Journal of Turbulence

ISSN: (Print) 1468-5248 (Online) Journal homepage: http://www.tandfonline.com/loi/tjot20

Drag reduction by herringbone riblet texture in

direct numerical simulations of turbulent channel
flow

H.O.G. Benschop & W.-P. Breugem

To cite this article: H.O.G. Benschop & W.-P. Breugem (2017) Drag reduction by herringbone
riblet texture in direct numerical simulations of turbulent channel flow, Journal of Turbulence, 18:8,
717-759, DOI: 10.1080/14685248.2017.1319951

To link to this article: https://doi.org/10.1080/14685248.2017.1319951

© 2017 The Author(s). Published by Informa

UK Limited, trading as Taylor & Francis
Group

Published online: 02 May 2017.

Submit your article to this journal

Article views: 581

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tjot20

Download by: [191.96.251.147] Date: 11 December 2017, At: 03:47

 JOURNAL OF TURBULENCE, 
VOL. , NO. , –
https://doi.org/./..

Drag reduction by herringbone riblet texture in direct

numerical simulations of turbulent channel flow
H.O.G. Benschop and W.-P. Breugem
Laboratory for Aero and Hydrodynamics, Delft University of Technology, Delft, The Netherlands

ABSTRACT ARTICLE HISTORY

A bird-feather-inspired herringbone riblet texture was investigated Received  November 
for turbulent drag reduction. The texture consists of blade riblets in a Accepted  April 
Downloaded by [191.96.251.147] at 03:47 11 December 2017

converging/diverging or herringbone pattern with spanwise wave-

KEYWORDS
length f . The aim is to quantify the drag change for this texture Drag reduction; riblets; direct
as compared to a smooth wall and to study the underlying mech- numerical simulations
anisms. To that purpose, direct numerical simulations of turbulent
flow in a channel with height Lz were performed. The Fukagata-
Iwamoto-Kasagi identity for drag decomposition was extended to
textured walls and was used to study the drag change mechanisms.
For f /Lz  O(10), the herringbone texture behaves similarly to a
conventional parallel-riblet texture in yaw: the suppression of turbu-
lent advective transport results in a slight drag reduction of 2%. For
f /Lz  O(1), the drag increases strongly with a maximum of 73%.
This is attributed to enhanced mean and turbulent advection, which
results from the strong secondary flow that forms over regions of
riblet convergence/divergence. Hence, the employment of conver-
gent/divergent riblets in the texture seems to be detrimental to tur-
bulent drag reduction.

1. Introduction
Drag-reducing techniques can be of great value for fuel consumption reduction, as a sig-
nificant part of the fuel used for transportation arises from drag in turbulent flows. Drag
reduction (DR) in fluid flows can be obtained by active and passive methods. Active meth-
ods include the use of additives [1], gas injection [2], wall mass transfer [3], wall cooling
or heating [4–6], wall motion [7], wall deformation [8] and electromagnetic forcing [9].
Passive methods include the change of surface chemistry (e.g. superhydrophobicity [10]),
elasticity (e.g. compliant walls [11]), shape (e.g. airfoil or ship hull shape) and texture (e.g.
riblets [12]). Surface roughness and surface textures have been successfully used for DR by
transition delay in laminar flow [13], separation delay in turbulent flow over curved sur-
faces [14,15] and turbulence modification in turbulent flow over flat surfaces.
Of the investigated flat plate textures – such as sifted sand grains [16], transverse square
grooves [17], dimples [18] and V-shaped protrusions [19] – the well-studied riblet texture
seems most promising for turbulent DR. This texture has been found on the scales of some
fast-shark skins [20] and consists of ridges or riblets aligned with the mean flow direction.

CONTACT H.O.G. Benschop [email protected]

©  The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License
(http://creativecommons.org/licenses/by-nc-nd/./), which permits non-commercial re-use, distribution, and reproduction in any medium,
provided the original work is properly cited, and is not altered, transformed, or built upon in any way.
 718 H. O. G. BENSCHOP AND W.-P. BREUGEM

With a simplified riblet geometry, a maximum DR of almost 10% has been obtained [21].
In the search for even higher values of DR, many variations on the standard riblet geome-
tries have been investigated [12], such as hierarchical or compound riblets [22], riblets
on a spanwise travelling surface wave [23], oscillating riblets [24–26], riblets in a wave-
like pattern (either in phase [27] or out of phase [28]) and riblets combined with drag-
reducing polymers [29]. The rationale behind these alternatives is to further reduce drag by
somehow incorporating other drag-reducing methods, such as oscillating walls or polymer
addition.
A little-studied alternative to the standard riblets is the herringbone riblet texture that
has been found on bird flight feathers. Feathers serve several functions, including flight,
thermal insulation, waterproofing and colouration (e.g. for camouflage or visual signals).
The following description of the feather morphology is derived from Chen et al. [30,31,32].
A feather consists of two vanes (anterior and posterior), separated by a central supporting
shaft. Each vane consists of parallel barbs attached obliquely to the shaft. The barbs are
Downloaded by [191.96.251.147] at 03:47 11 December 2017

linked together by a set of finer barbs, called barbules. A microgroove is formed between
neighbouring barbs. The angle between shaft and barbs is typically α  30°. The groove
spacing s remains approximately constant with s+  20. The groove depth decreases grad-
ually away from the shaft. A wing is formed when several feathers are positioned next to
each other with almost parallel shafts. The feather shafts are approximately parallel to the
flying direction for steady forward glide.
The study of Chen et al. [31] is one of the first to investigate the herringbone riblet tex-
ture for turbulent DR. Inspired by bird flight feathers, they designed a riblet texture with
two typical features that differ from the shark-skin riblets. First, the riblets were arranged in
a converging/diverging or herringbone pattern. Second, the riblet height or groove depth
changed gradually. Using laser engraving and replica moulding, the researchers manufac-
tured such bio-inspired herringbone-riblet skins. They covered the inner wall of a test pipe
with these skins and measured a DR of up to 20%, twice that of optimal standard riblets.
Although this seems promising, it has not been reproduced yet, and it is unclear for which
texture and flow parameters drag is maximally reduced.
Sagong et al. [19] investigated a rather comparable geometry, namely the V-shaped pro-
trusions that were found on the sailfish skin. In a comprehensive experimental and numer-
ical study, they found a few cases for which drag was decreased slightly (∼1%), although
within the experimental uncertainty. The herringbone riblet texture also resembles the vor-
tex generators that have been used to reduce drag by flow-separation delay [33] or transition
delay [34]. Furthermore, several studies confirm that roughness on a bird wing contributes
to separation control [35–37].
Herringbone riblet textures have been studied for several other reasons, such as for heat-
transfer enhancement [38,39] and mixing of laminar flows in microchannels [40]. The
hot-wire study by Koeltzsch et al. [41] was initiated by the observation of convergent and
divergent riblet patterns on the shark skin near sensory organs, possibly used for local flow-
noise reduction. Nugroho et al. [42] revisited the same texture and conducted a paramet-
ric study using the hot-wire technique. They were motivated by the potential use for pas-
sive flow control and disruption of large-scale coherent motions. They proposed that the
herringbone riblets induce large-scale counter-rotating vortices, giving rise to large-scale
spanwise periodicity in the boundary layer. The suggested streamwise counter-rotating
 JOURNAL OF TURBULENCE 719

vortices have recently been visualised in a flat-plate laminar boundary layer over con-
vergent riblets [43]. So, the motivation for herringbone riblet studies has been DR, heat-
transfer enhancement, mixing improvement, flow control and large-scale flow structures
manipulation.
The study by Nugroho et al. [42] is a manifestation of the recent interest in surfaces that
manipulate the whole boundary layer. There is much evidence that DR techniques that rely
on near-wall flow manipulation are less effective at higher Reynolds numbers: DR degrades
with increasing Re [44–47]. That has partially motivated the study of rough or textured
surfaces that also impact the outer part of the turbulent boundary layer by means of large-
scale secondary flows. These secondary flows have been observed over spanwise-varying
longitudinal bedforms [48], a regular array of cubes [49], a bed with two parallel lanes of
different roughnesses [50], the irregular surface of a replica of a damaged turbine blade
[51], streamwise strips of elevated roughness [52] and the aforementioned herringbone
riblet texture [41,42]. Interestingly, Schoppa and Hussain [53] obtained 20% DR in direct
Downloaded by [191.96.251.147] at 03:47 11 December 2017

numerical simulations (DNSs) of turbulent channel flow with imposed large-scale counter-
rotating streamwise vortices.
The aim of this paper is to quantify how and why drag is changed by a herringbone
texture as compared to a smooth wall, using DNSs. The numerical methods are described
in Section 2. Section 3 explains how DR is quantified. Section 4 validates the simulated
drag of smooth walls, parallel riblets aligned with the mean flow, and parallel riblets in
yaw. Section 5 examines the herringbone texture with use of a parametric study. A drag
decomposition is derived and applied in Section 6. A discussion of the results is presented
in Section 7, followed by the main conclusions and an outlook in Section 8.

2. Numerical methods
In this study, DNSs of incompressible turbulent flow were performed. An overview of all
simulations with the corresponding parameters can be found in Appendix C. This section
outlines the method that was used, describing successively the notation conventions, flow
domain, texture, grid, numerical code and averaging.
Notation: In what follows, dimensional variables are denoted by an asterisk ∗ . Vari-
ables without that asterisk are nondimensionalised using the domain height L∗z and the
bulk velocity Ub∗ , such that Lz = 1 and Ub = 1. Note that Ub∗ is a constant, as simulations
were performed at fixed mass flow rate. The constant bulk Reynolds number is defined as
Reb = Ub∗ L∗z /ν ∗ , with kinematic viscosity ν ∗ . The superscript + is used for nondimension-
√
alisation with ν and uτ = τw /ρ, with wall shear stress τ w and fluid density ρ. Nondimen-
sionalisation for textured walls uses the viscous wall units derived from the smooth-wall
simulation with the same Reb .
Domain: The flow domain is a plane channel, bounded by two horizontal walls. At the
channel walls, no-penetration and no-slip boundary conditions are applied, whereas peri-
odic boundary conditions are used in the streamwise and spanwise directions. The stream-
wise, spanwise and wall-normal coordinates are denoted by x, y and z with the correspond-
ing velocity components u, v and w.
The domain is specified by its length Lx , width Ly and height Lz . For a good comparison,
the domain size should ideally be the same for all simulations. However, slight size varia-
tion was needed to fit an integer number of texture periods in the streamwise and spanwise
 720 H. O. G. BENSCHOP AND W.-P. BREUGEM

directions, or to ensure that the number of grid cells complies with the parallel-computing
algorithm. In general, all domains are approximately of size (4.0 × 2.5 × 1), which is consid-
ered to be large enough to obtain reliable statistics for several reasons. First, it is comparable
to the domain size used by other researchers [54–57]. In addition, it is full-span, as opposed
to the recently reconsidered minimal-span channels [58]. Finally, it is larger than the mod-
erate box of size (π × π/2 × 1) that is large enough to reproduce the one-point statistics of
larger boxes [59].
Texture: Textures are applied to the inside of both channel walls to enforce symmetry in
the mean flow [60]. Unless stated otherwise, the top wall texture is the bottom wall texture
mirrored in the centreline plane. To simulate flow over a non-smooth surface, two methods
can be adopted: coordinate transformation or the immersed boundary method (IBM) [56].
Both methods have been applied to simulate turbulent flow over riblet walls [61,62]. We
used an IBM similar to the one employed by Breugem and Boersma [55] and Pourquie
et al. [63], which is based on Fadlun et al. [64]. The IBM forcing is direct, i.e. a forcing term
Downloaded by [191.96.251.147] at 03:47 11 December 2017

is added to the discretised equations. Appendix B provides case-specific details.

Grid: The chosen IBM allows the use of a simple staggered Cartesian grid. The number
of grid cells in the three Cartesian directions is denoted by Nx , Ny and Nz . The grid is uni-
form in the horizontal directions, so the grid spacings x and y are constant. In the wall-
normal direction, three zones are distinguished, namely the roughness regions near the two
walls and the remaining part of the channel. In the roughness region, which extends from
the wall to one grid cell above the maximum texture height, the vertical grid spacing zw
is constant. In the remaining part of the channel, grid stretching is applied using a cosine
function that is symmetrical with respect to the channel centreline. The maximum vertical
grid spacing occurs at the centreline and is called zc .
The grid-cell size is important to correctly resolve small-scale fluid motions. The recom-
mendations of Vreman and Kuerten [65] for finite difference codes were followed, namely
x+ = 4.4, y+ = 2.9, zw + = 0.49 and zc + = 2.2 as maximum grid spacings for smooth-
wall turbulent flows. For textured walls, the spanwise grid spacing was reduced to y+ 
1.0. For all investigated textures, one simulation at a double spanwise and/or streamwise res-
olution was performed. Particular attention was paid to the resulting drag change, which
was marginal in all cases. As only one simulation at a higher resolution was performed
for each texture, grid independence of the results cannot be claimed. However, as the grid
resolution is relatively high and about the same in all cases, comparison of results is still
justified.
Code: The incompressible Navier–Stokes equations and continuity equation were solved
at fixed bulk velocity:

∂ui ∂ui u j ∂p 1 ∂ 2 ui
+ =− + + γ f f driving δi1 + fiIBM , (1)
∂t ∂x j ∂xi Reb ∂x2j
∂u j
= 0, (2)
∂x j

where the Einstein summation convention for repeated indices is used. Here, ui represents one
component of the velocity vector, t time, xj a spatial coordinate, p the pressure and f driving
the spatially uniform forcing term to obtain a constant bulk velocity. The phase-indicator
function γ f is defined at grid points of the streamwise velocity. It equals 1 in fluid and 0
 JOURNAL OF TURBULENCE 721

in solid obstacle volume to ensure that only fluid experiences the bulk forcing that drives
the flow. The Kronecker-delta function δ i1 guarantees that fluid is driven in the streamwise
direction with i = 1. Appendix A describes how f driving is calculated. The IBM forcing fiIBM
is a body force that models the (drag) force that the texture exerts on the flow (see Appendix
B for more details).
These equations are discretised using the finite-volume method combined with a
pressure-correction scheme [66]. Fluxes or stresses at the cell faces are evaluated using lin-
ear interpolation, i.e. a central-differencing scheme is used. Time-integration is performed
using a fractional-step method that consists of three steps. For the pressure, which is stag-
gered in time with respect to the velocities, a Crank–Nicolson scheme is used. All other
terms are advanced in time using a three-step Runge–Kutta method [67]. This discreti-
sation procedure yields a Poisson equation, which is solved using a non-iterative solver
based on fast Fourier transforms (FFTs). Specifically, FFTs are applied to the horizontal
directions and the resulting tridiagonal system is solved using Gaussian elimination. More
Downloaded by [191.96.251.147] at 03:47 11 December 2017

details about the time advancement at fixed bulk velocity can be found in Appendix A.
Averaging: For computation of flow statistics, a unit-cell average was stored each 100
timesteps. Like in crystallography, a unit cell is the smallest unit of volume that builds up
the entire texture by translation. It extends vertically from bottom to top wall. For smooth
walls, its size in grid cells is 1 × 1 × Nz . In a unit-cell average, the data of all unit cells are
reduced to an average in one unit cell.
In addition to this unit-cell average during the computations, temporal and spatial aver-
ages were performed afterwards. Let φ = φ(x, y, z, t) represent an arbitrary flow variable.
The following averages were used:

 tstat +T
1
φ= φ dt, (3)
T tstat
 Lx
1 i
φxi = φ dxi , (4)
Lxi 0

1
φV = φ dV. (5)
V V

Here, φ is a time average over the statistically stationary part of the signal φ (which starts
at tstat and has duration T). An average over one spatial coordinate is denoted by φxi . For
instance, φx is a streamwise average. A similar notation is adopted for an average over two
spatial coordinates. For example, φxy is a streamwise and spanwise average. Finally, φV
represents an average over the entire volume V = Lx Ly Lz . For instance, the bulk velocity is
defined by Ub = uV .

3. Drag reduction quantification

This section describes the quantification of DR. It explains under what conditions the
smooth- and textured-wall flows are compared, with special attention for flow generation,
the definition of a reference case and the formulation of the benefit of DR.
 722 H. O. G. BENSCHOP AND W.-P. BREUGEM

In general, channel flow is generated by either a constant flow rate (CFR), a con-
stant pressure gradient (CPG) or a constant power input (CPI) [68]. The obtained DR
depends slightly on the choice of CFR, CPG or CPI. In this study, the flow was generated
by CFR.
The quantification of DR requires the definition of a reference case compared to which
drag is reduced or increased. Special attention should be paid to the Reynolds number and
channel height, as drag depends heavily on both of them. Conceptually, this study compares
two channels with the same fluid, the same flow rate per unit of spanwise width and the
same outer dimension L∗z . With Ub∗ L∗z being the volumetric flow rate per unit width, the first
two conditions imply that Reb is the same for both flows. The third condition guarantees
that both channels are geometrically identical, apart from the texture that is applied to the
inside of the channel walls in one case. It implies that the wall location is not adjusted to
compensate for the texture volume. The fluid volume for the case with textured walls is thus
slightly less than that of the smooth-wall case. This is a conservative choice: DR cannot
Downloaded by [191.96.251.147] at 03:47 11 December 2017

result from an increased fluid volume or a locally increased channel height [69,70].
This paper uses the drag change DC as a quantifying parameter. To account for pos-
sible differences in domain width and length, DC was computed from the time-averaged
drag force per unit volume. As the driving term balances the total drag, the instantaneous
 2  −1 −1 
drag force per unit volume fd∗ = ρ ∗ Ub∗ L∗z V V f
γ f driving dV . Since the dimen-
sional prefactor is equal for the smooth- and textured-wall channel flows, the drag change
is given by

  
γ f f driving V 
DC =   −1
textured

γf f driving
V  (6)
smooth
tot
dtextured
≡ tot
− 1.
dsmooth

It measures the increase of the driving force that is required to maintain a given flow rate.
The drag reduction DR = −DC, so drag is reduced in case DC is negative.
The thus computed drag change is supplemented by a 95% confidence interval. The
error in DC can be attributed to the uncertainty in the drag computed for both the tex-
tured and smooth wall. For ease of notation, define dt ≡ dtextured
tot
and ds ≡ dsmooth
tot
. Let uφ
for now denote the uncertainty in φ. Given the independence of dt and ds , the uncertainty
in DC follows from the law of error propagation:

2 2
udt dt uds
u2DC = + . (7)
ds ds2

The uncertainties in dt and ds were computed using the method outlined by Hoyas and
Jiménez [71]. It accounts for correlation in the drag time signal. The thus obtained error
bar only results from the finite simulation time. Errors of other origins (e.g. discretisation
errors) were not considered.
 JOURNAL OF TURBULENCE 723

4. Validation
The numerical methods were validated with simulations of smooth walls, parallel riblets,
and parallel riblets in yaw, as described in the following subsections.

4.1. Smooth wall

Smooth-wall DNSs were performed at three bulk Reynolds numbers, namely 5500, 11000
and 22000. Table 1 shows the corresponding friction Reynolds number Reτ = uτ δ/ν, where
δ = Lz /2 is the half-channel height. Compared to recent DNSs reaching Reτ = 4000 [72],
Reτ  4200 [59] and Reτ  5200 [73], the simulations in this study are considered to be
standard. Therefore, this subsection suffices to validate drag and flow statistics.
Drag is validated by comparing Reτ with the value predicted by an approximate relation
Pope Pope
Reτ = 0.09Re0.88 b [74]. The deviation of Reτ from Reτ is quantified using
Downloaded by [191.96.251.147] at 03:47 11 December 2017

Reτ
RePope
τ = Pope
− 1. (8)
Reτ

Pope Pope
Table 1 lists Reτ for all smooth-wall simulations. Although the relation for Reτ is
approximate, good agreement is obtained for all Reb .
Flow statistics of U, urms , v rms , wrms and prms were compared with simulations at Reτ =
180 [65] and Reτ = 590 [57]. Here, U = uxy , urms = ( u2 xy − u2xy )1/2 and similarly for
the other root-mean-square quantities. Let Q(z) be one of these statistics, then the root-
mean-square relative deviation

 
2
Q(z) − Qref (z)
δQ = (9)
Qref (z)
z

is used to quantify the difference between Q (current) and Qref (reference). It was computed
after piecewise cubic spline interpolation of Q and Qref to a uniform grid, z+ (k) = k for inte-
gers 1  k  kmax , with kmax = 175 at the lowest and kmax = 587 at the highest Reb . Table 1
shows δQ for the five flow quantities. The root-mean-square relative deviation is smaller
than 1% for the mean velocity and smaller than 2% for the root-mean-square fluctuations.
The somewhat larger value for prms at the lowest Reb is attributed to the difference between
Reτ and Reref
τ .

Table . Validation of drag and flow statistics of smooth-wall simulations. The relative deviation of Reτ
Pope
from that predicted by Pope’s relation is given by Reτ (Equation ()). Statistics of U, urms , v rms , w rms
and prms are compared with simulations by Vreman and Kuerten [,] at Rerefτ = 180 and . The rel-
ative difference between current and reference flow statistic Q is measured with the root-mean-square
relative deviation δQ (Equation ()).
Reb Reτ RePope
τ (%) Reref
τ δU (%) δurms (%) δv rms (%) δw rms (%) δprms (%)
 . −.  . . . . .
 . −. - - - - - -
 . −.  . . . . .
 724 H. O. G. BENSCHOP AND W.-P. BREUGEM

Figure . Parallel blade riblet texture. Left: Bird’s-eye view of the texture aligned with the mean flow,
showing six unit cells in the spanwise direction. Right: Bird’s-eye view of the texture in yaw with yaw
angle α, showing six unit cells in the spanwise and three in the streamwise direction.

4.2. Parallel riblets

Downloaded by [191.96.251.147] at 03:47 11 December 2017

A parallel riblet texture consists of riblets aligned with the mean flow direction and can
reduce turbulent drag up to almost 10% [21]. An overview of previous research can be
found elsewhere [12]. A thorough DNS study has been performed quite recently [60,76,77].
This paper investigates the blade riblet texture (see Figure 1). The blades have zero thick-
ness, spacing s and height h with h/s = 0.5. In a small parametric study, mainly s+ and Reb
were varied (see also Table C1). The grid resolution (specified in wall units) is about the
same for all cases.
Figure 2 (left) compares the simulated drag change as function of s+ with experiments
performed by Bechert et al. [21]. The top axis shows the experimental bulk Reynolds num-
ber ReBechert , which is based on the horizontal channel width and the average velocity
between the test plates. In the experiments only the Reynolds number was varied, whereas
in the present numerical study both the Reynolds number and the riblet spacing were var-
ied.
The drag change varies slightly with Reynolds number for fixed s+ : the drag at Reb =
5500 is higher than at Reb = 11000 and 22000. The approximate overlap of the data points
at the two highest Reynolds numbers (for s+ = 24) suggests a low-Reynolds-number effect,
which is underpinned by the observation that DR data below ReBechert  10000 deviated
more and more from previous high-Reynolds-number data [21]. The deviation of DC at
Reb = 5500 from that at higher Reynolds numbers is also larger at larger s+ , which might
be explained by riblet height increase. For Reb = 5500, an increase of s+ from 10 to 24
is accompanied by a decrease of δ/h from 35 to 15. The blades protrude farther into the
channel, which is presumably detrimental to DR. This is supported by the finding that, for
δ/h  50, the effect of roughness extends across the boundary layer and the original wall
flow dynamics is changed significantly [78].
The simulations capture the experimental trend quite well when Reb is close to ReBechert .
Especially good agreement is obtained around and below the optimum spacing. At s+ =
17, a maximum DR of 9.3% is achieved, which is very close to the 9.9% of the exper-
iments. The difference is slightly larger at s+ = 24: there the total drag is reproduced
within 5%. A higher reproduction accuracy probably requires a combined experimen-
tal/numerical study, a more extensive grid resolution study, the incorporation of blade
thickness and an analytical correction of momentum fluxes near the riblet tips to resolve
 JOURNAL OF TURBULENCE 725

ReBechert
6624 9992 13856 18006 22334 26843
4 Bechert et al. [1997] EXP 2001: trapezoidal 45° (7.8) 3.7
Reb = 5500 160 EXP 2001: trapezoidal 30° (8.2)
2 Reb = 11000 EXP 2001: triangular (6)
140
Reb = 22000 EXP 2001: semicircular (7.6)
0 120
SIM 2012: trapezoidal 30° (6.8)

DRloss [%]
-0.9
DC [%]

−2 100 SIM: blade (7.6)

−4 80
-4.0
60
−6
40
−8
-7.6
20
−10
0
−12
8 10 12 14 16 18 20 22 24 26 28 30 0 2 4 6 8 10 12 14 16 18 20
s+ α [°]

Figure . Validation of drag change for the parallel blade riblet texture. Left: Drag change as function
Downloaded by [191.96.251.147] at 03:47 11 December 2017

of riblet spacing in wall units (no yaw). Numerical results at three different Reynolds numbers are com-
pared with experimental data of Bechert et al. []. The experimental bulk Reynolds number is denoted by
ReBechert . Right: Drag reduction loss (Equation ()) as function of yaw angle for different riblet geometries
at s+ = . EXP  are experiments by Hage et al. [] at Reb  . SIM  are DNSs by Grüneberger
et al. [] at Reb = . SIM are the current simulations. The numbers in parentheses in the legend repre-
sent the maximum drag reduction DRmax in percent. The numbers in the figure represent the drag change
DC in percent for the simulated blade riblet geometry.

the high gradients there. However, the accuracy demonstrated here is sufficient for the her-
ringbone riblet simulations, as drag differences for these were found to be much larger
than 5%.

4.3. Parallel riblets in yaw

Among other factors, the performance of riblets deteriorates in yaw, i.e. when they are not
aligned with the mean flow direction. An overview of past research is given by Koeltzsch
et al. [41]. The study of Hage et al. [75] shows the significant influence of riblet geometry
and spacing s+ . Simulations of turbulent flow over parallel riblet textures in yaw are rare
in the literature. Grüneberger et al. [27] have performed DNSs at Reb = 5750 (Reτ = 180)
for trapezoidal grooves. The driving pressure gradient was rotated such that the streamwise
direction no longer coincides with the x-direction. Good agreement with experimental data
was obtained.
This paper investigates the blade riblet texture in yaw (see Figure 1) with h/s = 0.5 and
s+ = 17. The yaw angle α was varied for fixed Reb = 5500. The horizontal grid resolution was
comparable to that used for aligned parallel riblets: x+  4.1 and y+  1.1 for all cases.
As the trick of driving-pressure-gradient rotation cannot be applied for herringbone-riblets
simulations, the parallel riblet texture was rotated with respect to the grid. That required
a different IBM, as the blades are not anymore aligned with the Cartesian directions (see
Appendix B for details).
Figure 2 (right) shows DC as function of α (indicated by the numbers in the figure).
Clearly, the DR diminishes with increasing yaw angle, as is expected from the literature.
Unfortunately, to the best of the authors’ knowledge, no experimental data of blade riblets
 726 H. O. G. BENSCHOP AND W.-P. BREUGEM

in yaw is available. For comparison of the simulated results with experimental data for other
riblet geometries, the drag reduction loss DRloss is introduced [75]:

DRmax − DR(α, s+ )
DRloss (α, s+ ) = , (10)
DRmax

where DR = −DC, and DRmax is the maximum DR that can be obtained with a given geom-
etry: DRmax = DR(α = 0, s+ = s+ +
opt ). It was assumed that DRmax = DR(α = 0, s = 17) for the
simulations.
Figure 2 shows the DR losses as function of yaw angle for different riblet geometries at
s+ = 17. The experimental data of Hage et al. [75] (at Reb  14900) and the numerical
data of Grüneberger et al. [27] (at Reb = 5750) are included. The values of h/s are geometry
dependent, namely h/s = 0.5 for the trapezoidal and blade, h/s = 0.7 for the semicircular,
and h/s = 1 for the triangular geometries.
Downloaded by [191.96.251.147] at 03:47 11 December 2017

The dependence of DRloss on α is different for each geometry, which might be explained
by differences in h/s and riblet shape [75]. Of the two trapezoidal geometries, the one with
the sharpest tip (30°) is slightly more sensitive to misalignment. The triangular geometry
has the largest tip angle of about 54°, but its DR loss is nevertheless larger than for the trape-
zoidal geometries, presumably because of its larger height (h/s = 1). Of the experimental
data, the semicircular geometry is most sensitive to yaw, likely because of its larger height
(h/s = 0.7) and sharp tip.
The simulation results suggest that the blade riblet geometry is more susceptible to yaw
than any of the other geometries. This is likely not a low-Reynolds-number effect, as the
simulations at Reb = 5750 by Grüneberger et al. [27] reproduced well the experimental
data (see the figure). Instead, it might be ascribed to the sharper riblet tips (and the asso-
ciated pressure drag increase) and the broader riblet valleys (and the associated increase of
sloshing) [75].

5. Herringbone riblets

5.1. Texture description

The feather texture (described in Section 1) is modelled as shown in Figure 3, fairly similar
to the spatial three-dimensional (s-3D) texture proposed in Chen et al. [31]. It consists of
Nfeather feathers placed in parallel such that their edges touch. The modelled feather has
no physical shaft, although the term ‘shaft’ is employed to denote the symmetry axis of a
feather. The barbs are modelled as blade riblets with zero thickness. As opposed to the s-
3D texture that consists of sawtooth riblets, the present study uses blades because of their
superior drag-reducing performance in the conventional riblet texture. The angle between
the positive x-direction and the blades is called α, which is restricted to 0  α < 180°. The
shortest distance between the blades is s. Between two neighbouring blades a groove forms,
which has spanwise blade spacing sy = s/cos (α). Ngroove is the integer number of spanwise
blade spacings sy that fits in one feather half-width (see also Figure 3). The feather width
or spanwise texture wavelength f = 2Ngroove sy . The riblet height hr decreases linearly with
distance to the shaft. Let h represent the riblet height at the shaft and ys the y-coordinate of
 JOURNAL OF TURBULENCE 727

Figure . Herringbone blade riblet texture with  grooves per feather half (Ngroove = ), and angle
between shaft and riblets of  degrees (α = °). Left: Bird’s-eye view of the texture, showing five unit
cells in the streamwise and three unit cells (or feathers) in the spanwise direction. The edges and shaft of
Downloaded by [191.96.251.147] at 03:47 11 December 2017

the middle feather are shown. Backward flow over this texture is equivalent to forward flow over a texture
with α = °. Right: Top view of the texture, showing two unit cells in the streamwise and one unit cell
in the spanwise direction. The feather width is f . Four grooves on one feather half are numbered.

the shaft, then hr is given by

 
hr (y)  y − ys  1 y − ys 1

=1− , for − ≤ ≤ . (11)
h  f /2  2 f 2

Note that Koeltzsch et al. [41] and Nugroho et al. [42] used herringbone riblets of constant
height. The texture was implemented using the same IBM that was used for blade riblets in
yaw; only the texture indicator functions were different.
As Figure 3 shows, a difference is made between forward and backward flow. ‘Forward’
is used for bulk flow in the positive, ‘backward’ indicates flow in the negative x-direction.
Forward flow over a texture with angle α is the same as backward flow over a texture with
angle α backward = 180° − α. Using this trick, textures with angle α and α backward can be
compared to study the influence of mean flow direction on drag.
In addition to the standard herringbone texture, two texture variations were considered.
The first variation is called ‘shifted’. It differs from the standard herringbone texture by a
spanwise shift of the top wall texture by half a spanwise texture wavelength as compared
to the bottom wall (see Figure 4). This shifted texture was investigated for its ability to
generate the drag-reducing secondary flow that is described by Schoppa and Hussain [53],
namely one that extends from the bottom to the top wall. The second variation is a riblet
texture with α = 0° (see Figure 4), which results in a parallel blade riblet geometry with
blade height variation in the spanwise direction given by Equation (11). This texture does
not suffer from yaw and the resulting pressure drag, but it still might give rise to secondary
flows.
The herringbone texture and its variations have been subjected to a parametric study
(see also Table C1). According to Chen et al. [31], bird-feather parameters are typically s+
 20, h/s  0.5 and α  30°. Values for Ngroove or f were not given. Their SEM-pictures
show that the feather barbs are comparable to blades with finite thickness and rounded
tips. Although the present study was inspired by the bird-feather texture, it did not attempt
 728 H. O. G. BENSCHOP AND W.-P. BREUGEM

Figure . Herringbone riblet texture variations. Left: Bird’s-eye view of a shifted texture variation with
Ngroove =  and α = °. ‘Shifted’ refers to the spanwise shift of the top wall texture by half a spanwise
texture wavelength as compared to the bottom wall. The two vertical planes demarcate the middle unit
cell. Note that the distance between bottom and top wall is not to scale. Right: Bird’s-eye view of a par-
allel texture variation with Ngroove =  and α = °, showing three unit cells (or feathers) in the spanwise
Downloaded by [191.96.251.147] at 03:47 11 December 2017

direction.

to exactly reproduce that texture. Instead, texture parameters were chosen to allow a close
comparison with the parallel-riblet studies described in Subsections 4.2 and 4.3. Specifi-
cally, ridge spacing and maximum blade height were fixed at s+ = 17 and h/s = 0.5. Mainly
two parameters were varied, namely α and Ngroove . The angle α was either 15° or 165°, which
permits the direct comparison with the conventional riblet texture in 15 degrees yaw. Larger
angles (e.g. α = 30° or α = 150°) were not considered, as Figure 2 suggests that such angles
will result in a drag increase. The number of grooves Ngroove was varied between 1 and 128,
which resulted in a feather-width change from f = 0.10 to 12.9 in outer units or +f = 35
to 4506 in wall units.
Apart from texture parameters, the other simulation parameters were (almost) the same
for all herringbone simulations, which facilitates a fair comparison. Specifically, Table C1
shows that domain size, grid resolution, Reynolds number and simulation time do not
change much among the simulations. In addition, these parameters are also close to the
ones for parallel-riblet (without/in yaw) simulations. All herringbone simulations were per-
formed at Reb = 5500. Based on the validation of blade-riblet simulations (see Subsection
4.2), a low-Reynolds-number effect can be anticipated. However, as all simulations were
performed at the same Reb , a proper comparison can still be made. The number of grid
cells per groove Ncg = 16, so 16 grid cells span the groove width.
To substantiate the accuracy of the used numerical methods, two issues have to be
addressed. The first issue relates to the boundary conditions at the texture surface.
Appendix B.4 shows that the no-slip and no-penetration conditions are sufficiently satis-
fied. The second issue relates to the flow around the blade tip. The exact solution is singular
there, which might introduce errors in the numerical solution. The improvement of the
numerical accuracy for problems involving singularities is a research in itself [79], but it
is not the focus of the current study. For the present purpose, the used IBM is sufficiently
accurate. Very similar IBMs have been used in DNSs of flow around other obstacles with
sharp corners, such as cubes [55,80], square and triangular elements [56], and a flat plate
normal to the free stream [81,82]. The penultimate example shows that the IBM can deal
with obstacles that are not aligned with the Cartesian grid, whereas the last example demon-
strates that the IBM can also accurately capture separating flows at sharp corners.
 JOURNAL OF TURBULENCE 729

Λf (Λ+
f)
0.10 (35) 0.40 (141) 1.6 (563) 12.9 (4506)
80
α= 0°
70 α= 15°
α= 165°
60 α= 165° shift
50

DC [%]
40

30

20

10

−10
1 2 4 8 16 32 64 128
Ngroove

Figure . Drag change as function of feather width for the herringbone riblet geometry, including results
Downloaded by [191.96.251.147] at 03:47 11 December 2017

for the parallel (α = °) and shifted variants (see Figure ). The shown values of f and +
f
apply to the
textures with α = ° or °, and are approximate for α = ° textures. The data points on the right
vertical axis belong to conventional parallel riblets with yaw angle α = ° and α = °.

5.2. Drag reduction

Trend: Figure 5 shows the drag change as function of feather width for the herringbone
riblet geometry. Drag changes significantly with the spanwise texture wavelength or feather
width. For the texture with α = 165°, the drag increases by 61% for very narrow feathers,
while the drag decreases by 2% for very wide feathers. In general, wider feathers experience
less drag. Only the texture with α = 15° and Ngroove = 1 does not comply with this trend.
DR was only obtained when the texture approaches the parallel riblet texture in yaw, i.e.
in the limit of very wide feathers. Only 2% reduction was found in that limit, which is
indeed comparable to the 0.9% reduction that was obtained for parallel riblets at yaw angle
α = 15°. It is significantly less than the 7.6% reduction for parallel riblets aligned with the
mean flow.
Difference forward and backward flow: There is a clear drag difference between textures
with α = 15° (forward flow) and α = 165° (backward flow). For example, at Ngroove = 16
drag increases by 34% for α = 15°, and only by 7.5% for α = 165°. This shows that mean-
flow reversal has a considerable effect on drag, in agreement with Chen et al. [31]. Here,
forward flow experiences more drag than backward flow, except for the Ngroove = 1 texture.
The drag difference between forward and backward flow decreases for increasingly wide
feathers, as can be expected. The textures with α = 15° and 165° differ only because of riblet
convergence or divergence in the feather shaft or feather edge regions. When f → , only
a very small portion of the complete texture consists of converging or diverging riblets, so
their contribution to the total drag becomes negligible.
Standard texture variations: Figure 5 also shows the drag for the shifted and parallel riblet
(α = 0°) variations. The drag of textures with and without shift is about the same for all
feather widths. Hence, shifting of the top wall texture has almost no effect on drag, although
the next subsection will explain that the mean flow is different for some cases.
In contrast with most herringbone textures, the parallel riblet texture with span-
wise riblet height variation is able to reduce drag. The DR is the least for Ngroove = 1,
namely 0.5%. The texture with Ngroove = 1 and α = 0° is the conventional parallel-riblet
 730 H. O. G. BENSCHOP AND W.-P. BREUGEM

geometry with s+ = 17, but with every second blade removed. It is known that this
geometry is not optimal for DR. When Ngroove increases, DR increases as well. In the limit of
very large Ngroove , the texture approaches the standard parallel-riblet geometry, apart from
a very slight spanwise height variation. Therefore, in that limit one might expect the DR to
be close to that for standard riblets.

5.3. Flow description

Instantaneous streamwise velocity: To understand the DR results presented in the previous
subsection, a detailed flow analysis is indispensable. Figure 6 shows three cross sections
of the instantaneous streamwise velocity. The V-shaped contours in the horizontal cross
section reveal the presence of the herringbone texture. The plumes that appear in the yz-
plane represent up- or downdrafts that result from the converging/diverging riblets in the
texture.
Figure 7 shows part of the mean streamwise-averaged flow fields for textures with α =
Downloaded by [191.96.251.147] at 03:47 11 December 2017

165° and Ngroove = 1, 4, 16. Because of flow symmetry, the figures show only one feather
half. The feather shaft is located at the left side and the feather edge at the right side of the
figures.
α = 165◦ : For textures with α = 165°, the flow near the shaft converges, a local updraft
of fluid results, and an approximately square vortex appears. The vortex does not reach
the channel centreline, probably because of the small misalignment of the riblets with the
mean flow (α = 165° is relatively close to 180°). For Ngroove  4, it is limited in vertical
extent to approximately z = 0.2, as can be seen for Ngroove = 4 and 16 in the figure. Because
of spanwise confinement, for small feather widths the vortex decreases in size. For Ngroove =
1, it only reaches z = 0.05. Secondary flow is thus suppressed for small spanwise roughness
spacings. This accords with the experimental finding that secondary flow disappears for
decreasing spanwise spacing of roughness elements [52].
While mainly one vortex constitutes the secondary flow for Ngroove = 1 and 4, ter-
tiary flows appear for larger feather widths. For example, the mean flow for Ngroove =
16 (see Figure 7) shows a counterclockwise-rotating flow that extends to the channel
centreline. Less pronounced is the small clockwise-rotating vortex near the feather edge.
Several tertiary flows were also observed for even wider feathers (Ngroove = 32, 128). This

Figure . Three cross sections of the instantaneous streamwise velocity in the simulation domain that
belongs to the herringbone texture with Ngroove = , α = °. The horizontal plane is located at z =
. or z/h = ..
 JOURNAL OF TURBULENCE 731
Downloaded by [191.96.251.147] at 03:47 11 December 2017

Figure . Streamwise-averaged mean flow in a plane perpendicular to the streamwise direction for her-
ringbone textures with α = ° and Ngroove =  (top left),  (top centre and right),  (bottom). Vectors
exhibit in-plane secondary flow. Contours represent streamwise velocity (top left, top centre, bottom) or
streamwise velocity fluctuations (top right, urms = (u2 − u2 )1/2 ).

agrees with the experimental finding that tertiary flows appear when the spanwise spacing
of roughness elements increases above the boundary layer thickness [52]. These tertiary
flows are likely similar to the secondary flows that form over streamwise-aligned rough-
ness strips [83]. The latter are stress-induced (i.e. Prandtl’s secondary flows of the second
kind), as opposed to the flow-curvature-induced secondary flows (i.e. Prandtl’s secondary
flows of the first kind).
α = 15◦ : Reversing the flow direction from backward to forward also changes the sec-
ondary flow direction. For textures with α = 15°, riblets near the shaft diverge, resulting
in a local downdraft of fluid and the appearance of a counterclockwise-rotating secondary
flow to the right of the shaft. A tertiary flow again appears for Ngroove  16. However, this
time it is less well defined, as its rotation direction is also counterclockwise. So, the most
dominant tertiary flow does not change rotation direction by flow reversal.
 732 H. O. G. BENSCHOP AND W.-P. BREUGEM

Shifted variant: Shifting of the top-wall textures changes the mean-flow patterns only
slightly. For all cases, the dominant vortex near the top wall shifts in the spanwise direc-
tion. As this vortex does not reach the channel centreline, it has no noticeable effect on the
mean flow in the bottom channel half. For Ngroove = 1 and 4, this spanwise shift is the only
change of the mean flow. For Ngroove = 16, however, the pronounced counterclockwise-
rotating tertiary flow is also modified. It reaches the centreline for the not-shifted texture
(see Figure 7), but it occupies the whole channel for the shifted variant. It is almost square,
as for Ngroove = 16 the feather half-width (f /2 = 0.80) approximately equals the channel
height (Lz = 1). Tertiary flow for Ngroove = 32 also extends from bottom to top wall. How-
ever, these modified tertiary flows apparently have not much influence on drag (considering
Figure 5).
Spanwise modulation: The strong secondary-flow vortex near the shaft causes a span-
wise modulation of the boundary layer, as is clear from the streamwise-velocity contours
in Figure 7. Regions of updrafts (downdrafts) are characterised by low (high) streamwise
Downloaded by [191.96.251.147] at 03:47 11 December 2017

velocity. The trend for streamwise velocity fluctuations is opposite, as the top-right subfig-
ure shows: updrafts (downdrafts) are associated with increased (decreased) fluctuations.
The same was found in experimental studies of convergent/divergent riblets [41,42] and in
DNSs of turbulent boundary layers with uniform blowing/suction [3].
Instantaneous vortical structures and wall shear: Figure 8 shows the instantaneous vor-
tical structures and wall shear for three cases, namely smooth walls, parallel riblets in yaw
and herringbone riblets. The vortical structures are iso-surfaces of the second invariant of
the velocity gradient tensor (the Q-criterion, see e.g. [84]). The contours represent the shear
stress τw+ on the bottom wall.
The smooth-wall plot shows some well-known features, such as low-speed streaks and
hairpin-type vortical structures. The streaks are much less apparent in the second plot,
which is attributed to the parallel riblets that adjust the turbulence. The streaks are more
evident in the flow above the riblets. The vortices are comparable to the ones for the smooth
wall. The plot for parallel riblets without yaw is not shown, as it is very similar to the one
for parallel riblets in yaw.
The herringbone riblet texture exhibits the largest changes in vortical structures and
shear. High shear is associated with diverging riblets, and low shear with converging riblets.
The vortices seem to be ordered as well: they are abundant over regions with updrafts, but
almost absent over regions with downdrafts. The same trend was again found for blow-
ing/suction: vortices are enhanced by blowing in spite of the reduced wall shear stress, while
vortices are suppressed by suction despite the increase of wall shear stress [3].

5.4. Secondary flow strength

To quantify the strengths of the secondary flows described in the previous subsection, the
secondary flow strength s is introduced:

 
s = v 2 + w2 = 2
m + t,
2


m = v 2 + w2 , (12)

t = (v )2 + (w )2 ,
 JOURNAL OF TURBULENCE 733
Downloaded by [191.96.251.147] at 03:47 11 December 2017

Figure . Instantaneous vortical structures and wall shear in a part of the bottom channel half. Vortical
structures are iso-surfaces of the second invariant of the velocity gradient tensor with Q+ = .. The
coloured contours represent the wall shear stress τw+ on the bottom wall. Three cases are shown (from
top to bottom): smooth walls; parallel riblets in yaw of α = °; herringbone riblets with Ngroove = , α =
°. Reb =  for all cases and s+ =  for all textures.
 734 H. O. G. BENSCHOP AND W.-P. BREUGEM

0.015 0.020 Ngroove = 1

α = 0° Ngroove = 4
α = 15°
0.012 α = 165° 0.016 Ngroove = 16
α = 165° shift Ngroove = 128

0.009 0.012

xz
α = 15° α = 165°
Γm V

Γm
0.006 0.008

0.003 0.004

0 0
1 2 4 8 16 32 64 128 −0.4 −0.2 0 0.2 0.4
0.04 0.04 Ngroove = 1
α = 0° Ngroove = 4
α = 15°
Ngroove = 16

xz
0.03 α = 165° 0.03
V

α = 165° shift N groove = 128

− Γsmooth
− Γsmooth

0.02 0.02
t

t
Downloaded by [191.96.251.147] at 03:47 11 December 2017

0.01 0.01
α = 15° α = 165°

xz
Γt V

0
Γt 0

−0.01 −0.01
1 2 4 8 16 32 64 128 −0.4 −0.2 0 0.2 0.4
Ngroove (y − ys )/Λf

Figure . Strength of mean and turbulent secondary flow for the herringbone riblet geometry. Top left:
Volume-averaged mean-secondary-flow-strength as function of feather width (specified by Ngroove ). The
data points on the right vertical axis represent the conventional parallel blade riblets with yaw angle
α = ° and α = ° (not visible with  m V = .). Top right: Streamwise- and wall-normal-averaged
mean-secondary-flow-strength as function of spanwise distance for textures with α = ° (left part) and
α = ° (right part). Bottom left: turbulent-secondary-flow-strength plotted as in top-left subfigure. Bot-
tom right: turbulent-secondary-flow-strength plotted as in top-right
 subfigure.
  In the bottom figures, the
turbulent-secondary-flow-strength of the smooth wall tsmooth V = tsmooth xz = 0.069 is subtracted.

with v = v − v and w = w − w. The secondary flow strength is decomposed into the

mean-secondary-flow strength m and the turbulent-secondary-flow strength t . These
variables still depend on the spatial coordinates. A volume- and xz-average of m and t is
shown in Figure 9.
Influence of feather width: Figure 9 (left part) shows that both mean and turbulent sec-
ondary flow become stronger with decreasing feather width. This is attributed to the con-
verging or diverging riblets that trigger updrafts and downdrafts. When the feather width
decreases, the spanwise density of these texture-generated secondary flows increases. In
other words, the secondary flows that are created near the feather shaft and edges cover a
relatively larger portion of the total fluid volume. This yields a stronger volume-averaged
secondary flow for smaller feather widths.
Mean and turbulent contribution: The mean- and turbulent-secondary-flow strengths
follow the same trend for most textures (namely an increase with a decrease of f ). For
all Ngroove , the converging or diverging riblets near the shaft trigger updrafts and down-
drafts. As Figure 6 demonstrates, these are not steady flow patterns. Instead, they can be
considered as fluctuating ejections and sweeps that are generated by the texture. The result-
ing instantaneous secondary flow (v 2 + w2 ) projects both onto m and t (see Equation
(12)). A stronger instantaneous secondary flow yields in general an increase of both the
 JOURNAL OF TURBULENCE 735

mean and turbulent secondary flow. That explains why the mean and turbulent strength
follow the same trend for Ngroove  4. The results for Ngroove = 1, however, deviate in this
respect:  m V more than halves as compared to Ngroove = 4. This suppression of mean sec-
ondary flow is due to spanwise confinement (see previous subsection). However, the fluctu-
ating updrafts and downdrafts are not suppressed, so the turbulent secondary flow remains
strong.
Change with spanwise distance: Figure 9 (top right) shows how m varies with spanwise
distance. The mean secondary flow is clearly strongest near the shaft, which is due to the
counter-rotating vortices that form there. Such vortices are also generated near the feather
edges, but the riblets have a small height there, which yields only a relatively weak secondary
flow. Compared to α = 165°, textures with α = 15° have a stronger mean secondary flow
at the shaft, which is probably due to the downdraft of high-momentum fluid there.
Figure 9 (bottom right) shows how t varies with spanwise distance. The relatively uni-
form turbulence for the textures with Ngroove = 1, 4 shows that the flow is well mixed. In
Downloaded by [191.96.251.147] at 03:47 11 December 2017

contrast, the turbulence changes with y for the other textures. When approaching the shaft,
turbulence reduces when α = 15° and increases when α = 165°. The same was observed for
the streamwise velocity fluctuations in the previous subsection. When compared with the
smooth wall, however, the turbulence at the shaft is increased for all textures. Away from
the shaft, three textures exhibit local turbulence reductions.
Variants: Figure 9 also shows the volume-averaged secondary flow strength for the
shifted texture variation. The turbulent contribution is practically the same as for the not-
shifted texture. The same applies to the mean contribution for Ngroove = 1 and 4. This agrees
with the observation that, apart from the spanwise shift of the main vortex near the top wall,
the mean secondary flow does not change. For Ngroove = 16 and 32, however, the mean
secondary flow is stronger for the shifted texture, which is ascribed to the formation of a
tertiary flow that extends from bottom to top wall.
The secondary flow for the texture variation with α = 0° is much weaker than for the
herringbone riblet geometries, which is attributed to alignment of the riblets with the mean
flow. Both  m V and  t V are still larger than the values for the parallel riblet geometry
with constant blade height, although they seem to approach those values in the limit of
large Ngroove . The figure shows that parallel riblet geometries can reduce turbulence.

5.5. Spanwise transport of streamwise momentum

To confirm that advective transport is most important around the shaft, this subsection con-
siders the streamwise momentum balance as function of spanwise distance. The Navier–
Stokes equation for streamwise momentum (see Equation (1)) is rewritten as u/t = f,
where f is one of the terms in that equation. A time, streamwise and wall-normal average
(denoted by  f xz ) is applied to this equation. Assuming statistical stationarity and using
the boundary conditions, the balance becomes:

         
∂uv 1 ∂ 2u 1 ∂ 2u
0 =− + + + γ f f driving + f IBM . (13)
∂y Reb ∂y2 Reb ∂z2 x
  xz   xz   xz   xz   xz
driving IBM
adv y dif y shear
 736 H. O. G. BENSCHOP AND W.-P. BREUGEM

Ngroove = 4, α = 15° Ngroove = 128, α = 15° No yaw

2.0 adv y
dif y
shear
driving
1.0
smooth

IBM
/dtot

0
xz
f

-1.0

-2.0

-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4 -5 0 5

(y − ys )/Λf (y − ys )/Λf (y − yb )+
Ngroove = 4, α = 165° Ngroove = 128, α = 165° Yaw, α = 15°

2.0
Downloaded by [191.96.251.147] at 03:47 11 December 2017

1.0
/dsmooth
tot

0
xz
f

-1.0

-2.0

-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4 -5 0 5

(y − ys )/Λf (y − ys )/Λf y+

Figure . Streamwise momentum balance as function of the spanwise coordinate (Equation ()). Six
textures are shown: narrow herringbone feathers (Ngroove = ) for α = ° (top left) and α = ° (bottom
left); wide herringbone feathers (Ngroove = ) for α = ° (top centre) and α = ° (bottom centre);
parallel riblets aligned with the mean flow (top right) and in yaw of α = ° (bottom right). The blade
location is denoted by yb . For all cases, Reb =  and s+ = .

The terms in this equation only depend on the spanwise coordinate y. The third term is
called ‘shear’, since it equals the sum of bottom and top-wall shear as a result of the wall-
tot
normal integration. All terms are divided by dsmooth and shown in Figure 10 for four her-
ringbone textures. The abscissa represents the spanwise distance to the feather shaft. Most
curves exhibit approximately equidistant wiggles that result from staircasing: the riblet
height increases in 17 steps of size zw from 0 at the feather edges to h at the shaft.
Before the differences between Ngroove = 4 and 128 are highlighted, the general behaviour
of the different terms is clarified. The driving term is almost constant with y, because γ f
changes only marginally from the feather shaft to the edges. For Ngroove = 4, this term is
clearly larger than 1, indicating a significant drag increase.
The shear term is negative, as it tends to decelerate the fluid. Its magnitude quantifies
how the flat-wall streamwise shear changes with the spanwise coordinate. When moving
from the feather edges towards the shaft, the wall-shear term first decreases (in magnitude),
which is attributed to shielding of the flat wall by blades which increase in height. This trend
continues for textures with α = 165°. However, the shear magnitude peaks near the shaft for
α = 15°, which is ascribed to the local downdraft that transports high momentum towards
 JOURNAL OF TURBULENCE 737

the wall. For very wide feathers (Ngroove = 128), the shear term equals −1 at the feather
edges. There the blade height is zero and the smooth-wall result is recovered.
The IBM term represents the streamwise drag force on the riblets. It is negative, as it
is responsible for a velocity decrease (like the shear term). It equals zero at the feather
edges, because the riblets have no height there. When approaching the feather shaft from
the edges, the IBM-force magnitude first increases, which is due to riblet-height increase.
Near the shaft, it has a local minimum for α = 165°. In contrast, it is very large there for
α = 15°, likely due to the downdraft of high-speed fluid.
The balance for Ngroove = 4 shows significant contributions from spanwise advective
and diffusive transport, especially near the shaft. For α = 15°, advection is on average
responsible for an increase of streamwise momentum near the shaft. This is attributed to
the secondary-flow vortex that transports low momentum away from the shaft (near the
wall) and high momentum towards the shaft (closer to the channel centreline). In contrast,
diffusion transports high momentum away from the shaft. These trends are opposite to that
Downloaded by [191.96.251.147] at 03:47 11 December 2017

of textures with α = 165°, for which spanwise advection causes a streamwise momentum
decrease and diffusion an increase near the shaft.
Compared to Ngroove = 4, the relative importance of the terms in Equation (13) is very
different for Ngroove = 128. The prominent peaks near the shaft are very narrow. Spanwise
transport by advection and diffusion is close to zero for the largest part of the feather. Away
from the shaft, the texture behaves as parallel riblets in yaw with a local balance between
the driving force on the one hand, and the IBM and wall-shear force on the other hand.

5.6. Wall-normal transport of streamwise momentum

The previous subsections show that drag increase is accompanied by a strong secondary
flow, which suggests that enhanced advection is responsible for the drag augmentation. To
underpin this suggestion, this subsection considers the streamwise momentum balance as
function of the wall-normal coordinate. A time, streamwise, and spanwise average (denoted
by  f xy ) is applied to the Navier–Stokes equation u/t = f. Assuming statistical station-
arity and using the boundary conditions, the balance reads:
      
∂uw 1 ∂ 2u
0=− + + γ f f driving + f IBM . (14)
∂z xy Reb ∂z2 xy xy
x
xy
           
driving IBM
adv z dif z

The terms in this equation only depend on the wall-normal coordinate z. All terms are
tot
divided by dsmooth and displayed in Figure 11 for four cases. The profiles for parallel riblets
without yaw (not shown) are very similar to that for riblets in yaw. In addition, the profiles
for herringbone textures with α = 15° (not shown) are similar to the ones for α = 165°.
The behaviour of the different terms is clarified below. The driving term is almost constant,
as before.
The IBM term represents the drag force on the texture, so it is only present near the
wall and it is responsible for a velocity decrease. The drag force is especially large near the
blade tips. That explains the peak at blade height for riblets in yaw, for which all blades
have the same height. The large IBM force near the blade tips is in Figure 11 not evident for
herringbone textures. Instead, the peak is smeared out due to the spanwise blade-height
 738 H. O. G. BENSCHOP AND W.-P. BREUGEM

Smooth Yaw, α = 15° Ngroove = 4, α = 165° Ngroove = 128, α = 165°

0.25
adv z
dif z
0.20 driving
IBM

0.15
z

0.10

0.05

0
-20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20
f xy
/dtot
smooth f xy
/dtot
smooth f xy
/dtot
smooth f xy
/dtot
smooth

Figure . Streamwise momentum balance as function of the wall-normal coordinate (Equation ()). Four
cases are shown (from left to right): smooth walls; parallel riblets in yaw of α = °; narrow herringbone
feathers (Ngroove = ) for α = °; wide herringbone feathers (Ngroove = ) for α = °. Reb =  for
Downloaded by [191.96.251.147] at 03:47 11 December 2017

all cases and s+ =  for all textures.

variation between 0 and h (see Equation (11)). For the two herringbone textures shown
in the figure, the narrower feather clearly experiences a larger IBM force than the wider
feather.
Diffusive transport is especially important near the wall. It has a negative tendency for
smooth walls, but for textured walls it becomes positive in between the riblets. The latter is
associated with an inflection point in the mean streamwise-velocity profile.
Advective transport is significant throughout the whole channel. Near the channel cen-
treline, it balances the driving force (for all cases). Near the wall, turbulent advective trans-
port is responsible for a velocity increase and the associated drag augmentation. Parallel
riblets suppress turbulent advection, as is clear from the second sub-figure. Compared to
smooth walls, the peak of advection shifts upwards and shrinks. The herringbone case with
Ngroove = 128 shows the same trend, although the upward shift is less pronounced. Turbu-
lent transport below z = h is less suppressed as compared to parallel riblets in yaw, which is
ascribed to the blade height decrease with spanwise distance to the shaft. Still, weakening
and lifting of advective transport is quite apparent. In contrast, advective transport is much
stronger for Ngroove = 4 as compared to a smooth wall. This reinforces the suggestion that
drag augmentation is caused by enhanced advection, which will be confirmed in the next
section with a quantitative analysis.

6. Drag change decomposition

6.1. Previous research

Fukagata et al. [85] derived an equation (called Fukagata-Iwamoto-Kasagi (FIK)-identity)
that decomposes the frictional-drag coefficient into contributions from different dynam-
ical effects. They distinguished four contributions, namely laminar (or bulk), turbulent,
 JOURNAL OF TURBULENCE 739

inhomogeneous and transient. For homogeneous and steady plane channel flow, the FIK-
identity reads (in our notation):

1 driving 1 1
1  
f = + − z −u w xy dz, (15)
12 Reb 0 2

where it is used that the domain height Lz = 1. Note that the skin-friction coefficient C f ≡
 
2τw∗ / ρ ∗Ub∗2 = f driving for smooth-wall channel flow.
This relation has been extended to geometrically more complex surfaces by Peet
and Sagaut [86]. They derived analytical relations for streamwise-, spanwise- and
quasistreamwise-homogeneous surfaces. To our knowledge, however, these relations can-
not be used for the herringbone textures. That asks for a second extension of the FIK-
identity.
Downloaded by [191.96.251.147] at 03:47 11 December 2017

6.2. Derivation of extended FIK-identity

The extended FIK-identity is derived from the slab-averaged streamwise momentum bal-
ance (Equation (14)). For ease of notation, that equation is written as 0 = l f (l) |z , where f (l)
represents one of the terms in that equation. The subscript z expresses that f (l) only depends
on the wall-normal coordinate. In what follows, the
 global
 1 (l)drag
 balance is needed. It follows

from a single integration of Equation (14): 0 = l 0 f z dz or
 
d tot = γ f f driving
 V      
1 ∂u  1 ∂u  (16)
= − − f IBM .
Reb ∂z z=0 xy Reb ∂z z=1 xy x
V

Next, Equation (14) is rewritten as 0 = l (1/2)(f (l) |z + f (l) |1 − z ) to explicitly account for
symmetry in the mean flow. Triple integration is applied to this equation, such that the
extended FIK-identity in condensed form reads:

  1  z  z 1    
0= f (l) ẑ + f (l) 1−ẑ dẑ d
z dz. (17)
l 0 0 0 2

Using
1 the boundary conditions at the channel walls, the definition of the bulk velocity (i.e.
0
u xy dz = 1), integration by parts to transform multiple to single integrations and the
global drag balance (i.e. Equation (16)), the last equation becomes:
 ! "
1 1  
f driving z (1 − z) γ f xy dz =
0 2
 1 !  1 !  (18)
1 1 1
+ − z −uwxy dz + z (1 − z) − fxIBM dz.
Reb 0 2 0 2 xy

1
To arrive at an equation for d tot = γ f f driving V = f driving 0 γ f xy dz, Equation (18) is
1
divided by the prefactor in square brackets and multiplied by 0 γ f xy dz, which yields the
 740 H. O. G. BENSCHOP AND W.-P. BREUGEM

final extended FIK-identity:

d tot = d bulk + d mean advection + d turbulent advection + d IBM


= d (l) . (19)
l

Notice that the total advection term is split up into contributions from mean and turbu-
lent flow, using that uw = u w + u w . The symmetry or antisymmetry with respect to the
channel centreline at z = 1/2 is evident from the factors in braces in Equation (18). As the
multiplier of f driving in that equation depends on γ f , the bulk term dbulk is slightly texture
dependent, as was also found by Peet and Sagaut [86]. For smooth walls, γ f = 1 and that
multiplier equals 1/12, such that Equation (15) is recovered.
As Equation (19) applies to both smooth and textured walls, the drag change can be
Downloaded by [191.96.251.147] at 03:47 11 December 2017

decomposed in a similar way:

d tot − dsmooth
tot
DC = tot
dsmooth
# $ (20)
 d (l) − d (l)  (l)
= tot
smooth
≡ dc .
l
dsmooth l

The term within braces is abbreviated as dc(l) . It represents the change of a certain term for
textured walls as compared to that term for smooth walls. This decomposition quantifies
which terms contribute to DR or increase.
The identity presented here is somewhat different from the one derived by Peet and
Sagaut [86]. First, their decomposition only applies to skin friction, whereas the drag
decomposition in Equation (19) also includes the pressure drag. Second, the IBM-term is
not present in their identity. They used a body-fitted coordinate system in their derivation.
As a result, the shear stress on the texture directly derives from integration of the viscous
diffusion term. That approach has the added advantage that the skin-friction coefficient
for simple textures in a laminar flow can be computed exactly based on purely geometri-
cal considerations without performing the flow calculations [86]. However, their relation
applies to quasi-homogeneous surfaces only and adopts a more complicated integration
using a body-fitted grid. In contrast, the Cartesian integration that is employed here is not
restricted to certain geometries.
The extended FIK-identity (Equation (19)) is not only useful when an IBM is used. In
the present work, fiIBM is a body force that models the shear and pressure forces that the
texture exerts on the flow. However, the FIK-identity applies to any body force. Further-
more, the current drag decomposition is also applicable to body-fitted calculations. In that
case, the obstacles should be considered as part of the domain, because the identity is based
on integration over the entire rectangular channel volume. Three steps are required for a
successful use of the decomposition in this situation. (1) The geometry should be trans-
lated into a 3D phase-indicator function γ f . (2) A zero-flow condition should be used for
the obstacle volume. (3) The drag force on the obstacle surface should be translated into a
3D body force or IBM force.
 JOURNAL OF TURBULENCE 741

Table . Contributions to the total skin friction for smooth-wall turbulent channel flow.
The terms are as given in Equation () and divided by dtot . The rest term equalises the
left-hand and right-hand side of that equation.
Reb dtot (%) dbulk (%) dmean advection (%) dturbulent advection (%) dIBM (%) drest (%)
  . . · − .  − .
  . . · − .  .
  . .· − .  .

In view of the factors between braces in Equation (18), advection and the IBM force
contribute differently to the total drag. The weighing factor for advection is largest near the
wall, so significant advective transport near the wall contributes most to drag augmentation.
In contrast, the weighing factor for the IBM force is largest near the channel centreline.
Obstructing the flow there is for two reasons more detrimental than an obstruction near
the wall: drag increases due to a larger flow velocity (so a larger IBM force) and a larger
weighing factor.
Downloaded by [191.96.251.147] at 03:47 11 December 2017

6.3. Smooth walls

Table 2 shows the decomposition of the total drag into the different contributions for
smooth-wall turbulent channel flow. The results for three Reynolds numbers are shown.
The total drag is taken as a reference. The rest term equalises the left-hand and right-hand
side of Equation (19). It results from the finite simulation time. It becomes slightly larger
for higher Reb , which is attributed to a relatively shorter simulation time. Its magnitude is
typical for all the other simulations.
As the table shows, only bulk transport and turbulent advection contribute to the total
drag in smooth-wall channel flows. The mean advection term is zero to machine precision,
and dIBM = 0. At Reb = 5500, the total drag comes for 27% from the bulk and for 73%
from the turbulent advection term. These numbers are close to the 26% and 74% reported
previously for Reb = 5460 [86]. As the Reynolds number increases, the relative contribution
of turbulence to the total drag increases. This stresses once more the need for reduction of
turbulent drag in high-Reynolds-number flows.

6.4. Parallel riblets

Using Equation (20), the drag change for parallel riblets is decomposed into contributions
from the bulk, mean advection, turbulent advection and IBM terms. Figure 12 exhibits this
decomposition as function of s+ and α in a bar graph. The total term equals the total drag
change, which was shown already in Figure 2. The bulk term is in both graphs not visible,
as it is very close to zero.
Change with blade spacing: The IBM term, which represents the drag force on the blades,
is responsible for an increase of the total drag. Apart from being positive, it also increases
with s+ . When the riblet spacing increases, the riblet height increases as well (h/s = 0.5
is fixed), but the total blade area per unit spanwise width remains constant. Therefore, the
increase of the IBM term with s+ cannot be due to an increased wetted area, but it is ascribed
to blades that further protrude into the flow. Although it remains small as compared to the
other terms, the contribution from mean advection also increases with blade spacing. That
is attributed to a stronger mean secondary flow for larger s+ (not shown here, see also e.g.
 742 H. O. G. BENSCHOP AND W.-P. BREUGEM

No yaw Yaw
20 20
total total
15 bulk 15 bulk
mean advection mean advection
turbulent advection turbulent advection
10 10
IBM IBM
5 5
dc [%]

0 0

−5 −5

−10 −10

−15 −15

−20 −20
9.9 16.9 23.9 0 10 15 20
s + α [°]

Figure . Different contributions to the total drag change (Equation ()) for parallel blade riblets. Note
that five bars belong to only one abscissa. Left: Decomposition as function of riblet spacing in wall units
Downloaded by [191.96.251.147] at 03:47 11 December 2017

for riblets aligned with the mean flow (α = °). Right: Decomposition as function of yaw angle for fixed
riblet spacing s+ = .
[61,87]). Finally, the turbulent drag contribution is negative, which indicates suppression of
turbulent streamwise momentum transport. For the three s+ values shown here, the max-
imum turbulent-drag suppression is almost 15% at s+ = 17. The figure demonstrates that
the optimum s+ is a trade-off between an additional drag force on the blades and reduced
turbulent transport because of the blades.
Change with yaw angle: Figure 12 shows a similar decomposition for parallel riblets in
yaw at fixed s+ = 17. The IBM term increases for increasing yaw angle α, which might be
due to the additional pressure drag. The change in the mean advection term is very small.
The contribution from turbulent advection is increasingly less negative when α increases.
The figure thus indicates that deterioration of riblets in yaw is both due to an increased drag
force on the blades and reduced suppression of turbulent transport.

6.5. Herringbone riblets

Drag decomposition: Figure 13 shows the decomposition of the total drag change into its
contributions from the IBM force, turbulent advection and mean advection. The bulk and
rest terms are not presented. They never exceed 0.04% (bulk) and 0.08% (rest) in magni-
tude. The top left graph is the same as Figure 5 and is shown here again for ease of com-
parison. Apart from the bulk contribution, the IBM term is the least important term for
most textures. In the worst case, it causes an 8.1% increase of drag, which is comparable to
results for standard parallel riblets (7.0% for α = 0°, 8.6% for α = 15°). Advection by the
mean flow can be a significant contribution to drag increase, up to 28% for Ngroove = 4, α
= 165°. The turbulent advection term exhibits the largest changes: it varies between a 6%
reduction and a 50% increase.
The variation of the mean and turbulent drag contributions with feather width very
much resembles that of the mean and turbulent secondary flow strengths (compare
Figure 13 and Figure 9). It thus appears that advection and secondary flow go hand in hand.
The same can be observed from the instantaneous flow field in Figure 14, which shows both
contours of advection and vectors of secondary flow. Regions with strong secondary flow
(in particular strong w) are also regions with strong advection. Both secondary flow and
 JOURNAL OF TURBULENCE 743

total IBM
80 60
α= 0° α = 0°
α= 15° α = 15°
α= 165° α = 165°
60 α= 165° shift α = 165° shift
40

40
dc [%]

20
20

0
0

−20 −20
1 2 4 8 16 32 64 128 1 2 4 8 16 32 64 128

turbulent advection mean advection

60 60
α= 0° α = 0°
α= 15° α = 15°
α= 165° α = 165°
α= 165° shift α = 165° shift
Downloaded by [191.96.251.147] at 03:47 11 December 2017

40 40
dc [%]

20 20

0 0

−20 −20
1 2 4 8 16 32 64 128 1 2 4 8 16 32 64 128
Ngroove Ngroove

Figure . Decomposition of the total drag change (top left) for herringbone riblets in contributions from
the IBM-force (top right), turbulent advection (bottom left) and mean advection (bottom right) according
to Equation (). All figures show the contribution change relative to a smooth wall as function of feather
width. The data points on the right vertical axes represent the conventional parallel blade riblets with
yaw angle α = ° and α = °.

the advective flux are split into mean and turbulent contributions in view of s2 = m2 + t2
and uw = u w + u w , respectively. In summary, the results clearly indicate an intimate
connection between increased advective transport and stronger secondary flow.
Given the close correspondence between secondary flow and advection, the change of
the secondary flow strength with feather width as clarified in Subsection 5.4 also explains
the trend of the advective drag contributions. In particular, the increased advective drag for
smaller feather widths is due to the higher spanwise density of the converging or diverging
riblets that generate the secondary flows. Also, reduction of the mean advective drag for
Ngroove = 1 is ascribed to a weaker mean secondary flow due to spanwise confinement.
There is, however, one major difference between Figure 13 and Figure 9, namely the effect
of shifting of the top-wall texture. For 16 and 32 grooves, the mean secondary flow for the
shifted textures is clearly stronger than for the not-shifted textures. In contrast, the drag due
to mean advection is about the same, independent of the shift. This is explained by the fact
that a stronger secondary flow near the channel centreline does not contribute much to drag
because of the factor (1/2 − z) in Equation (18). This demonstrates that mean-secondary-
flow strength is a good indicator for the contribution of mean advection to drag, provided
that the mean secondary flow near the centreline is weak.
 744 H. O. G. BENSCHOP AND W.-P. BREUGEM
Downloaded by [191.96.251.147] at 03:47 11 December 2017

Figure . Advective transport in a plane perpendicular to the streamwise direction for the herringbone
texture with Ngroove = , α = °. Contours represent the instantaneous advection uw (left), the averaged
mean advection (centre) or the averaged turbulent advection (right). Vectors depict the instantaneous
(left) or averaged (centre, right) in-plane flow velocity.

Spanwise decomposition of drag change: Figure 13 shows that DRs are possible in the
limit of large Ngroove . Like for the conventional riblet texture, these reductions originate
from weakened turbulent advective transport. To reveal the origin of this weakening, the
spanwise dependence of the advective FIK-terms is investigated with a spanwise decompo-
sition. Let  f xy be one term in Equation (14) and d(l) the corresponding drag contribution.
The calculation of this FIK-term can be rewritten as:
 1  
(l)
d = g f dz
0 xy
 1      (21)
(l)
= g f dz ≡ dsp .
0 x y y

The function g = g(z) results from conversion of a triple to a single integral, and normalisa-
(l) (l)
tion. The function dsp = dsp (y) represents the spanwise decomposition of the FIK-term,
as indicated by the subscript sp. The drag change contribution dc(l) can be decomposed in
a similar way:

(l)
d (l) − dsmooth
dc(l) =
d tot
 smooth(l)  (22)
(l)
dsp − dsmooth  
(l)
= tot
≡ dcsp ,
dsmooth y
y

(l)
where dcsp quantifies how the drag change depends on the spanwise coordinate. Note that
(l) (l) (l) tot
dsp and dcsp have a spanwise dependence, whereas dsmooth and dsmooth have not. The span-
wise decomposition of the mean and turbulent advective terms is shown in Figure 15 for
α = 15° and 165°, and four feather widths. Small asymmetries with respect to y = ys are
 JOURNAL OF TURBULENCE 745

α = 15° , mean advection α = 165° , mean advection

800
Ngroove = 1 (0.5%) Ngroove = 1 (6.4%)
Ngroove = 4 (14%) 400 Ngroove = 4 (28%)
600 Ngroove = 16 (11%) Ngroove = 16 (5.6%)
Ngroove = 128 (2.1%) 200 Ngroove = 128 (1.5%)
dcsp [%]

400
0
200
−200
0
−400
−200
−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4

α = 15° , turbulent advection α = 165° , turbulent advection

100
Ngroove = 1 (46%) 100
Ngroove = 1 (50%)
Ngroove = 4 (51%) Ngroove = 4 (15%)
Ngroove = 16 (17%) Ngroove = 16 (−1.8%)
Downloaded by [191.96.251.147] at 03:47 11 December 2017

80 80
Ngroove = 128 (−3.4%) Ngroove = 128 (−6.6%)
60 60
dcsp [%]

40 40

20 20

0 0

−20 −20

−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4

(y − ys )/Λf (y − ys )/Λf

Figure . Spanwise decomposition of advective contributions to the total drag change (Equation ()).
The total advection term is split into a mean (top) and turbulent (bottom) part. Results for herringbone
textures with α = ° (left) and α = ° (right) are shown. The number in the legend quantifies dc, the
spanwise average of each curve (Equation ()), which was shown already in Figure .

attributed to slow convergence of weak secondary flows. The small oscillations in the curves
for Ngroove = 128 probably appear for the same reason.
The effect of convergent/divergent riblets can be isolated from that of parallel riblets in
yaw for the textures with Ngroove = 128. The parallel riblets dominate the flow in a region suf-
ficiently far away from the shaft. In that region, the mean advection term fluctuates around
zero. The turbulent term is zero at the feather edges. Riblets have no height there and the
smooth-wall result is recovered. Away from the edges, the term decreases more or less lin-
early, which is attributed to riblet-height increase and the associated suppression of tur-
bulent transport. It approaches approximately −10%, which belongs to full-height parallel
riblets with yaw angle α = 15° or 165°. These favourable trends of mean and turbulent
advective drag stop near the shaft because of the strong secondary flow there.
The converging/diverging riblets dominate the flow in a region around the shaft. That
is especially evident from the drag change due to mean advection, which shows a clear sig-
nature of the mean flow described before, in particular the strong secondary-flow vortices
near the shaft. For α = 165° and any Ngroove , the updraft of fluid around the shaft results
in a local DR, as is apparent from the dip in dcsp at y = ys . The peak next to this dip is
associated with that part of the vortex that transports momentum towards the wall. For
Ngroove = 16, a second dip appears, which is ascribed to the tertiary flow shown in Figure 7.
 746 H. O. G. BENSCHOP AND W.-P. BREUGEM

The central dip is narrower for larger Ngroove , because the secondary-flow vortex near the
shaft is smaller compared to the total feather width. A similar (but opposite) description
applies to textures with α = 15°.
Local turbulent DR can be obtained by diverging riblets. For α = 15° and Ngroove  16,
the strong wall-directed mean flow is accompanied by reduced turbulent transport. The
reverse is true for α = 165°, namely that the strong wall-leaving mean flow is accompa-
nied by increased wall-directed turbulent transport, as can also be seen from Figure 14.
These observations fully agree with the findings for uniform blowing or suction. In DNSs,
it has been found that uniform blowing reduces mean advective drag and enhances tur-
bulent drag, while uniform suction enhances mean advective drag and reduces turbulent
drag [3].
The local contribution of advection to drag might be very different from its global
(or volume-averaged) contribution. For instance, mean advection might seem much more
important than turbulent advection in view of the scales of Figures 14 and 15. However, the
Downloaded by [191.96.251.147] at 03:47 11 December 2017

volume-averaged turbulent term is often at least as important as the mean term. As a second
example, the texture with α = 165° and Ngroove = 4 exhibits locally a DR as high as 400%
due to a strong wall-leaving flow. However, one should realise that the favourable updraft of
low-momentum fluid is compensated by an adverse downdraft of high-momentum fluid.
The influence of the whole vortex on the drag should be considered. In the case of α = 165°
and Ngroove = 4, the vortex covers one feather half. The corresponding spanwise-averaged
mean advective drag is 28%, which establishes again an unfavourable effect of the vortices
near the shaft on the drag. Therefore, one should be careful to judge the performance of this
(or any) texture based on a local drag determination. The mean or turbulent advective drag
might be reduced locally. However, when the flow is dominated by strong advection (such
as near the shaft), the volume-averaged drag generally increases due to an overall increase
of both the mean and turbulent advective contributions.
In summary, the present study confirms two effects of the herringbone riblet texture on
the turbulent drag. The first effect relates to the texture-generated secondary flows around
the shaft. Although turbulent drag might be reduced locally due to a wall-directed mean
flow, the overall trend is an increased turbulent drag because of the fluctuating secondary
flows that are generated by the converging/diverging riblets near the shaft. The second effect
relates to the parallel riblets sufficiently far away from the shaft. Those riblets suppress tur-
bulent transport, which results in turbulent DR.

7. Discussion
Detrimental effect of convergent/divergent riblets: The results presented in this paper clarify
the influence of the herringbone riblets on the drag. The drag increase for f /Lz  O(1) is
attributed to an increase of advective transport (Figure 13), which in turn is associated with
the secondary-flow vortex near the shaft (Figure 15). Although the vortex might locally be
responsible for a DR, as a whole it is responsible for a drag increase. That vortex originates
from the herringbone riblets near the shaft (Figures 7, 9, 10). That strongly suggests that
these converging/diverging riblets are detrimental to DR. That is confirmed by the finding
that f →  is most beneficial for DR (Figure 5). In that limit, the herringbone texture
approaches the conventional parallel-riblet texture in yaw. Hence, the presence of conver-
gent/divergent riblets in the texture seems unfavourable for DR.
 JOURNAL OF TURBULENCE 747

Research on spanwise-periodic spanwise forcing: Research on the drag-reducing spanwise

forcing further underpins the detrimental effect of the herringbone riblets on the drag. The
spanwise travelling wave of spanwise body force is described by fy = A(z)sin {(2π/f )y −
ωt}. Here, fy represents the spanwise body force, A(z) the forcing amplitude (only nonzero
in the vicinity of the wall), f the spanwise wavelength of the forcing, and ω the angular
frequency of the forcing. The herringbone texture is similar to this forcing with respect
to its spanwise periodicity, though the texture is static (ω = 0). Research shows that the
best-performing spanwise travelling wave (in terms of DR or net energy saving) is the one
with infinite wavelength, i.e. the spanwise wall oscillation [88,89]. In other words, f →
 is most beneficial for DR, like for the herringbone texture. Hence, the addition of span-
wise periodicity seems detrimental to the drag-reducing performance of both the spanwise
forcing and the riblet texture in yaw.
Alternative passive riblet textures: As the herringbone texture with a spanwise variation
of the forcing appears to be unprofitable for DR, a streamwise variation of the forcing might
Downloaded by [191.96.251.147] at 03:47 11 December 2017

be more advantageous. Indeed, streamwise travelling waves of spanwise forcing are superior
to spanwise travelling waves of spanwise forcing, presumably because of the unaltered wall-
normal gradient of the wall-normal velocity component at the wall [89]. Both experimen-
tally and numerically, DR has been obtained with a stationary streamwise variation of span-
wise forcing [27,90]. DR by streamwise variation of wall-normal forcing (e.g. suction and
blowing) has been reported as well, although the net energy saving has been small [91,92].
The drag increasing/reducing trends in the current study are similar to what has been found
for uniform suction/blowing (Figures 8, 15), but a net drag-reducing effect could not be
confirmed. It is difficult (if not impossible) to eliminate the unfavourable effects of down-
drafts, as mass conservation dictates that updrafts need to be compensated by downdrafts.
Also, textures with converging/diverging riblets are not fully comparable to uniform blow-
ing/suction. For instance, uniform blowing originates from a nonzero mass flux through
the wall, in contrast to the texture-generated updrafts. Still, textures with streamwise vari-
ation of the forcing seem more promising than the herringbone texture.
Comparison with experiments: The conclusion that the herringbone texture seems detri-
mental to turbulent DR apparently contrasts with the experimental study of Chen et al.
[31]. Although it has not been the aim of the present study to reproduce their experiments
numerically, a comparison might still be illuminating. Chen et al. [31] obtained DR in for-
ward flow for f /D  1 (D being the pipe diameter), with a maximum of 20%. In con-
trast, the present study only achieved DR for f /Lz > 10, with a maximum of 2% in back-
ward flow. These contrasting results might be ascribed to differences in riblet texture and
Reynolds number.
First, the textures were different, particularly the riblet shape and angle. The feather
width might have been different too, but it is unclear what value for Ngroove was used in
the experiments. The experimental texture consisted of sawtooth riblets at an angle of 30°
with the flow direction. The numerical texture was composed of blades at an angle of 15°.
Although blades seem to be more sensitive to yaw (see Figure 2), they were studied at a
smaller yaw angle. Therefore, there is currently no clear indication that the use of blade
riblets in the numerical study contributed to a lesser drag-reducing performance of the
herringbone texture.
Second, the Reynolds numbers differed significantly. The 20% DR was obtained at
Reb  2.6 · 105 (based on pipe diameter and bulk velocity), while the numerical study
 748 H. O. G. BENSCHOP AND W.-P. BREUGEM

was performed at Reb = 5500. For DR techniques that rely on near-wall flow manipula-
tion, the attainable DR is approximately independent of the Reynolds number (although
not fully [44–47]) when near-wall scaling is applied. However, herringbone riblets (in con-
trast to conventional riblets) cannot be regarded as viscous-region modifiers, because they
generate a secondary flow throughout a large part of the channel (Figure 7). Therefore, vis-
cous scaling is probably inappropriate in this case. Because of the unknown scaling and
the Reynolds number difference, the texture parameters (e.g. f ) and the DR results of the
experiments and the simulations cannot be directly compared.

8. Conclusions and outlook

The drag-reducing performance of a herringbone riblet texture was studied with DNSs
of turbulent flow in a channel with height Lz . The FIK-identity for drag decomposition
Downloaded by [191.96.251.147] at 03:47 11 December 2017

was extended to textured walls and was used to study the underlying drag change mech-
anisms. For validation, simulations with smooth walls, parallel blade riblets and parallel
blade riblets in yaw were performed, which showed good agreement with literature. The
parallel-riblet simulations exhibited an expected but small low-Reynolds-number effect. A
maximum DR of 9.3% was obtained, close to the 9.9% that has been found experimentally
[21].
The herringbone texture can both increase or reduce the drag, depending on the span-
wise texture wavelength f . For f /Lz  O(1) (i.e. narrow feathers), the drag increases with
a maximum of 73% for f /Lz = 0.4. This increase is ascribed to the convergent/divergent
riblets. They generate a fluctuating secondary flow, which on average consists of two
counter-rotating vortices centred above the regions of riblet convergence/divergence. The
strong secondary flow increases both mean and turbulent advective transport, which in
turn results in the significant drag increase.
A slight DR of 2% was found for f /Lz  O(10) (wide feathers). Due to the large feather
width, the secondary flow generated by the converging/diverging riblets now influences
only a relatively small part of the whole texture. Its drag-increasing contribution is there-
fore small. The largest part of the texture behaves similarly to a conventional parallel-riblet
texture in yaw. Specifically, suppression of turbulent advective transport is responsible for
the small DR that was obtained.
As was found by other researchers for spanwise travelling waves of spanwise forcing, the
current study confirms that f →  is most beneficial for DR. In that limit, the texture
approaches the conventional parallel-riblet texture in yaw. Therefore, the presence of con-
vergent/divergent riblets in the texture seems detrimental to turbulent DR, which appar-
ently contrasts with the experiments of Chen et al. [31]. However, differences in Reynolds
number and texture parameters (riblet shape, feather width, angle between riblets and flow
direction) hindered a one-to-one comparison between the present simulations and the
experiments.
More elaborate experiments and simulations are required to further investigate the drag-
reducing potential of the herringbone texture. First, the maximum DR of 20% should be
reproduced and the optimum texture parameters (f , s, h/s, α) should be determined. In
numerical simulations, the more realistic sawtooth riblet geometry should be implemented.
The Reynolds-number influence and the parameter scaling also need further attention. As
 JOURNAL OF TURBULENCE 749

the herringbone riblets generate a secondary flow throughout a large part of the channel,
the scaling of DR with the texture and flow parameters is presumably non-trivial.
Whether feather riblets have an aerodynamic function remains an open question. Chen
et al. [31] claimed that feather riblets greatly impact flight performance because of DR. The
DR was ascribed to suppression of turbulent momentum transport, but evidence has been
inconclusive so far. One should realise that wings of birds are not flat and operate at a rela-
tively low Reynolds number, so flow-separation delay seems a more plausible aerodynamic
function of feather riblets. Indeed, several studies confirm that roughness on a bird wing
contributes to separation control [35–37]. Furthermore, separation delay has been obtained
with vortex generators that resemble the herringbone texture [33]. So, future studies might
investigate the potential of the herringbone texture for flow-separation control.

Acknowledgments
Downloaded by [191.96.251.147] at 03:47 11 December 2017

The research leading to these results has received funding from the European Union Seventh Frame-
work Programme in the SEAFRONT project under grant agreement nr. 614034. The simulations
were performed on the Dutch national supercomputer Cartesius at SURFsara, Amsterdam, The
Netherlands. This work was sponsored by NWO Physical Sciences for the use of supercomputer
facilities. H.O.G.B. would like to thank his colleague Pedro Costa for his supercomputational sup-
port.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This work was sponsored by the European Union Seventh Framework Programme in the
SEAFRONT project [grant agreement 614034]; and the Netherlands Organisation for Scientific
Research (NWO) [project number SH-023-15].

References
[1] White CM, Mungal MG. Mechanics and prediction of turbulent drag reduction with polymer
additives. Annu Rev Fluid Mech. 2008;40:235–256.
[2] Ceccio SL. Friction drag reduction of external flows with bubble and gas injection. Annu Rev
Fluid Mech. 2010;42:183–203.
[3] Kametani Y, Fukagata K. Direct numerical simulation of spatially developing turbulent bound-
ary layers with uniform blowing or suction. J Fluid Mech. 2011;681:154–172.
[4] Yoon HS, El-Samni OA, Chun HH. Drag reduction in turbulent channel flow with periodically
arrayed heating and cooling strips. Phys Fluids (1994-present). 2006;18(2):025104.
[5] Kametani Y, Fukagata K. Direct numerical simulation of spatially developing turbulent
boundary layer for skin friction drag reduction by wall surface-heating or cooling. J Turbul.
2012;13(34):1–20.
[6] Vakarelski IU, Chan DYC, Thoroddsen ST. Leidenfrost vapour layer moderation of the drag
crisis and trajectories of superhydrophobic and hydrophilic spheres falling in water. Soft Mat-
ter. 2014;10:5662–5668.
[7] Quadrio M. Drag reduction in turbulent boundary layers by in-plane wall motion. Phil Tran
R Soc A Mat Phys Eng Sci. 2011;369(1940):1428–1442.
 750 H. O. G. BENSCHOP AND W.-P. BREUGEM

[8] Tomiyama N, Fukagata K. Direct numerical simulation of drag reduction in a turbulent chan-
nel flow using spanwise traveling wave-like wall deformation. Phys Fluids (1994-present).
2013;25(10):105115.
[9] Shatrov V, Gerbeth G. Magnetohydrodynamic drag reduction and its efficiency. Phys Fluids
(1994-present). 2007;19(3):035109.
[10] Rothstein JP. Slip on superhydrophobic surfaces. Annu Rev Fluid Mech. 2010;42:89–109.
[11] Choi K-S, Yang X, Clayton BR, et al. Proc R Soc A Mat Phys Eng Sci. 1997;453(1965):2229–
2240.
[12] Dean B, Bhushan B. Shark-skin surfaces for fluid-drag reduction in turbulent flow: a review.
Phil Trans R Soc A: Mat Phys Eng Sci. 2010;368(1929):4775–4806.
[13] Fransson JHM, Talamelli A, Brandt L, et al. Delaying transition to turbulence by a passive
mechanism. Phys Rev Lett. 2006;96(6):064501.
[14] Choi Jin, Jeon WP, Choi H. Mechanism of drag reduction by dimples on a sphere. Phys Fluids
(1994-present). 2006;18(4):041702.
[15] Son K, Choi J, Jeon WP, et al. Mechanism of drag reduction by a surface trip wire on a sphere.
J Fluid Mech. 2011;672:411–427.
[16] Abe K, Matsumoto A, Munakata H, et al. Drag reduction by sand grain roughness. In: Gyr
Downloaded by [191.96.251.147] at 03:47 11 December 2017

A, editor. Structure of turbulence and drag reduction. Berlin Heidelberg: Springer; 1990.
p. 341–348.
[17] Wahidi R, Chakroun W, Al-Fahed S. The behavior of the skin-friction coefficient of a turbulent
boundary layer flow over a flat plate with differently configured transverse square grooves. Exp
Ther Fluid Sci. 2005;30(2):141–152.
[18] Abdulbari HA, Yunus RM, Abdurahman NH, Charles A. Going against the flow—A review of
non-additive means of drag reduction. J Ind Eng Chem. 2013;19(1):27–36.
[19] Sagong W, Kim C, Choi S, et al. Does the sailfish skin reduce the skin friction like the shark
skin? Phys Fluids (1994-present). 2008;20(10):101510.
[20] Díez G, Soto M, Blanco JM. Biological characterization of the skin of shortfin mako shark Isu-
rus oxyrinchus and preliminary study of the hydrodynamic behaviour through computational
fluid dynamics. J Fish Biol. 2015;87(1):123–137.
[21] Bechert DW, Bruse M, Hage W, et al. Experiments on drag-reducing surfaces and their opti-
mization with an adjustable geometry. J Fluid Mech. 1997;338(5):59–87.
[22] Wilkinson SP, Anders JB, Lazos BS, et al. Turbulent drag reduction research at NASA Langley:
progress and plans. Int J Heat Fluid Flow. 1988;9(3):266–277.
[23] Li W, Jessen W, Roggenkamp D, et al. Turbulent drag reduction by spanwise traveling ribbed
surface waves. Eur J Mech-B/Fluids. 2015;53:101–112.
[24] Wassen E., Kramer F, Thiele F, et al. Turbulent drag reduction by oscillating riblets. In Pro-
ceedings of the 4th AIAA Flow Control Conference, volume 2; Jun 23–26; Seattle, WA; 2008.
p. 740–754.
[25] Grüneberger R, Kramer F, Hage W, et al. Experimental investigation of oscillating riblets for
turbulent drag reduction. In: Dillmann A, Heller G, Kreplin H-P, Nitsche W, Peltzer I, edi-
tors. New results in numerical and experimental fluid mechanics VIII, volume 121 of notes on
numerical fluid mechanics and multidisciplinary design. Springer Berlin Heidelberg; 2013. p.
193–200. ISBN 978-3-642-35679-7.
[26] Vodop’yanov IS , Nikitin NV, Chernyshenko SI. Turbulent drag reduction by spanwise oscil-
lations of a ribbed surface. Fluid Dyn. 2013;48(4):461–470.
[27] Grüneberger R, Kramer F, Wassen E, et al. Influence of wave-like riblets on turbulent friction
drag. In: Tropea C, Bleckmann H, editors. Nature-inspired fluid mechanics. Berlin Heidelberg:
Springer; 2012. p. 311–329.
[28] Sasamori M, Mamori H, Iwamoto K, et al. Experimental study on drag-reduction effect due to
sinusoidal riblets in turbulent channel flow. Exp Fluids. 2014;55(10):1–14.
[29] Chen H, Che D, Zhang X, et al. UV grafting process for synthetic drag reduction of biomimetic
riblet surfaces. J Appl Poly Sci. 2015;132(33):1–8.
[30] Chen H, Rao F, Shang X, et al. Biomimetic drag reduction study on herringbone riblets of bird
feather. J Bionic Eng. 2013;10(3):341–349.
 JOURNAL OF TURBULENCE 751

[31] Chen H, Rao F, Shang X, et al. Flow over bio-inspired 3D herringbone wall riblets. Exp Fluids.
2014;55(3):1–7.
[32] Chen HW, Rao FG, Zhang DY, et al. Drag reduction study about bird feather herringbone
riblets. Appl Mech Mat. 2014; 461:201–205.
[33] Lin JC. Review of research on low-profile vortex generators to control boundary-layer separa-
tion. Prog Aerosp Sci. 2002;38(4):389–420.
[34] Shahinfar S, Sattarzadeh SS, Fransson JHM, et al. Revival of classical vortex generators now for
transition delay. Phys Rev Lett. 2012;109(7):074501.
[35] Bushnell DM, Moore KJ. Drag reduction in nature. Annu Rev Fluid Mech. 1991;23(1):65–79.
[36] Lilley G. A study of the silent flight of the owl. AIAA paper. 1998;2340(1998):l–6.
[37] van Bokhorst E, de Kat R, Elsinga GE, et al. Feather roughness reduces flow separation during
low Reynolds number glides of swifts. J Exp Biol. 2015;218(20):3179–3191. ISSN 0022-0949.
[38] Gao X, Sunden B. Heat transfer and pressure drop measurements in rib-roughened rectangular
ducts. Exp Therm Fluid Sci 2001;24(1):25–34.
[39] Fang X, Yang Z, Wang B-C, et al. Highly-disturbed turbulent flow in a square channel with
v-shaped ribs on one wall. Int J Heat Fluid Flow. 2015;56:182–197.
[40] Stroock AD, Dertinger SKW, Ajdari A, et al. Chaotic mixer for microchannels. Science
Downloaded by [191.96.251.147] at 03:47 11 December 2017

2002;295(5555):647–651.
[41] Koeltzsch K, Dinkelacker A, Grundmann R. Flow over convergent and divergent wall riblets.
Exp Fluids. 2002;33(2):346–350.
[42] Nugroho B, Hutchins N, Monty JP. Large-scale spanwise periodicity in a turbulent bound-
ary layer induced by highly ordered and directional surface roughness. Int J Heat Fluid Flow.
2013;41:90–102.
[43] Nadesan T, Mitsudharmadi H, Lee TS, et al. Quasi-streamwise counter-rotating vortices gen-
erated by convergent riblets in flat plate boundary layer. J Vis. 2014;17(4):319–325.
[44] Iwamoto K, Suzuki Y, Kasagi N. Reynolds number effect on wall turbulence: toward effective
feedback control. Int J Heat Fluid Fl. 2002;23(5):678–689.
[45] Iwamoto K, Fukagata K, Kasagi N, et al. Friction drag reduction achievable by near-
wall turbulence manipulation at high Reynolds numbers. Phys Fluids (1994-present).
2005;17(1):011702–011702.
[46] Spalart PR, McLean JD. Drag reduction: enticing turbulence, and then an industry. Phil Trans
R Soc London A Math Phys Eng Sci. 2011;369(1940):1556–1569.
[47] Gatti D, Quadrio M. Performance losses of drag-reducing spanwise forcing at moderate values
of the reynolds number. Phys Fluids (1994-present). 2013;25(12):125109.
[48] Wang Z-Q, Cheng N-S. Time-mean structure of secondary flows in open channel with longi-
tudinal bedforms. Adv Water Resou. 2006;29(11):1634–1649.
[49] Reynolds RT, Hayden P, Castro IP, et al. Spanwise variations in nominally two-dimensional
rough-wall boundary layers. Exp Fluids. 2007;42(2):311–320.
[50] Vermaas DA, Uijttewaal WSJ, Hoitink AJF. Lateral transfer of streamwise momentum caused
by a roughness transition across a shallow channel. Water Resources Res. 2011;47(2):1–12.
[51] Barros JM, Christensen KT. Observations of turbulent secondary flows in a rough-wall bound-
ary layer. J Fluid Mech. 2014;748:R1, 1–13.
[52] Vanderwel C, Ganapathisubramani B. Effects of spanwise spacing on large-scale secondary
flows in rough-wall turbulent boundary layers. J Fluid Mech. 2015;774:R2, 1–12.
[53] Schoppa W, Hussain F. A large-scale control strategy for drag reduction in turbulent boundary
layers. Phys Fluids (1994-present). 1998;10(5):1049–1051.
[54] Moser RD, Kim J, Mansour NN. Direct numerical simulation of turbulent channel flow up to
re= 590. Phys Fluids. 1999;11(4):943–945.
[55] Breugem WP, Boersma BJ. Direct numerical simulations of turbulent flow over a permeable
wall using a direct and a continuum approach. Phys Fluids (1994-present). 2005;17(2):025103.
[56] Orlandi P, Leonardi S, Antonia RA. Turbulent channel flow with either transverse or longitu-
dinal roughness elements on one wall. J Fluid Mech. 2006;561:279–305.
[57] Vreman AW, Kuerten JGM. Statistics of spatial derivatives of velocity and pressure in turbulent
channel flow. Phys Fluids (1994-present). 2014;26(8):085103.
 752 H. O. G. BENSCHOP AND W.-P. BREUGEM

[58] Chung D, Chan L, MacDonald M, et al. A fast direct numerical simulation method for char-
acterising hydraulic roughness. J Fluid Mech. 2015;773:418–431.
[59] Lozano-Durán A, Jiménez J. Effect of the computational domain on direct simulations of tur-
bulent channels up to Reτ = 4200. Phys Fluids (1994-present). 2014;26(1):011702.
[60] García-Mayoral R, Jiménez J. Hydrodynamic stability and breakdown of the viscous regime
over riblets. J Fluid Mech. 2011;678:317–347.
[61] Choi H, Moin P, Kim J. Direct numerical simulation of turbulent flow over riblets. J Fluid
Mech. 1993;255:503–539.
[62] Goldstein D, Handler R, Sirovich L. Direct numerical simulation of turbulent flow over a mod-
eled riblet covered surface. J Fluid Mech. 1995;302:333–376.
[63] Pourquie MBJM, Breugem WP, Boersma BJ. Some issues related to the use of immersed
boundary methods to represent square obstacles. Int J Multiscale Comput Eng. 2009;7(6):509–
522.
[64] Fadlun EA, Verzicco R, Orlandi P, et al. Combined immersed-boundary finite-difference
methods for three-dimensional complex flow simulations. J Comput Phys. 2000;161(1):35–60.
[65] Vreman AW, Kuerten JGM. Comparison of direct numerical simulation databases of turbulent
channel flow at Reτ = 180. Phys Fluids (1994-present). 2014;26(1):015102.
Downloaded by [191.96.251.147] at 03:47 11 December 2017

[66] Ferziger J.H., Perić M. Computational methods for fluid dynamics. Berlin Heidelberg: Springer;
2002.
[67] Wesseling P. Principles of computational fluid dynamics. Springer Series in Computational
Mathematics. Berlin Heidelberg: Springer; 2001.
[68] Frohnapfel B, Hasegawa Y, Quadrio M. Money versus time: evaluation of flow control in terms
of energy consumption and convenience. J Fluid Mech. 2012;700:406–418.
[69] Daschiel G, Baier T, Saal J, et al. On the flow resistance of wide surface structures. PAMM.
2012;12(1):569–570.
[70] Mohammadi A, Floryan JM. Groove optimization for drag reduction. Phys Fluids (1994-
present). 2013;25(11):113601.
[71] Hoyas S, Jiménez J. Reynolds number effects on the Reynolds-stress budgets in turbulent chan-
nels. Phys Fluids (1994-present). 2008;20(10):101511.
[72] Bernardini M, Pirozzoli S, Orlandi P. Velocity statistics in turbulent channel flow up to Reτ =
4000. J Fluid Mech. 2014;742:171–191.
[73] Lee M, Moser RD. Direct numerical simulation of turbulent channel flow up to Reτ  5200. J
Fluid Mech. 2015;774:395–415.
[74] Pope SB. Turbulent flows. Cambridge: Cambridge University Press; 2000.
[75] Hage W, Bechert DW, Bruse M. Yaw angle effects on optimized riblets. In: Thiede P, editor.
Aerodynamic drag reduction technologies. Berlin Heidelberg: Springer; 2001. p. 278–285.
[76] García-Mayoral R, Jiménez J. Drag reduction by riblets. Phil Trans R Soc A: Math Phys Eng
Sci. 2011;369(1940):1412–1427.
[77] García-Mayoral R, Jiménez J. Scaling of turbulent structures in riblet channels up to Reτ  550.
Phys Fluids (1994-present). 2012;24(10):105101.
[78] Jiménez J. Turbulent flows over rough walls. Annu Rev Fluid Mech. 2004;36:173–196.
[79] Shi J-M, Breuer M, Durst F. A combined analytical–numerical method for treating corner
singularities in viscous flow predictions. Int. J Numer Meth Fluids. 2004;45(6):659–688.
[80] Orlandi P, Leonardi S. DNS of turbulent channel flows with two-and three-dimensional rough-
ness. J Turbul. 2006;7(73):1–22.
[81] Saha AK. Far-wake characteristics of two-dimensional flow past a normal flat plate. Phys Fluids
(1994-present). 2007;19(12):128110.
[82] Narasimhamurthy VD, Andersson HI. Numerical simulation of the turbulent wake behind a
normal flat plate. Int J Heat Fluid Flow. 2009;30(6):1037–1043.
[83] Anderson W, Barros JM, Christensen KT, et al. Numerical and experimental study of mech-
anisms responsible for turbulent secondary flows in boundary layer flows over spanwise het-
erogeneous roughness. J Fluid Mech. 2015;768:316–347.
[84] Dubief Y, Delcayre F. On coherent-vortex identification in turbulence. J Turbul. 2000;1(1):011–
011.
 JOURNAL OF TURBULENCE 753

[85] Fukagata K, Iwamoto K, Kasagi N. Contribution of Reynolds stress distribution to the skin
friction in wall-bounded flows. Phys Fluids (1994-present). 2002;14(11):L73–L76.
[86] Peet Y, Sagaut P. Theoretical prediction of turbulent skin friction on geometrically complex
surfaces. Phys Fluids (1994-present). 2009;21(10):105105.
[87] Goldstein DB, Tuan T-C. Secondary flow induced by riblets. J Fluid Mech. 1998;363:115–151.
[88] Du Y, Symeonidis V, Karniadakis GE. Drag reduction in wall-bounded turbulence via a trans-
verse travelling wave. J Fluid Mech. 2002;457:1–34.
[89] Quadrio M, Xie W. Turbulent drag reduction by traveling waves of spanwise forcing. In: Pro-
ceedings of the 15th European Turbulence Conference; Aug 25–28; Delft, The Netherlands;
2015.
[90] Viotti C, Quadrio M, Luchini P. Streamwise oscillation of spanwise velocity at the wall of a
channel for turbulent drag reduction. Phys Fluids (1994-present). 2009;21(11):115109.
[91] Quadrio M, Floryan JM, Luchini P. Effect of streamwise-periodic wall transpiration on turbu-
lent friction drag. J Fluid Mech. 2007;576:425–444.
[92] Mamori H, Fukagata K. Drag reduction effect by a wave-like wall-normal body force in a tur-
bulent channel flow. Phys Fluids (1994-present). 2014;26(11):115104.
Downloaded by [191.96.251.147] at 03:47 11 December 2017

Appendix A. Time advancement at fixed bulk velocity

For simplicity, Equation (1) is rewritten as:

∂ui ∂p
=− + ri + γ f f driving δi1 , (A1)
∂t ∂xi

where ri contains the advection, diffusion and IBM terms. Let the integer n denote the
time steps. The used Runge–Kutta scheme (RK3) employs three sub-steps, which are num-
bered by the integer s. RK3 introduces intermediate velocities ui(s) , where ui(0) = uni and
ui(3) = un+1
i . Similarly, intermediate pressures p(s) and driving forces f driving(s) are intro-
duced. Because of the Crank–Nicolson scheme for pressure, p(s) = p(s−1) +  p(s) with cor-
rection pressure  (s)
p . The time advancement is illustrated here for an arbitrary sub-step s:

∂p(s) ∂ p(s−1)  
ui(s) + α(s) t = ui(s−1) − α(s) t + t γ(s) ri(s−1) + ζ(s−1) ri(s−2)
∂xi ∂xi
+ α(s) tγ f f driving(s) δi1
≡ u∗(s)
i + α(s) tγ f f driving(s) δi1 (A2)
≡ u∗∗(s)
i ,

where u∗(s)
i is the first and u∗∗(s)
i the second prediction velocity. Note that the asterisk here
is not used to denote dimensional quantities. The parameters α (s) , γ (s) and ζ (s − 1) are RK3
parameters (see e.g. [67]). To obtain the driving force, the equation for the streamwise veloc-
ity is volume-averaged, which yields:
 (s) 
 (s)  ∂
p    
u V + α(s) t = u∗(s) V + α(s) t γ f V f driving(s) . (A3)
∂x V

The first term represents the intermediate bulk velocity in sub-step s, which is set equal
to one to obtain the constant bulk flow. The second term disappears because of periodic
 754 H. O. G. BENSCHOP AND W.-P. BREUGEM

boundary conditions. The first prediction velocity is known, so the third term can be com-
puted. The resulting equation can be solved for f driving(s) , which yields:
 
1 − u∗(s) V
f driving(s)
=   . (A4)
α(s) t γ f V

Next, u∗∗(s)
i is computed. With use of the continuity equation, the divergence of Equation
(A2) yields a Poisson equation for the correction pressure:
∗∗(s)
∂ 2
p(s) 1 ∂u j
= . (A5)
∂x2j α(s) t ∂x j

When this is solved, the updated velocity and pressure are computed:
Downloaded by [191.96.251.147] at 03:47 11 December 2017

∂p(s)
ui(s) = u∗∗(s) − α(s) t , p(s) = p(s−1) + 
p(s) . (A6)
i
∂xi

This procedure guarantees that the bulk velocity equals one in each RK3 sub-step, which
results in three values of the driving force. The total forcing (i.e. one per time step) is cal-
culated as follows:


3
f driving = α(s) f driving(s) . (A7)
s=1

Appendix B. Immersed boundary method for blade riblet textures

The IBM adjusts diffusive and advective fluxes around blades with help of indicator func-
tions, as is described below for parallel riblets without and in yaw. The third subsection
explains how the fiIBM term in Equation (1) is computed. The final subsection describes
how well the IBM approximates the no-slip and no-penetration conditions at the texture
surface.

B.. Parallel riblets without yaw

Figure B1 shows the numerical grid used for parallel riblets without yaw. As a staggered grid
is used, the velocity vectors are located at the faces of the grid cells. A few of these vectors
are shown in the right figure. The blade coincides with a few grid points of the spanwise
velocity v. The right figure helps to explain the IBM. The dash-dotted boxes in that figure
indicate three staggered grid cells that belong to the velocity components uleft , v (j, k) and
wleft . The IBM implements an adjustment of the advective and diffusive fluxes at the cell
faces marked with a thick solid line, as is explained below.
The streamwise velocity u was only adjusted in grid cells next to the blades (both
on the left and right side). For the grid cell of uleft , the two spanwise fluxes through its
thick solid face were changed: the advective flux uv = 0 and the diffusive flux u/y =
−2uleft /y.
The spanwise velocity v was changed for two reasons, namely to enforce no-penetration
at the blades and to adjust the fluxes above the blade tip. To enforce no-penetration, the
 JOURNAL OF TURBULENCE 755

0.07 v(j,k) v(j+1,k)

Δz
0.06
φ left φ right
0.05

wleft v(j,k−1) v(j+1,k−1)

0.04
z

0.03
vblade
0.02

h uleft
0.01
z z
0
s Δy
y y

Figure B. Numerical grid and IBM used for parallel riblets without yaw. Left: Part of the numerical grid
used for s+ =  simulations at Reb = , showing two unit cells in the spanwise direction. Both for
Downloaded by [191.96.251.147] at 03:47 11 December 2017

blade spacing and blade height,  grid cells were used. Right: Small part of the numerical grid ( grid
cells) around the blade tip. The text explains the IBM with use of this subfigure. The three dash-dotted
boxes represent staggered grid cells that belong to the velocity components shown in their centres. Each
of these cells has one face indicated with a thick and solid line. At that cell face, the IBM adjusts the
advective and diffusive fluxes.

prediction velocity was set to zero at the grid points that coincide with a blade. Let v blade
represent a spanwise velocity component that coincides with a blade (see Figure B1) and
let ∗ represent the first prediction velocity (as in Equation (A2)), then v blade ∗
= 0. The actual
velocity v blade follows from the correction step (Equation (A6)) and is very close to zero,
although not exactly zero (details follow in Subsection B.4).
The second change to v comprises the adjustment of advective and diffusive fluxes in the
grid cell just above the blade tip (shown as a dash-dotted box in Figure B1). The vertical
fluxes of v at the bottom face of that cell were adjusted. The diffusive flux v/z was split
into two contributions, namely from the left and right side of the blade (indicated by φ left
and φ right in the figure). It accounts for the fact that the thin blade does not inhibit vertical
transport. Specifically, v/z = 0.5φ left + 0.5φ right . The flux φ right was computed by linear
interpolation of the four velocity components labelled with indices in the figure: φ right =
(v t − v b )/z with v t = 0.75v (j, k) + 0.25v (j + 1, k) and v b = 0.75v (j, k − 1) + 0.25v (j + 1, k − 1) .
The flux φ left was computed in a similar way. As advection near the blade tips is likely less
important than diffusion, the advective flux was not split but simply set to zero (i.e. vw =
0 at the thick solid face).
The wall-normal velocity w was only adjusted in grid cells next to the blades, similarly
to what was done for u. For the grid cell of wleft , two spanwise fluxes at the thick solid face
were changed: the advective flux vw = 0 and the diffusive flux w/y = −2wleft /y. These
adjustments were also applied for the w-cell next to the blade tip, so this diffusive flux was
not split into two contributions (although the blade covers only half of the cell face). To
justify this choice, the simulation with s+ = 24 at Reb = 5500 was repeated. The diffusive
flux near the blade tip was separated into two contributions, namely from above and below
the blade tip. No significant difference in drag was found.
 756 H. O. G. BENSCHOP AND W.-P. BREUGEM

utop
Δz

wleft
s

ublade
y z

Δx
x x

Figure B. Numerical grid and IBM used for parallel riblets in yaw. Left: Top view of the texture with
α = °, showing one unit cell in the streamwise and two in the spanwise direction. The numerical grid
Downloaded by [191.96.251.147] at 03:47 11 December 2017

consists of Ncg =  grid cells per groove. The markers on one blade indicate that the blades intersect the
grid cells at the locations of the staggered velocity vectors u (circle) and v (square). Right: Small part of
the numerical grid ( grid cells) around the blade tip. The text explains the IBM with use of this subfigure,
see also Figure B.

B.. Parallel riblets in yaw

To simulate the turbulent flow over blades in yaw, the riblet texture was rotated with respect
to the grid, see Figure B2 (left). That required a different immersed boundary method
(IBM), as the surfaces were not anymore aligned with the Cartesian directions. As the left
figure shows, the grid was generated in such a way that the obstacle surface still coincides
with part of the staggered velocity vectors. The main disadvantage of this approach is that
the streamwise and spanwise grid spacings cannot be chosen independently: they depend
on α. However, the major advantage is a relatively simple IBM. Due to the specific align-
ment of grid and texture, the number of grid cells per groove Ncg is equal for the streamwise
and spanwise directions. The grid shown in Figure B2 has Ncg = 16.
The IBM is explained with the use of Figure B2 (right). Staggered grid cells for utop and
wleft are shown as dash-dotted boxes. The IBM implements an adjustment of the advective
and diffusive fluxes at the cell faces marked with a thick solid line, as is explained below.
The streamwise and spanwise velocities were adjusted in a similar way, so only the change
to u is described here. The changes to u were almost identical to that for v in the previous
subsection. To enforce no-penetration, the streamwise prediction velocity was set to zero
(i.e. u∗blade = 0) at the grid points that coincide with a blade. At the bottom cell face of the
grid cell for utop , both the advective and diffusive fluxes were modified. The advective flux
was set to zero: uw = 0. The diffusive flux was not split into two contributions: u/z =
2utop /z.
For the wall-normal velocity, two indicator functions were used, namely one for stream-
wise and another one for spanwise transport of wall-normal momentum. As the treatment
of streamwise and spanwise transport of w is analogous, only the former is described here.
The wall-normal prediction velocity was not changed directly, as w never coincides with
blades. Therefore, only the fluxes in w-cells next to the blades were modified. Specifically,
 JOURNAL OF TURBULENCE 757

Figure B. Spanwise profile of streamwise velocity. Left: Bird’s-eye view of one unit cell of a herringbone
texture with Ngroove = , α = °. The thick line parallel to the y-axis cuts through the centre of the
first grid cell above the bottom wall. Right: Time-averaged streamwise velocity as function of spanwise
distance along the thick line in the left subfigure.
Downloaded by [191.96.251.147] at 03:47 11 December 2017

the fluxes at the right cell face of the grid cell for wleft were adjusted as follows: uw = 0 and
w/x = −2w left /x.

B.. Computation of fiIBM

Although the IBM forcing fiIBM appears as a separate term in Equation (1) and following
equations, it is not implemented as an explicit relation. The two previous subsections have
shown instead that the numerical code for smooth walls was adjusted to obtain the zero
velocity at the blades. However, the IBM term was needed as a separate term for the post-
processing. Its computation is illustrated with reference to Appendix A. The first prediction
velocity u∗(s)
i that appears there contains all adjustments that are needed to obtain the zero
velocity at the blades. Next, all these adjustments are left out to compute the prediction
velocity ui∗(s) no IBM in absence of the IBM forcing. The intermediate IBM forcing fiIBM(s)
then follows from u∗(s) i = ui∗(s) no IBM + α(s) t fiIBM(s) . Finally, the total IBM forcing fiIBM is
obtained from summing α(s) fiIBM(s) , similar to Equation (A7).

B.. Boundary conditions at the texture surface

The no-slip and no-penetration conditions at the texture surface need to be sufficiently
satisfied for the IBM to be accurate. Let a penetration velocity denote the absolute value of a
texture-collocated velocity component. As ublade and v blade denote two velocity components
that coincide with the texture, then the corresponding penetration velocities are upen =
|ublade | and v pen = |v blade |. Note that wpen does not exist, as w does not coincide with the
textures studied in the present work. To check whether the no-penetration condition was
sufficiently satisfied, each tenth simulation time-step three penetration-related velocities
were computed, namely upen , v pen  and max (upen , v pen ), where ... denotes an average.
The resulting time series were time averaged, yielding upen , v pen  and max (upen , v pen ).
The magnitude of these penetration velocities (normalised with the bulk velocity) is similar
for all the simulations, namely a mean on the order of 10−6 to 10−5 and a maximum on the
order of 10−4 to 10−3 . This shows that the penetration velocities are very small, so the no-
slip and no-penetration conditions are sufficiently satisfied.
 758 H. O. G. BENSCHOP AND W.-P. BREUGEM

Figure B3 presents a streamwise velocity profile as function of the spanwise distance for
a herringbone riblet texture with four grooves per feather half (so Ngroove = 4). The left sub-
figure shows the texture together with a line parallel to the y-axis. The right subfigure shows
the time-averaged streamwise velocity profile that was extracted along that line. Clearly, the
velocity is zero at the eight riblet locations. This illustrates that the boundary conditions at
the riblet surfaces are satisfied.

Appendix C. Simulation parameters

An overview of all simulations with the corresponding parameters can be found in Table C1.
Downloaded by [191.96.251.147] at 03:47 11 December 2017
 Downloaded by [191.96.251.147] at 03:47 11 December 2017

Table C. Parameters of all direct numerical simulations presented in this paper. The simulations are grouped based on the type of texture. The short
name indicates which parameters have been varied for a certain texture. The addition (sp. res.) refers to a case with double spanwise resolution; (res.)
indicates a double streamwise and spanwise resolution. The parallel riblet variant with Ngroove = . refers to the conventional parallel-riblet texture.
The averaging time T is normalised with δ/uτ derived from smooth-wall flow at the same bulk Reynolds number.
Short name s+ α (°) h/s Ngroove Lx Ly Nx Ny Nz x+ y+ z+
w
z+
c
Ncg Reb Tuτ /δ dtot ·  DC (%)
Smooth wall
Re - - - - . .    . . . . -   . ± . . ± .
Re - - - - . .    . . . . -   . ± . . ± .
Re - - - - . .    . . . . -   . ± . . ± .
Parallel riblets
splus .  . - . .    . . . .    . ± . −. ± .
splus .  . - . .    . . . .    . ± . −. ± .
splus .  . - . .    . . . .    . ± . . ± .
splus (sp. res.) .  . - . .    . . . .    . ± . . ± .
splus Re .  . - . .    . . . .    . ± . −. ± .
splus Re .  . - . .    . . . .    . ± . −. ± .
splus Re .  . - . .    . . . .    . ± . −. ± .
Parallel riblets in yaw
alpha .  . - . .    . . . .    . ± . −. ± .
alpha .  . - . .    . . . .    . ± . −. ± .
alpha (res.) .  . - . .    . . . .    . ± . −. ± .
alpha .  . - . .    . . . .    . ± . . ± .
Herringbone riblets: α = 15◦
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove (res.) .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
Herringbone riblets: α = 165◦
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . −. ± .
Herringbone riblets: α = 165◦ , shifted variant
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
ngroove .  .  . .    . . . .    . ± . . ± .
JOURNAL OF TURBULENCE

Herringbone riblets: α = 0◦ , parallel variant

ngroove. .  . . . .    . . . .    . ± . −. ± .
ngroove .  .  . .    . . . .    . ± . −. ± .
ngroove .  .  . .    . . . .    . ± . −. ± .
759

ngroove .  .  . .    . . . .    . ± . −. ± .