Pith. sign in

REVIEW 3 major objections 6 minor 59 references

Information-theoretic machine learning for time-varying mode decomposition of separated aerodynamic flows

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 Informative mode decomposition of separated airfoil wakes, built on a bijective neural map from future lift to local vorticity, yields time-varying modes that trace the causal chain from flow structure to lift response.

desk verdict A useful application of a known decomposition to new flow problems, but the causal interpretation is not supported by Eq. (9): the bijective network makes the informative component trivially carry all information about the future lift. read the letter →

arxiv 2505.24132 v2 pith:3SUJPC5V submitted 2025-05-30 physics.flu-dyn cs.LGphysics.comp-ph

Kai Fukami , Ryo Araki This is my paper
classification physics.flu-dyncs.LGphysics.comp-ph
keywords informativemodedecompositiondeepsigmoidalflowmutualinformationentropyseparatedaerodynamicflowsliftcoefficientvortex-gustinteractionturbulentwake
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 aims to turn the fluid-dynamic question of which vortical structures cause the lift response into an information-theoretic computation. It applies a neural-network-based informative mode decomposition to separated wakes around a wing, splitting each vorticity snapshot into an informative part that carries all information about the lift coefficient at a later time and a residual part that carries none. Across four test cases — laminar post-stall wakes, numerical and experimental gust-wing encounters, and a turbulent spanwise-periodic wake — the method produces modes whose spatial location shifts with time, the lift fluctuation level, and the chosen time lag. The claim is that these time-varying modes expose the causal relationship between flow structure and aerodynamic force, with no prior aerodynamic or length-scale knowledge.

What carries the argument

The machinery is the informative mode decomposition implemented with a deep sigmoidal flow, a fully-connected neural network whose activations are bijective sigmoid-like functions and whose weights are positive, making the whole map $F$ bijective and forcing $H(C_L|q_I)=0$. The decomposition is local: each spatial point $x$ is processed by the network using the global lift coefficient at $t+\Delta t$ and the coordinate $x$ as inputs, so no joint distribution over the full flow field is needed. The weights are optimized by minimizing the reconstruction error of the vorticity field plus a weighted penalty on the mutual information between residual and informative parts, with mutual information estimated by a kernel density estimator, the weight $\beta$ chosen by L-curve analysis, and training performed with the Adam optimizer. The dependence of the modes on the time lag $\Delta t$ and on $\beta$ is what reveals the underlying convective causality.

What would settle it

On held-out snapshots from the periodic airfoil cases, compute an independent estimate of the conditional entropy $H(C_L(t+\Delta t)|q_I)$; if it is measurably positive, or if two snapshots with the same lift yield visibly different informative fields, then the bijective-map assumption fails and the causal reading of the modes is falsified.

Watch

Extended reading notes

Core claim

The central claim is that, for a separated flow around a wing, the spanwise-vorticity field at time $t$ can be decomposed as $q_I(x,t)=F(C_L(t+\Delta t),x)$, where $F$ is a bijective map learned by a deep sigmoidal flow, so that the conditional entropy $H(C_L(t+\Delta t)|q_I)$ is zero by construction. This means the informative component is, by construction, the only part of the local flow that contains information about the future lift. Applied to four separated-flow datasets, the extracted $q_I$ fields are time-varying modes: they concentrate near the wing when the lift fluctuation is small, migrate to downstream vortex cores when the fluctuation is large, track the gust during the interaction rather than appearing before it, and in the turbulent case highlight vortex cores and near-wing structures without being given any length-scale input. The paper presents this as evidence that the method captures causality as a time-varying modal structure.

Load-bearing premise

The whole construction rests on assuming that a single pointwise bijective map from the future lift coefficient to the local flow variable exists, so that the informative component provably determines the lift; if different flow states can give the same lift, the apparent informative structures are a regression artifact rather than a causal mode.

Editorial extensions

If this is right

  • For laminar post-stall wakes at angles of attack of 40 degrees and above, the informative structures switch from near the leading and trailing edges at small lift fluctuation to downstream vortex cores at large lift fluctuation.
  • For vortex-gust interactions, the gust itself is not marked as informative until its influence on lift begins, and before that the informative mode contains only the pre-existing near-wing structures.
  • Increasing the time window $\Delta t$ shifts the informative mode upstream, consistent with the convective delay between the gust and the lift response.
  • For the experimental gust encounter, the informative component is smooth even though the raw data are noisy, so the decomposition acts as a causal filter that removes components with no effect on lift.
  • For the turbulent wake, the method highlights near-wing structures and vortex cores with no length-scale or aerodynamic prior, and POD of the informative modes shows a single dominant mode instead of the usual shedding pair.

Reading between the lines

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

  • If the bijective-map assumption is relaxed to a probabilistic mapping, the method could distinguish different flow states that produce the same lift, which the current formulation cannot do.
  • The observed dependence of the modes on $\Delta t$ suggests a data-driven way to infer convective time scales: the time lag at which a structure switches from informative to uninformative should match its travel time to the wing.
  • Because the residual field is what remains after removing all lift-relevant information, it offers a natural target for flow-control studies that seek to modify the flow without changing the lift response.
  • Applying the same decomposition to drag or moment as the target variable would test whether the causal structure of aerodynamic force production is the same for each force component.
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. This paper applies the informative/non-informative decomposition of Arranz and Lozano-Durán to separated aerodynamic flows. The key methodological step is Eq. (9), q_I = F(C_L(t+Δt), x), with F a deep sigmoidal flow constrained to be bijective in its scalar argument. The informative field is trained to reconstruct the instantaneous spanwise vorticity while penalizing mutual information between q_I and the residual q_R. The method is applied to four cases: a periodic laminar post-stall wake, numerically simulated and experimentally measured gust-wing interactions, and a turbulent separated wake. The authors report that the extracted modes vary in time and with the time lag Δt, identify structures consistent with the Γ∝C_L relationship, and compare the turbulent-wake modes with POD modes. They interpret the results as revealing causal vortical structures for the future lift response.

Significance. The proposed extension to a spatially global target is a useful contribution, and the four test cases, especially the experimental and three-dimensional turbulent cases, broaden the evidence base beyond the original channel-flow study. The parameter studies for β and Δt and the POD comparison are informative. However, the central causal claim is not supported by the mathematical construction: because F is bijective in C_L(t+Δt), the condition H(C_L|q_I)=0 in Eq. (5) holds identically, independent of the data. The method is better characterized as a nonlinear predictive regression that identifies structures correlated with the future lift. If the authors reframe the claims accordingly and add residual-information diagnostics, the work could be a valuable descriptive tool for unsteady aerodynamics.

major comments (3)
  1. [Section II, Eq. (9)] Because the deep sigmoidal flow F is bijective in its scalar input, for each fixed x the informative field q_I(x,t) is an invertible function of C_L(t+Δt). The conditional entropy H(C_L(t+Δt) | q_I(x,t)) is therefore zero by construction, regardless of the data. As the paper itself notes in Section II, this constraint is what enables the solution of q_I to satisfy Eq. (5). Consequently, the maximization I(C_L;q_I)=H(C_L) is not a data-dependent criterion and cannot serve as evidence that the extracted structures are causally informative. The informative label is enforced by the architecture, not learned from the flow.
  2. [Section III.A] In the periodic shedding case the lift coefficient is not a one-to-one function of the flow state: on the limit cycle the same value of C_L(t+Δt) occurs at two different phases, with different vorticity fields (e.g., the leading-edge and trailing-edge vortex configurations in Fig. 2). Since q_I is a single-valued function of C_L for each x, it cannot reproduce both branch states; the residual q_R must carry the phase information. The paper does not report the mutual information I(C_L;q_R) or otherwise test whether the residual remains informative about the future lift. Without such a check, the decomposition does not establish that q_I captures all (or even the dominant) information about C_L(t+Δt), and the reported agreement with Γ∝C_L may simply reflect the regression of q onto C_L.
  3. [Abstract, Section III.B, Section IV] The causal interpretation is not justified by the evidence presented. The quantity I(C_L(t+Δt); q_I(t)) measures statistical dependence at a time lag; temporal precedence is necessary but not sufficient for causation. The Δt-dependence shown in Fig. 6 demonstrates that the modes vary with the lag, but this is expected for any predictive regression and does not establish that the identified structures are the cause of the lift response. To support the claims of capturing the causality (Abstract) and causally informative flow structures (Conclusions), the authors would need to use a genuinely causal measure (e.g., transfer entropy or Granger causality) or include a control experiment with shuffled or independent future-lift time series. As written, the causal claims outstrip the evidence.
minor comments (6)
  1. [Section III.A] The text states that snapshots are shown for phases θ=tan^{-1}(C_L_dot/C_L) of 0, π/2, π, and 3π/2, but later refers to θ=π/4; π/4 is not among the listed phases. Please correct the phase labels or the description.
  2. [Section II] There is a typo: 'refereed to as' should be 'referred to as' in the description of the deep sigmoidal flow, and the special character in 'bijective' should be a standard 'bijective'.
  3. [Section III.B] The sentence 'primally affects to the sharp fluctuation of the lift response' should read 'primarily affects the sharp fluctuation of the lift response.'
  4. [Section III.C] The phrase 'within sole two convective time' should be 'within only two convective times.'
  5. [Section II, Eq. (10)] In Eq. (10), the mutual information I(q_R;q_I) is a scalar, so the double norm ||·||^2 is unnecessary; please clarify the notation.
  6. [Section II and III] The manuscript does not report the network architecture (e.g., number of hidden layers and neurons, activation function), training hyperparameters (learning rate, number of epochs, batch size), or the bandwidth of the Gaussian kernel density estimator used for the mutual information estimates. These details are needed for reproducibility.

Circularity Check

2 steps flagged · score 8.0 of 10

Eq. (9) defines the 'informative' component as a bijective function of the future lift, so H(C_L|q_I)=0 holds by construction and the claimed causal mode is an input–output relabeling rather than a data-driven extraction.

  1. self definitional [Section II, Eqs. (5), (8), (9) and the paragraph following Eq. (10)]
    "With these constraints, the deep sigmoidal flow achieves a bijective transformation, enabling the solution of𝑞𝐼 produced by the mode extractorF to satisfy the constraint of the discounted Shannon information𝐻(𝐶𝐿(𝒙,𝑡+Δ𝑡),𝑞𝐼(𝒙,𝑡)) = 0inEq.5."

    Equation (9) defines q_I = F(C_L(t+Δt), x) and F is constrained to be bijective in its scalar argument. Therefore q_I contains the exact value of C_L(t+Δt) by construction, so H(C_L|q_I) is identically zero for every data set. The condition I(λ; q_I)=H(λ) in Eq. (5) is thus not an optimization outcome that selects informative flow structures; it is a property of the chosen parameterization. The only data-dependent content in Eq. (10) is the least-squares reconstruction of q from C_L and x, so the 'informative' label is assigned by definition rather than discovered from the snapshot.

  2. fitted input called prediction [Section II, Eq. (9) and Section III.A (periodic laminar wake results)]
    "A modified version of approximated informative and non-informative mode decomposition for the present aerodynamic problems is hence expressed as 𝑞𝐼 =F(𝐶𝐿(𝑡+Δ𝑡), 𝒙), with the optimization, ... which learns the causality from the informative component𝒒𝐼(𝒙,𝑡) to the lift coefficient𝐶𝐿(𝒙,𝑡+Δ𝑡) over the time intervalΔ𝑡."

    The network input is the future lift value, not the snapshot: at evaluation, q_I(x,t) is obtained by evaluating F at C_L(t+Δt). Thus the 'informative structure at time t' is not predicted from q(x,t); it is a regression of q(x,t) against the already-known target. The agreement with Γ∝C_L reported in the periodic case is then a consequence of feeding C_L(t+Δt) into the extractor, not a validation that the method discovered the lift–circulation relation from information metrics.

full rationale

The central causal claim reduces to a construction: the paper sets q_I = F(C_L(t+Δt), x) with F bijective, so the zero-conditional-entropy condition in Eq. (5) is satisfied identically. The mutual-information criterion that the method claims to optimize is therefore not an independent selection principle; any bijective function of the target would satisfy it. What is learned is only a reconstruction of the instantaneous field from the future lift, together with a tunable independence penalty. The time-varying modal behavior and the physical interpretations, while plausible and visually consistent with known aerodynamics, are not evidence that informative structures were extracted from snapshots on the basis of information. No load-bearing self-citation or imported uniqueness theorem is involved; the paper's method follows the earlier non-self-cited framework of Arranz and Lozano-Durán. The circularity is internal to the present formulation: the 'prediction' of causal modes uses the target as input by definition, so the score is high but not maximal because the reconstruction loss and β-penalty do provide some data-dependent content beyond pure tautology.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The central result depends on the bijective parametrization guaranteeing maximal informativeness, on per-case choices of Delta_t and beta, and on treating time-lagged mutual information as causality. The network hyperparameters and MI bandwidth are undisclosed, so the contribution is not self-contained.

free parameters (3)
  • Weighting coefficient beta = 0.01 baseline; 1e-5 and 0.1 in sensitivity study
    Balances reconstruction loss and mutual information penalty in Eq. (10). Fig. 5 shows the informative structures change qualitatively with beta, and the baseline is chosen per case.
  • Time lag Delta_t = 0.0085, 0.005, 0.05 for the four examples; sensitivity up to 0.425
    The informative structures depend strongly on Delta_t (Fig. 6), yet no objective criterion is given for choosing it; different choices highlight different structures.
  • Neural network hyperparameters = not reported
    Number of layers, neurons, learning rate, epochs, and the bandwidth of the Gaussian kernel density estimator for mutual information are not specified. Reproduction requires guessing these.
assumptions (7)
  • standard math Shannon entropy and mutual information are valid measures of statistical dependence.
    Used in Eqs. (2)-(4); standard information theory.
  • domain assumption Source and target variables can be restricted to scalars via a local decomposition.
    Relaxation from [22] to avoid the curse of dimensionality; q and lambda are treated pointwise.
  • ad hoc to paper The deep sigmoidal flow is bijective, so H(C_L|q_I)=0 holds by construction.
    Section II after Eq. (9); this design choice guarantees the informative component is maximally informative about the target.
  • domain assumption The Gaussian kernel density estimator yields a faithful mutual information estimate from finite samples.
    Section II states the MI is computed with a Gaussian KDE; bandwidth and convergence are not reported.
  • ad hoc to paper Temporal precedence plus mutual information is sufficient for causal claims.
    Abstract and Section II use 'causality' while only computing time-lagged MI; no deconfounding or conditional independence tests are performed.
  • domain assumption The known relation Gamma proportional to C_L applies in these separated flows.
    Section II invokes prior knowledge of the lift-circulation relation to motivate the target choice.
  • domain assumption The numerical and experimental datasets are accurate and representative of the flows.
    The paper relies on cited simulations and the Towne et al. database; no independent verification is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Information-theoretic machine learning for time-varying mode decomposition of separated aerodynamic flows." pith.science (2026). https://pith.science/paper/3SUJPC5V

@misc{pith2026250524132,
  author       = {Pith},
  title        = {Pith review of: Information-theoretic machine learning for time-varying mode decomposition of separated aerodynamic flows},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3SUJPC5V}},
  note         = {Machine review of arXiv:2505.24132}
}
read the original abstract

We perform an information-theoretic mode decomposition for separated aerodynamic flows. The current data-driven approach based on a neural network referred to as deep sigmoidal flow enables the extraction of an informative component from a given flow field snapshot with respect to a target variable at a future time stamp, thereby capturing the causality as a time-varying modal structure. We consider four examples of separated flows around a wing, namely, 1. laminar periodic wake at post-stall angles of attack, strong gust-wing interactions of 2. numerical and 3. experimental measurements, and 4. a turbulent wake in a spanwise-periodic domain. The present approach reveals informative vortical structures associated with a time-varying lift response. For the periodic shedding cases, the informative structures vary in time corresponding to the fluctuation level from their mean values. With the examples of gust-wing interactions, how the effect of gust on a wing emerges in the lift response over time is identified in an interpretable manner. Furthermore, for the case of turbulent wake, the present model highlights structures near the wing and vortex cores as informative components based solely on the information metric without any prior knowledge of aerodynamics and length scales. This study provides causality-based insights into a range of unsteady aerodynamic problems.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 28 canonical work pages

  1. [1]

    D.,Fundamentals of Aerodynamics, 6th ed., McGraw Hill, 2016

    Anderson, J. D.,Fundamentals of Aerodynamics, 6th ed., McGraw Hill, 2016

    work page 2016

  2. [2]

    13, Cambridge university press, 2001

    Katz, J., and Plotkin, A.,Low-speed aerodynamics, Vol. 13, Cambridge university press, 2001

    work page 2001

  3. [3]

    J.,Principles of helicopter aerodynamics, Cambridge university press, 2006

    Leishman, G. J.,Principles of helicopter aerodynamics, Cambridge university press, 2006

    work page 2006

  4. [4]

    Modal analysis of fluid flows: An overview,

    Taira, K., Brunton, S. L., Dawson, S. T. M., Rowley, C. W., Colonius, T., McKeon, B. J., Schmidt, O. T., Gordeyev, S., Theofilis, V., and Ukeiley, L. S., “Modal analysis of fluid flows: An overview,”AIAA J., Vol. 55, No. 12, 2017, pp. 4013–4041. https://doi.org/https://doi.org/10.2514/1.J056060

    doi:10.2514/1.j056060 2017

  5. [5]

    The structure of inhomogeneous turbulent flows,

    Lumley, J. L., “The structure of inhomogeneous turbulent flows,”Atmospheric turbulence and radio wave propagation, edited by A. M. Yaglom and V. I. Tatarski, Nauka, 1967

    work page 1967

  6. [6]

    Dynamic mode decomposition of numerical and experimental data,

    Schmid, P. J., “Dynamic mode decomposition of numerical and experimental data,”J. Fluid Mech., Vol. 656, 2010, pp. 5–28. https://doi.org/https://doi.org/10.1017/S0022112010001217

    doi:10.1017/s0022112010001217 2010

  7. [7]

    Guide to spectral proper orthogonal decomposition,

    Schmidt, O. T., and Colonius, T., “Guide to spectral proper orthogonal decomposition,”AIAA J., Vol. 58, No. 3, 2020, pp. 1023–1033. https://doi.org/https://doi.org/10.2514/1.J058809

    doi:10.2514/1.j058809 2020

  8. [8]

    Model reduction for flow analysis and control,

    Rowley, C. W., and Dawson, S. T. M., “Model reduction for flow analysis and control,”Annu. Rev. Fluid Mech., Vol. 49, No. 1, 2017, pp. 387–417. https://doi.org/https://doi.org/10.1146/annurev-fluid-010816-060042. 18

    doi:10.1146/annurev-fluid-010816-060042 2017

Show all 59 references
  1. [9]

    Modal Analysis of Fluid Flows: Applications and Outlook,

    Taira, K., Hemati, M. S., Brunton, S. L., Sun, Y., Duraisamy, K., Bagheri, S., Dawson, S., and Yeh, C.-A., “Modal Analysis of Fluid Flows: Applications and Outlook,”AIAA J., Vol. 58, No. 3, 2020, pp. 998–1022. https://doi.org/https: //doi.org/10.2514/1.J058462

    2020 doi

  2. [10]

    A critical-layer framework for turbulent pipe flow,

    Mckeon, B. J., and Sharma, A. S., “A critical-layer framework for turbulent pipe flow,”J. Fluid Mech., Vol. 658, 2010, pp. 336–382. https://doi.org/https://doi.org/10.1017/S002211201000176X

    2010 doi

  3. [11]

    Spectral proper orthogonal decomposition and its relationship to dynamic mode decomposition and resolvent analysis,

    Towne, A., Schmidt, O. T., and Colonius, T., “Spectral proper orthogonal decomposition and its relationship to dynamic mode decomposition and resolvent analysis,”J. Fluid Mech., Vol. 847, 2018, pp. 821–867. https://doi.org/https://doi.org/10.1017/jfm. 2018.283

    2018 doi

  4. [12]

    Resolvent-analysis-based design of airfoil separation control,

    Yeh, C.-A., and Taira, K., “Resolvent-analysis-based design of airfoil separation control,”J. Fluid Mech., Vol. 867, 2019, pp. 572–610. https://doi.org/https://doi.org/10.1017/jfm.2019.163

    2019 doi

  5. [13]

    Extracting dominant dynamics about unsteady base flows,

    Linot, A., Lopez-Doriga, B., Zhong, Y., and Taira, K., “Extracting dominant dynamics about unsteady base flows,”Fluid Dyn. Res., Vol. 57, 2025, p. 031401. https://doi.org/https://doi.org/10.1088/1873-7005/ade338

    2025 doi

  6. [14]

    Network broadcast analysis and control of turbulent flows,

    Yeh, C.-A., Gopalakrishnan Meena, M., and Taira, K., “Network broadcast analysis and control of turbulent flows,”J. Fluid Mech., Vol. 910, 2021, p. A15. https://doi.org/https://doi.org/10.1017/jfm.2020.965

    2021 doi

  7. [15]

    A minimization principle for the description of modes associated with finite-time instabilities,

    Babaee, H., and Sapsis, T. P., “A minimization principle for the description of modes associated with finite-time instabilities,” Proc. Roy. Soc. A, Vol. 472, No. 2186, 2016, p. 20150779. https://doi.org/https://doi.org/10.1098/rspa.2015.0779

    2016

  8. [16]

    Onset of absolute instability on a pitching aerofoil,

    Kern, J. S., Negi, P. S., Hanifi, A., and Henningson, D. S., “Onset of absolute instability on a pitching aerofoil,”J. Fluid Mech., Vol. 988, 2024, p. A8. https://doi.org/https://doi.org/10.1017/jfm.2024.407

    2024 doi

  9. [17]

    Optimally time-dependent modes of vortex gust-airfoil interactions,

    Zhong, Y., Amiri-Margavi, A., Babaee, H., and Taira, K., “Optimally time-dependent modes of vortex gust-airfoil interactions,” J. Fluid Mech., Vol. 1006, 2025, p. A18. https://doi.org/https://doi.org/10.1017/jfm.2025.9

    2025 doi

  10. [18]

    Sparsespace–timeresolventanalysisforstatisticallystationary and time-varying flows,

    Lopez-Doriga,B.,Ballouz,E.,Bae,H.J.,andDawson,S.T.M.,“Sparsespace–timeresolventanalysisforstatisticallystationary and time-varying flows,”J. Fluid Mech., Vol. 999, 2024, p. A87. https://doi.org/https://doi.org/10.1017/jfm.2024.955

    2024 doi

  11. [19]

    Grasping extreme aerodynamics on a low-dimensional manifold,

    Fukami, K., and Taira, K., “Grasping extreme aerodynamics on a low-dimensional manifold,”Nat. Commun., Vol. 14, No. 6480, 2023. https://doi.org/https://doi.org/10.1038/s41467-023-42213-6

    2023 doi

  12. [20]

    Data-driven transient lift attenuation for extreme vortex gust-airfoil interactions,

    Fukami, K., Nakao, H., and Taira, K., “Data-driven transient lift attenuation for extreme vortex gust-airfoil interactions,”J. Fluid Mech., Vol. 992, 2024, p. A17. https://doi.org/https://doi.org/10.1017/jfm.2024.592

    2024 doi

  13. [21]

    Compressingfluidflowswithnonlinearmachinelearning: modedecomposition,latentmodeling, and flow control,

    Fukagata,K.,andFukami,K.,“Compressingfluidflowswithnonlinearmachinelearning: modedecomposition,latentmodeling, and flow control,”Fluid Dyn. Res.,in press, 2025. https://doi.org/https://doi.org/10.1088/1873-7005/ade8a2

    2025 doi

  14. [22]

    Informative and non-informative decomposition of turbulent flow fields,

    Arranz, G., and Lozano-Durán, A., “Informative and non-informative decomposition of turbulent flow fields,”J. Fluid Mech., Vol. 1000, 2024, p. A95. https://doi.org/https://doi.org/10.1017/jfm.2024.1007. 19

    2024

  15. [23]

    Information-theoretic formulation of dynamical systems: causality, modeling, and control,

    Lozano-Durán, A., and Arranz, G., “Information-theoretic formulation of dynamical systems: causality, modeling, and control,” Phys. Rev. Res., Vol. 4, No. 2, 2022, p. 023195. https://doi.org/https://doi.org/10.1103/PhysRevResearch.4.023195

    2022 doi

  16. [24]

    Linear and nonlinear Granger causality analysis of turbulent duct flows,

    Lopez-Doriga, B., Atzori, M., Vinuesa, R., Bae, H. J., Srivastava, A., and Dawson, S. T. M., “Linear and nonlinear Granger causality analysis of turbulent duct flows,”J. Phys. Conf. Ser., Vol. 2753, IOP Publishing, 2024, p. 012017. https://doi.org/https://doi.org/10.1088/1742-...

    2024 doi

  17. [25]

    Information-thermodynamic bound on information flow in turbulent cascade,

    Tanogami, T., and Araki, R., “Information-thermodynamic bound on information flow in turbulent cascade,”Phys. Rev. Res., Vol. 6, No. 1, 2024, p. 013090. https://doi.org/https://doi.org/10.1103/PhysRevResearch.6.013090

    2024 doi

  18. [26]

    Forgetfulness of turbulent energy cascade associated with different mechanisms,

    Araki, R., Vela-Martín, A., and Lozano-Durán, A., “Forgetfulness of turbulent energy cascade associated with different mechanisms,”J. Phys. Conf. Ser., Vol. 2753, IOP Publishing, 2024, p. 012001. https://doi.org/https://doi.org/10.1088/1742- 6596/2753/1/012001

    2024 doi

  19. [27]

    Scale-to-scale information flow amplifies turbulent fluctuations,

    Tanogami, T., and Araki, R., “Scale-to-scale information flow amplifies turbulent fluctuations,”Phys. Rev. Res., Vol. 7, No. 2, 2025, p. 023078. https://doi.org/https://doi.org/10.1103/PhysRevResearch.7.023078

    2025 doi

  20. [28]

    Decomposing causality into its synergistic, unique, and redundant components,

    Martínez-Sánchez, Á., Arranz, G., and Lozano-Durán, A., “Decomposing causality into its synergistic, unique, and redundant components,”Nat. Commun., Vol. 15, No. 1, 2024, p. 9296. https://doi.org/https://doi.org/10.1038/s41467-024-53373-4

    2024 doi

  21. [29]

    Neural autoregressive flows,

    Huang, C.-W., Krueger, D., Lacoste, A., and Courville, A., “Neural autoregressive flows,”Proceedings of the 35th International Conference on Machine Learning, Proceedings of Machine Learning Research, PMLR, 2018, pp. 2078–2087

    2018

  22. [30]

    A mathematical theory of communication,

    Shannon, C. E., “A mathematical theory of communication,”Bell Syst. Tech. J., Vol. 27, No. 3, 1948, pp. 379–423. https://doi.org/https://doi.org/10.1002/j.1538-7305.1948.tb01338.x

    1948

  23. [31]

    M., and Joy, A

    Cover, T. M., and Joy, A. T.,Elements of information theory,2nd ed., John Wiley & Sons, 2005

    2005

  24. [32]

    Synthetic turbulent inflow generator using machine learning,

    Fukami, K., Nabae, Y., Kawai, K., and Fukagata, K., “Synthetic turbulent inflow generator using machine learning,”Phys. Rev. Fluids, Vol. 4, 2019, p. 064603. https://doi.org/https://doi.org/10.1103/PhysRevFluids.4.064603

    2019 doi

  25. [33]

    The use of the L-curve in the regularization of discrete ill-posed problems,

    Hansen, P. C., and O’Leary, D. P., “The use of the L-curve in the regularization of discrete ill-posed problems,”SIAM J. Sci. Comput., Vol. 14, No. 6, 1993, pp. 1487–1503. https://doi.org/https://doi.org/10.1137/0914086

    1993 doi

  26. [34]

    Learning representations by back-propagation errors,

    Rumelhart, D. E., Hinton, G. E., and Williams, R. J., “Learning representations by back-propagation errors,”Nature, Vol. 322, 1986, pp. 533–536. https://doi.org/https://doi.org/10.1038/323533a0

    1986 doi

  27. [35]

    Adam: A method for stochastic optimization,

    Kingma, D. P., and Ba, J., “Adam: A method for stochastic optimization,” arXiv:1412.6980, 2014

    2014 arXiv

  28. [36]

    Energy conservation in collocated discretization schemes on unstructured meshes,

    Ham, F., and Iaccarino, G., “Energy conservation in collocated discretization schemes on unstructured meshes,”Annual Research Briefs, Center for Turbulence Research, 2004, pp. 3–14

    2004

  29. [37]

    Accurate and stable finite volume operators for unstructured flow solvers,

    Ham, F., Mattsson, K., and Iaccarino, G., “Accurate and stable finite volume operators for unstructured flow solvers,”Annual Research Briefs, Center for Turbulence Research, 2006, pp. 243–261. 20

    2006

  30. [38]

    Sparse sensor reconstruction of vortex-impinged airfoil wake with machine learning,

    Zhong, Y., Fukami, K., An, B., and Taira, K., “Sparse sensor reconstruction of vortex-impinged airfoil wake with machine learning,”Theor. Comput. Fluid Dyn., Vol. 37, 2023, pp. 269–287. https://doi.org/https://doi.org/10.1007/s00162-023-00657-y

    2023 doi

  31. [39]

    OntheunsteadybehavioroftheflowaroundNACA0012airfoilwithsteadyexternalconditionsat 𝑅𝑒 = 1000,

    Kurtulus,D.F.,“OntheunsteadybehavioroftheflowaroundNACA0012airfoilwithsteadyexternalconditionsat 𝑅𝑒 = 1000,” Int. J. Micro Air Veh., Vol. 7, No. 3, 2015, pp. 301–326. https://doi.org/https://doi.org/10.1260/1756-8293.7.3.301

    2015 doi

  32. [40]

    Numerical bifurcation analysis of static stall of airfoil and dynamic stall under unsteady perturbation,

    Liu, Y., Li, K., Zhang, J., Wang, H., and Liu, L., “Numerical bifurcation analysis of static stall of airfoil and dynamic stall under unsteady perturbation,”Commun. Nonlinear Sci. Numer. Simul., Vol. 17, No. 8, 2012, pp. 3427–3434. https://doi.org/https://doi.org/10.1016/j.cns...

    2012 doi

  33. [41]

    Fluid flow around NACA 0012 airfoil at low-Reynolds numbers with hybrid lattice Boltzmann method,

    Di Ilio, G., Chiappini, D., Ubertini, S., Bella, G., and Succi, S., “Fluid flow around NACA 0012 airfoil at low-Reynolds numbers with hybrid lattice Boltzmann method,”Comput. Fluids, Vol. 166, 2018, pp. 200–208. https://doi.org/https: //doi.org/10.1016/j.compfluid.2018.02.014

    2018 doi

  34. [42]

    Adjoint-based phase reduction analysis of incompressible periodic flows,

    Kawamura, Y., Godavarthi, V., and Taira, K., “Adjoint-based phase reduction analysis of incompressible periodic flows,”Phys. Rev. Fluids, Vol. 7, No. 10, 2022, p. 104401. https://doi.org/https://doi.org/10.1103/PhysRevFluids.7.104401

    2022 doi

  35. [43]

    Optimal waveform for fast synchronization of airfoil wakes,

    Godavarthi, V., Kawamura, Y., and Taira, K., “Optimal waveform for fast synchronization of airfoil wakes,”J. Fluid Mech., Vol. 976, No. R1, 2023. https://doi.org/https://doi.org/10.1017/jfm.2023.929

    2023 doi

  36. [44]

    Optimal external forces of the lock-in phenomena for flow past an inclined plate in uniform flow,

    Iima, M., “Optimal external forces of the lock-in phenomena for flow past an inclined plate in uniform flow,”Phys. Rev. E, Vol. 109, No. 4, 2024, p. 045102. https://doi.org/https://doi.org/10.1103/PhysRevE.109.045102

    2024 doi

  37. [45]

    Physics and modeling of large flow disturbances: Discrete gust encounters for modern air vehicles,

    Jones, A. R., Cetiner, O., and Smith, M. J., “Physics and modeling of large flow disturbances: Discrete gust encounters for modern air vehicles,”Annu. Rev. Fluid Mech., Vol. 54, 2022, pp. 469–493. https://doi.org/https://doi.org/10.1146/annurev- fluid-031621-085520

    2022 doi

  38. [46]

    Overviewof unsteadyaerodynamicresponse ofrigidwingsingustencounters,

    Jones, A.R., andCetiner, O., “Overviewof unsteadyaerodynamicresponse ofrigidwingsingustencounters,” AIAAJ., Vol.59, No. 2, 2021, pp. 731–736. https://doi.org/https://doi.org/10.2514/1.J059602

    2021 doi

  39. [47]

    On the dissipation of eddies,

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

    1918

  40. [48]

    A database for reduced-complexity modeling of fluid flows,

    Towne, A., Dawson, S., Brès, G. A., Lozano-Durán, A., Saxton-Fox, T., Parthasarathy, A., Jones, A. R., Biler, H., Yeh, C.-A., Patel, H. D., and Taira, K., “A database for reduced-complexity modeling of fluid flows,”AIAA J., Vol. 61, No. 7, 2023, pp. 2867–2892. https://doi.org/...

    2023 doi

  41. [49]

    Experimentalinvestigationoftransverseandvortexgustencounters at low Reynolds numbers,

    Biler,H.,Sedky,G.,Jones,A.R.,Saritas,M.,andCetiner,O.,“Experimentalinvestigationoftransverseandvortexgustencounters at low Reynolds numbers,”AIAA J., Vol. 59, No. 3, 2021, pp. 786–799. https://doi.org/https://doi.org/10.2514/1.J059658

    2021 doi

  42. [50]

    Effect of transverse gust velocity profiles,

    Andreu-Angulo, I., Babinsky, H., Biler, H., Sedky, G., and Jones, A. R., “Effect of transverse gust velocity profiles,”AIAA J., Vol. 58, No. 12, 2020, pp. 5123–5133. https://doi.org/https://doi.org/10.2514/1.J059665. 21

    2020 doi

  43. [51]

    Robust principal component analysis for modal decomposition of corrupt fluid flows,

    Scherl, I., Strom, B., Shang, J. K., Williams, O., Polagye, B. L., and Brunton, S. L., “Robust principal component analysis for modal decomposition of corrupt fluid flows,”Phys. Rev. Fluids, Vol. 5, No. 5, 2020, p. 054401. https://doi.org/https: //doi.org/10.1103/PhysRevFluids...

    2020 doi

  44. [52]

    An invitation to resolvent analysis,

    Rolandi, L. V., Ribeiro, J. H. M., Yeh, C.-A., and Taira, K., “An invitation to resolvent analysis,”Theor. Comput. Fluid Dyn., Vol. 38, No. 5, 2024, pp. 603–639. https://doi.org/https://doi.org/10.1007/s00162-024-00717-x

    2024 doi

  45. [53]

    Biglobal resolvent analysis of separated flow over a NACA0012 airfoil,

    Rolandi, L. V., Smith, L., Amitay, M., Theofilis, V., and Taira, K., “Biglobal resolvent analysis of separated flow over a NACA0012 airfoil,” arXiv:2501.04255, 2025

    2025

  46. [54]

    Extreme vortex-gust airfoil interactions at Reynolds number 5,000,

    Fukami, K., Smith, L., and Taira, K., “Extreme vortex-gust airfoil interactions at Reynolds number 5,000,”Phys. Rev. Fluids,in Revision, 2025

    2025

  47. [55]

    An eddy-viscosity subgrid-scale model for turbulent shear flow: Algebraic theory and applications,

    Vreman, A. W., “An eddy-viscosity subgrid-scale model for turbulent shear flow: Algebraic theory and applications,”Phys. Fluids, Vol. 16, No. 10, 2004, pp. 3670–3681. https://doi.org/https://doi.org/10.1063/1.1785131

    2004 doi

  48. [56]

    Hierarchy of coherent vortices in turbulence behind a cylinder,

    Fujino, J., Motoori, Y., and Goto, S., “Hierarchy of coherent vortices in turbulence behind a cylinder,”J. Fluid Mech., Vol. 975, 2023, p. A13. https://doi.org/https://doi.org/10.1017/jfm.2023.824

    2023 doi

  49. [57]

    Identifyingregionsofimportanceinwall-boundedturbulencethroughexplainabledeeplearning,

    Cremades, A., Hoyas, S., Deshpande, R., Quintero, P., Lellep, M., Lee, W. J., Monty, J. P., Hutchins, N., Linkmann, M., Marusic,I., andVinuesa,R.,“Identifyingregionsofimportanceinwall-boundedturbulencethroughexplainabledeeplearning,” Nat. Commun., Vol. 15, No. 1, 2024, p. 3864...

    2024 doi

  50. [58]

    Additive-feature-attribution methods: a review on explainable artificial intelligence for fluid dynamics and heat transfer,

    Cremades, A., Hoyas, S., and Vinuesa, R., “Additive-feature-attribution methods: a review on explainable artificial intelligence for fluid dynamics and heat transfer,”Int. J. Heat Fluid Flow, Vol. 112, 2025, p. 109662. https://doi.org/https://doi.org/10.1016/ j.ijheatfluidflow....

    2025

  51. [59]

    Improving turbulence control through explainable deep learning,

    Beneitez, M., Cremades, A., Guastoni, L., and Vinuesa, R., “Improving turbulence control through explainable deep learning,” arXiv:2504.02354, 2025. 22

    2025 arXiv

Pith tools

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