pith. sign in

arxiv: 2606.00949 · v1 · pith:OM7JPQ3Enew · submitted 2026-05-31 · 💻 cs.LG · cs.AI· physics.flu-dyn

Explainable deep reinforcement learning reveals energy-efficient control strategies for turbulent drag reduction

Federica Tonti , Ricardo Vinuesa This is my paper

Pith reviewed 2026-06-28 17:57 UTC · model grok-4.3

classification 💻 cs.LG cs.AIphysics.flu-dyn
keywords turbulent drag reductiondeep reinforcement learningSHAP explainabilitymulti-agent RLwall-bounded turbulenceskin-friction coefficientwall-pressure fluctuationsenergy-efficient control
0
0 comments X

The pith

SHAP attributions from U-net predictors shape MARL rewards to discover a pressure-gated policy that reduces turbulent drag by 34% while using only 0.43% input power.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that incorporating SHAP values from two U-nets—one predicting skin-friction coefficient and one predicting wall-pressure fluctuations—into the reward function of multi-agent deep reinforcement learning produces control policies superior to both direct wall-shear targeting and opposition control. A sympathetic reader would care because the resulting strategy simultaneously raises drag reduction and net energy saving while cutting normalized actuation cost by more than an order of magnitude. Analysis of the learned policy shows it activates mainly at near-zero wall pressure and on timescales comparable to the lifetime of near-wall turbulent structures.

Core claim

The combined SHAP strategy based on skin-friction coefficient and wall-pressure fluctuations achieves the best overall performance, achieving a DR of 34.44% and a NES of 34.01% with only 0.43% normalized input power. Relative to opposition control, drag reduction and net energy saving increase by 49.41% and 48.52%, respectively. Compared with the direct wall-shear-stress baseline, the proposed strategy simultaneously improves performance while reducing the normalized actuation cost from 5.90% to 0.43%. The energetically efficient policy is consistent with pressure-gated actuation, activating predominantly at near-zero wall pressure, and operates on a temporal timescale comparable to the life

What carries the argument

SHAP attributions from U-nets predicting future skin-friction coefficient and wall-pressure fluctuations, used as the reward signal for multi-agent reinforcement learning agents.

If this is right

  • The pressure-gated policy achieves 34.44% drag reduction and 34.01% net energy saving at 0.43% normalized input power.
  • Actuation occurs predominantly when instantaneous wall pressure is near zero.
  • The temporal scale of effective control matches the lifetime of near-wall turbulent structures.
  • Normalized actuation cost drops from 5.90% in the direct wall-shear baseline to 0.43% while performance improves.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same SHAP-reward construction could be tested on other canonical wall-bounded flows such as pipe or boundary-layer turbulence to check transferability.
  • Laboratory experiments that implement only pressure-based actuation at near-zero crossings would provide a direct physical test of the simulated energy savings.
  • Future sensor designs for active flow control might prioritize wall-pressure measurements over shear-stress measurements if the pressure-gated mechanism holds.

Load-bearing premise

The U-nets accurately predict future skin-friction and pressure values from the current flow state so that their SHAP attributions produce a reward signal that genuinely improves the agents' long-term control policy rather than rewarding spurious correlations.

What would settle it

Deploying the trained agents in an independent direct numerical simulation at the same Reynolds number and observing whether the measured drag reduction falls below 30% or the normalized actuation power rises above 1% would falsify the performance claims.

Figures

Figures reproduced from arXiv: 2606.00949 by Federica Tonti, Ricardo Vinuesa.

Figure 1
Figure 1. Figure 1: Drag reduction (left) and net energy saving (right) as a function of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Instantaneous near-wall fields for the five control methods. Top row: streamwise velocity fluctuation [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Joint probability density of the wall-pressure fluctuation [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Temporal character of the wall-normal actuation in the stationary regime ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

We propose a method combining Multi-Agent Deep Reinforcement Learning (MARL) and eXplainable Deep Learning (XDL) to reduce drag in wall-bounded turbulent flows. Taking as a baseline the results of training agents directly targeting wall-shear stress and opposition control, three SHAP-guided approaches are compared. In the first, the reward is computed from SHAP attributions of a U-net predicting the future velocity field; in the second, from SHAP attributions of a U-net predicting the skin-friction coefficient; in the third, from a combination of SHAP attributions of two U-nets predicting the skin-friction coefficient and the wall pressure fluctuations, respectively. The combined SHAP strategy based on skin-friction coefficient and wall-pressure fluctuations achieves the best overall performance, achieving a DR of 34.44% and a NES of 34.01% with only 0.43% normalized input power. Relative to opposition control, drag reduction and net energy saving increase by 49.41% and 48.52%, respectively. Compared with the direct wall-shear-stress baseline, the proposed strategy simultaneously improves performance while reducing the normalized actuation cost from 5.90% to 0.43%. Analysis of the results reveals that the energetically efficient policy is consistent with pressure-gated actuation, activating predominantly at near-zero wall pressure, and operates on a temporal timescale comparable to the lifetime of the near-wall turbulent structures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a hybrid MARL+XDL framework for turbulent drag reduction in wall-bounded flows. Three SHAP-guided reward formulations are derived from U-nets that predict future velocity, skin-friction coefficient, or wall-pressure fluctuations; the best-performing variant (combined skin-friction and pressure SHAP) is reported to achieve 34.44% drag reduction and 34.01% net energy saving at 0.43% normalized input power, outperforming both opposition control and a direct wall-shear-stress baseline while lowering actuation cost.

Significance. If the U-net predictions generalize and the resulting SHAP rewards produce policies that improve the true Navier-Stokes dynamics, the work would supply both a high-performance, low-power control law and an interpretable link between near-wall pressure and actuation timing. The explicit comparison against two external baselines and the reported energy metrics would constitute a concrete advance in data-driven flow control.

major comments (2)
  1. [Abstract / Methods (U-net training)] The central performance claims (34.44% DR, 34.01% NES) rest on SHAP attributions obtained from U-nets whose predictive accuracy is never quantified. No test-set MSE, correlation coefficient, or horizon-dependent error is reported for any of the three U-nets, making it impossible to determine whether the attributions identify causally relevant features or merely exploit model error.
  2. [Results (SHAP-guided strategies)] No ablation is presented that replaces the SHAP-derived reward with the ground-truth future skin-friction or pressure values. Without this control, it remains unclear whether the reported gains over the direct wall-shear baseline arise from the explainability step or simply from using a different reward formulation.
minor comments (2)
  1. [Abstract] The abstract states relative improvements of 49.41% and 48.52% over opposition control; these percentages should be accompanied by absolute values and statistical uncertainty to allow direct comparison.
  2. [Abstract] Notation for normalized input power and net energy saving is introduced without an explicit equation; a short definitions subsection would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the opportunity to respond to the referee's comments on our manuscript. We appreciate the constructive feedback and address each major comment below. We propose revisions to strengthen the presentation of the U-net models and clarify the role of the SHAP step.

read point-by-point responses

  1. Referee: [Abstract / Methods (U-net training)] The central performance claims (34.44% DR, 34.01% NES) rest on SHAP attributions obtained from U-nets whose predictive accuracy is never quantified. No test-set MSE, correlation coefficient, or horizon-dependent error is reported for any of the three U-nets, making it impossible to determine whether the attributions identify causally relevant features or merely exploit model error.

    Authors: We agree that the predictive accuracy of the U-nets must be quantified to support the reliability of the SHAP attributions. In the revised manuscript we will add a dedicated subsection (or appendix) reporting test-set MSE, Pearson correlation coefficients, and horizon-dependent error curves for all three U-nets. These metrics will confirm that the models achieve sufficient accuracy for the attributions to reflect physically relevant features rather than model artifacts. revision: yes

  2. Referee: [Results (SHAP-guided strategies)] No ablation is presented that replaces the SHAP-derived reward with the ground-truth future skin-friction or pressure values. Without this control, it remains unclear whether the reported gains over the direct wall-shear baseline arise from the explainability step or simply from using a different reward formulation.

    Authors: The direct wall-shear-stress baseline already employs instantaneous measurements as the reward. Using ground-truth future skin-friction or pressure as a reward would require an oracle unavailable in any practical online control setting and would therefore not constitute a fair ablation of the SHAP component. The SHAP formulation is deliberately chosen to extract interpretable, predictive features from the U-net outputs. In the revision we will expand the discussion to explicitly contrast the instantaneous baseline with the predictive SHAP approach and will add a limited comparison that replaces SHAP with the raw U-net predictions (without attribution) to isolate the contribution of the explainability step. revision: partial

Circularity Check

0 steps flagged

No significant circularity; results measured against external simulation baselines

full rationale

The derivation trains separate U-nets on flow data to predict future skin-friction or pressure, extracts SHAP attributions as a reward signal, and optimizes MARL policies against that reward. Final DR and NES percentages are obtained by running the resulting policies in the true Navier-Stokes solver and comparing to independent external controls (opposition control, direct wall-shear targeting). No equation or performance metric is defined in terms of the fitted reward itself, no self-citation supplies a uniqueness theorem, and no prediction is statistically forced by construction. The chain therefore remains externally falsifiable.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that the U-nets produce sufficiently accurate future predictions for their SHAP attributions to serve as a useful reward; no free parameters or invented entities are stated in the abstract.

axioms (1)
  • domain assumption U-nets trained on flow data can produce accurate short-term predictions of skin-friction coefficient and wall-pressure fluctuations from instantaneous wall measurements.
    The reward signal is constructed directly from SHAP values of these predictions; if the networks are inaccurate the attributions lose meaning.

pith-pipeline@v0.9.1-grok · 5793 in / 1501 out tokens · 25698 ms · 2026-06-28T17:57:29.488651+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 31 canonical work pages · 1 internal anchor

  1. [1]

    and Brunton, S

    Vinuesa, R. and Brunton, S. L. , title =. Nat. Comput. Sci. , volume =. 2022 , doi =

    2022

  2. [2]

    The autonomous cycle of near-wall turbulence , journal =

    Jim. The autonomous cycle of near-wall turbulence , journal =. 1999 , doi =

    1999

  3. [3]

    , title =

    Waleffe, F. , title =. Phys. Fluids , volume =. 1997 , doi =

    1997

  4. [4]

    Coherent structures in wall-bounded turbulence , journal =

    Jim. Coherent structures in wall-bounded turbulence , journal =. 2018 , doi =

    2018

  5. [5]

    and Moin, P

    Choi, H. and Moin, P. and Kim, J. , title =. J. Fluid Mech. , volume =. 1994 , doi =

    1994

  6. [6]

    and Rabault, J

    Guastoni, L. and Rabault, J. and Schlatter, P. and Azizpour, H. and Vinuesa, R. , title =. Eur. Phys. J. E , volume =. 2023 , doi =

    2023

  7. [7]

    and Liu, Z

    Sonoda, T. and Liu, Z. and Itoh, T. and Hasegawa, Y. , title =. J. Fluid Mech. , volume =. 2023 , doi =

    2023

  8. [8]

    and Rabault, J

    Vignon, C. and Rabault, J. and Vasanth, J. and Alc. Effective control of two-dimensional. Phys. Fluids , volume =. 2023 , doi =

    2023

  9. [9]

    Font, B. and Alc. Deep reinforcement learning for active flow control in a turbulent separation bubble , journal =. 2025 , doi =

    2025

  10. [10]

    Flow control of three-dimensional cylinders transitioning to turbulence via multi-agent reinforcement learning , journal =

    Su. Flow control of three-dimensional cylinders transitioning to turbulence via multi-agent reinforcement learning , journal =. 2025 , doi =

    2025

  11. [11]

    and Hoyas, S

    Cremades, A. and Hoyas, S. and Deshpande, R. and Quintero, P. and Lellep, M. and Lee, W. J. and Monty, J. P. and Hutchins, N. and Linkmann, M. and Marusic, I. and Vinuesa, R. , title =. Nat. Commun. , volume =. 2024 , doi =

    2024

  12. [12]

    and Hoyas, S

    Cremades, A. and Hoyas, S. and Vinuesa, R. , title =. Nat. Commun. , volume =. 2025 , doi =

    2025

  13. [13]

    and Hoyas, S

    Cremades, A. and Hoyas, S. and Vinuesa, R. , title =. Int. J. Heat Fluid Flow , volume =. 2025 , doi =

    2025

  14. [14]

    and Cremades, A

    Molina-Casino, S. and Cremades, A. and Hoyas, S. and Cardesa, J. I. and Chedevergne, F. and Vinuesa, R. , title =. Eur. J. Mech. B/Fluids , year =

  15. [15]

    Lundberg, S. M. and Lee, S.-I. , title =. Adv. Neural Inf. Process. Syst. , volume =

  16. [16]

    and Brunton, S

    Vinuesa, R. and Brunton, S. L. and Mengaldo, G. , title =. arXiv preprint arXiv:2601.05525 , year =

    work page arXiv

  17. [17]

    and Rabault, J

    Tonti, F. and Rabault, J. and Vinuesa, R. , title =. J. Comput. Phys. , pages =. 2025 , doi =

    2025

  18. [18]

    and Vinuesa, R

    Tonti, F. and Vinuesa, R. , title =. arXiv preprint arXiv:2510.25679 , year =

    work page arXiv

  19. [19]

    and Rabault, J

    Tonti, F. and Rabault, J. and Vinuesa, R. , title =. Advances in Heat Transfer , publisher =

  20. [20]

    and Matteucci, L

    Amico, E. and Matteucci, L. and Cafiero, G. , title =. Phys. Fluids , volume =. 2025 , doi =

    2025

  21. [21]

    An Explainable Deep Learning for Data-Driven Turbulence Model Feature Discovery , year=

    Alhafiz, Muhammad Ridho and Zuhal, Lavi Rizki and Dung, Duong Viet and Palar, Pramudita Satria , journal=. An Explainable Deep Learning for Data-Driven Turbulence Model Feature Discovery , year=

  22. [22]

    and Benedikt, N

    Hoyas, S. and Benedikt, N. and Cremades, A. and Vinuesa, R. , title =. Phys. Rev. Fluids , volume =. 2025 , doi =

    2025

  23. [23]

    Explainable deep learning reveals the physical mechanisms behind the turbulent kinetic energy equation , journal =

    Alc. Explainable deep learning reveals the physical mechanisms behind the turbulent kinetic energy equation , journal =

  24. [24]

    The minimal flow unit in near-wall turbulence , journal =

    Jim. The minimal flow unit in near-wall turbulence , journal =. 1991 , doi =

    1991

  25. [25]

    and Cremades, A

    Beneitez, M. and Cremades, A. and Guastoni, L. and Vinuesa, R. , title =. 2025 , journal =

    2025

  26. [26]

    and Fukagata, K

    Kametani, Y. and Fukagata, K. , title =. J. Fluid Mech. , volume =. 2011 , doi =

    2011

  27. [27]

    Burns, K. J. and Vasil, G. M. and Oishi, J. S. and Lecoanet, D. and Brown, B. P. , title =. Phys. Rev. Research , volume =. 2020 , doi =

    2020

  28. [28]

    Ascher, U. M. , title =. Numer. Algorithms , volume =

  29. [29]

    and van Hoof, H

    Fujimoto, S. and van Hoof, H. and Meger, D. , title =. Proc. 35th Int. Conf. Machine Learning , pages =. 2018 , publisher =

    2018

  30. [30]

    and Hill, A

    Raffin, A. and Hill, A. and Gleave, A. and Kanervisto, A. and Ernestus, M. and Dormann, N. , title =. J. Mach. Learn. Res. , volume =

  31. [31]

    and Janizek, J

    Erion, G. and Janizek, J. D. and Sturmfels, P. and Lundberg, S. M. and Lee, S.-I. , title =. Nat. Mach. Intell. , volume =. 2021 , doi =

    2021

  32. [32]

    , title =

    Bewley, Thomas R. , title =. J. Fluid Mech. , volume =. 2009 , doi =

    2009

  33. [33]

    and Wine, David and Holloway, Brian and Chung, Daniel and Smits, Alexander J

    Marusic, Ivan and Chandran, Dileep and Rouhi, Amirreza and Fu, Matt K. and Wine, David and Holloway, Brian and Chung, Daniel and Smits, Alexander J. , title =. Nat. Commun. , volume =. 2021 , doi =

    2021

  34. [34]

    Gatti, Davide and Quadrio, Maurizio , title =. J. Fluid Mech. , volume =. 2016 , doi =

    2016

  35. [35]

    Fukagata, Koji and Iwamoto, Kaoru and Hasegawa, Yosuke , title =. Annu. Rev. Fluid Mech. , volume =. 2024 , doi =

    2024

  36. [36]

    Causal features in turbulent channel flow , journal =

    Osawa, Kosuke and Jim. Causal features in turbulent channel flow , journal =. 2024 , doi =

    2024

  37. [37]

    Methods for interpreting and understanding deep neural networks , journal =

    Montavon, Gr. Methods for interpreting and understanding deep neural networks , journal =. 2018 , doi =

    2018

  38. [38]

    Multiscale analysis of the topological invariants in the logarithmic region of turbulent channels at a friction

    Lozano-Dur. Multiscale analysis of the topological invariants in the logarithmic region of turbulent channels at a friction. J. Fluid Mech. , volume =. 2021 , doi =

    2021

  39. [39]

    and Noack, Bernd R

    Brunton, Steven L. and Noack, Bernd R. and Koumoutsakos, Petros , title =. Annu. Rev. Fluid Mech. , volume =. 2020 , doi =

    2020

  40. [40]

    Artificial neural networks trained through deep reinforcement learning discover control strategies for active flow control , journal =

    Rabault, Jean and Kuchta, Miroslav and Jensen, Atle and R. Artificial neural networks trained through deep reinforcement learning discover control strategies for active flow control , journal =. 2019 , doi =

    2019

  41. [41]

    Quadrio, Maurizio and Ricco, Pierre , title =. J. Fluid Mech. , volume =. 2004 , doi =

    2004

  42. [42]

    Quadrio, Maurizio , title =. Phil. Trans. R. Soc. A , volume =. 2011 , doi =

    2011

  43. [43]

    and Monkewitz, Peter A

    Marusic, Ivan and McKeon, Beverley J. and Monkewitz, Peter A. and Nagib, Hassan M. and Smits, Alexander J. and Sreenivasan, Katepalli R. , title =. Phys. Fluids , volume =. 2010 , doi =

    2010

  44. [44]

    Hasegawa, Yosuke and Quadrio, Maurizio and Frohnapfel, Bettina , title =. J. Fluid Mech. , volume =. 2014 , doi =

    2014

  45. [45]

    and Namkoong, Hongseok and Farhadi, Ali and Carmon, Yair and Kornblith, Simon and Schmidt, Ludwig , title =

    Wortsman, Mitchell and Ilharco, Gabriel and Gadre, Samir Ya and Roelofs, Rebecca and Gontijo-Lopes, Raphael and Morcos, Ari S. and Namkoong, Hongseok and Farhadi, Ali and Carmon, Yair and Kornblith, Simon and Schmidt, Ludwig , title =. Proceedings of the 39th International Conference on Machine Learning , series =. 2022 , url =

    2022

  46. [46]

    Vinuesa and S

    R. Vinuesa and S. L. Brunton. Enhancing computational fluid dynamics with machine learning. Nat. Comput. Sci., 2: 0 358--366, 2022. doi:10.1038/s43588-022-00264-7

    work page doi:10.1038/s43588-022-00264-7 2022

  47. [47]

    F. Waleffe. On a self-sustaining process in shear flows. Phys. Fluids, 9: 0 883--900, 1997. doi:10.1063/1.869185

    work page doi:10.1063/1.869185 1997

  48. [48]

    Jim é nez and A

    J. Jim é nez and A. Pinelli. The autonomous cycle of near-wall turbulence. J. Fluid Mech., 389: 0 335--359, 1999. doi:10.1017/S0022112099005066

    work page doi:10.1017/s0022112099005066 1999

  49. [49]

    Jim é nez

    J. Jim é nez. Coherent structures in wall-bounded turbulence. J. Fluid Mech., 842: 0 P1, 2018. doi:10.1017/jfm.2018.144

    work page doi:10.1017/jfm.2018.144 2018

  50. [50]

    H. Choi, P. Moin, and J. Kim. Active turbulence control for drag reduction in wall-bounded flows. J. Fluid Mech., 262: 0 75--110, 1994. doi:10.1017/S0022112094000431

    work page doi:10.1017/s0022112094000431 1994

  51. [51]

    McKeon, Peter A

    Ivan Marusic, Beverley J. McKeon, Peter A. Monkewitz, Hassan M. Nagib, Alexander J. Smits, and Katepalli R. Sreenivasan. Wall-bounded turbulent flows at high Reynolds numbers: recent advances and key issues. Phys. Fluids, 22 0 (6): 0 065103, 2010. doi:10.1063/1.3453711

    work page doi:10.1063/1.3453711 2010

  52. [52]

    Thomas R. Bewley. A fundamental limit on the balance of power in a transpiration-controlled channel flow. J. Fluid Mech., 632: 0 443--446, 2009. doi:10.1017/S0022112008004886

    work page doi:10.1017/s0022112008004886 2009

  53. [53]

    Critical assessment of turbulent drag reduction through spanwise wall oscillations

    Maurizio Quadrio and Pierre Ricco. Critical assessment of turbulent drag reduction through spanwise wall oscillations. J. Fluid Mech., 521: 0 251--271, 2004. doi:10.1017/S0022112004001855

    work page doi:10.1017/s0022112004001855 2004

  54. [54]

    Drag reduction in turbulent boundary layers by in-plane wall motion

    Maurizio Quadrio. Drag reduction in turbulent boundary layers by in-plane wall motion. Phil. Trans. R. Soc. A, 369: 0 1428--1442, 2011. doi:10.1098/rsta.2010.0366

    work page doi:10.1098/rsta.2010.0366 2011

  55. [55]

    Turbulent drag reduction by streamwise traveling waves of wall-normal forcing

    Koji Fukagata, Kaoru Iwamoto, and Yosuke Hasegawa. Turbulent drag reduction by streamwise traveling waves of wall-normal forcing. Annu. Rev. Fluid Mech., 56: 0 69--90, 2024. doi:10.1146/annurev-fluid-120720-021445

    work page doi:10.1146/annurev-fluid-120720-021445 2024

  56. [56]

    Reynolds -number dependence of turbulent skin-friction drag reduction induced by spanwise forcing

    Davide Gatti and Maurizio Quadrio. Reynolds -number dependence of turbulent skin-friction drag reduction induced by spanwise forcing. J. Fluid Mech., 802: 0 553--558, 2016. doi:10.1017/jfm.2016.485

    work page doi:10.1017/jfm.2016.485 2016

  57. [57]

    Su á rez, F

    P. Su á rez, F. Alc á ntara- Á vila, J. Rabault, A. Mir o , B. Font, O. Lehmkuhl, and R. Vinuesa. Flow control of three-dimensional cylinders transitioning to turbulence via multi-agent reinforcement learning. Commun. Eng., 4: 0 113, 2025. doi:10.1038/s44172-025-00446-x

    work page doi:10.1038/s44172-025-00446-x 2025

  58. [58]

    Vignon, J

    C. Vignon, J. Rabault, J. Vasanth, F. Alc á ntara- Á vila, M. Mortensen, and R. Vinuesa. Effective control of two-dimensional R ayleigh-- B \'enard convection: invariant multi-agent reinforcement learning is all you need. Phys. Fluids, 35: 0 065146, 2023. doi:10.1063/5.0153181

    work page doi:10.1063/5.0153181 2023

  59. [59]

    B. Font, F. Alc á ntara- Á vila, J. Rabault, R. Vinuesa, and O. Lehmkuhl. Deep reinforcement learning for active flow control in a turbulent separation bubble. Nat. Commun., 16: 0 1422, 2025. doi:10.1038/s41467-025-56408-6

    work page doi:10.1038/s41467-025-56408-6 2025

  60. [60]

    Tonti, J

    F. Tonti, J. Rabault, and R. Vinuesa. Navigation in a simplified urban flow through deep reinforcement learning. J. Comput. Phys., page 114194, 2025 a . doi:10.1016/j.jcp.2025.114194

    work page doi:10.1016/j.jcp.2025.114194 2025

  61. [61]

    Tonti, J

    F. Tonti, J. Rabault, and R. Vinuesa. Deep reinforcement learning for combustion applications. In Advances in Heat Transfer. Elsevier, 2025 b

    2025

  62. [62]

    Guastoni, J

    L. Guastoni, J. Rabault, P. Schlatter, H. Azizpour, and R. Vinuesa. Deep reinforcement learning for turbulent drag reduction in channel flows. Eur. Phys. J. E, 46: 0 27, 2023. doi:10.1140/epje/s10189-023-00285-8

    work page doi:10.1140/epje/s10189-023-00285-8 2023

  63. [63]

    Sonoda, Z

    T. Sonoda, Z. Liu, T. Itoh, and Y. Hasegawa. Reinforcement learning of control strategies for reducing skin friction drag in a fully developed turbulent channel flow. J. Fluid Mech., 960: 0 A30, 2023. doi:10.1017/jfm.2023.147

    work page doi:10.1017/jfm.2023.147 2023

  64. [64]

    S. M. Lundberg and S.-I. Lee. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst., 30, 2017

    2017

  65. [65]

    Cremades, S

    A. Cremades, S. Hoyas, R. Deshpande, P. Quintero, M. Lellep, W. J. Lee, J. P. Monty, N. Hutchins, M. Linkmann, I. Marusic, and R. Vinuesa. Identifying regions of importance in wall-bounded turbulence through explainable deep learning. Nat. Commun., 15: 0 3864, 2024. doi:10.1038/s41467-024-47954-6

    work page doi:10.1038/s41467-024-47954-6 2024

  66. [66]

    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: 0 10189, 2025 a . doi:https://doi.org/10.1038/s41467-025-65199-9

    work page doi:10.1038/s41467-025-65199-9 2025

  67. [67]

    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: 0 109662, 2025 b . doi:10.1016/j.ijheatfluidflow.2024.109662

    work page doi:10.1016/j.ijheatfluidflow.2024.109662 2025

  68. [68]

    Multiscale analysis of the topological invariants in the logarithmic region of turbulent channels at a friction Reynolds number of 932

    Adri a n Lozano-Dur á n, Markus Holzner, and Javier Jim é nez. Multiscale analysis of the topological invariants in the logarithmic region of turbulent channels at a friction Reynolds number of 932 . J. Fluid Mech., 911: 0 A48, 2021. doi:10.1017/jfm.2020.1007

    work page doi:10.1017/jfm.2020.1007 2021

  69. [69]

    Causal features in turbulent channel flow

    Kosuke Osawa and Javier Jim é nez. Causal features in turbulent channel flow. J. Fluid Mech., 1000: 0 A4, 2024. doi:10.1017/jfm.2024.760

    work page doi:10.1017/jfm.2024.760 2024

  70. [70]

    Hoyas, N

    S. Hoyas, N. Benedikt, A. Cremades, and R. Vinuesa. Deep-learning-based assessment of skin friction in wall-bounded turbulence. Phys. Rev. Fluids, 10 0 (6): 0 L062601, 2025. doi:10.1103/b36b-m5hd

    work page doi:10.1103/b36b-m5hd 2025

  71. [71]

    Improving turbulence control through explainable deep learning

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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  72. [72]

    Kametani and K

    Y. Kametani and K. Fukagata. Direct numerical simulation of spatially developing turbulent boundary layers with uniform blowing or suction. J. Fluid Mech., 681: 0 154--172, 2011. doi:10.1017/jfm.2011.219

    work page doi:10.1017/jfm.2011.219 2011

  73. [73]

    Fu, David Wine, Brian Holloway, Daniel Chung, and Alexander J

    Ivan Marusic, Dileep Chandran, Amirreza Rouhi, Matt K. Fu, David Wine, Brian Holloway, Daniel Chung, and Alexander J. Smits. An energy-efficient pathway to turbulent drag reduction. Nat. Commun., 12: 0 5805, 2021. doi:10.1038/s41467-021-26128-8

    work page doi:10.1038/s41467-021-26128-8 2021

  74. [74]

    Jim é nez and P

    J. Jim é nez and P. Moin. The minimal flow unit in near-wall turbulence. J. Fluid Mech., 225: 0 213--240, 1991. doi:10.1017/S0022112091002033

    work page doi:10.1017/s0022112091002033 1991

  75. [75]

    K. J. Burns, G. M. Vasil, J. S. Oishi, D. Lecoanet, and B. P. Brown. Dedalus: a flexible framework for numerical simulations with spectral methods. Phys. Rev. Research, 2 0 (2): 0 023068, 2020. doi:10.1103/PhysRevResearch.2.023068

    work page doi:10.1103/physrevresearch.2.023068 2020

  76. [76]

    U. M. Ascher. Stabilization of invariants of discretized differential systems. Numer. Algorithms, 14: 0 1--24, 1997

    1997

  77. [77]

    Fujimoto, H

    S. Fujimoto, H. van Hoof, and D. Meger. Addressing function approximation error in actor-critic methods. In Proc. 35th Int. Conf. Machine Learning, pages 1587--1596. PMLR, 2018

    2018

  78. [78]

    Raffin, A

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann. Stable- B aselines3: reliable reinforcement learning implementations. J. Mach. Learn. Res., 22 0 (268): 0 1--8, 2021

    2021

  79. [79]

    Numerical simulation of turbulent duct flows with constant power input

    Yosuke Hasegawa, Maurizio Quadrio, and Bettina Frohnapfel. Numerical simulation of turbulent duct flows with constant power input. J. Fluid Mech., 750: 0 191--209, 2014. doi:10.1017/jfm.2014.269

    work page doi:10.1017/jfm.2014.269 2014

  80. [80]

    Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on...

    2022