WorksPhysics and Astronomy
-optimal riblet profiles for turbulence suppression in channel flows
Abstract
The drag-reducing properties of streamwise-elongated surface corrugations, commonly referred to as riblets, have been extensively studied through experiments, simulations, and reduced-order modeling. Notably, the linearized Navier-Stokes framework introduced by Ran, Zare, and Jovanovic (J. Fluid Mech., vol. 906, 2021) employs volume penalization to model riblet-induced boundary conditions and predict drag reduction. These approaches typically quantify drag reduction as a function of a single geometric parameter, such as spanwise spacing or groove cross-sectional area, evaluated over fixed riblet shapes. As a result, identifying optimal riblet geometries across a broader design space often necessitates exhaustive parametric studies. To address this problem, we propose an optimization framework for the systematic design of optimal riblet shapes. Accordingly, we reformulate the volume penalization term in the linearized dynamics as a structured state-feedback operator. Targeting maximal suppression of turbulence, we pose a composite optimization problem that minimizes the steady-state variance of velocity fluctuations and riblet-induced dynamical augmentations while maintaining a constant bulk-flux. For a fixed spanwise wavelength (riblet spacing), this framework yields optimal wall-normal profiles that achieve maximal turbulence attenuation.
Introduction
The influence of engineered surface roughness on skin-friction drag has been studied for decades. Early work in this area was motivated by the distinctive texture of shark skin and its apparent role in facilitating efficient locomotion through water [1, 2, 3]. These observations prompted further investigation into the physical mechanisms underlying this biological phenomenon [4, 5] and, subsequently, the development of biomimetic surfaces designed to reproduce its drag-reducing characteristics.
Among such surfaces, riblets—streamwise-aligned surface corrugations used for passive flow control—have received considerable attention. Their capacity to reduce turbulent skin-friction drag has been demonstrated analytically [6], numerically [7, 8, 9], and experimentally [10, 11]. Nevertheless, identifying optimal riblet geometries and spacings remains challenging. A principal difficulty lies in accurately modeling turbulent flow over rough surfaces, while the resulting design problems are often nonconvex and computationally demanding. Consequently, advances in riblet design have historically relied heavily on improved physical understanding of the mechanisms responsible for drag reduction [12, 13], rather than on systematic optimization alone.
Most studies of riblet design focus on analyzing the flow physics associated with prescribed geometries rather than systematically optimizing the riblet shape [7, 14, 15]. Several studies have nevertheless addressed the design of riblets for drag reduction. These approaches often rely on high-fidelity simulations, such as direct numerical or large-eddy simulations, which provide detailed flow information but impose substantial computational costs. For example, Lee et al. [16] combined Bayesian optimization with large-eddy simulations to develop a design-by-morphing framework. Friedmann [17], by contrast, formulated an objective functional based on wall shear stress and used it to determine the optimal periodicity of sinusoidal surface roughness.
In this study, we develop a riblet shape-optimization framework that accommodates a broad class of geometries rather than restricting the design space to a single prescribed shape of riblets, e.g., triangular, blade, or scalloped riblets. Unlike approaches that rely exclusively on computationally intensive nonlinear flow simulations, our framework employs low-complexity models obtained by linearizing the flow dynamics about the long-time-averaged velocity field over riblets. This formulation enables systematic exploration of the riblet design space while retaining the influence of surface geometry on the underlying turbulent dynamics.
Linearized Navier-Stokes (NS) equations have been widely used to characterize the structural and statistical properties of turbulent wall-bounded shear flows [18, 19, 20, 21, 22, 23]. In this class of models, the governing equations are cast in evolution form and driven by deterministic or stochastic forcing to excite physically relevant flow responses. Such models have also enabled physical analysis and discovery [24, 25, 26], in addition to the development of flow estimation [27] and control strategies, including sensor-free techniques for turbulence mitigation based on surface-normal blowing and suction in the form of streamwise traveling waves [28, 29], transverse wall oscillations [30], and riblets [15, 31].
Here, we adopt the modeling framework developed in [15] to represent both the rough-wall boundary conditions and the statistical signatures of turbulent channel flow over riblets. Following Ref. [32], this framework employs a volume-penalization technique [33] to approximate the influence of a spatially periodic surface on the flow. Rather than deforming the computational mesh to conform to the rough geometry, volume penalization introduces a spatially varying, static feedback term into the momentum equations that suppresses the flow within the solid region and restricts the dynamics to the fluid domain surrounding the riblet surface. The framework further incorporates the turbulence-modeling approach of Moarref and Jovanović [30] to account for the influence of background turbulence on the mean velocity and, consequently, on skin-friction drag. Specifically, the turbulent eddy viscosity for channel flow over riblets is inferred from the second-order statistics of the eddy-viscosity-enhanced linearized NS equations. The resulting corrections to the turbulent viscosity are then used to determine the riblet-induced modifications to the mean flow and drag.
Using this linear low-complexity model, we formulate an unconventional optimal-control problem in which the state variance quantifies the energy of velocity fluctuations as a surrogate measure of drag [15], while the control effort quantifies the magnitude of the dynamical augmentation induced by the riblets. The optimization is subject to both structural and physical constraints. The structural constraints prescribe the sparsity pattern of the feedback gain matrix required to represent harmonic interactions among the mean flow, the volume-penalization field, and the velocity fluctuations. The physical constraints enforce geometric connectivity, consistency of the dynamical penalization, and preservation of a constant bulk flux when quantifying the influence of riblets on drag.
In unconstrained state-feedback design, an optimal stabilizing feedback gain can often be obtained from the solution of an algebraic Riccati equation. Imposing structural constraints on the gain matrix substantially complicates the problem and has motivated customized solution methods, including projected-gradient algorithms [34]. More general formulations that impose additional restrictions directly on the feedback gain, such as linear equality constraints, remain an active area of research [35, 36].
We address the physical constraints in our formulation by parameterizing admissible solutions using a library of riblet geometries represented by signed-distance functions. These are level-set functions that encode the distance to a parameterized boundary curve; see, e.g., [37]. We show that linear combinations of these functions, embedded within the structured feedback matrix that augments the linearized dynamics, generate morphed riblet geometries that satisfy the prescribed shape constraints. This parameterization also yields a reformulated optimization problem that is amenable to gradient-based methods.
To solve the resulting problem, we develop an augmented-Lagrangian-based algorithm. Because the formulation is nonconvex, optimization trajectories initialized at nearby points may converge to distinct local minima with different fluctuation-energy and drag-reduction characteristics. We therefore employ an ensemble of randomly generated initial conditions selected to ensure that the objective and its gradients are well defined. For a prescribed riblet spacing and under constant-bulk-flux conditions, the optimized solutions reveal a recurring preference for scallop-shaped riblet geometries that effectively suppress turbulent fluctuations.
The main contributions of this paper are as follows:
Formulation of riblet shape optimization as a state-feedback synthesis problem subject to application-specific structural and physical constraints, in which
riblet geometry is represented through structured dynamical augmentations of the linearized NS equations.
Development of a customized augmented-Lagrangian algorithm for solving the resulting constrained optimization problem.
Demonstration that a level-set representation of riblet geometries based on signed-distance functions enforces physical constraints associated with geometric connectivity and consistent volume penalization.
The remainder of the paper is organized as follows. In § 2, we formulate the riblet shape-optimization problem as a constrained state-feedback synthesis problem based on the linearized NS model and volume-penalization framework, and summarize the associated structural and physical constraints. In § 4, we introduce the signed-distance function parameterization and reformulate the optimization problem in terms of parameterization weights. In Section IV, we develop a customized method of multipliers (MM) algorithm for solving the reformulated problem. In § 5, we construct the library of riblet geometries used in the morphing procedure and analyze solutions obtained from both deterministic and randomized initializations. Finally, Section VI summarizes the main findings and outlines future directions.
Problem Formulation
The dynamics of a pressure-driven turbulent channel flow over a riblet-mounted surface (Figure 1(a)) are given by the NS equations,
Figure 1. (a) Channel flow with riblets on the lower wall together with the mean velocity in blue. (b) Semi-circular riblets of spanwise frequency , height , groove cross-sectional area , and spacing . The parameter is the proportion of the riblet height below .
subject to no-slip/no-penetration boundary conditions,
that capture the effect of the streamwise-constant corrugation of the lower wall through the shape function . Here, is the three-dimensional velocity vector, is pressure, is the gradient operator, is the Laplacian, is time, and the Reynolds number results from non-dimensionalization of length by the channel half-height , velocity by the friction velocity , time by , and pressure by . Here, is the wall shear stress (averaged over horizontal directions and time), is the fluid density, and is the kinematic viscosity.
The effect of engineered surface corrugations (e.g., riblets) is typically quantified via their impact on skin-friction drag, which is governed by the near-wall gradient of the turbulent mean velocity, and by the energy contained in turbulent fluctuations. Here, we focus on the latter and optimize the surface geometry to reduce the energy of the turbulent velocity fluctuations. The dynamics of zero-mean velocity and pressure fluctuations, and , about the turbulent mean , which completes the decomposition of the flow field in Eq. (1a), i.e., and , are described by the linearized NS equations,
In these equations, we follow [15] in accounting for riblets not by enforcing boundary conditions through mesh deformation, but by penalizing the fluctuation dynamics using a permeability field that conforms to the lower-wall geometry. This field damps fluctuations within the riblet grooves: inside the riblet structures, which forces , while in the fluid region, recovering the unmodified equations. Due to this, we can now impose standard homogeneous Dirichlet and Neumann boundary conditions on at both the upper and lower walls. In Eq. (3), is a zero-mean stochastic forcing that is added to the momentum equations to excite a stochastic response from the linearized dynamics. We model this source of excitation as a white-in-time process with covariance , i.e., , where denotes the Dirac delta distribution and the tensor (outer) product. Finally, the mean velocity is itself obtained as the solution to the steady-state equations,
which result from application of the turbulent viscosity hypothesis [38] to the Reynolds-averaged NS equations. Here, is the turbulent eddy viscosity normalized by molecular viscosity; see [15] for additional details. For the small-size riblets considered in this paper, we consider the flow in the grooved region of the riblet structures (Figure 1(b)), which is dominated by viscosity, to be laminar, i.e., for in both Eqs. (3) and (4).
For streamwise-constant spanwise-periodic surface corrugation, can be expanded using the Fourier transform,
Here, is the spatial frequency of the surface corrugation and the Fourier coefficients (functions of ) specify their shape and height in the vertical direction. Based on this, the solution to Eqs. (4) adopts a similar harmonic form, i.e.,
For small riblets, the solution to the nonlinear equations (4) obtained via Newton’s method has a single streamwise component, , which significantly simplifies the equations by bringing them into a linear form.
The solution to Eqs. (3) shares the harmonic structure of and , while the operators in Eqs. (3) are invariant in , they are spatially periodic in . By adopting Bloch wave theory [39], [40] to parameterize the spatially periodic dynamics in the Fourier domain and applying a standard pressure-elimination conversion [41], the PDEs (3) can be recast in evolution form as
where the state consists of the wall-normal velocity and vorticity . Here, , and are bi-infinite column vectors parameterized by the wavenumber pair , with and denoting the streamwise wavenumber and spanwise wavenumber offset, e.g., for every , with for any integer . As a result of this parameterization, the spanwise wavenumber is given by [15]. At both the top and the bottom boundaries, homogeneous Dirichlet and Neumann conditions are imposed on , and homogeneous Dirichlet conditions are imposed on . The adopted volume penalization method should automatically enforce immersed boundary conditions on the non-smooth lower wall without the need for additional boundary conditions. However, to avoid numerical oscillations in practice, the computational domain is divided into two parts using block operators [42] with smoothness enforced at the interface; see Eqs. (9).
In Eq. (7), the system matrices, and , are bi-infinite with operator-valued elements in [15 Appendix A]. We employ a change of variables to obtain a state-space representation in which the kinetic energy is determined by the Euclidean norm of the state vector in a finite-dimensional approximation of the evolution model [20 Appendix A]. A pseudospectral scheme is used for discretizing the wall-normal operators in both the mean flow equations (1a) and the linearized dynamics (7) [43]. This yields the finite-dimensional approximation,
where and are now vectors whose dimensions are determined by the adopted wall-normal discretization; if and collocation points are used to discretize the lower and upper wall-normal regimes, respectively, and harmonics are used in the Fourier Series expansion (4), is a vector of complex-values entries. Moreover, matrices and , with , are discretized versions of the operators in Eq. (7) after incorporating the aforementioned change of coordinates. The state matrix can be decomposed into a nominal component associated with the dynamics of a channel flow with smooth walls () and a component that represents the riblet-induced deviation from the smooth-wall channel flow dynamics (), i.e., . Here, is a
structured matrix in which the Fourier coefficients of and appear as diagonals of submatrices; see Appendix A for the description of .
The permeability function , whose inverse appears as a coefficient both directly in the fluctuation dynamics (2) and indirectly through the mean velocity (cf. Eq. (4)), can be tailored to capture the structure of arbitrarily shaped riblets defined by roughness functions [15]. On the other hand, the appearance of as a multiplicative gain in the dynamics of velocity fluctuations (i.e., term in Eqs. (3)) highlights an opportunity to design optimally rough surfaces for suppressing energy via optimal control theory.
Constrained optimal control problem
We examine the wall-normal structure of for a streamwise-constant, spanwise-periodic roughness pattern of prescribed height and spanwise frequency . Our goal is to determine the feedback gain matrix that minimizes the quadratic cost functional
where is the state autocorrelation matrix, and specify penalties on the state and riblet-induced dynamical modification, respectively, denotes the set of structured stabilizing feedback matrices that satisfy all physical requirements levied on the system (§ 2-B), and is the complex-conjugate transpose. Both the stability criterion and consistency with an admissible state autocorrelation reduce to an algebraic constraint on and , namely
In a standard unstructured state-feedback optimal control problem, the optimal gain matrix can be obtained by solving a single algebraic Riccati equation (equivalently, the LQR ARE); see, e.g., [44 Sec. 9.2, Eqs. (9.12)–(9.13)]. However, as described in the following subsection, the structural constraints imposed on our problem render it nonconvex, precluding solution by standard Riccati-based methods.
Optimization constraints
There are a variety of physical constraints that comprise the feasibility set . We classify these constraints into three categories: Harmonic-consistency constraints; constant bulk and equal spacing; consistent penalization; and structural connectivity.
1) Harmonic-consistency constraints: As noted above, the Fourier coefficients and appear on the diagonal of the submatrices that comprise . Figure 2 illustrates the sparsity pattern of the feedback matrix if the harmonic expansions (4) and (5) are truncated at 3 harmonics. Each submatrix encodes the dynamical coupling between (i) harmonics of the state fluctuations , and (ii) harmonics of the penalization term and the base flow , which are coefficients in Eqs. (3); see [15] for details. Consequently, any optimization-based construction of must be posed at the level of the Fourier profiles of the penalization function . This is necessary not only to preserve the intended harmonic interactions, but also to enforce consistency with the base-flow harmonics that satisfy

Figure 2. Sparsity pattern of entries in the feedback gain matrix in the riblet dynamical augmentation. Gray lines separate submatrices corresponding to various harmonic interactions when and .
Here, is a bi-infinite vector of the streamwise base-flow harmonics. Since these harmonics appear in the structured feedback matrix , a consistency constraint must be enforced to ensure that Eq. (12) holds whilst adhering to the structure shown in Figure 2.
2) Constant bulk and equal spacing: To obtain a tractable design process with reduced degrees of freedom, we assume that the riblet spacing is fixed, i.e., is fixed. Nevertheless, an additional constraint is necessary to prevent the optimization from reducing kinetic energy simply by decreasing the mass flux, which would inadvertently increase surface drag. To this end, we enforce the constant bulk-flow constraint,
where is the bulk flux in a turbulent channel flow with smooth walls. Here, denotes the riblet height, while represents the ratio of the riblet height to the thickness of the laminar subregion; see Figure 1(b).
3) Consistent penalization: A property of the permeability function that is not inherently encoded in the Fourier coefficients is binary masking. Specifically, this property requires to assume only one of two values at any location in the – plane, i.e.,
Therefore, throughout the optimization process, the Fourier profiles must belong to a class of functions that preserve this property.
If this penalization constraint is satisfied, the riblet structure will be physically realizable as it will have a well-defined edge. (14) describes an instantaneous jump at the edge of the riblet. In practice, smooth transitions are necessary to maintain physical consistency with wall-bounded turbulent flows. If there is too sharp of a transition, the feedback configuration may become unstable. On the other hand, in the case that the edge of the riblet is too smoothed out, spurious velocity fluctuation values can appear within the riblet structures. This limitation of immersed boundary methods can be seen by describing the gradient at the edge of the roughness structure using as a hyperbolic tangent function [15 Eq. (2.9)].
4) Structural connectivity: Finally, we must maintain vertical support (connectivity) to the wall: every point within the riblet must have a fully enclosed path back to the wall. Consider a finely discretized domain, refined to grid convergence. Because each cell in the discretization is permitted to take values of either 0 or , isolated (disconnected) solid cells can arise. The left panel of Figure 3(a) illustrates a hollow yet fully connected riblet: every point within the solid phase admits a continuous, enclosed path to the wall. By contrast, the right panel shows a disconnected configuration, which is not manufacturable. The admissibility of local cell patterns depends on proximity to riblet edges, implying that connectivity cannot be enforced purely through local rules; it is inherently a global constraint.
Figure 3. Sparsity pattern of resistance function for (a) a fully connected manufacturable design; and (b) a vertically disconnected design.
To impose this requirement, prior work in additive manufacturing has introduced pointwise topological constraints. One common approach injects support structures for unsupported regions: when a “floating island” appears, the constraint violation triggers the creation of a support beam [45]. In our setting, however, enforcing such a pointwise condition in Fourier space would introduce a large number of nonlinear constraints (one per grid point) and may also generate extraneous structures that are irrelevant outside the additive-manufacturing context. As an alternative, Li et al. [46] proposed a virtual-temperature formulation that guarantees connectivity by solving an auxiliary Laplace-type problem at each iteration. While it is possible to impose this virtual temperature constraint within an optimization loop, doing so introduces numerous embedded constraints; translating each of these into a corresponding constraint on the feedback matrix can be cumbersome.
We instead adopt parameter-gradient methods applied to the cost, which have primarily been employed outside of structured control problems [47, 48]. Within this framework, we pursue a shape-optimization approach that preserves wall connectivity by construction. In particular, the design-by-morphing method introduced in the next section uses a change of variables that maps the baseline optimization field to a lower-dimensional set of morphing parameters . Whether enforcing connectivity through a topologically constrained, grid-based formulation or via shape optimization—the central modeling decision is identical: how the riblet boundary (edge) is defined and parameterized.
Approach: signed-distance function tuning
We enforce uniform penalization (§ 2-B3) and compactness (§ 2-B4) by employing a variable transformation that brings the problem into an equivalent shape optimization form. This approach produces a morphed riblet geometry using a linear combination of signed-distance functions. To solve the resulting optimization problem, we develop a customized algorithm, which uses a parameter-gradient method to minimize the norm of the feedback synthesis.
Surface representation using signed-distance functions
Design-by-morphing has recently been combined with large-eddy simulations and Bayesian optimization to identify riblet geometries that are optimal for drag reduction [16]. The simulation-free method we develop in this section adapts existing tools in design by morphing to optimal control under physical constraints. In this method, a morphed riblet shape is achieved by parameterizing the riblet geometry over a library of curves (Figure 4). In order to find an optimal morphed shape from this library, one might first consider applying Here, are tuning weights, is a sigmoid function with controlling the transition rate between fluid and solid interfaces, and denotes the discrete Fourier transform in the spanwise direction. If all weights sum to one, the convex combination yields riblet shapes obtained by interpolating among curves in the library. Note that the wall-normal discretization of the signed-distance functions is chosen to match that of the linearized flow model.

Figure 4. Visualization of the library of nine riblet shapes over which the optimization is performed.
Problem reformulation
The expression of optimization variables provided by the weighted sum of signed-distance functions Eq. (17) transforms the constrained optimal control problem described in Sec. II into the following form
where the optimization variable denotes the vector of tuning weights , denotes riblet-induced changes to the streamwise component of the base flow , is the vector of ones of the same length as and the dependence of the quadratic cost function on weights follows from its dependence on (Eq. (10)) and Eq. (17); see Appendix A for details. In this formulation, all constraints highlighted in Sec. II-B are implicitly satisfied but the constant bulk constraint (Sec. II-B2). The linear equality constraint on ensures a convex combination and the constant bulk flux constraint is nonlinear. The problem is nonconvex due to the nonconvexity of the sigmoid function (Eq. (17)) through which the weights are passed onto the Fourier coefficients in and nonlinearity of the constraints on the bulk flux. As a result, the optimization landscape does not necessarily have a global minimizer. This reflects the dependence of solutions on initial conditions; see Sec. V-D. We next develop a customized optimization algorithm to solve problem (15).
a standard linear combination to the library of resistance functions corresponding to these curves, i.e.,
However, using this method to combine the basis shapes violates the consistent penalty constraint. This is because the intersection of any two resistance functions would contain twice the penalty. To resolve this issue, we use signed-distance functions, which were first developed in the computer graphics community for morphing binary masks through linear combinations [37].
Any parameterized curve can be converted into a two-dimensional signed-distance function that describes its boundary. For any point , the signed-distance function gives the distance to the nearest point on a shape’s boundary: it is negative inside the shape and positive outside. Figure 5(a) and Figure 5(b) illustrates this construction for a representative curve. Using this representation, the Fourier coefficients in Eq. (5) can be computed from the Fourier transform of a sigmoid applied to a linear combination of signed-distance functions , i.e.,

Figure 5. Illustration of (a) the 2D shape function of a riblet structure; and (b) the corresponding signed-distance function.
Optimization algorithm
We employ the MM algorithm [49 Algorithm 17.4] to solve the constrained optimization algorithm outlined in the previous section. This algorithm is a dual-ascent optimization scheme which has wide application in solving constrained nonlinear programming problems [49, 50, 51, 52]. The MM algorithm for solving problem (15) is given by
where is the augmented Lagrangian associated with linear constraints at the th iteration, i.e.,
Here, () are Lagrange multipliers, () are positive scalars, are entries of the vector of equality constraints,
The algorithm terminates when the primal and dual residuals are small enough.
Solution to the -minimization problem
For fixed , minimizing the augmented Lagrangian with respect to amounts to finding the minimizer of subject to the box constraint . Since is smooth, a projected gradient method can be used to find a minimizer of Eq. (20) iteratively:
where denotes the projection of vector entries onto the feasible domain, is the inner gradient descent iteration counter, and is the step size. The expression for the gradient of is provided in Appendix B.
At each iteration of the projected gradient method, we determine the step size via an adaptive Barzilai–Borwein (BB) initial step size selection [53], i.e.,
followed by backtracking to ensure sufficient descent, i.e.,
In Eq. (21), and are the “steepest descent” and “minimum residual” step sizes, which are given by,
When or , we use the following expression developed by Dai et. al. [54]:
For additional details, see [53 Sec. 4.1]. Finally, we use a combined stopping criterion that terminates the projected gradient iterations when either the relative residual
or the normalized residual
meet a desired tolerance. In the equations above, and are small positive constants and the residual is defined as
where . Note that the denominator of nearly vanishes when the gradient is small. In such cases, the first order optimality condition () and provide an appropriate stopping criteria; see [55 Section 4.6] for additional information.
Lagrange multiplier and penalty parameter updates
Customized MM for optimal riblet design is summarized in Algorithm 1. The Lagrange multipliers and penalty parameters are updated based on a slight variation of the procedure proposed in [49 Algorithm 17.4], which incorporates rules for the independent adjustment of the augmented Lagrangian penalty parameters [56, 57]. This procedure permits inexact solutions to subproblem (17a) and refined updates of the Lagrange multipliers and penalty parameters based on the adjustment of convergence tolerances on and .
Computational complexity
Each iteration of Algorithm 1 requires solving Lyapunov equations, performing matrix multiplications, evaluating traces, and carrying out other standard linear-algebraic operations. Since solving a single Lyapunov equation has worst-case computational complexity and dominates the remaining operations, the overall computational complexity is , where denotes the number of outer-loop iterations and denotes the average number of inner-loop iterations required for convergence.
Results
In this section, we apply Algorithm 1 to determine the optimal shape of riblets with spanwise frequency on the lower wall of a turbulent channel flow with . The optimization algorithm is initialized from a selection of weights that satisfy and ensure stability of the augmented dynamics (). We exploit the spatial parameterization of our model by first conducting optimization based on the linearized dynamics of individual turbulent length scales, i.e., a single wavenumber pair . We then compare results to those obtained based on multi-scale dynamics. Finally, given the nonconvexity of problem (16), we conduct a random seed experiment to analyze the dependence of solutions on initial conditions. The optimization algorithm was developed in MATLAB and was executed on a 4.50 GHz AMD Ryzen 7 7700X 8-Core Processor with 32 GB of RAM.
Riblet library synthesis
Our shape-optimization framework relies on constructing a library of comparable riblet profiles that satisfy widely accepted geometric criteria for optimal drag reduction while
| Riblet Type | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
| 10.8 | 9.3 | 11.1 | 10.7 | 11.2 | 10.8 | 9.8 | 9.7 | 10.2 | |
| 0.057 | 0.056 | 0.051 | 0.054 | 0.055 | 0.057 | 0.056 | 0.051 | 0.053 |
Table 1. AND FOR EACH LIBRARY MEMBER
Algorithm 1 Customized MM algorithm
input: , , , , , and positive constants , , , ,
initialize:
for do
while or or do
Calculate using method in Appendix B
Initialize with Barzilai-Borwein scheme
while fails sufficient descent do
end while
compute and ,
end while
if then
if and then
stop with approximate solution
end if
, ,
,
else
if then
,
,
end if
end if
end for
maintaining a constant bulk flux. Specifically, we constrain the viscous-scaled groove size and the ratio of riblet height to channel half-height according to
where and denotes the viscous-scaled groove area. The prescribed range of encompasses optimally drag-reducing riblets of various shapes [8], [15]. The height constraint limits interference with the buffer-layer viscous cycle and mitigates blockage effects [58], [59]. Finally, to reduce the influence of secondary motions and spanwise dispersion, and thereby support the validity of linearization about , we restrict attention to riblets with mid-height spacing , as suggested by [14]. Figure 4 presents the library of riblet profiles with spacing considered in the optimization, with the corresponding geometric specifications summarized in Table I.
Optimization setup and performance
Following the approach described in § 2, we partition the wall-normal domain into turbulent and laminar subregions located above and below , respectively. We discretize these subregions using collocation points above and collocation points below it. Consistent with the methodology of [15] for an optimal triangular riblet with a tip angle, we set the height of the laminar subregion to and the magnitude of the resistance field in Eq. (14)) to .
We retain the minimum number of harmonics, , in the Fourier expansion of in Eq. (5) required to accurately capture the energy of flow fluctuations over the riblets. Figure 6 shows the relative error introduced by harmonic truncation across several turbulent length scales. Based on these results, we use for the riblet-optimization computations presented in the remainder of this section and employ a higher-fidelity model with to evaluate the flow physics associated with the optimized riblets.
Figure 6. Convergence of energy by number of harmonics for riblets (−) 1, (−) 2, (−) 3, (−) 4, (−) 5, (−) 6, (−) 7, (−) 8, and (−) 9 as identified in Figure 4 for (a) , (b) (c) . rel. err. (%) := where is the energy at any number of harmonics and corresponds to the energy with 25 harmonics.
Figure 7 shows the convergence histories of the customized MM algorithm introduced in Sec. IV when applied to problem (16) at . This wavenumber pair corresponds to the characteristic length scales of streamwise-elongated streaks that dominate the near-wall cycle. The MM algorithm was initialized from a riblet profile resulting from an equal distribution of weights among library members (Figure 9(a)). The primal and dual stopping tolerances are set to and , respectively. To assess the asymptotic convergence behavior, we estimate the rate from

Figure 7. Convergence curves of Algorithm 1 (—) versus (a) iteration number; and (b) cumulative solve time for a single length scale optimization with . Here, denotes the final iterate.
Figure 9. Resistance function for corresponding to the initial weight distribution (left column) and final solution of problem (16) (right column). (a) single-scale initial condition; (b) single-scale solution; (c) multi-scale initial condition; (d) multi-scale solution; (e) the initial condition for the best result of the random seed optimization; (f) the best result of the random seed optimization. Solutions were obtained with the optimal tolerances and described in subsection V-C.
for sufficiently large . Here is the value obtained at final iteration. The resulting values do not satisfy for sufficiently large , indicating sub-linear convergence.
Multi-scale optimization
Figure 8 compares the weight distributions obtained from shape optimization using single-scale and multi-scale cost functions. As evident from this figure and Figure 9, although the single-scale optimization targets the dynamically dominant structures associated with the near-wall cycle, accounting for the energy costs of two additional dynamically significant modes leads to a different distribution of weights and, consequently, to different optimal riblet profiles. Specifically, the multi-scale formulation includes the energy cost associated with an energetically amplified mode, , and the most strongly suppressed mode, , resulting from the introduction of triangular riblets with a tip angle. Although the resulting geometric differences are relatively small, they produce noticeable changes in the achievable energy suppression and drag reduction; see Table II. The skin-friction drag associated with the ribleted surfaces is computed using the turbulence-modeling framework of Ran et al. [15], in which riblet-induced modifications to the mean flow are obtained from the corrections to the turbulent eddy viscosity in Eqs. (4).

Figure 8. Optimal weight magnitudes resulting from single-scale and multi-scale optimization, in addition to the best result of the random seed optimization. The initial weights for each optimization has also been included.
induced modifications to the mean flow are obtained from the corrections to the turbulent eddy viscosity in Eqs. (4).
The energy suppression and drag reduction values reported in Table II are computed using a higher-fidelity model of the flow over the optimized riblets, with and spectral ranges and . These results demonstrate the importance of accounting for a broader range of dynamically relevant scales. In particular, optimizing the riblet shape for a single dominant scale does not necessarily yield improved overall performance. Finally, while the resulting performance is comparable to that reported for high-performance riblets in the existing literature, the riblet profiles identified by our optimization framework provide a substantial improvement.
Random-seed experiments
Due to the nonconvexity of problem (16), solutions obtained using the MM algorithm can depend on the initial condition. To elucidate this dependence, we conduct a random-seed experiment in which multi-scale optimization is conducted from 100 different initial weight distributions. Figure 10 shows the final distribution of weights among the 9-member library in addition to the final value of the objective function resulting from each initial seed. We observe a variety of local optima. However, we observe a clustering of results around indices 3 and 8, which correspond to blade- and scallop-shaped members of our library, respectively.
Figure 10. Randomized initialization results. (a) Distribution of optimal weights vs seed number; and (b) the cost functional by seed index number.
The best result obtained from this experiment is compared against the single-scale and multi-scale computations presented in the previous subsection, which were initialized from an initial condition that results from an equal distribution of weights. As seen in Table II, the riblet shape resulting from the best solution of the random-seed experiment corresponds to a scallop-shaped riblet and outperforms both single-scale and multi-scale minimizers generated from deterministic initial conditions. Critically, the of all three shapes falls within the previously established optimal range of [8, 15]. Importantly, small variations in the energy of fluctuations results in meaningful improvements in drag reduction when post-processing calculations across the full spectrum of horizontal scales.
| (%) | (%) | |||
| Triangular [15] | ||||
| Single-scale | ||||
| Multi-scale | ||||
| Randomized |
Table 2. Comparison of energy suppression and drag reduction resulting from the optimal riblet shapes obtained from single-scale, multi-scale, and random seed optimizations.
Concluding Remarks
In this study, we developed a model-based framework for optimizing riblet geometries by examining their influence on the energy of flow fluctuations. The framework formulates riblet design as a constrained optimal control problem, enabled by the evolution form of the linearized Navier–Stokes equations and the use of volume penalization to represent the effects of riblets on the flow dynamics. Signed-distance functions were employed to parameterize the riblet geometries, allowing the original feedback-gain optimization problem to be reformulated as a weight-selection problem. We then developed a customized algorithm to solve the resulting optimization problem. Although the random-seed optimization results suggest that scallop-shaped riblets may be optimal within the considered design space, we also showed that appropriately weighted combinations of blade- and scallop-shaped riblets constitute competitive alternatives.
The present approach uses the energy of flow fluctuations as a surrogate measure for drag reduction. Further analysis is therefore needed to determine whether the identified riblet geometries are also optimal with respect to skin-friction drag. To reduce the computational cost, the optimization was restricted to a prescribed riblet spacing, and the geometry library was limited to nine commonly studied riblet shapes from the literature. Future work should expand this library to include broader variations in riblet height, width, and spacing, particularly geometries that lie outside the convex hull defined by the current parameterization.
The proposed low-complexity modeling framework relies on linearization about a base-flow profile containing only a streamwise velocity component. More sophisticated models that account for dispersive flow motions could provide more accurate predictions of riblet performance and enable the consideration of additional geometric families that are currently excluded by the modeling assumptions. Ultimately, a comprehensive evaluation of the optimal riblet geometries will require verification of the model-based predictions through high-fidelity numerical simulations.
in the decomposition of the state matrix
The structured bi-infinite matrix ,
represents riblet-induced changes to the dynamics of velocity fluctuations with operator representing the influence of the th harmonic of the state on the dynamics of the th harmonic . These operators themselves have the block structure
where
Here, , , , and .
Computation of the gradient
As shown in § 2, the objective function contains an integration over all scales . This can be discretized such that the objective function of problem (15) can be expressed as
where is the number of scales considered in approximating the cost function and the covariance corresponding to the th scale, , solves the Lyapunov equation
where , and are defined over the set of stabilizing augmentation matrices for , i.e., .
Because the objective gradients associated with individual scales can be computed independently, the total gradient can be evaluated in parallel across the scales. For notational simplicity, throughout the remainder of this appendix we omit the subscript , which identifies the particular -pair associated with each dynamical system.
As shown in [60 Lemma 3.1], the objective function is indeed differentiable over the set of stabilizing gain matrices. Building on [60 Theorem 3.2], the following theorem provides an expression for the derivative of the objective function with respect to weighting parameters .
Theorem 1: The derivative of the objective function with respect to the th entry of is given by
where solves the Lyapunov equation
Proof: The derivative of the objective function with respect to the weight is given by
Let
The function and its derivative are zero, i.e.,
which yields
This equation is obtained by substituting the partial derivatives of , which are denoted by the subscripts, i.e.,
and
into Eq. (27). Multiplying Eq. (28) by from the left and the dual of the Lyapunov equation (25), i.e.,
by from the left and taking the trace of both sides of each equation yields the identity
Comparing the right-hand-side of this identity with that of Eq. (26) yields
and completes the proof.
In addition to the matrices and , which can be obtained by solving two Lyapunov equations (23) and (27), respectively, the computation of the derivative of the cost function () requires the derivatives , which can be expressed analytically. The dependence of the augmentation operator on the weights can be decomposed into two contributions: one mediated by the streamwise mean (), and another arising from the direct dependence on the Fourier coefficients , which themselves depend on through Eq. (17). Specifically,
Since scalar differentiation commutes with linear and diagonal operators, we obtain
Thus, computing reduces to evaluating and .
To compute , we exploit the linearity of the Fourier transform in the expression for , which allows the derivative to be moved inside the transform. Applying the chain rule yields
To compute , we first express the mean , which solves Eq. (12), as
where . Here, and denote the dynamical contributions from the smooth wall and the riblets, respectively, and is the forcing vector associated with the nominal back pressure. Using the derivative of an inverse operator, together with the linearity of , gives the closed-form expression
Acknowledgments
The Office of Information Technology Cyberinfrastructure Research Computing (CIRC) at The University of Texas at Dallas and the Texas Advanced Computing Center (TACC) are acknowledged for providing computing resources. Modern LLM’s were utilized in accordance with the IEEE GenAI guidelines for brainstorming, reference identification, and grammatical editing for enhanced textual clarity.
References
- [1]W. E. Reif and A. Dinkelacker, “Hydrodynamics of the squamation in fast swimming sharks,” Neues Jahrbuch für Geologie und Paläontologie-Abhandlungen, pp. 184–187, 1982.DOI
- [2]D. Bechert and W. Reif, “On the drag reduction of the shark skin,” in 23rd Aerospace sciences meeting, 1985, p. 546.DOI
- [3]D. Bechert, M. Bartenwerfer, G. Hoppe, and W. E. Reif, “Drag reduction mechanisms derived from shark skin,” 15th Congr. of the Int. Counc. of the Aeronautical Sciences, vol. 2, pp. 1044–1068, 1986.
- [4]D. W. Bechert, M. Bruse, and W. Hage, “Experiments with three-dimensional riblets as an idealized model of shark skin,” Experiments in fluids, vol. 28, no. 5, pp. 403–412, 2000.DOI
- [5]Y. Fu, C. Yuan, and X. Bai, “Marine drag reduction of shark skin inspired riblet surfaces,” Biosurface Biotribology, vol. 3, no. 1, pp. 11–24, 2017.DOI
- [6]D. W. Bechert and M. Bartenwerfer, “The viscous flow on surfaces with longitudinal ribs,” J. Fluid Mech., vol. 206, pp. 105–129, 1989.DOI
- [7]R. García-Mayoral and J. Jiménez, “Drag reduction by riblets,” Phil. Trans. R. Soc. A, vol. 369, no. 1940, pp. 1412–1427, 2011.
- [8]R. García-Mayoral and J. Jiménez, “Hydrodynamic stability and breakdown of the viscous regime over riblets,” J. Fluid Mech., vol. 678, pp. 317–347, 2011.
- [9]O. El-Samni, H. H. Chun, and H. S. Yoon, “Drag reduction of turbulent flow over thin rectangular riblets,” Int. J. Eng. Sci., vol. 45, no. 2-8, pp. 436–454, 2007.
- [10]D. Bechert, M. Bruse, W. Hage, H. Van, J. Th, and G. Hoppe, “Experiments on drag-reducing surfaces and their optimization with an adjustable geometry,” J. Fluid Mech., pp. 59–87, 1997.DOI
- [11]M. Sasamori, H. Mamori, K. Iwamoto, and A. Murata, “Experimental study on drag-reduction effect due to sinusoidal riblets in turbulent channel flow,” Experiments in Fluids, vol. 55, no. 10, p. 1828, 2014.DOI
- [12]M. Walsh, “Turbulent boundary layer drag reduction using riblets,” in 20th aerospace sciences meeting, 1982, p. 169.DOI
- [13]M. J. Walsh, “Riblets as a viscous drag reduction technique,” AIAA J., vol. 21, no. 4, pp. 485–486, 1983.DOI
- [14]D. Modesti, S. Endrikat, N. Hutchins, and D. Chung, “Dispersive stresses in turbulent flow over riblets,” J. Fluid Mech., vol. 917, p. A55, 2021.DOI
- [15]W. Ran, A. Zare, and M. R. Jovanović, “Model-based design of riblets for turbulent drag reduction,” J. Fluid Mech., vol. 906, p. A7 (38 pages), January 2021.DOI
- [16]S. Lee, S. H. Moazam, D. D. Lim, G. X. Gu, and P. S. Marcus, “Bayesian-optimized riblet surface design for turbulent drag reduction via design-by-morphing with large eddy simulation,” J. of Mech. Des., vol. 146, no. 8, p. 081701, 2024.DOI
- [17]E. Friedmann, “The optimal shape of riblets in the viscous sublayer,” J. of Math. Fluid Mech., vol. 12, no. 2, pp. 243–265, 2010.
- [18]B. J. McKeon and A. S. Sharma, “A critical-layer framework for turbulent pipe flow,” J. Fluid Mech., vol. 658, pp. 336–382, 2010.DOI
- [19]Y. Hwang and C. Cossu, “Amplification of coherent streaks in the turbulent Couette flow: an input-output analysis at low Reynolds number,” J. Fluid Mech., vol. 643, pp. 333–348, 2010.DOI
- [20]A. Zare, M. R. Jovanovic, and T. T. Georgiou, “Colour of turbulence,” J. Fluid Mech., vol. 812, pp. 636–680, 2017.
- [21]A. Zare, T. T. Georgiou, and M. R. Jovanovic, “Stochastic dynamical modeling of turbulent flows,” Annu. Rev. Control Robot. Auton. Syst., vol. 3, pp. 195–219, 2020.DOI
- [22]M. R. Jovanović, “From bypass transition to flow control and data-driven turbulence modeling: An input-output viewpoint,” Annu. Rev. Fluid Mech., vol. 53, no. 1, pp. 311–345, January 2021.DOI
- [23]A. H. Bhatt, M. Rodrigues, F. Bernardoni, S. Leonardi, and A. Zare, “Stochastic dynamical modeling of wind farm turbulence,” Energies, vol. 16, no. 19, p. 6908 (24 pages), September 2023.DOI
- [24]M. R. Jovanovic and B. Bamieh, “Componentwise energy amplification in channel flows,” J. Fluid Mech., vol. 534, pp. 145–183, July 2005.DOI
- [25]D. B. Hewawaduge and A. Zare, “Input-output analysis of stochastic base flow uncertainty,” Phys. Rev. Fluids, 2022, to appear; also arXiv:2012.14918.DOI
- [26]S. Abootorabi and A. Zare, “Model-based spectral coherence analysis,” J. Fluid Mech., vol. 958, p. A16 (32 pages), March 2023.DOI
- [27]S. Abootorabi, S. Leonardi, M. Rotea, and A. Zare, “Short-term wind forecasting via surface pressure measurements: stochastic modeling and sensor placement,” J. Renew. Sustain. Energy, vol. 17, no. 1, 2025.arxiv.org/abs/2412.14403
- [28]R. Moarref and M. R. Jovanovic, “Controlling the onset of turbulence by streamwise traveling waves. part 1: Receptivity analysis,” J. Fluid Mech., vol. 663, pp. 70–99, November 2010.DOI
- [29]B. K. Lieu, R. Moarref, and M. R. Jovanovic, “Controlling the onset of turbulence by streamwise traveling waves. part 2: Direct numerical simulations,” J. Fluid Mech., vol. 663, pp. 100–119, November 2010.DOI
- [30]R. Moarref and M. R. Jovanovic, “Model-based design of transverse wall oscillations for turbulent drag reduction,” J. Fluid Mech., vol. 707, pp. 205–240, September 2012.DOI
- [31]M. Naseri and A. Zare, “Perturbation analysis of turbulent channel flow over riblets,” in Proceedings of the 2025 American Control Conference, 2025, pp. 5007–5012.DOI
- [32]A. Chavarin and M. Luhar, “Resolvent analysis for turbulent channel flow with riblets,” AIAA Journal, pp. 1–11, 2019.DOI
- [33]Y. Khalighi, A. Mani, F. Ham, and P. Moin, “Prediction of sound generated by complex flows at low Mach numbers,” AIAA J., vol. 48, no. 2, pp. 306–316, 2010.DOI
- [34]F. Lin, M. Fardad, and M. R. Jovanovic, “Augmented lagrangian approach to design of structured optimal state feedback gains,” IEEE Trans. Automat. Control, vol. 56, no. 12, pp. 2923–2929, December 2011.
- [35]Z. Cheng, J. Ma, X. Li, M. Tomizuka, and T. H. Lee, “Second-order nonconvex optimization for constrained fixed-structure static output feedback controller synthesis,” IEEE Trans. Automat. Control, vol. 67, no. 9, pp. 4854–4861, 2022.DOI
- [36]L. Yuan, S. Chen, C. Zhang, and G. Yang, “Structured controller synthesis through block-diagonal factorization and parameter space optimization,” Automatica, vol. 147, p. 110709, 2023.DOI
- [37]R. Malladi, J. A. Sethian, and B. C. Vemuri, “Shape modeling with front propagation: A level set approach,” IEEE T. Patt. Anal. Mach. Intel., vol. 17, no. 2, pp. 158–175, 2002.
- [38]W. D. McComb, The Physics of Fluid Turbulence. Oxford University Press, 1991.
- [39]F. Odeh and J. B. Keller, “Partial differential equations with periodic coefficients and Bloch waves in crystals,” J. Math. Phys., vol. 5, pp. 1499–1504, 1964.DOI
- [40]A. Bensoussan, J. L. Lions, and G. Papanicolaou, Asymptotic Analysis for Periodic Structures. North Holland, 1978.
- [41]P. J. Schmid and D. S. Henningson, Stability and Transition in Shear Flows. New York: Springer-Verlag, 2001.DOI
- [42]J. L. Aurentz and L. N. Trefethen, “Block operators and spectral discretizations,” SIAM Rev., vol. 59, no. 2, pp. 423–446, 2017.DOI
- [43]J. A. C. Weideman and S. C. Reddy, “A MATLAB differentiation matrix suite,” ACM Trans. Math. Software, vol. 26, no. 4, pp. 465–519, December 2000.
- [44]S. Skogestad and I. Postlethwaite, Multivariable feedback control: analysis and design. Wiley New York, 2007, vol. 2.
- [45]A. T. Gaynor and J. K. Guest, “Topology optimization considering overhang constraints: Eliminating sacrificial support material in additive manufacturing through design,” Struct Multidisc Optim, vol. 54, no. 5, pp. 1157–1172, 2016.DOI
- [46]Q. Li, W. Chen, S. Liu, and L. Tong, “Structural topology optimization considering connectivity constraint,” Struct Multidisc Optim, vol. 54, no. 4, pp. 971–984, 2016.DOI
- [47]D. Petersson, “A nonlinear optimization approach to H₂-optimal modeling and control,” Ph.D. dissertation, Linkopings Universitet (Sweden), 2013.
- [48]M. A. Gomez, A. V. Egorov, S. Mondié, and W. Michiels, “Optimization of the H₂ norm for single-delay systems, with application to control design and model approximation,” IEEE Trans. Automat. Control, vol. 64, no. 2, pp. 804–811, 2018.DOI
- [49]J. Nocedal and S. J. Wright, Numerical Optimization. Springer, 2006.
- [50]D. P. Bertsekas, Constrained optimization and Lagrange multiplier methods. New York: Academic Press, 1982.
- [51]D. P. Bertsekas, Nonlinear programming. Belmont, MA: Athena Scientific, 1999.
- [52]A. Zare, H. Mohammadi, N. K. Dhingra, T. T. Georgiou, and M. R. Jovanović, “Proximal algorithms for large-scale statistical modeling and sensor/actuator selection,” IEEE Trans. Automat. Control, vol. 65, no. 8, pp. 3441–3456, 2020.DOI
- [53]T. Goldstein, C. Studer, and R. Baraniuk, “A field guide to forward-backward splitting with a FASTA implementation,” arXiv preprint arXiv:1411.3406, 2014.arxiv.org/abs/1411.3406
- [54]Y.-H. Dai, M. Al-Baali, and X. Yang, A positive Barzilai–Borwein-like stepsize and an extension for symmetric linear systems. Springer, 2015.DOI
- [55]N. Parikh and S. Boyd, “Proximal algorithms,” Found. Trends Optim., vol. 1, no. 3, pp. 123–231, 2013.DOI
- [56]H. Antil, D. P. Kouri, and D. Ridzal, “Alesqp: An augmented lagrangian equality-constrained sqp method for optimization with general constraints,” SIAM Journal on Optimization, vol. 33, no. 1, pp. 237–266, 2023.
- [57]S. Basir and I. Senocak, “An adaptive augmented lagrangian method for training physics and equality constrained artificial neural networks,” arXiv, 2023.arxiv.org/abs/2306.04904
- [58]J. Jiménez, “Turbulent flows over rough walls,” Annu. Rev. Fluid Mech., vol. 36, no. 1, pp. 173–196, 2004.
- [59]I. P. Castro, “Rough-wall boundary layers: mean flow universality,” J. Fluid Mech, vol. 585, pp. 469–485, 2007.DOI
- [60]T. Rautert and E. W. Sachs, “Computational design of optimal output feedback controllers,” SIAM J. Optim., vol. 7, no. 3, pp. 837–852, 1997.




