subspaces, Journal of statistical physics 87, 211 (1997)

    R. de La Llave, Invariant manifolds associated to nonres- onant spectral subspaces, Journal of statistical physics 87, 211 (1997)

    1997

  24. [32]

    R. S. Ellis and M. A. Pinsky, The first and second fluid approximations to the linearized Boltzmann equation, J. Math. Pures Appl54, 125 (1975)

    1975

  25. [33]

    Kogelbauer and I

    F. Kogelbauer and I. Karlin, Rigorous hydrodynamics from linear Boltzmann equations and viscosity-capillarity balance, Phys. Rev. E110, 055105 (2024)

    2024

  26. [34]

    X.-H. Zhou, J. Han, and H. Xiao, Learning nonlocal con- stitutive models with neural networks, Computer Meth- ods in Applied Mechanics and Engineering384, 113927 (2021)

    2021

  27. [35]

    J. Hana, C. Ma, Z. Ma, and E. Weinan, Uniformly accurate machine learning-based hydrodynamic mod- els for kinetic equations, Proceedings of the National Academy of Sciences of the United States of America 10.1073/pnas.1909854116 (2019)

    2019 doi

  28. [36]

    Zheng, Y

    C. Zheng, Y. Wang, and S. Chen, Data-driven constitu- tive relation reveals scaling law for hydrodynamic trans- port coefficients, Physical Review E107, 015104 (2023)

    2023

  29. [37]

    Q. Lou, X. Meng, and G. E. Karniadakis, Physics- informed neural networks for solving forward and in- verse flow problems via the Boltzmann-BGK formula- tion, Journal of Computational Physics447, 110676 (2021)

    2021

  30. [38]

    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 physics378, 686 (2019)

    2019

  31. [39]

    Cercignani,The Boltzmann Equation(Springer New York, 1988)

    C. Cercignani,The Boltzmann Equation(Springer New York, 1988)

    1988

  32. [40]

    Risken,The Fokker–Planck Equation(Springer New York, 1996)

    H. Risken,The Fokker–Planck Equation(Springer New York, 1996)

    1996

  33. [41]

    Kogelbauer and I

    F. Kogelbauer and I. Karlin, Exact hydrodynamic mani- folds for the linear Boltzmann BGK equation I: spectral theory, Continuum Mechanics and Thermodynamics , 1 (2024)

    2024

  34. [42]

    C. D. Levermore, Moment closure hierarchies for kinetic theories, Journal of statistical Physics83, 1021 (1996)

    1996

  35. [43]

    I. V. Karlin, A. N. Gorban, G. Dukek, and T. Nonnen- macher, Dynamic correction to moment approximations, Physical Review E57, 1668 (1998)

    1998

  36. [44]

    Karlin, M

    I. Karlin, M. Colangeli, and M. Kr¨ oger, Exact linear hy- drodynamics from the Boltzmann equation, Physical re- view letters100, 214503 (2008)

    2008

  37. [45]

    R. B. Miles, W. R. Lempert, and J. N. Forkey, Laser Rayleigh scattering, Measurement science and technology 12, R33 (2001)

    2001

  38. [46]

    Y. Wang, Z. Gu, K. Liang, and W. Ubachs, Rayleigh– Brillouin light scattering spectroscopy of air; experiment, predictive model and dimensionless scaling, Molecular Physics119, e1804635 (2021)

    2021

  39. [47]

    V. E. Ambru¸ s, F. Sharipov, and V. Sofonea, Comparison of the Shakhov and ellipsoidal models for the Boltzmann equation and DSMC for ab initio-based particle interac- tions, Computers & Fluids211, 104637 (2020)

    2020

  40. [48]

    Kogelbauer and I

    F. Kogelbauer and I. Karlin, Spectral analysis and hy- drodynamic manifolds for the linearized Shakhov model, Physica D: Nonlinear Phenomena459, 134014 (2024)

    2024

  41. [49]

    P. L. Bhatnagar, E. P. Gross, and M. Krook, A model for collision processes in gases. I. Small amplitude processes in charged and neutral one-component systems, Physical Review94, 511 (1954). 10

    1954

  42. [50]

    Hansen and I

    J.-P. Hansen and I. R. McDonald,Theory of simple liq- uids: with applications to soft matter(Academic press, 2013)

    2013

  43. [51]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville,Deep learning (MIT press, 2016)

    2016

  44. [52]

    D. P. Kingma, Adam: A method for stochastic optimiza- tion, arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  45. [53]

    Palczewski, Spectral properties of the space nonhomo- geneous linearized Boltzmann operator, Transport The- ory and Statistical Physics13, 409 (1984)

    A. Palczewski, Spectral properties of the space nonhomo- geneous linearized Boltzmann operator, Transport The- ory and Statistical Physics13, 409 (1984)

    1984

  46. [54]

    Wiener, Generalized harmonic analysis, Acta mathe- matica55, 117 (1930)

    N. Wiener, Generalized harmonic analysis, Acta mathe- matica55, 117 (1930)

    1930

  47. [55]

    L. D. Landau, J. S. Bell, M. Kearsley, L. Pitaevskii, E. Lifshitz, and J. Sykes,Electrodynamics of continuous media, Vol. 8 (elsevier, 2013)

    2013

  48. [56]

    E. M. Lifshitz and L. P. Pitaevskii,Statistical physics: theory of the condensed state, Vol. 9 (Elsevier, 2013)

    2013

  49. [57]

    D. G. Anderson, Iterative procedures for nonlinear inte- gral equations, J. ACM12, 547–560 (1965)

    1965

  50. [58]

    Kogelbauer and I

    F. Kogelbauer and I. Karlin, Exact hydrodynamic mani- folds for the linear Boltzmann BGK equation II: Spectral closure, arXiv preprint arXiv:2306.07103 (2023). 11 Appendix A: General Spectral Properties of Kinetic Operators and Spectral Closure In this appendix, we recall the ...

    2023 arXiv

  51. [59]

    A parameter-predicting sub-network,NetNN, which maps the input wave numberkto a set of interme- diate parameters

  52. [60]

    Parameter-Predicting Network (NetNN) The networkNetNNis defined by a feedforward archi- tecture: •Input:A single scalar input, the wave numberk

    Aspectral solver, which uses these parameters to compute the full set of generalized transport co- efficients{τ i}1≤j≤6 and thereafter the fluctuation spectra. Parameter-Predicting Network (NetNN) The networkNetNNis defined by a feedforward archi- tecture: •Input:A single scal...

  53. [61]

    Larger spectral magnitudes do not disproportion- ately dominate the training loss, as we normalize byM(k) 2

  54. [62]

    The factor 1/(1 + 0.2k) avoids overemphasis on largek-values, leading to a more balanced training across scales. Appendix G: Computing Time Evolution from F requency–W ave number Spectra In this appendix, we describe how to compute the time evolution of density and temperature...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.