Pith. sign in

REVIEW 3 major objections 6 minor 52 references

Extracting informative vortical structures of turbulent wake-extreme vortex gust interactions with machine learning

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read A convolutional information-theoretic network splits extreme gust–airfoil flow snapshots into the vortical pieces that carry future lift or energy transfer and the residual that does not.

desk verdict Solid extension of the authors’ causal-decomposition tool to 3D extreme gusts at Re=5000, with useful dual-target physics; the “causal” label still outruns the evidence. read the letter →

arxiv 2607.26683 v1 pith:UBI4ERLM submitted 2026-07-29 physics.flu-dyn physics.comp-ph

Ryo Koshikawa , Kai Fukami This is my paper · ORCID
classification physics.flu-dynphysics.comp-ph
keywords extremeaerodynamicsvortexgustinformativemodedecompositioninformation-theoreticlearningliftcoefficientscale-dependentenergytransferconvolutionalneuralnetworksturbulentwake
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

Extreme vortex gusts hitting an airfoil at Reynolds number 5000 produce fast, multiscale wakes in which it is hard to tell which structures actually set later aerodynamic forces or energy exchange. This paper trains a three-dimensional convolutional network with an information-theoretic objective so that any single Q-criterion snapshot is decomposed into an informative part that maximises mutual information with a chosen future scalar and a residual that is independent of that part. When the scalar is future lift, the network first highlights large vortex cores and, after massive separation, also the emerging shear layers, in qualitative agreement with instantaneous force-element maps. When the scalar is scale-dependent energy transfer, the extracted structures differ from the lift-based ones before impingement yet become similar afterward. The method therefore lets a researcher ask, from data alone, which pieces of a transient turbulent field are tied to the physics of interest.

What carries the argument

Informative mode decomposition: a deep-sigmoidal-flow 3-D CNN that realises q = q_I + q_R by maximising mutual information I(λ; q_I) while enforcing I(q_R; q_I) = 0 through non-negative weights and bijective activations, with the future target λ and time lag Δt supplied as inputs.

What would settle it

In a new simulation, locally suppress or perturb only the extracted informative structures and measure whether the subsequent lift (or integrated energy transfer) changes far more than when the residual structures are perturbed by the same amplitude; comparable effects would falsify the causal claim.

Watch

Extended reading notes

Core claim

Convolutional information-theoretic learning decomposes a Q-criterion snapshot of an extreme vortex-gust–airfoil interaction into informative and residual components with respect to an arbitrary future target. For future lift the extracted modes are mainly vortex cores before impingement and additionally shear layers after separation, matching force-element locations; for scale-dependent energy transfer the modes are distinct early yet converge (spatial cosine similarity rising toward 0.8) after impingement, showing a shared structural basis across the two mechanisms.

Load-bearing premise

That maximising mutual information between a network-extracted field piece and a future scalar, under the stated architectural constraints, isolates structures that dynamically cause that scalar rather than merely correlating with it.

Editorial extensions

If this is right

  • Lift-driving structures can be read from experimental snapshots without constructing an auxiliary potential for force-element analysis.
  • The same snapshot can be re-decomposed for different targets (lift versus energy transfer) to compare mechanisms side by side.
  • Varying the time lag Δt reveals which structures matter on short versus longer horizons.
  • Spatial maps of informative modes can indicate where and when to actuate for gust-load control.
  • The framework applies in principle to any transient aerodynamic flow that supplies snapshots and a scalar target.

Reading between the lines

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

  • Because only observables are required, the same extractor could be trained on multi-camera PIV of laboratory gust encounters and still flag the post-separation shear layers.
  • The post-impingement rise in similarity between lift and energy modes suggests a single set of structures whose suppression might reduce both force spikes and cascade activity.
  • Replacing the mutual-information loss with an interventional (do-operator) loss could further shrink the extracted set to the true causal skeleton.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. The manuscript applies convolutional information-theoretic mode decomposition to LES of extreme vortex-gust–airfoil interactions at Re=5000 (G=±{2,5}). A 3D CNN (deep sigmoidal flow with non-negative weights) decomposes each Q-criterion snapshot into informative and residual fields by maximizing mutual information with a future scalar λ while enforcing residual independence (Eqs. 1–3). Two choices of λ are examined: lift coefficient CL and domain-integrated scale-dependent energy transfer T. Lift-targeted modes are compared visually to instantaneous volume force elements; energy-targeted modes are compared to lift-targeted modes via a spatial cosine-similarity time series that rises from ~0.4 pre-impingement to ~0.8 after. The authors conclude that the method selectively extracts structures responsible for the chosen future physics and can support causal, data-driven study of transient aerodynamics.

Significance. If the extracted fields are accepted as physically meaningful, the work usefully extends information-theoretic decomposition from prior 2D/separated-wake settings to three-dimensional turbulent extreme-gust encounters and, importantly, demonstrates target-dependence across two distinct aerodynamic mechanisms (force vs inter-scale transfer). Strengths include a well-documented LES setup, multi-gust-ratio coverage, an external physical check against force elements (independent of the MI objective), Q–R topology diagnostics, and an explicit cross-target comparison. The framework is in principle applicable to experimental data where auxiliary potentials are unavailable. These are genuine contributions to data-driven analysis of extreme aerodynamics, provided the causal reading is either validated or carefully scoped.

major comments (3)
  1. [§2.1, Eqs. (2)–(3); Abstract; §3.1] §2.1, Eqs. (2)–(3) and the abstract/title framing: the extractor F is conditioned on the realized future scalar λ(t+Δt), and the loss only balances reconstruction against residual independence; H(λ|q_I)=0 is imposed by non-negative weights and bijective activations rather than by an explicit out-of-sample predictive objective. Among MI-sufficient subsets, the L2 term therefore favors large-norm features. The manuscript equates these fields with structures “causally responsible” for future λ, yet provides no check that q_I alone retains predictive skill for λ while q_R does not, nor any interventional test. The force-element comparison (Eq. 7) is instantaneous and only qualitative (isosurface overlap in Figs. 4–5), so it does not validate the future-oriented causal claim. Either (i) add a simple hold-out prediction or ablation (forecast λ from q_I vs q_R) or (ii) systematically replace “c
  2. [§3.1, Figs. 4–5, Eq. (7)] §3.1, Figs. 4–5 and Eq. (7): agreement with lift elements is asserted solely by visual isosurface correspondence. Because force elements diagnose instantaneous force generation while q_I is defined with respect to CL(t+Δt), a quantitative spatial metric (e.g., cosine similarity or overlap of thresholded supports between |Le| and q_I as functions of Δt and t) is needed to make the comparison falsifiable and to show that the time-delay effect in Fig. 6 is not merely a change in visualization. Without it, the external physical check remains impressionistic.
  3. [§3.2, Eq. (13), Fig. 9] §3.2, Eq. (13) and Fig. 9: the claimed analogy between lift- and energy-targeted structures rests on a single cosine-similarity time series for G=2 (rising to ~0.8). The manuscript does not report the same metric for G=5 or the negative-gust cases, nor sensitivity to the band-pass scales defining T (σ_max, σ1) or to Δt. Given that free parameters (β via L-curve, Δt, filter scales, Q_th) are numerous, at least one additional gust ratio and a brief sensitivity check on the similarity curve are required before the cross-mechanism “analogy after impingement” can be treated as a robust result rather than a single-case observation.
minor comments (6)
  1. [§2.1] The L-curve used to fix β is cited (§2.1) but never shown; a brief inset or appendix panel would make the regularization choice reproducible.
  2. [§3.1] POD citation is broken in the text (“proper orthogonal decomposition [?, POD;]]lumley1967structure”). Restore a proper Lumley/Sirovich reference.
  3. [§2.1, Fig. 2] Fig. 2 architecture: filter sizes and channel counts are given, but training set size (number of snapshots × cases), batch size, learning rate, and early-stopping patience are not stated. Add a short paragraph or table for reproducibility.
  4. [Figs. 4–6, 8–9] Q_th = 0.1 is used uniformly for visualization; state whether informative/residual fields are thresholded identically and whether conclusions change under modest threshold variation.
  5. [Acknowledgements] Acknowledgements list R.A. and Q.L. who are not on the author line; clarify contributions or authorship.
  6. [Introduction; §2.1] Minor prose: “asextreme” (p.1), “that decomposes” fragment in §2.1 opening, and “R.A.” vs author list consistency. A careful copy-edit pass is needed.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild self-definitional labeling: ‘informative/causal’ structures are those that maximize MI with λ by construction; physical checks (force elements, cross-target similarity) remain independent.

  1. self definitional [§2.1 Eqs. (1)–(3); Abstract; §3.1 Eq. (8)]
    "This extraction is achieved by decomposing a given vortical flow snapshot into its informative and residual components based on the contribution to an arbitrary future target variable with convolutional information-theoretic learning. ... Specifically, the informative component q_I is extracted throughout the optimization that maximizes the mutual information I(λ;q_I)=H(λ)−H(λ|q_I) ... q_I(t)=F(λ(t+Δt),q(t);w)"

    Structures labeled ‘informative’ or ‘causally important’ for future λ are exactly the subset the network is trained to retain under the MI objective (plus residual independence). Asserting that the extracted modes are informative for that same λ therefore holds by the definition of the loss, not by an independent derivation. The paper’s external force-element comparison and cross-target cosine-similarity trends are not forced by this definition and remain non-circular empirical content.

full rationale

The paper’s extractor defines q_I by maximizing mutual information with a chosen future scalar λ (lift or scale-dependent energy transfer) subject to residual independence, implemented via a 3D CNN conditioned on λ(t+Δt). Consequently, calling the output ‘informative with respect to λ’ restates the training objective and is partly tautological. That is the only clear circular step. The load-bearing scientific claims—which spatial structures appear (cores vs shear layers), qualitative overlap with instantaneous force-element maps (an independent continuum identity), the Δt dependence, Q–R topology shifts, and the empirical rise of cosine similarity between lift-targeted and energy-targeted modes after impingement—are data-dependent outcomes of applying the method to LES snapshots, not algebraic reductions of the loss or of self-cited uniqueness theorems. Self-citations to the authors’ prior method papers and to Arranz & Lozano-Durán supply the tool; they do not force the flow-physics conclusions. No fitted parameter is renamed a prediction, and no uniqueness result is imported to forbid alternatives. Score 2 reflects one minor definitional circularity that does not collapse the central empirical content.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the information-theoretic decomposition framework (imported), the adequacy of LES Q-fields as the state, the identification of MI with causality, several hand-chosen hyperparameters (β, Δt, architecture, filter scales), and qualitative visual validation against force elements. No new physical entity is postulated; the ‘informative component’ is an operational output of the trained map F.

free parameters (6)
  • β (MI vs reconstruction loss weight) = not numerically reported; L-curve selected
    Balances ||q−q_I||² against ||I(q_R;q_I)||² in Eq. 3; chosen by L-curve analysis, not derived.
  • Δt (cause–effect time delay) = 0.05 (primary); 0.5 (comparison)
    User-defined horizon in q_I(t)=F(λ(t+Δt),q(t)); results shown for 0.05 and 0.5; changes which structures are kept.
  • CNN architecture (filter sizes H, channel counts N, depth) = as in Fig. 2(b)
    Fully specified in Fig. 2 but chosen by the authors; capacity is asserted to be sufficient to reconstruct the full field if desired.
  • Q_th visualization threshold = 0.1
    Iso-surface level Q_th=0.1 used throughout figures; affects what the reader sees as ‘extracted’.
  • σ_max and band-pass scales for energy transfer T = σ_max from undisturbed St
    σ_max=c/(2π St) from baseline shedding; bands [σ1,2σ1] with σ1∈{σ_max,σ_max/4}; defines the target scalar T.
  • POD mode count for baseline comparison = 12
    12 leading POD modes used to argue that energy-dominant structures differ from informative ones.
assumptions (6)
  • domain assumption Mutual information I(λ;q_I) with I(q_R;q_I)=0 is a valid operational definition of causal informativeness for fluid structures.
    Imported from Arranz & Lozano-Durán and Shannon; invoked in §2.1 as the optimization target without independent causal proof in this flow.
  • domain assumption The second invariant Q of the velocity-gradient tensor is an adequate given state for capturing the vortical motions of interest.
    Stated in §2.1; standard but not unique (vorticity, λ2, etc. could differ).
  • domain assumption LES at Re=5000 with the stated Cliff discretization and Taylor-vortex gust adequately represents the extreme interaction physics being claimed.
    Data curation §2.2; resolution and subgrid model details deferred to Fukami et al. [18].
  • domain assumption Volume lift element dominates surface contribution at Re=5000, so Le is a fair instantaneous benchmark for lift-related structures.
    Stated in §3.1 citing prior work; used to validate informative modes visually.
  • ad hoc to paper Non-negative CNN weights plus bijective activations enforce H(λ|q_I)=0 sufficiently for the extracted q_I to be information-theoretically valid.
    §2.1 / deep sigmoidal flow construction; practical surrogate for the information constraint.
  • standard math Standard calculus and Shannon entropy definitions hold for the discrete, interpolated field samples used in training.
    Background for Eqs. 1–4.
invented entities (1)
  • Informative mode extractor F (3D convolutional deep sigmoidal flow)
    purpose: Maps (q(t), λ(t+Δt)) to q_I(t) realizing the informative/residual split for arbitrary future targets.
    Operational neural architecture realizing the Arranz–Lozano-Durán decomposition in 3D; not a new physical object, but the paper’s central constructed tool. No independent evidence outside the trained MI objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extracting informative vortical structures of turbulent wake-extreme vortex gust interactions with machine learning." pith.science (2026). https://pith.science/paper/UBI4ERLM

@misc{pith2026260726683,
  author       = {Pith},
  title        = {Pith review of: Extracting informative vortical structures of turbulent wake-extreme vortex gust interactions with machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UBI4ERLM}},
  note         = {Machine review of arXiv:2607.26683}
}
abstract

This study considers extracting causally important vortical structures from the extreme vortex gust-airfoil interaction at a chord-based Reynolds number of $5000$. This extraction is achieved by decomposing a given vortical flow snapshot into its informative and residual components based on the contribution to an arbitrary future target variable with convolutional information-theoretic learning. For the current vortex-airfoil interactions that exhibit transient and multiscale flow characteristics, we first examine the important vortical structures with respect to a future lift coefficient. While the vortex cores are primarily highlighted before vortex impingement, the emerging shear layers are additionally captured after the massive separation, which is evident from a comparison to an instantaneous force-element analysis. We further take the scale-dependent energy transfer as a future variable of interest to examine its impact on the extracted informative structures compared to the lift-associated structures. They are distinct from the lift-based structures in the early stage of the gust encounter yet become similar after impingement, revealing an analogy between informative structures across different transient aerodynamic mechanisms. The present data-driven approach selectively extracts the specific important flow structures responsible for the physics of interest, which can support studying a range of transient aerodynamic flows from the causal, data-driven perspective.

Figures

Figures reproduced from arXiv: 2607.26683 by the authors.

Figure 1
Figure 1. An example of the given state q and the informative component qI decomposed by the present informative mode extractor F. specific aerodynamic force components. Furthermore, scale decomposition [16, 17] isolates flow features at a specific characteristic length scale, which allows the extraction of the hierarchy of coherent vortices and the quantification of scale-dependent energy transfer. This scale-decomposition e… view at source ↗
Figure 2
Figure 2. (a) The operation of three-dimensional convolutional layer. (b) The architecture of the convolutional deep sigmoidal flow used in the present study. The convolutional layer is denoted as ‘Conv3D.’ The size of the filter H and its number N are shown for each layer in the form of (H, H, H, N). 2.1 Informative mode decomposition To extract time-varying causal vortical structure from extreme aerodynamic flows, we consid… view at source ↗
Figure 3
Figure 3. Extreme vortex gust-airfoil interaction at [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Informative mode decomposition for extreme positive vortex gust encounter at [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Informative mode decomposition for extreme negative vortex gust encounter at [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The dependence of informative modes on the value of the time window [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Time-dependent Q-R distributions of extreme vortex-gust airfoil interaction wake for gust ratios of G = 2 and 5, colored by input and informative Q-criterion. The flow structures extensively modified by the gust-airfoil interaction show a peak for the shorter ∆t = 0.05…
Figure 8
Figure 8. Figure 8: Scale-decomposition and informative mode decomposition with respect to the scale-dependent energy [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Comparison of informative modes extracted with respect to distinct target variables. Shown above is the time [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 1 linked inside Pith

  1. [1]

    Fukami and K

    K. Fukami and K. Taira. Grasping extreme aerodynamics on a low-dimensional manifold.Nat. Commun., 14(1):6480, 2023

    2023

  2. [2]

    A. R. Jones, O. Cetiner, and M. J. Smith. Physics and modeling of large flow disturbances: discrete gust encounters for modern air vehicles.Annu. Rev. Fluid Mech., 54(1):469–493, 2022

    2022

  3. [3]

    Lozano-Durán and G

    A. Lozano-Durán and G. Arranz. Information-theoretic formulation of dynamical systems: causality, modeling, and control.Phys. Rev. Res., 4(2):023195, 2022

    2022

  4. [4]

    K. Taira. Extreme aerodynamics: A data-driven perspective.Phys. Rev. Fluids, 11(1):014702, 2026

    2026

  5. [5]

    Fukami, H

    K. Fukami, H. Nakao, and K. Taira. Data-driven transient lift attenuation for extreme vortex gust–airfoil interactions.J. Fluid Mech., 992:A17, 2024

    2024

  6. [6]

    Lopez-Doriga, A

    B. Lopez-Doriga, A. R. M. Jones, and K. Taira. On the effect of airfoil geometry on extreme vortex-gust encounters. J. Fluid Mech., 1037:A61, 2026

    2026

  7. [7]

    C. J. Barnes and M. R. Visbal. Clockwise vortical-gust/airfoil interactions at a transitional Reynolds number. AIAA J., 56(10):3863–3874, 2018

    2018

  8. [8]

    C. J. Barnes and M. R. Visbal. Counterclockwise vortical-gust/airfoil interactions at a transitional Reynolds number.AIAA J., 56(7):2540–2552, 2018

    2018

Show all 52 references
  1. [9]

    Odaka, L

    H. Odaka, L. Smith, and K. Taira. Extreme vortex gust encounters by a square wing.J. Fluid Mech., 1037:A67, 2026

    2026

  2. [10]

    Odaka, B

    H. Odaka, B. Lopez-Doriga, and K. Taira. V ortical similarities across laminar and turbulent extreme gust encounters.J. Fluid Mech., 1031:R3, 2026

    2026

  3. [11]

    Biler, G

    H. Biler, G. Sedky, A. R. Jones, M. Saritas, and O. Cetiner. Experimental investigation of transverse and vortex gust encounters at low Reynolds numbers.AIAA J., 59(3):786–799, 2021

    2021

  4. [12]

    Andreu-Angulo, H

    I. Andreu-Angulo, H. abinsky, H. Biler, G. Sedky, and A. R. Jones. Effect of transverse gust velocity profiles. AIAA J., 58(12):5123–5133, 2020

    2020

  5. [13]

    C.-C. Chang. Potential flow and forces for incompressible viscous flow.Proc. Roy. Soc. A, 437(1901):517–525, 1992

    1901

  6. [14]

    M. Howe. On the force and moment on a body in an incompressible fluid, with application to rigid bodies and bubbles at high and low Reynolds numbers.Q. J. Mech. Appl. Math., 48(3):401–426, 1995

    1995

  7. [15]

    ¯Otomo, P

    S. ¯Otomo, P. Gehlert, H. Babinsky, and J. Li. V ortex force map method to estimate unsteady forces from snapshot flowfield measurements.Exp. Fluids, 66(3):1–13, 2025

    2025

  8. [16]

    S. Goto, Y . Saito, and G. Kawahara. Hierarchy of antiparallel vortex tubes in spatially periodic turbulence at high Reynolds numbers.Phys. Rev. Fluids, 2(6):064603, 2017

    2017

  9. [17]

    Fujino, Y

    J. Fujino, Y . Motoori, and S. Goto. Hierarchy of coherent vortices in turbulence behind a cylinder.J. Fluid Mech., 975:A13, 2023

    2023

  10. [18]

    Fukami, L

    K. Fukami, L. Smith, and K. Taira. Extreme vortex-gust airfoil interactions at Reynolds number 5000.Phys. Rev. Fluids, 10(8):084703, 2025. 14 APREPRINT- JULY30, 2026

    2025

  11. [19]

    S. L. Brunton, B. R. Noack, and P. Koumoutsakos. Machine learning for fluid mechanics.Annu. Rev. Fluid Mech., 52(1):477–508, 2020

    2020

  12. [20]

    Fukami, K

    K. Fukami, K. Fukagata, and K. Taira. Assessment of supervised machine learning methods for fluid flows: K. fukami et al.Theor . Comput. Fluid Dyn., 34(4):497–519, 2020

    2020

  13. [21]

    Taira, G

    K. Taira, G. Rigas, and K. Fukami. Machine learning in fluid dynamics: A critical assessment.Phys. Rev. Fluids, 10(9):090701, 2025

    2025

  14. [22]

    Smith, K

    L. Smith, K. Fukami, G. Sedky, A. Jones, and K. Taira. A cyclic perspective on transient gust encounters through the lens of persistent homology.J. Fluid Mech., 980:A18, 2024

    2024

  15. [23]

    Zamani Ashtiani and K

    S. Zamani Ashtiani and K. Fukami. Data-driven time-dependent bases for turbulent airfoil wake-extreme vortex gust interactions.AIAA J.in press, 2026

    2026

  16. [24]

    Linot, B

    A. Linot, B. Lopez-Doriga, Y . Zhong, and K. Taira. Extracting dominant dynamics about unsteady base flows. Fluid Dyn. Res., 57(3):031401, 2025

    2025

  17. [25]

    Arranz and A

    G. Arranz and A. Lozano-Durán. Informative and non-informative decomposition of turbulent flow fields.J. Fluid Mech., 1000:A95, 2024

    2024

  18. [26]

    Cremades, S

    A. Cremades, S. Hoyas, and R. Vinuesa. Additive-feature-attribution methods: A review on explainable artificial intelligence for fluid dynamics and heat transfer.Int. J. Heat Fluid Flow, 112:109662, 2025

    2025

  19. [27]

    Cremades, S

    A. Cremades, S. Hoyas, and R. Vinuesa. Classically studied coherent structures only paint a partial picture of wall-bounded turbulence.Nat. Commun., 16(1):10189, 2025

    2025

  20. [28]

    Cremades, R

    A. Cremades, R. Freibergs, S. Hoyas, A. Ianiro, S. Discetti, and R. Vinuesa. Assessment of non-intrusive sensing in wall-bounded turbulence through explainable deep learning.J. Fluid Mech., 1028:A19, 2026

    2026

  21. [29]

    Fukami and R

    K. Fukami and R. Araki. Information-theoretic machine learning for time-varying mode decomposition of separated aerodynamic flows.AIAA J., 64(2):605–613, 2026

    2026

  22. [30]

    Koshikawa, R

    R. Koshikawa, R. Araki, Q. Liu, and K. Fukami. Convolutional causal learning for aerodynamic flows.J. Fluid Mech., 1037:A6, 2026

    2026

  23. [31]

    Koshikawa, M

    R. Koshikawa, M. Kishimoto, and K. Fukami. Extracting the causal relationship of flow around an automobile with information-theoretic machine learning.AIAA paper, 2027

    2027

  24. [32]

    C. E. Shannon. A mathematical theory of communication.Bell Syst. Tech. J., 27(3):379–423, 1948

    1948

  25. [33]

    J. C. R. Hunt, A. A. Wray, and P. Moin. Eddies, streams, and convergence zones in turbulent flows. In P. Moin, W. C. Reynolds, and J. Kim, editors,Studying Turbulence Using Numerical Simulation Databases - II. Proceedings of the 1988 Summer Program, pages 193–208. Stanford Uni...

    1988

  26. [34]

    Huang, D

    C.-W. Huang, D. Krueger, A. Lacoste, and A. Courville. Neural autoregressive flows. InInternational conference on machine learning, pages 2078–2087. PMLR, 2018

  27. [35]

    D. P. Kingma. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014

    2014 arXiv

  28. [36]

    P. C. Hansen and D. P. O’Leary. The use of the L-curve in the regularization of discrete ill-posed problems.SIAM J. Sci. Comput., 14(6):1487–1503, 1993

    1993

  29. [37]

    Fukami, Y

    K. Fukami, Y . Iwatani, S. Maejima, H. Asada, and S. Kawai. Compact representation of transonic airfoil buffet flows with observable-augmented machine learning.J. Fluid Mech., 1021:A39, 2025

    2025

  30. [38]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition.Proc. IEEE, 86(11):2278–2324, 1998

    1998

  31. [39]

    Morimoto, K

    M. Morimoto, K. Fukami, K. Zhang, A. G. Nair, and K. Fukagata. Convolutional neural networks for fluid flow analysis: toward effective metamodeling and low dimensionalization.Theor . Comput. Fluid Dyn., 35(5):633–658, 2021

    2021

  32. [40]

    Ham and G

    F. Ham and G. Iaccarino. Energy conservation in collocated discretization schemes on unstructured meshes.Annu. Res. Briefs, 2004(3-14):118, 2004

    2004

  33. [41]

    F. Ham, K. Mattsson, and G. Iaccarino. Accurate and stable finite volume operators for unstructured flow solvers. Annu. Res. Briefs, 243, 2006

    2006

  34. [42]

    G. I. Taylor. On the dissipation of eddies.Meteorology, Oceanography and Turbulent Flow, pages 96–101, 1918

    1918

  35. [43]

    H. Hornung. V orticity generation and transport. InTenth Australasian Fluid Mechanics Conference, pages KS3.1–KS3.7. University of Melbourne, 1989

    1989

  36. [44]

    Wu, H.-Y

    J.-Z. Wu, H.-Y . Ma, and M.-D. Zhou.V orticity and V ortex Dynamics. Springer, Berlin, Heidelberg, 2012. 15 APREPRINT- JULY30, 2026

    2012

  37. [45]

    J. M. Ribeiro, J. Neal, A. Burtsev, M. Amitay, V . Theofilis, and K. Taira. Laminar post-stall wakes of tapered swept wings.J. Fluid Mech., 976:A6, 2023

    2023

  38. [46]

    Davidson.Turbulence: an introduction for scientists and engineers

    P. Davidson.Turbulence: an introduction for scientists and engineers. Oxford University Press, 2015

    2015

  39. [47]

    A. Ooi, J. Martin, J. Soria, and M. S. Chong. A study of the evolution and characteristics of the invariants of the velocity-gradient tensor in isotropic turbulence.J. Fluid Mech., 381:141–174, 1999

    1999

  40. [48]

    Fukami, S

    K. Fukami, S. Goto, and K. Taira. Data-driven nonlinear turbulent flow scaling with buckingham pi variables.J. Fluid Mech., 984:R4, 2024

    2024

  41. [49]

    Fukami and K

    K. Fukami and K. Taira. Observable-augmented manifold learning for multi-source turbulent flow data.J. Fluid Mech., 1010:R4, 2025

    2025

  42. [50]

    J. Pearl. Causal inference in statistics: an overview.Stat. Surv., 3:95–146, 2009

    2009

  43. [51]

    B. J. McKeon and A. S. Sharma. A critical-layer framework for turbulent pipe flow.J. Fluid Mech., 658:336–382, 2010

    2010

  44. [52]

    J. H. Marques Ribeiro and K. Taira. Triglobal resolvent-analysis-based control of separated flows around low-aspect-ratio wings.J. Fluid Mech., 995:A13, 2024. 16

    2024

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.