betaFeedback

WorksPhysics and Astronomy

High-lift Wing Separation Control via Bayesian Optimization and Deep Reinforcement Learning

Abstract

This study investigates active flow control (AFC) of a 30P30N high-lift wing at a Reynolds number Rec_c = 450,000 and angle of attack αα = 23∘^\circ using wallresolved large-eddy simulations (LES). Two optimization strategies are explored: open-loop Bayesian optimization (BO) and closed-loop deep reinforcement learning (DRL), both targeting the mitigation of stall and the improvement of aerodynamic efficiency via synthetic jets on the slat, main, and flap elements. The uncontrolled configuration was validated against literature data, confirming the reliability of the LES setup. The BO framework successfully identified steady jet velocities that increased efficiency by +10.9% through a -9.7% drag reduction while maintaining lift. In contrast, the DRL agent, despite leveraging instantaneous flow information from distributed sensors, achieved only minor improvements in lift and drag, with negligible efficiency gain. Training analysis indicated that the penalty-dominated reward constrained exploration. These results highlight the need for carefully designed rewards and computational acceleration strategies in DRL-based flow control at high Reynolds numbers.

Introduction

High-lift wing configurations are essential for modern transport aircraft during take-off and landing, when multi-element devices such as slats and flaps are deployed to generate lift at low airspeeds. Their deployment, however, introduces complex flow phenomena, including separation in slat and cove regions and wake interactions between wing elements. Aircraft performance metrics such as runway distance, climb rate, payload capacity, and noise emissions are therefore strongly influenced by high-lift aerodynamics [1], making accurate understanding and control of these flows critical for efficient and reliable aircraft design.

The 30P30N high-lift configuration, originally developed by McDonnell Douglas, is a widely used benchmark for the aerodynamics and aeroacoustics of multi-element wings. Introduced in the NASA high-lift prediction workshop [2], early studies combined experiments and RANS simulations to assess predictive capabilities. The configuration later became a reference for aeroacoustic investigations within the AIAA BANC-III workshop [3], focusing on tonal noise generated by shear-layer impingement in the slat cove at the Reynolds number Rec=1.7×106Re_c = 1.7 \times10^{6} and angle of attack α=5.5∘\alpha= 5.5^\circ [4, 5]. The Reynolds number is defined as Rec=U∞c/νRe_c = U_\infty c/\nu, with ν\nu being the fluid kinematic viscosity, U∞U_\infty the freestream velocity, and cc the chord of the wing. For high-lift wings, cc refers to the nested chord.

More recently, the 30P30N has been used to evaluate advanced numerical methods, including high-order schemes [6], mesh refinement techniques [7], and hybrid RANS/LES approaches [8]. Recent high-fidelity LES studies have further shown that, at high angles of attack, stall originates from shear-layer separation between the main and flap elements [9, 10], providing a detailed baseline for flow-control studies.

To mitigate stall, active flow control (AFC) has emerged as a promising technique to manipulate wing flows and enhance aerodynamic performance. Among various AFC methods, synthetic jets, characterized by zero net mass flux, have received significant attention due to their effectiveness in controlling flow separation. Numerical studies have shown that synthetic jets can delay stall and improve efficiency at high angles of attack by mitigating large-scale separation [11, 12]. Their applicability has been extended from single-element airfoils to high-lift wing configurations, both experimentally [13] and numerically [14]. More recently, Lehmkuhl et al. [15] performed wall-modeled LES on a complete aircraft (JSM model) under stall conditions, demonstrating the feasibility of AFC at full scale.

Nevertheless, selecting optimal AFC parameters remains a significant challenge due to the complex, nonlinear nature of flow dynamics. In recent years, machine learning (ML) has attracted increasing attention, and its combination with AFC has opened new possibilities. In particular, deep reinforcement learning (DRL) has shown potential as a methodology capable of discovering complex actuation strategies to enhance aerodynamic performance. Rabault et al. [16] pioneered this approach, achieving drag reduction on a two-dimensional cylinder at ReD=100Re_D = 100 (based on the cylinder diameter DD). Building on this, Rabault and Kuhnle [17] introduced multi-environment training strategies to improve learning efficiency and scalability to higher Reynolds numbers. More recently, Suárez et al. [18, 19] extended DRL to three-dimensional cylinders, achieving significant drag reduction at ReD=3,900Re_D = 3,900. Beyond bluff body control, DRL has also been applied to reduce skin friction in wall-bounded turbulence [20] and to three-dimensional Rayleigh-Bénard flows [21], the former representing the first successful DRL study to control turbulence. However, applications to airfoils and wings remain limited, mostly focusing on two-dimensional geometries at low Reynolds numbers [22, 23]. More recently, our group has demonstrated the potential of combining DRL with a GPU-accelerated CFD solver [24]. This approach achieved improved performance for a NACA 0012 wing at Rec=1,000Re_c = 1,000 [25] and an SD7003 wing at Rec=60,000Re_c = 60,000 [26], highlighting the promise of this methodology for more complex configurations.

Another data-driven methodology to tackle optimization problems is Bayesian optimization (BO). This approach has proven robust in fluid mechanics. Morita et al. [27] applied BO to a range of shape optimization problems, Mahfoze et al. [28] used it to reduce skin-friction drag in a turbulent boundary layer via optimum blowing amplitude, Han et al. [29] reduced the drag coefficient of a cylinder at ReD=200Re_D = 200 by selecting appropriate steady tangential velocities along the surface, and Li et al. [30] enhanced the jet mixing by optimizing multiple actuation inputs in both the bulk flow and nozzle periphery. Although less advanced than DRL a priori, as it typically relies on fixed control parameters rather than producing a closed-loop control strategy, BO remains an efficient and reliable tool for identifying high-performance configurations.

The present study focuses on the 30P30N geometry at α=23∘\alpha= 23^\circ and Rec=450,000Re_c = 450,000, conditions where stall separation dominates the flow. The objective is to apply an AFC strategy that mitigates this separation and thus improves the wing performance. Two optimization approaches are employed: BO, which systematically explores the actuation parameter space to identify optimal configurations, and DRL, which enables the autonomous discovery of complex, time-dependent control strategies that adapt to the evolving flow dynamics.

Methodology

In this study, wall-resolved LES are performed for the flow over the 30P30N high-lift wing at a Reynolds number of Rec=450,000Re_c = 450{,}000 and an angle of attack of α=23∘\alpha= 23^\circ. The objective is to evaluate the effectiveness of synthetic jets in mitigating stall conditions. Two optimization strategies are employed to determine the optimal jet mass flow rate: (i) a classical BO approach, and (ii) a more novel method based on DRL. The Reynolds number has been slightly reduced compared to values reported in the literature to decrease the computational cost of the simulations without altering the flow physics.

CFD solver

The CFD simulations are performed by solving the filtered incompressible Navier-Stokes equations:

∂u‾i∂xi=0(1)\frac{\partial\overline{u}_i}{\partial x_i}=0 \tag*{(1)}
∂u‾i∂t+∂∂xj(u‾iu‾j)−ν∂2u‾i∂xj∂xj+1ρ∂p‾∂xi=−∂τij∂xj(2)\frac{\partial\overline{u}_i}{\partial t}+\frac{\partial}{\partial x_j}\left(\overline{u}_i\overline{u}_j\right)-\nu\frac{\partial^2\overline{u}_i}{\partial x_j\partial x_j}+\frac{1}{\rho}\frac{\partial\overline{p}}{\partial x_i}=-\frac{\partial\tau_{ij}}{\partial x_j} \tag*{(2)}

where u‾i\overline{u}_i (i=1,2,3i=1,2,3) and p‾\overline{p} denote the filtered velocity and pressure fields, respectively, ρ\rho is the fluid density, and τij\tau_{ij} is the subgrid-scale (SGS) stress tensor. The latter represents the effect of unresolved turbulent motions and must therefore be modeled. Its deviatoric part is here modeled through a subgrid-scale viscosity νSGS\nu_{\mathrm{SGS}} as

τij−13τkkδij=−2νSGSS‾ij(3)\tau_{ij}-\frac{1}{3}\tau_{kk}\delta_{ij}=-2\nu_{\mathrm{SGS}}\overline{S}_{ij} \tag*{(3)}

where S‾ij=1/2(∂u‾i/∂xj+∂u‾j/∂xi)\overline{S}_{ij}=1/2(\partial\overline{u}_i/\partial x_j+\partial\overline{u}_j/\partial x_i) is the rate-of-strain tensor of the resolved velocity field, and δij\delta_{ij} is the Kronecker delta. The subgrid-scale viscosity is here evaluated using the Vreman model [31]. The simulations are carried out using SOD2D [32], an in-house GPU-enabled CFD solver developed at the Barcelona Supercomputing Center, which is based on the spectral element method (SEM).

The three-dimensional computational domain extends Lx/c=20L_x/c=20, Ly/c=20L_y/c=20, and Lz/c=0.1L_z/c=0.1 in the streamwise, cross-stream, and spanwise directions, respectively. The 30P30N wing is positioned approximately at the center of the domain and spans the entire width. The domain size is selected based on previous studies [9, 10]. A uniform freestream velocity U∞U_\infty is imposed at the inlet, with its xx- and yy-components determined by the angle of attack α\alpha. At the outlet, zero-gradient boundary conditions are applied to the velocity field, while the pressure is fixed. A no-slip condition is enforced on all solid surfaces, whereas Dirichlet conditions are used at the jet locations to prescribe jet velocities when AFC is activated.

One synthetic jet is positioned on each wing element. These are located at xjet,slat/c=−0.050x_{\mathrm{jet,slat}}/c=-0.050, xjet,main/c=0.725x_{\mathrm{jet,main}}/c=0.725, and xjet,flap/c=1.0x_{\mathrm{jet,flap}}/c=1.0. These positions are selected to provide the optimization algorithms sufficient flexibility to influence the flow over all elements. The slat and main element jets are positioned to interact with their respective wakes, whereas the flap jet targets the upper-surface recirculation region. The jet angles are set to ϕjet=90∘\phi_{\mathrm{jet}}=90^\circ for suction and ϕjet=20∘\phi_{\mathrm{jet}}=20^\circ for blowing, measured with respect to the wall-tangential direction. The streamwise jet width is hjet/c=0.01h_{\mathrm{jet}}/c=0.01, and each jet spans the full domain width along the spanwise direction.

In this study, the velocities of these synthetic jets are determined through optimization. Specifically, BO is used to select a fixed, steady jet velocity, while DRL provides closed-loop control, allowing the agent to adjust the jet velocity instantaneously based on the evolving flow state. In both cases, the optimization determines the appropriate velocity for each jet, but DRL offers dynamic control, whereas BO yields a single optimal steady velocity.

Bayesian optimization

For the BO process, fixed jet velocities are applied to the slat and main jets, while the velocity of the flap jet is constrained to ensure mass conservation, being set to produce an opposite mass flow rate equal to the sum of the other two jets:

Ujet,flap=−(Ujet,slat+Ujet,main).(4)U_{\mathrm{jet,flap}} = -\left(U_{\mathrm{jet,slat}} + U_{\mathrm{jet,main}}\right). \tag*{(4)}

Therefore, two control parameters are optimized during the BO process: the slat and main jet velocities. In each simulation, fixed values of jet velocities are prescribed, where the exploration range is set to Ujet,slat/U∞∈[−1.22,1.22]U_{\mathrm{jet,slat}}/U_{\infty} \in[-1.22,1.22] and Ujet,main/U∞∈[−1.22,1.22]U_{\mathrm{jet,main}}/U_{\infty} \in[-1.22,1.22] to prevent impractical actuations in which the jet velocities would significantly exceed the freestream velocity. Additionally, this maximum magnitude is based on the results by Shmilovich and Yadlin [14], who used a momentum coefficient of Cμ=(hjet/c)(Ujet,max2/U∞2)=0.015C_{\mu} = (h_{\mathrm{jet}}/c)(U_{\mathrm{jet,max}}^{2}/U_{\infty}^{2}) = 0.015 in their AFC numerical simulations.

The objective (or target) function to be optimized in the BO framework is defined as

r=E−EbaseEbase−WClPCl−WCdPCd,(5)r = \frac{E-E_{\mathrm{base}}}{E_{\mathrm{base}}} - W_{C_l}P_{C_l} - W_{C_d}P_{C_d}, \tag*{(5)}

where

PCl={Cl,base−ClCl,base,if Cl<Cl,base,0,otherwise.(6)P_{C_l} = \begin{cases} \frac{C_{l,\mathrm{base}}-C_l}{C_{l,\mathrm{base}}}, & \text{if } C_l < C_{l,\mathrm{base}},\\ 0, & \text{otherwise.} \end{cases} \tag*{(6)}

and

PCd={Cd−Cd,baseCd,base,if Cd>Cd,base,0,otherwise.(7)P_{C_d} = \begin{cases} \frac{C_d-C_{d,\mathrm{base}}}{C_{d,\mathrm{base}}}, & \text{if } C_d > C_{d,\mathrm{base}},\\ 0, & \text{otherwise.} \end{cases} \tag*{(7)}

The objective rr aims to maximize the aerodynamic efficiency E=Cl/CdE = C_l/C_d relative to the non-actuated baseline case EbaseE_{\mathrm{base}}. However, improvements in EE may lead to undesirable trade-offs, such as reduced lift or increased drag. To penalize such effects, penalty terms PClP_{C_l} and PCdP_{C_d} are introduced whenever the lift coefficient decreases, or the drag coefficient increases relative to the uncontrolled scenario. These penalties are expressed as relative deviations and are weighted by WClW_{C_l} and WCdW_{C_d} in the objective function. In this work, both weights are set to WCl=100W_{C_l} = 100 and WCd=100W_{C_d} = 100 to impose strong constraints, ensuring that neither lift reduction nor drag increase occurs in the optimized configuration.

The relationship between the control parameters v=[Ujet,slat,Ujet,main]\mathbf{v} = [U_{\mathrm{jet,slat}}, U_{\mathrm{jet,main}}] and the objective rr is modeled using Gaussian process (GP) surrogate models. A GP is a probabilistic model used to approximate an unknown function, here representing aerodynamic quantities such as Cl=fl(v)C_l = f_l(\mathbf{v}), Cd=fd(v)C_d = f_d(\mathbf{v}), and E=fE(v)E = f_E(\mathbf{v}), which are combined to construct the objective function rr defined above. Then, at any new control input v∗=[Ujet,slat,Ujet,main]\mathbf{v}^{*} = [U_{\mathrm{jet,slat}}, U_{\mathrm{jet,main}}], the GP provides a probabilistic prediction of the form

f(v∗)∼N(μ(v∗),σ2(v∗)),(8)f(\mathbf{v}^{*}) \sim\mathcal{N}\left(\mu(\mathbf{v}^{*}),\sigma^{2}(\mathbf{v}^{*})\right), \tag*{(8)}

for each aerodynamic quantify, where μ(v∗)\mu(\mathbf{v}^{*}) is the predicted mean (surrogate estimate) and σ(v∗)\sigma(\mathbf{v}^{*}) is the standard deviation, representing the model uncertainty. These quantities are computed as

μ(v∗)=κT(K+σn2I)−1y,σ2(v∗)=k(v∗,v∗)−κT(K+σn2I)−1κ.(9)\mu(\mathbf{v}^{*}) = \boldsymbol{\kappa}^{T}(K+\sigma_{n}^{2}I)^{-1}\mathbf{y}, \qquad\sigma^{2}(\mathbf{v}^{*}) = k(\mathbf{v}^{*},\mathbf{v}^{*})-\boldsymbol{\kappa}^{T}(K+\sigma_{n}^{2}I)^{-1}\boldsymbol{\kappa}. \tag*{(9)}

Here y=[y1,…,yN]T\mathbf{y} = [y_1,\ldots,y_N]^T are the outputs of the previous NN CFD simulations, corresponding to the training set {(vi,yi)}i=1N\{(\mathbf{v}_i,y_i)\}_{i=1}^{N}. The covariance matrix K∈RN×NK \in\mathbb{R}^{N \times N} contains the covariances between all training points Kij=k(vi,vj)K_{ij}=k(\mathbf{v}_i,\mathbf{v}_j), where k(vi,vj)k(\mathbf{v}_i,\mathbf{v}_j) is the kernel function. In this work, a radial basis function (RBF) kernel is employed,

k(vi,vj)=exp⁡(−∥vi−vj∥22l2),(10)k(\mathbf{v}_i,\mathbf{v}_j)=\exp\left(-\frac{\lVert\mathbf{v}_i-\mathbf{v}_j\rVert^2}{2l^2}\right), \tag*{(10)}

with ∥vi−vj∥\lVert\mathbf{v}_i-\mathbf{v}_j\rVert denoting the Euclidean distance between the two input vectors and ll the kernel lengthscale controlling the decay of correlation. The vector κ=[k(v1,v∗),…,k(vN,v∗)]T\boldsymbol{\kappa}=[k(\mathbf{v}_1,\mathbf{v}^{*}),\ldots,k(\mathbf{v}_N,\mathbf{v}^{*})]^{T} represents the covariance between the new input v∗\mathbf{v}^{*} and each training point vi\mathbf{v}_i, while k(v∗,v∗)k(\mathbf{v}^{*},\mathbf{v}^{*}) denotes the self-covariance of the new input. The noise variance σn2\sigma_n^2 accounts for numerical uncertainty in the CFD outputs, and both σn2\sigma_n^2 and ll are fitted during GP training. Each newly sampled configuration v∗\mathbf{v}^{*} updates the GP model at each BO iteration. The predicted mean guides the optimization toward promising regions of the control space, while the associated uncertainty promotes exploration of under-sampled regions, ensuring efficient convergence toward the optimal jet configuration.

DRL optimization

The DRL framework, illustrated in Figure 1, consists of two main components: the CFD environment, which was presented in § 2.1, and the DRL agent. The latter is implemented as a neural network (NN) and maps flow states to control actions. In this setup, the NN receives pressure measurements from sensors within the CFD environment and outputs the slat and main jet velocities (Ujet,slatU_{\mathrm{jet,slat}} and Ujet,mainU_{\mathrm{jet,main}}) to be applied back to the system. The flap jet is constrained to operate at an opposing mass flow rate relative to the other jets, as defined in Eq. (4). The agent is implemented using the TF-Agents Python library [33] and communicates with the Fortran-based CFD solver via a Redis in-memory database managed by SmartSim [34], enabling efficient coupling and online training of the agent.

CFD-DRL set-up

Figure 1. CFD-DRL set-up

To enable three-dimensional actuation without excessive computational cost, a multi-agent reinforcement learning (MARL) framework is employed [35, 18, 19, 24]. The computational domain is then divided into multiple pseudo-environments, each containing its own set of jets (slat, main, and flap), while all pseudo-environments share a single agent. Thus, based on local flow states, the agent outputs distinct jet velocities for each pseudo-environment, allowing spatially distributed control with manageable training complexity. In this study, three pseudo-environments are used (nMARL=3n_{\mathrm{MARL}}=3), each spanning Lz,MARL/c=0.1L_{z,\mathrm{MARL}}/c = 0.1, so that the DRL domain is three times larger (Lz/c=0.3L_z/c = 0.3) than in the baseline configuration.

To accelerate data collection and reduce wall-clock time, ten independent CFD simulations run concurrently (nCFD=10n_{\mathrm{CFD}} = 10). At the end of each episode, a total of 3×10=303 \times10 = 30 state-action-reward trajectories are collected for policy updates. Each episode lasts TepsU∞/c=8.57T_{\mathrm{eps}} U_{\infty}/c = 8.57, corresponding to twelve vortex-shedding cycles in the uncontrolled case. A total of 120 actions are applied per pseudo-environment, yielding an action duration of Tact=Teps/120T_{\mathrm{act}} = T_{\mathrm{eps}}/120.

Each pseudo-environment contains 78 sensors in the zz-midplane (see Figure 1). To provide three-dimensional flow information, the NN input includes sensor slices from adjacent pseudo-environments, producing a total state vector of size 78×3=23478 \times3 = 234.

The proximal policy optimization (PPO) algorithm [36] is used to refine the agent’s policy. Both the actor and critic networks are implemented as multi-layer perceptrons (MLPs) with two hidden layers of 512 neurons each. The actor maps observed states to actions, while the critic estimates the advantage function, quantifying the relative benefit of an action compared to the current policy. The critic is trained using trajectory data at the end of each episode, and the actor is subsequently updated based on these advantage estimates, with PPO’s clipped objective ensuring stable policy updates.

The reward function for DRL is based on the same target quantity used in the BO optimization, as defined in Eq. (5). Within the MARL framework, each pseudo-environment computes a local reward rir_i, which is then combined with the rewards from the other pseudo-environments to form the final training reward RiR_i:

Ri=γri+1−γnMARL∑j=1nMARLrj,(11)R_i = \gamma r_i + \frac{1-\gamma}{n_{\mathrm{MARL}}} \sum_{j=1}^{n_{\mathrm{MARL}}} r_j, \tag*{(11)}

where the weighting factor γ\gamma balances the relative importance of local versus averaged global performance. In this work, this is set to γ=0.8\gamma= 0.8.

Results

This section first validates the non-actuated scenario to ensure the reliability of the subsequent BO and DRL results (§ 3.1). It also establishes a reference case to facilitate comparison and quantify the performance improvements achieved by the optimization processes. Following this, the AFC results obtained using the BO and DRL methodologies are analyzed in § 3.2 and § 3.3, respectively.

Baseline validation

In this section, the baseline case without AFC is validated against numerical results available in the literature. Since no studies at the same Reynolds number are present in the literature, the results from Montañá et al. [9, 10] obtained at a slightly higher Reynolds number (Rec=750,000Re_c = 750{,}000) and the same angle of attack (α=23∘\alpha= 23^\circ) are used for comparison. Although minor discrepancies may arise due to the difference in Reynolds number, the overall flow physics are expected to be similar, and thus only small variations in the results are anticipated.

In this work, two mesh refinement levels are employed. Since both optimization strategies (BO and DRL) require running multiple CFD simulations with varying parameters, a coarse mesh is used during model training, while the final evaluations are performed on the fine mesh. This approach significantly reduces the overall computational cost of both methodologies. This mesh refinement is achieved through pp-refinement, i.e., by increasing the polynomial order of the high-order basis functions used in SEM. For the coarse mesh, a polynomial order of p=2p = 2 is used, while for the fine mesh this is increased to p=4p = 4. Consequently, the topology and geometry of the elements remain identical between the two meshes, and only the number of grid points within each element increases from the coarse to the fine mesh.

The mesh characteristics are summarized in Table 1. The table also includes the dimensionless wall distances of the mesh. These quantities are evaluated in the fully turbulent regions of the flow along the suction sides of the elements, and both the average and maximum values are reported. Although the coarse mesh slightly exceeds the recommended limits for wall-resolved LES, particularly the Δy+\Delta y^{+} distances, the fine mesh lies well within the accepted ranges [37], ensuring that Δx+<150\Delta x^{+} < 150, Δy+<2\Delta y^{+} < 2, and Δz+<40\Delta z^{+} < 40 across the entire wall surface where turbulence is present.

NxyN_{xy}NzN_zNxyzN_{xyz}Δxmax⁡+\Delta x^{+}_{\max}Δymax⁡+\Delta y^{+}_{\max}Δzmax⁡+\Delta z^{+}_{\max}Δxavg+\Delta x^{+}_{\mathrm{avg}}Δyavg+\Delta y^{+}_{\mathrm{avg}}Δzavg+\Delta z^{+}_{\mathrm{avg}}
Coarse Mesh2.10×1052.10 \times10^{5}4910 M1253.570722.248
Fine Mesh8.38×1058.38 \times10^{5}9781 M851.540430.928

Table 1. Mesh information for the two mesh refinements considered, including the number of grid points in the xx–yy plane NxyN_{xy}, along the zz-direction NzN_z, and the total number of points NxyzN_{xyz} (in millions). Also, the maximum and average wall-tangential Δx+\Delta x^{+}, wall-normal Δy+\Delta y^{+}, and spanwise Δz+\Delta z^{+} distances are reported.

To assess differences with the literature, Figure 2 shows the distributions of the pressure coefficient CpC_p and skin friction coefficient CfC_f over the surfaces of each element. From this figure, it is evident that the fine mesh exhibits better agreement with the reference results than the coarse mesh, even though some discrepancies are anticipated due to the differences in the Reynolds number. Differences are especially pronounced for the skin friction coefficient, which is more sensitive to Reynolds number and near-wall resolution. The coarse mesh exhibits larger deviations due to its higher Δy+\Delta y^{+}, although both meshes correctly capture overall trends and transition locations. Since the flow is dominated by pressure forces at this high angle of attack, discrepancies in skin friction have only a minor effect on the overall aerodynamic coefficients.

Pressure coefficient and skin friction coefficient distributions for the non-actuated case

Figure 2. Pressure coefficient CpC_p (left) and skin friction coefficient CfC_f (right) distributions for the non-actuated case over the walls of the wing. The skin friction coefficient is only shown on the suction side of the main and flap elements.

The time-averaged aerodynamic coefficients obtained with each mesh are reported in Table 2, including the lift ClC_l and drag CdC_d coefficients, along with the root-mean-square values of their fluctuations, Cl,rms′C'_{l,\mathrm{rms}} and Cd,rms′C'_{d,\mathrm{rms}}, respectively. From this table, it is evident that, despite the considerably lower resolution of the coarse mesh, it still predicts the lift coefficient reasonably well. Since the lift is mainly dominated by the pressure contribution, the difference between the coarse and fine meshes is less than 1%. The drag coefficient, however, shows a much larger discrepancy, with a difference of approximately 48%. This is because the skin friction contribution, which is more sensitive to mesh resolution, is comparatively higher in the drag coefficient, leading to greater differences between the coarse and fine meshes. Additionally, the offset in the pressure coefficient distribution downstream of the suction peaks further contributes to this discrepancy (see Figure 2). When comparing the rms values, a similar behavior is observed: Cl,rms′C'_{l,\mathrm{rms}} is similar for both the coarse and fine meshes, whereas the difference in Cd,rms′C'_{d,\mathrm{rms}} is considerably larger. Nevertheless, this is not considered critical, as the coarse mesh is used only for training the models and not for their final evaluation.

CASEClC_lCl,rms′C'_{l,\mathrm{rms}}CdC_dCd,rms′C'_{d,\mathrm{rms}}EE
Baseline Coarse3.95550.02650.36270.006910.91
Baseline Fine3.91870.02660.24480.011816.01
Montalà et al. [9, 10]4.0627-0.1983-20.49

Table 2. Time-averaged aerodynamic coefficients for the non-actuated case are shown for both the coarse and fine meshes, compared against values reported in the literature. Each column displays the lift coefficient ClC_l, the root-mean-square of the lift coefficient fluctuations Cl,rms′C'_{l,\mathrm{rms}}, the drag coefficient CdC_d, and the root-mean-square of the drag coefficient fluctuations Cd,rms′C'_{d,\mathrm{rms}}.

Figure 3 shows instantaneous flow visualizations. In these images, the higher resolution of the fine mesh is evident, with many more small-scale details captured compared to the coarse mesh. Nevertheless, the same physical features are present in both grids despite the differences in accuracy. The recirculation bubbles in the main and slat cove regions have roughly the same size, the laminar-to-turbulent transition regions occur at the same positions on the leading edges of each element, and similar wakes are observed. All these flow phenomena were thoroughly explained in the work of Montalà et al. [9, 10], and the results here are consistent with their observations.

Instantaneous velocity magnitude fields and vortical structures for coarse and fine meshes

Figure 3. Instantaneous fields from the coarse (left panel) and fine meshes (right panel) in the uncontrolled scenario. The top rows show the velocity magnitude field in the mid-zz plane, while the bottom rows depict the vortical structures visualized using Q-criterion iso-contours, colored by velocity magnitude.

The main difference between the coarse and fine meshes appears in the low-velocity wake region above the flap. In the fine mesh, the wake exhibits a pronounced curvature, which is a signature of vortex shedding. However, this is not captured by the coarse mesh and, although the main statistical quantities are accurately captured by both meshes, as discussed in Table 2, the coarse mesh does not fully resolve the instantaneous dynamics. Consequently, the characteristic peak associated with vortex shedding in the Fourier transform of the lift signal is absent in the coarse mesh. Nevertheless, this limitation is not considered important, as the time-averaged coefficients used for training the models are still well represented.

Bayesian optimization

Initially, 15 cases were evaluated to construct the GP models, which were iteratively refined as the BO algorithm proposed new sampling points. The optimization proceeded until convergence, defined as the situation in which a newly proposed point was very close to the previous one and the associated uncertainty in the optimization space was nearly eliminated. In total, 4 additional simulations were required beyond the initial samples, resulting in 19 evaluated cases. All simulations were performed using the coarse mesh described in § 3.1.

Figure 4 shows the aerodynamic efficiency EE, lift coefficient ClC_l, and drag coefficient CdC_d isocontours as functions of the slat and main jet velocities (Ujet,slatU_{\mathrm{jet,slat}} and Ujet,mainU_{\mathrm{jet,main}}) predicted by the GP models. The black points indicate the data used to train the model, while the red point represents the optimum identified by the BO algorithm. White dashed lines are also included. These lines indicate the combinations of slat and flap jet actuation that yield the same aerodynamic coefficients as the non-actuated case (see Table 2). These plots confirm that the BO approach successfully increases aerodynamic efficiency EE while remaining in the region of the design space where Cl>Cl,baseC_l > C_{l,\mathrm{base}} and Cd<Cd,baseC_d < C_{d,\mathrm{base}}, which is consistent with the constraints imposed in the target function defined in Eq. (5).

Efficiency, lift coefficient, and drag coefficient manifolds

Figure 4. Efficiency EE (left), lift coefficient ClC_l (middle), and drag coefficient CdC_d (right) manifolds predicted by the GP models as a function of the slat and main jet velocities.

To further illustrate the uncertainty across the design space, Figure 5 presents two slices (A and B) highlighted in Figure 4 as red dashed lines, corresponding to cross-sections through the optimal point along each direction. It is evident that the uncertainty at the optimal point is nearly zero, while regions with fewer data points and less exploration exhibit noticeably higher uncertainty. Moreover, even in these high-uncertainty regions, the aerodynamic efficiency is lower with 95% confidence. These slices also show that the maximum efficiency is not located exactly at the optimal point; however, the selected optimum satisfies the constraints imposed by the target function. In the region of maximum EE, the drag coefficient is lower than at the optimal point, but the lift coefficient is also significantly reduced, even below the baseline value, violating the constraint. Therefore, the BO algorithm successfully balances efficiency improvement with the lift and drag constraints.

Figure 5. Slice A (left panel) and B (right panel) of the efficiency EE (top), lift coefficient ClC_l (middle), and drag coefficient CdC_d (bottom) manifolds predicted by the GP models, corresponding to cross-sections through the optimal point in Figure 4.

The results for the optimal point are reported in Table 3. Once the optimal point was identified, the results were also evaluated on the fine mesh to ensure the reliability of the optimum. This table therefore presents the results for both the coarse and fine meshes, evaluated at the optimal point shown in Figure 4 and Figure 5, i.e. Ujet,slat/U∞=−1.22U_{\mathrm{jet,slat}}/U_{\infty}=-1.22 and Ujet,main/U∞=0.16U_{\mathrm{jet,main}}/U_{\infty}=0.16

CASEClC_lCl,rms′C'_{l,\mathrm{rms}}CdC_dEE
Opt. solution (Coarse)3.9654 (+0.25%)0.0299 (+12.83%)0.3157 (-12.97%)12.56 (+15.19%)
Opt. solution (Fine)3.9244 (+0.15%)0.0289 (+8.30%)0.2210 (-9.69%)17.76 (+10.89%)

Table 3. Time-averaged aerodynamic coefficients for the coarse and fine meshes in the BO-optimized AFC scenario. Each column displays the lift coefficient ClC_l, the root-mean-square of the lift coefficient fluctuations Cl,rms′C'_{l,\mathrm{rms}}, the drag coefficient CdC_d, and the aerodynamic efficiency EE, along with the relative changes with respect to the baseline scenario for each mesh.

Despite the expected differences between the coarse and fine meshes, the coarse mesh is able to provide reasonable predictions that translate into an improvement in wing performance on the fine mesh, confirming the reliability of the BO study conducted with the coarse grid. Although the final efficiency achieved on the fine mesh is slightly lower than that on the coarse mesh, compared to the respective baseline scenarios, the results are still promising: the aerodynamic efficiency increases by up to 10.89% compared with the baseline scenario, while the lift is maintained (+0.15%) and the drag is substantially reduced (-9.69%), showing trends similar to those observed on the coarse mesh. Therefore, the improvement in aerodynamic efficiency is primarily driven by drag reduction.

Figure 6 shows the flow streamlines for the two analyzed cases, the non-actuated and BO-optimized AFC scenarios. The application of suction on the slat suction side reduces the boundary layer thickness, causing the slat wake to nearly disappear downstream. This is then a signature of drag reduction and hence, the slat element primarily drives the decrease in drag. However, the recirculation area above the flap is larger than in the baseline scenario, increasing the drag of the main and flap elements. Nevertheless, the net effect of the slat compensates for these increases, resulting in an overall drag reduction compared to the baseline scenario, and, consequently, an improvement in aerodynamic efficiency in the AFC configuration.

Streamlines of the time-averaged flow field with velocity magnitude in the background for the baseline scenario and the BO-optimized AFC scenario using the fine mesh

Figure 6. Streamlines of the time-averaged flow field with velocity magnitude in the background for the baseline scenario (left) and the BO-optimized AFC scenario (right) using the fine mesh.

DRL optimization

In this section, the slat and main jet velocities are determined through an optimization process based on the training of a DRL agent, as described in § 2.3. The objective is to demonstrate that a closed-loop control strategy, where actions are selected according to the instantaneous flow-field state, can outperform a conventional open-loop optimization with a fixed actuation law, such as the one presented in § 3.2. This richer agent-environment interaction is expected to exploit more complex flow dynamics, potentially leading to greater improvements in aerodynamic performance, especially considering the MARL framework used, which enables three-dimensional actuations.

Figure 7 illustrates the evolution of the lift coefficient ClC_l, drag coefficient CdC_d, and reward rr during the DRL training. In total, 990 MARL episodes are performed. This value considers all the environments nenv=30n_{\mathrm{env}}=30 executed in parallel as individual episodes, which translates into 990/30=33990/30=33 policy updates. As training progresses, ClC_l increases until reaching a plateau in the final episodes where, although a few episodes slightly outperform the baseline case, the average ClC_l remains below the baseline value. The drag coefficient CdC_d, which initially matches the baseline level, shows a gradual reduction toward the end of the training.

Lift coefficient, drag coefficient and local reward evolution during DRL training

Figure 7. Lift coefficient ClC_l (left), drag coefficient CdC_d (middle) and local reward rr (right) evolution during the DRL training. All quantities are averaged over the last 4U∞/c4U_{\infty}/c of each episode and values from each MARL episode are concatenated sequentially.

Consequently, the reward rr increases over the course of training, indicating a certain degree of learning. However, this improvement primarily stems from the reduction of penalty terms in the reward (Eqs. (6) and (7)) rather than from the improvement of aerodynamic efficiency EE, which is the main goal of the selected reward. Specifically, as the drag constraint (Cd<Cd,baseC_d < C_{d,\mathrm{base}}) is satisfied at roughly the middle of the training, the subsequent reward growth is mainly due to a reduction in the lift penalty term. Nevertheless, since the lift coefficient asymptotically approaches but never surpasses the baseline value, the penalty term remains nonzero. As a result, no significant improvement in aerodynamic efficiency is achieved, which stays nearly constant throughout the training. The overall trend thus suggests that the penalty cannot be eliminated, preventing any increase in the efficiency term.

The hard constraint imposed within the reward function may limit the agent’s ability to explore and progress toward regions of the optimization space that could enhance efficiency. This might create a local region from which the agent cannot escape once it converges there. A more continuous reward formulation might be a better selection for DRL agents and consequently, the reward design may not be the optimal in this case. As observed in other DRL studies, the final performance is highly sensitive to the reward formulation [25]. Small modifications, such as adjusting the weighting factors of the penalty terms, can substantially alter the exploration landscape, leading to distinct solution manifolds and training outcomes.

Nevertheless, the practical computational limits of the current framework are being reached. While previous wing-based trainings required approximately 10 minutes per episode at Rec=1,000Re_c = 1{,}000 [25] and 120 minutes per episode at Rec=60,000Re_c = 60{,}000 [26], the present simulations require about 800 minutes per episode. This represents an exponential increase in wall-clock time with the Reynolds number. Considering the total number of 990 MARL episodes and the nenv=30n_{\mathrm{env}} = 30 parallel environments being used, the overall training duration is approximately two weeks. The increase in computational cost primarily arises from the smaller time steps required at higher Reynolds numbers. The training was conducted in 30 nodes of the MN5 supercomputer cluster located at the Barcelona Supercomputing Center, each equipped with four NVIDIA H100 GPUs and two Intel XEON processors (80 CPU cores in total), yielding a total computational cost of approximately 48,000 GPU-h.

Given this computational expense, there is little room for hyperparameter tuning or testing alternative reward functions, and strategies to accelerate the training process must be considered. Several approaches can be envisioned. Even though the computational mesh cannot be further coarsened as it already lies within the acceptable resolution limits for LES, RANS-based training could be explored, as only time-averaged aerodynamic coefficients are required for learning.

Another possibility is to increase the number of domain partitions (to allocate more GPUs per CFD simulation) or to increase the number of parallel CFD simulations. Regarding the former, previous experience indicates that the minimum wall-clock time in SOD2D is achieved with approximately one million degrees of freedom (DOF) per GPU, whereas the present setup employs 2.5 million DOF per GPU. Concerning the latter, the DRL framework exhibits excellent scalability with the number of parallel CFD simulations. Although the upper limit of this scalability is still unknown, increasing from the current nCFD=10n_{\mathrm{CFD}} = 10 to around 30 parallel CFD runs is expected to be feasible. These two strategies could yield an approximate 6× reduction in wall-clock time while maintaining the overall computational cost.

Another potential strategy is the incorporation of surrogate models, which could replace high-fidelity CFD simulations in selected episodes. These surrogate-based episodes could be complemented by periodic high-fidelity simulations to refine both the surrogate model and the DRL agent simultaneously. Finally, transfer learning could also be employed by initially training at lower Reynolds numbers. This would significantly reduce the mesh size and time-step constraints, which otherwise increase exponentially with ReRe.

Conclusions

Wall-resolved large-eddy simulations (LES) of the 30P30N high-lift wing at Rec=450,000Re_c = 450{,}000 and α=23∘\alpha= 23^\circ were performed to evaluate active flow control (AFC) strategies using synthetic jets. Two optimization approaches were considered: Bayesian optimization (BO) for open-loop control and deep reinforcement learning (DRL) for closed-loop control, both aiming to mitigate stall and improve aerodynamic efficiency by adjusting the slat, main, and flap jet velocities.

The uncontrolled baseline was validated against literature data, showing good agreement in pressure distributions and aerodynamic coefficients. The fine mesh provided wall-resolved LES accuracy, while the coarser mesh offered reliable predictions at lower computational cost, enabling efficient optimization.

BO successfully identified an optimal steady jet configuration, increasing aerodynamic efficiency by +10.9% relative to the baseline, maintaining lift, and reducing drag by -9.7%. Flow analysis showed that suction on the slat element primarily drove the improvement by thinning the boundary layer and reducing wake intensity, while drag on the main and flap elements increased slightly. These results demonstrate BO’s ability to efficiently explore the control space and identify interpretable near-optimal AFC strategies with few CFD evaluations.

In contrast, the DRL-based closed-loop control, despite using instantaneous flow information from distributed sensors, did not significantly improve efficiency. The agent slightly reduced drag and nearly maintained lift, but the reward was dominated by penalty terms, limiting gains in aerodynamic performance. This highlights the importance of reward design in DRL and suggests smoother formulations could enhance learning. However, high computational cost remains a bottleneck, restricting hyperparameter tuning and reward exploration. Future work should focus on accelerating training through surrogate modeling, transfer learning, or improved parallelization to allow more extensive exploration of the control space.

Acknowledgments

This research was partially funded by the Ministerio de Ciencia e Innovación of Spain (PID2023-150408OB-C21/C22). Work has also been co-funded by the European High Performance Computing Joint Undertaking (EuroHPC JU) and Germany, Italy, Slovenia, Spain, Sweden, and France under grant agreement No 101092621; and under grant PCI2022-135085-2 funded by MICIU/AEI/10.13039/501100011033 and by the “European Union NextGenerationEU/PRTR”. Simulations were supported by the Red Española de Supercomputación (IM-2024-2-0004 and IM-2024-3-0005) and EuroHPC JU (EHPC-REG-2024R01-038), granting access to MN5 at the Barcelona Supercomputing Center. We also acknowledge AGAUR for supporting the LS/CFD (2021 SGR 00902) and TUAREG (2021 SGR 01051) research groups. R. M. thanks AGAUR for the FI-SDUR grant (2022 FISDU 00066), and R. V. acknowledges financial support from ERC grant no. 2021-CoG-101043998, DEEPCONTROL. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council.

References

References

  1. [1]Thibert T J, Reneaux J, Moens F and Preist J 1995 The Aeronautical Journal 99 395–411
  2. [2]Klausmeyer S M and Lin J C 1997 Comparative results from a CFD challenge over a 2D three-element high-lift airfoil Tech. Rep. NASA TM-112858 NASA Langley Research Center NASA Technical Memorandum
  3. [3]Choudhari M and Lockard D 2015 Assessment of slat noise predictions for 30P30N high-lift configuration from BANC-III workshop 21st AIAA/CEAS Aeroacoustics ConferenceDOI
  4. [4]Pascioni K, Cattafesta L and Choudhari M 2014 An experimental investigation of the 30P30N multi-element high-lift airfoil 20th AIAA/CEAS Aeroacoustics ConferenceDOI
  5. [5]Ashton N, West A and Mendonça F 2016 AIAA Journal 54 3657–3667
  6. [6]Gao J, Li X and Lin D 2020 AIAA Journal 58 2517–2532
  7. [7]Ueno Y and Ochi A 2019 Airframe noise prediction using Navier-Stokes code with cartesian and boundary-fitted layer meshes 25th AIAA/CEAS Aeroacoustics ConferenceDOI
  8. [8]Shur M, Strelets M, Spalart P and Travin A 2023 Journal of Turbulence 24 554–576
  9. [9]Montalà R, Lehmkuhl O and Rodriguez I 2024 Physics of Fluids 36 025125 ISSN 1070-6631
  10. [10]Montalà R, Lehmkuhl O and Rodriguez I 2025 Flow, Turbulence and Combustion 115 51–77
  11. [11]You D and Moin P 2008 Journal of Fluids and Structures 24 1349–1357 ISSN 0889-9746 unsteady Separated Flows and their Control
  12. [12]Rodriguez I, Lehmkuhl O and Borrell R 2020 Flow, Turbulence and Combustion 105 607–626
  13. [13]Melton L, Yao C and Seifert A 2006 AIAA Journal 44 012017
  14. [14]Shmilovich A and Yadlin Y 2009 Journal of Aircraft 46 1354–1364
  15. [15]Lehmkuhl O, Lozano-Durán A and Rodriguez I 2020 Journal of Physics: Conference Series 1522 012017
  16. [16]Rabault J, Kuchta M, Jensen A, Réglade U and Cerardi N 2019 J. Fluid Mech. 865 281–302
  17. [17]Rabault J and Kuhnle A 2019 Phys. Fluids 31 094105 ISSN 1070-6631
  18. [18]Suárez P, Alcántara-Ávila F, Rabault J, Miró A, Font B, Lehmkuhl O and Vinuesa R 2025 Commun. Eng. 4
  19. [19]Suárez P, Alcántara-Ávila F, Rabault J, Miró A, Font B, Lehmkuhl O and Vinuesa R 2025 Flow Turbul. Combust. 115 3–27
  20. [20]Guastoni L, Rabault J, Schlatter P, Azizpour H and Vinuesa R 2023 Eur. Phys. J. E 46 27
  21. [21]Vasanth J, Rabault J, Alcántara-Ávila F, Mortensen M and Vinuesa R 2024 Flow Turbul. Combust.
  22. [22]Wang Y, Mei Y, Aubry N, Chen Z, Wu P and Wu W 2022 Phys. Fluids 34 033606
  23. [23]Garcia X, Miró A, Suárez P, Alcántara-Ávila F, Rabault J, Font B, Lehmkuhl O and Vinuesa R 2025 Int. J. Heat Fluid Flow 116 109913 ISSN 0142-727X
  24. [24]Font B, Alcántara-Ávila F, Rabault J, Vinuesa R and Lehmkuhl O 2025 Nat. Commun. 16 1422
  25. [25]Montalà R, Font B, Suárez P, Rabault J, Lehmkuhl O, Vinuesa R and Rodriguez I 2025 Deep reinforcement learning for active flow control around a three-dimensional flow-separated wing at Re = 1,000 https://doi.org/10.48550/arXiv.2509.10195
  26. [26]Montalà R, Font B, Suárez P, Rabault J, Lehmkuhl O, Vinuesa R and Rodriguez I 2025 Discovering flow separation control strategies in 3D wings via deep reinforcement learning https://doi.org/10.48550/arXiv.2509.10185
  27. [27]Morita Y, Rezaeiravesh S, Tabatabaei N, Vinuesa R, Fukagata K and Schlatter P 2022 Journal of Computational Physics 449 110788 ISSN 0021-9991
  28. [28]Mahfoze O A, Moody A, Wynn A, Whalley R D and Laizet S 2019 Phys. Rev. Fluids 4(9) 094601
  29. [29]Han B Z, Huang W X and Xu C X 2023 Physics of Fluids 35 115144 ISSN 1070-6631
  30. [30]Li Y, Noack B R, Wang T, Cornejo Maceda G Y, Pickering E, Shaqarin T and Tyliszczak A 2024 Journal of Fluid Mechanics 991 A5
  31. [31]Vreman A W 2004 Physics of Fluids 16 3670–3681
  32. [32]Gasparino L, Spiga F and Lehmkuhl O 2024 Computer Physics Communications 297 109067 ISSN 0010-4655
  33. [33]Guadarrama S, Korattikara A, Ramirez O, Castro P, Holly E, Fishman S, Wang K, Gonina E, Wu N, Kokiopoulou E, Sbaiz L, Smith J, Bartók G, Berent J, Harris C, Vanhoucke V and Brevdo E 2018 TF-Agents: A library for reinforcement learning in TensorFlow URL https://github.com/tensorflow/agents
  34. [34]Partee S, Ellis M, Rigazzi A, Shao A, Bachman S, Marques G and Robbins B 2022 J. Comput. Sci. 62 101707 ISSN 1877-7503
  35. [35]Belus V, Rabault J, Viquerat J, Che Z, Hachem E and Reglade U 2019 AIP Advances 9 125014 ISSN 2158-3226
  36. [36]Schulman J, Wolski F, Dhariwal P, Radford A and Klimov O 2017 Proximal policy optimization algorithms https://doi.org/10.48550/arXiv.1707.06347
  37. [37]Piomelli U and Chasnov J R 1996 Large-Eddy Simulations: Theory and Applications (Dordrecht: Springer Netherlands) pp 269–336 ISBN 978-94-015-8666-5DOI

Paper details

vlnc:W_zsh5886d
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents