Discovering Flow Separation Control Strategies in 3D Wings via Deep Reinforcement Learning

B. Font; I. Rodriguez; J. Rabault; O. Lehmkuhl; P. Su\'arez; R. Montal\`a; R. Vinuesa

REVIEW 4 major objections 4 minor 5 references

Discovering Flow Separation Control Strategies in 3D Wings via Deep Reinforcement Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper demonstrates that deep reinforcement learning can discover a closed-loop active-flow-control policy that reattaches massively separated flow over a three-dimensional wing, increasing lift by 79% and reducing drag by 65% in high-f

desk verdict A genuine step forward for DRL-based flow control on a 3D wing, but the headline numbers outrun the evidence: no error bars, no energy cost, and a sign flip in lift fluctuations between training and evaluation meshes. read the letter →

arxiv 2509.10185 v1 pith:3WVCFGPJ submitted 2025-09-12 cs.CE

R. Montalà , B. Font , P. Suárez , J. Rabault , O. Lehmkuhl , R. Vinuesa , I. Rodriguez This is my paper · ORCID

classification cs.CE

keywords deepreinforcementlearningactiveflowcontrolseparationthree-dimensionalwinglarge-eddysimulationmulti-agentliftenhancementdragreduction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The authors set out to show that a deep reinforcement learning agent can learn useful active flow control for a fully separated, fully turbulent flow over a three-dimensional wing—a step beyond the usual two-dimensional or low-Reynolds benchmarks. Training on a coarse mesh with multiple parallel spanwise subdomains, the agent learns to apply leading-edge suction and rear blowing. When the learned policy is evaluated on a much finer mesh, it reattaches the separated shear layer, increasing mean lift by 79%, reducing mean drag by 65%, and raising lift-to-drag efficiency by 408%. The results support the claim that DRL can discover effective, physically interpretable, closed-loop control strategies for complex aerodynamic configurations.

What carries the argument

The key mechanism is a multi-agent reinforcement-learning loop: the three-dimensional spanwise-periodic domain is split into three subdomains, each with its own jet pair, and a single shared neural-network policy chooses jet velocities from pressure measurements at 90 sensor locations (with adjacent-span information giving a 270-dimensional state). The reward couples each subdomain's drag reduction, lift increase, and lift-fluctuation penalty to the mean reward of all subdomains. During training the agent interacts with a coarse-mesh large-eddy simulation; the same policy is then frozen and evaluated on a fine mesh. The mass-conserving jet pair (front suction, rear blowing) is the actuator c

What would settle it

Evaluate the learned policy on a yet finer mesh (or in a wind-tunnel experiment at the same Re and AoA) and measure mean lift and drag. If the 79% lift increase and 65% drag reduction shrink substantially or reverse—or if the policy's action ranking changes when rewards are recomputed on the finer grid—then the coarse-mesh training signal was not representative and the transfer claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that a single DRL policy, shared across three spanwise pseudo-environments, can control a pair of synthetic jets (front suction, rear blowing) on an SD7003 wing at Re=60,000 and AoA=14° so that a massively separated turbulent flow reattaches. After a short transient, the controlled flow reaches a statistically steady state with substantially better aerodynamic coefficients than the uncontrolled baseline. The learned actuation is not a fixed-frequency forcing: its energy is spread across the turbulent spectrum with only weak peaks, indicating the controller continuously adjusts to instantaneous flow structures. The strategy is physically interpretable—mean suction nea

Load-bearing premise

The load-bearing premise is that the 16-million-degree-of-freedom coarse mesh, although it does not resolve all turbulent scales, captures the large-scale separated-flow dynamics well enough that reward signals during training guide the agent toward a policy that also works on the fine mesh; the paper notes this directly and offers the fine-mesh evaluation as the check.

Editorial extensions

If this is right

If the coarse-mesh-to-fine-mesh transfer is accepted, DRL-based AFC training becomes affordable for 3D turbulent configurations, since the expensive training can be done on a coarse mesh and the result checked on a high-fidelity mesh.
The learned strategy—leading-edge suction paired with rear blowing—is a physically interpretable control law, not just an opaque black-box action, which may help aerodynamicists understand and port the mechanism.
The absence of a dominant actuation frequency suggests closed-loop DRL control can outperform fixed-frequency harmonic forcing in broadband turbulent flows.
The reported 408% efficiency improvement, if robust, makes fully separated high-angle-of-attack wings a viable target for flow control, with implications for gust resistance and manoeuvrability.

Reading between the lines

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

Beyond the paper: the same pipeline could be tested on swept or tapered wings, where spanwise variation breaks the translational invariance the multi-agent setup exploits; success there would extend the method beyond infinite-wing geometries.
Beyond the paper: a direct head-to-head comparison with optimal steady suction/blowing or harmonic forcing on the same fine mesh would quantify how much of the 79%/65% gain comes from closed-loop adaptivity rather than from the mean actuation alone.
Beyond the paper: the coarse-mesh vs fine-mesh disagreement on lift fluctuations (coarse +31%, fine -39%) suggests the coarse reward may mis-rank control policies on secondary metrics; an even finer validation mesh or experiment would test whether the primary lift/drag gains survive.

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.

Desk Editor's Note

read the letter

You should know two things: this is the most ambitious DRL active flow control case yet published, and the headline numbers (79% lift, 65% drag, 408% efficiency) are plausible but not yet solid. The paper is a natural extension of the authors' own prior work—multi-agent DRL on a 3D wing at Re=60k with a fully turbulent separated wake. That is genuinely new; prior DRL AFC has been stuck at 2D airfoils or 3D cylinders at lower Reynolds numbers.

What it does well: the training infrastructure is serious (SOD2D GPU spectral-element solver, 10 parallel CFD environments, 3 MARL subdomains, PPO with a 270-sensor state). Evaluating the trained policy on a 418M-DoF mesh after training on 16M is a legitimate transfer test, and that is more than most papers in this area do. The learned strategy—leading-edge suction with trailing-edge blowing—is physically interpretable and consistent with the reattachment seen in the Q-criterion and velocity fields. If the numbers are right, it is a meaningful step.

Now the soft spots, in order of seriousness.

First, the C'_l,rms inconsistency between meshes. In Table 1, the controlled case shows lift fluctuations increasing by 31% on the coarse mesh (0.0226 to 0.0296) but decreasing by 39% on the fine mesh (0.0334 to 0.0205). Since the reward explicitly penalizes lift fluctuations, the agent was optimized in an environment where that reward component pointed the wrong way relative to the evaluation. The paper simply says the results are "highly consistent," but they are not for this metric. This does not kill the paper—mean lift and drag transfer cleanly—but it does mean the coarse-mesh reward signal was not fully representative, and the transfer claim needs to be stated more carefully.

Second, there are no error bars anywhere. The drag reduction of 65% is presented without any measure of statistical uncertainty from a single trajectory, no ensemble over initial conditions or policy seeds. Given how noisy separated turbulent flows are, that is a real omission.

Third, the efficiency gain is C_l/C_d only; the jet power required is not accounted for. At a mean jet velocity of about 0.11 U∞ with mass conservation constraints, the actuation cost may be small, but it should be quantified before calling a 408% efficiency gain an aerodynamic outcome.

Fourth, there is no comparison to simpler control baselines—steady suction, periodic forcing, or a fixed jet velocity. Without that, it is hard to know what the DRL closed-loop policy adds beyond the mean actuation.

None of these are fatal, and the central qualitative result—that a learned closed-loop policy can reattach a massively separated 3D wing at moderate Reynolds number—is likely to survive. The paper deserves serious peer review; the authors are honest about the coarse mesh approximation but need to address the fluctuation sign flip, add uncertainty estimates, and include the energy cost before this can be fully trusted.

I would bring it to a reading group focused on DRL for fluid mechanics, and I would cite it as evidence of progress in realistic configurations, with a caveat on the quantitative claims.

Referee Report

4 major / 4 minor

Summary. This manuscript applies deep reinforcement learning (DRL) to active flow control over a three-dimensional SD7003 wing at Re = 60,000 and AoA = 14 degrees, where the baseline flow is massively separated. Training is performed on a coarse mesh (16M DoF) with a multi-agent PPO setup, and the resulting policy is evaluated on a fine mesh (418M DoF). The authors report a 79% increase in mean lift, a 65% decrease in mean drag, and a 408% improvement in the lift-to-drag ratio relative to the uncontrolled baseline, together with flow visualization indicating reattachment of the separated shear layer. The paper argues that the coarse-mesh training environment captures the large-scale flow structures and that the fine-mesh evaluation provides a genuine transfer test of the learned control strategy.

Significance. If the reported results are robust, this is a noteworthy demonstration: it extends DRL-based active flow control to a fully turbulent, three-dimensional separated flow over an airfoil with spanwise-periodic geometry, and it includes a coarse-to-fine mesh transfer evaluation that goes beyond many prior studies. The fine-mesh evaluation is a genuine strength, as is the use of multiple concurrent CFD environments to make training tractable. The flow visualizations and control-signal analysis also provide physically interpretable support for reattachment. However, the paper's central quantitative claims currently rest on single-point statistics without uncertainty quantification, on a coarse-mesh training signal that is inconsistent with the fine-mesh result for lift fluctuations, and on an efficiency metric that omits the energy cost of actuation. These issues must be addressed before the quantitative claims can be considered established.

major comments (4)

[Table 1, Section 3] The reported improvements (79% lift, 65% drag reduction, 408% efficiency) are single-point estimates with no error bars, confidence intervals, or statistical convergence metrics. For a fully turbulent separated flow at Re = 60,000, the time-averaged coefficients and especially the rms lift fluctuations can exhibit significant low-frequency variability. The paper should report block-averaged means with standard errors, the evaluation-time window, and convergence checks, so that the reader can assess whether the improvements are statistically significant and not within the natural variability of the baseline or controlled flows.
[Section 3, first paragraph; Table 1; Eq. (4)] The coarse-to-fine transfer claim is undermined by the sign flip in C'_l,rms: Table 1 shows +31% relative to baseline on the coarse mesh (0.0226 to 0.0296) but -39% on the fine mesh (0.0334 to 0.0205). Since Eq. (4) explicitly penalizes lift fluctuations through -α|C_l - C_l,avg| with α = 0.3, the agent was trained in an environment where this reward component was qualitatively different from the evaluation environment. The paper's assertion that the coarse mesh 'effectively captures the large-scale structures that dominate the flow dynamics' is not supported by this statistic. The authors should either explain why this sign disagreement does not affect the policy, or revise the transfer claim and provide additional coarse-mesh validation of the controlled-flow response.
[Section 4, Eq. (4)] The aerodynamic efficiency is computed as E = C_l/C_d, which excludes the energy input of the actuators. The learned policy uses mean suction at the front jets and blowing at the rear jets with U_jet/U∞ ≈ 0.11, and the reward function in Eq. (4) contains no actuation cost. For active flow control, the net benefit must account for the momentum or power coefficient of the jets; otherwise the 408% efficiency improvement is an upper bound that may be substantially reduced or even reversed when actuation energy is included. The authors should report a net efficiency metric, such as C_l/(C_d + C_μ) or an equivalent power budget, and state the jet momentum coefficient.
[Section 4, Fig. 4] The paper attributes the control benefit to the real-time, closed-loop responsiveness of the DRL policy, noting that the action spectrum is broadband with no dominant frequency. However, the mean action is nonzero (suction at the front, blowing at the rear), so the improvement could in principle be achieved by a steady or simple sinusoidal actuation at the same mean amplitude. To substantiate the claim that the DRL policy's closed-loop adaptation is beneficial, the authors should compare against a steady-blowing/suction baseline with the same mean jet velocity. Without such a comparison, the incremental value of the learned unsteady policy over a much simpler open-loop controller is not established.

minor comments (4)

[Figure 2] The x-axis is described as training steps, but the text states that the number of episodes is divided by n_env to obtain the number of training steps. Please clarify the definition of 'training step' and ensure the axis label is consistent with the text.
[References] The reference to Partee et al. contains typos: 'oceans cimate' should likely be 'ocean climate'. Also 'an Applications' should be 'an application' or similar.
[Section 2, Fig. 4] The notation C'_l,rms and C_l,rms is used interchangeably; please standardize. Also, the Strouhal number St is used in Fig. 4b but not defined in the text.
[Section 2, Agent Configuration] Minor wording: 'neuronal network' should be 'neural network'.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the lift/drag gains are reward objectives, but the learned control policy and coarse-to-fine transfer evaluation are independent.

full rationale

The paper trains a DRL policy with the reward in Eq. 4, which explicitly includes drag reduction and lift increase terms, and then reports those same quantities as achieved improvements. At first glance this looks like the outcome is the reward objective. However, the reward specifies what to optimize, not how: the policy is a 270-dimensional pressure-sensor to jet-velocity mapping learned through PPO, and the reported values are empirical results of evaluating that policy on a different, 418M-DoF fine mesh. The fine-mesh evaluation is a genuine transfer test of a policy trained on a 16M-DoF coarse mesh, and the flow visualizations provide an independent check that the actuation changes the separation topology. Thus the central claim, that DRL can discover a closed-loop reattachment strategy for a 3D separated turbulent wing flow, does not reduce by construction to the reward. The paper does rely heavily on prior work by the same authors for the SOD2D/DRL/MARL framework, but those citations support methodology rather than substitute for the new fine-mesh result. The acknowledged coarse-mesh limitation (e.g., C'_l,rms has opposite sign between meshes in Table 1) is a correctness/fidelity concern, not circularity. Overall no load-bearing circular step was found; score reflects only the minor self-citation density of the framework description.

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

The central claim rests on the reward design (alpha, beta, gamma), the actuator geometry, and the assumption that coarse-mesh training transfers to fine-mesh physics. No physical entities are introduced. The free parameters are chosen by hand with no sensitivity analysis.

free parameters (6)

reward coefficient alpha = 0.3
Eq 4, chosen by hand to penalize lift fluctuations; directly shapes the learned policy and hence the reported improvements.
reward coefficient beta = 0.5
Eq 4, chosen by hand to reward lift increase; the reported +79% lift is an objective of this term.
reward coefficient gamma = 0.8
Eq 3, chosen by hand to weight local vs global reward; affects multi-agent coordination.
actuator locations = x/c=0.01 and 0.4
Front and rear jet slots chosen by design; no sensitivity study. The discovered strategy relies on this geometry.
episode length / action count = T_eps=10.52 U_infinity/c, 120 actions per pseudo-env
Chosen by design; affects training dynamics and final policy.
jet velocity bounds
The policy output is likely bounded but the limit is unspecified; action scale affects the mean Ujet/U_infinity=-0.11 result.

assumptions (5)

domain assumption Large-eddy simulation with ILSA SFS model accurately represents the turbulent separation at Re=60k and AoA=14.
Eqs. 1-2 and Piomelli et al. 2015; the entire study rests on LES fidelity, but no validation against experiments or DNS is provided.
domain assumption Periodic spanwise boundary conditions with Lz=3c model an infinite wing with sufficient spanwise extent for the relevant 3D structures.
Section 2, environment configuration; no spanwise-convergence study.
ad hoc to paper Coarse mesh (16M DoF) provides reward feedback that leads to a policy transferable to fine mesh.
Section 3 first paragraph; the paper asserts this, and the fine-mesh evaluation is the only check, showing mixed consistency (C_l,rms discrepancy).
ad hoc to paper The reward function in Eqs. 3-4 captures the intended aerodynamic objectives; maximizing it yields physically useful control.
No principled derivation; the reported improvements are the direct objectives.
domain assumption Opposite jet velocities at front and rear slots ensure mass conservation and are a valid actuation model.
Section 2; assumes equal slot areas and incompressibility.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Discovering Flow Separation Control Strategies in 3D Wings via Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/3WVCFGPJ

@misc{pith2026250910185,
  author       = {Pith},
  title        = {Pith review of: Discovering Flow Separation Control Strategies in 3D Wings via Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WVCFGPJ}},
  note         = {Machine review of arXiv:2509.10185}
}

read the original abstract

In this work, deep reinforcement learning (DRL) is applied to active flow control (AFC) over a threedimensional SD7003 wing at a Reynolds number of Re = 60,000 and angle of attack of AoA = 14 degrees. In the uncontrolled baseline case, the flow exhibits massive separation and a fully turbulent wake. Using a GPU-accelerated CFD solver and multi-agent training, DRL discovers control strategies that enhance lift (79%), reduce drag (65%), and improve aerodynamic efficiency (408%). Flow visualizations confirm reattachment of the separated shear layer, demonstrating the potential of DRL for complex and turbulent flows.

Figures

Figures reproduced from arXiv: 2509.10185 by the authors.

**Figure 1.** CFD-DRL set-up. Environment Configuration The environment is modelled using the SOD2D solver (Gasparino et al., 2024), a GPU-enabled spectral element method (SEM) code developed at the Barcelona Supercomputing Center (BSC). Its scalability and computational efficiency on distributed accelerator-based architectures make it well-suited for DRL applications, where rapid CFD data generation is crucial for the training… view at source ↗

**Figure 2.** Evolution of the local reward during the training [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Temporal signals of (a) lift Cl and (b) drag Cd coefficients for both uncontrolled and DRLcontrolled cases, evaluated using the coarse and fine meshes. The lift and drag coefficient signals for both meshes in the uncontrolled scenario are shown in black (fine mesh) and grey (coarse mesh) in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: (a) Jet velocity signals Ujet/U∞ for the front jets and (b) their corresponding energy spectra (vertically offset for clarity) for the fine mesh case. The instantaneous actions applied by the agent (for the fine mesh) are shown in Fig. 4a, along with the correspondin…

**Figure 5.** Figure 5: The top row shows vortical structures using [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 5.** Figure 5: On the top row, iso-contours of the Q-criterion colored by the velocity magnitude [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references

[1]

Cerutti, C

J.J. Cerutti, C. Sardu, G. Cafiero and G. Iuso (2020). Active Flow Control on a Square-Back Road Vehicle.Fluids5,

2020
[27]

Vasanth, J

https://doi.org/10.1140/epje/s10189-023-00285-8 J. Vasanth, J. Rabault, F. Alc ´antara- ´Avila, M. Mortensen and R. Vinuesa (2024). Multi-agent Reinforcement Learning for the Control of Three-Dimensional Rayleigh–B´enard Convection.Flow Turbul. Combust. https://doi.org/10.1007/s10494-024-00619-2 Y . Z. Wang, Y . F. Mei, N. Aubry, Z. Chen, P. Wu and W. T. ...

arXiv 2024
[55]

Minelli, S

https://doi.org/10.3390/fluids5020055 G. Minelli, S. Krajnovi´c, B. Basara and B.R. Noack (2016). Numerical Investigation of Active Flow Control Around a Generic Truck A-Pillar.Flow Turbul. Combust.97, 1235–1254. https://doi.org/10.1007/s10494-016-9760-3 I. Rodriguez, O. Lehmkuhl and R. Borrell (2020). Effects of the Actuation on the Boundary Layer of an ...

arXiv 2016
[113]

Su´arez, F

https://doi.org/10.1038/s44172-025-00446-x P. Su´arez, F. ´Alcantara- ´Avila, A. Mir ´o, J. Rabault, B. Font, O. Lehmkuhl and R. Vinuesa (2025b). Active flow control for drag reduction through multi-agent reinforcement learn- ing on a turbulent cylinder atRe D = 3900.Flow Turbul. Combust. https://doi.org/10.1007/s10494-025-00642-x L. Guastoni, J. Rabault,...

doi:10.1038/s44172-025-00446-x 2023
[1422]

Montal `a, B

https://doi.org/10.1038/s41467-025-56408-6 R. Montal `a, B. Font, P. Su ´arez, J. Rabault, O. Lehmkuhl, R. Vinuesa and I. Rodr ´ıguez (2024). Towards Active Flow Control Strategies Through Deep Reinforcement Learn- ing, in: 9th ECCOMAS, 3rd-7th June, Lisbon (Portugal). https://doi.org/10.23967/eccomas.2024.115 R. Montal `a, B. Font, P. Su ´arez, J. Rabaul...

arXiv 2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Cerutti, C

J.J. Cerutti, C. Sardu, G. Cafiero and G. Iuso (2020). Active Flow Control on a Square-Back Road Vehicle.Fluids5,

2020

[2] [27]

Vasanth, J

https://doi.org/10.1140/epje/s10189-023-00285-8 J. Vasanth, J. Rabault, F. Alc ´antara- ´Avila, M. Mortensen and R. Vinuesa (2024). Multi-agent Reinforcement Learning for the Control of Three-Dimensional Rayleigh–B´enard Convection.Flow Turbul. Combust. https://doi.org/10.1007/s10494-024-00619-2 Y . Z. Wang, Y . F. Mei, N. Aubry, Z. Chen, P. Wu and W. T. ...

arXiv 2024

[3] [55]

Minelli, S

https://doi.org/10.3390/fluids5020055 G. Minelli, S. Krajnovi´c, B. Basara and B.R. Noack (2016). Numerical Investigation of Active Flow Control Around a Generic Truck A-Pillar.Flow Turbul. Combust.97, 1235–1254. https://doi.org/10.1007/s10494-016-9760-3 I. Rodriguez, O. Lehmkuhl and R. Borrell (2020). Effects of the Actuation on the Boundary Layer of an ...

arXiv 2016

[4] [113]

Su´arez, F

https://doi.org/10.1038/s44172-025-00446-x P. Su´arez, F. ´Alcantara- ´Avila, A. Mir ´o, J. Rabault, B. Font, O. Lehmkuhl and R. Vinuesa (2025b). Active flow control for drag reduction through multi-agent reinforcement learn- ing on a turbulent cylinder atRe D = 3900.Flow Turbul. Combust. https://doi.org/10.1007/s10494-025-00642-x L. Guastoni, J. Rabault,...

doi:10.1038/s44172-025-00446-x 2023

[5] [1422]

Montal `a, B

https://doi.org/10.1038/s41467-025-56408-6 R. Montal `a, B. Font, P. Su ´arez, J. Rabault, O. Lehmkuhl, R. Vinuesa and I. Rodr ´ıguez (2024). Towards Active Flow Control Strategies Through Deep Reinforcement Learn- ing, in: 9th ECCOMAS, 3rd-7th June, Lisbon (Portugal). https://doi.org/10.23967/eccomas.2024.115 R. Montal `a, B. Font, P. Su ´arez, J. Rabaul...

arXiv 2024