Computers & Fluids 88 (2013) 136–155

Contents lists available at ScienceDirect

Computers & Fluids

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m p fl u i d

Turbulent Navier–Stokes analysis of an oscillating wing

in a power-extraction regime using the shear stress transport
turbulence model
M. Sergio Campobasso a,⇑, Andreas Piskopakis b, Jernej Drofelnik b, Adrian Jackson c
a
Department of Engineering, Lancaster University, Engineering Building, Bailrigg, Lancaster LA1 4YR, United Kingdom
b
School of Engineering, University of Glasgow, James Watt Building South, University Avenue, Glasgow G12 8QQ, United Kingdom
c
EPCC, University of Edinburgh, James Clerk Maxwell Building, Mayfield Road, Edinburgh EH9 3JZ, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: A wing that is simultaneously heaving and pitching may extract energy from an oncoming air or water
Received 19 December 2012 stream. The aerodynamic performance of this device in terms of extracted power and energy conversion
Received in revised form 26 July 2013 efficiency is here investigated by means of time-dependent turbulent flow simulations performed with a
Accepted 30 August 2013
compressible Reynolds-averaged Navier–Stokes research solver using the K  x shear stress transport
Available online 11 September 2013
model of Menter for the turbulence closure. Previous studies of this device have focused primarily on
laminar flow regimes, and only recently systematic turbulent flow analyses of this device have appeared.
Keywords:
This paper presents comparative computational fluid dynamics analyses of the energy extraction process
Energy-extracting oscillating wing
Turbulent Navier–Stokes multigrid solvers
in a fully turbulent and a fully laminar flow regime. Presented results highlight that (a) substantial dif-
Strongly-coupled integration for ferences of the flow aerodynamics exist between the two cases, (b) the efficiency of the device in the con-
time-dependent problems sidered turbulent and laminar regimes achieves values of about 40% and 34% respectively, in line with the
findings of previous independent studies, and (c) further improvement of the efficiency observed in the
turbulent regime may be achieved by optimizing the kinematic characteristics of the device including
turbulent Reynolds number effects in the flow analyses used for the optimization. On the algorithmic
and modeling sides, the analyses make use of a computationally efficient method for the fully coupled
semi-implicit integration of the time-dependent Navier–Stokes equations and the two equations of the
K  x shear stress transport model. The paper also provides a systematic assessment of the impact of
the turbulent wall boundary condition for the specific dissipation rate on the computed flow field. It is
highlighted that the solution variations due to particular choices of this boundary condition may be
higher than those caused by the use of different turbulence models.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction ultimately, electrical energy is receiving increasing attention from

the scientific and industrial communities. A promising concept re-
Large-scale electricity production from the wind is primarily lies on the use of oscillating wings simultaneously heaving and
based on the use of multi-megawatt horizontal axis turbines. An pitching as power extraction devices. This concept was initially
increasing number of new small horizontal and vertical axis wind proposed by McKinney and DeLaurier [26] in 1981, and was further
turbines is currently being developed and commercialized for ‘dis- investigated by Jones et al. [16] more recently. A thorough
tributed’ wind energy harvesting. These machines have rated parametric computational fluid dynamics (CFD) investigation into
power of the order 10 kW, and they are typically installed both the effects of kinematic parameters (heaving amplitude, pitching
in rural and urban areas. Small horizontal and, to a minor extent, amplitude and frequency) and geometric parameters (airfoil shape
vertical axis wind turbines have so far dominated the market of and location of pitching axis) on the efficiency of the power
small-scale wind machines. However, the use of significantly dif- extraction achievable by the oscillating wing has been performed
ferent devices for renewable energy production based on the con- by Kinsey and Dumas [18]. The simulations reported in that study
version of the kinetic energy of fluid streams into mechanical and, have been performed using the incompressible solver of the com-
mercial CFD package FLUENT, and the majority of these analyses
⇑ Corresponding author. refer to a laminar flow regime with a Reynolds number based on
E-mail addresses: m.s.campobasso@lancaster.ac.uk (M. Sergio Campobasso),
the airfoil chord and the freestream velocity of 1100. The main
apiskop@gmail.com (A. Piskopakis), jernej.drofelnik@gmail.com (J. Drofelnik), conclusions of the article [18] are that the oscillating wing can
adrian.jackson@ed.ac.uk (A. Jackson).

0045-7930/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compfluid.2013.08.016
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 137

extract energy from an oncoming fluid stream with efficiencies as The time-dependent 2D laminar flow analyses presented in this
high as 34%, and the main aerodynamic feature responsible for paper have been performed with the compressible solver of the
such a relatively high efficiency is the unsteady leading edge vortex COSA code [8,7], and have been partly discussed in [9]. The time-
shedding associated with dynamic stall. These conclusions have dependent 2D turbulent analyses have been performed with the
also been confirmed by a later independent study performed by newly implemented turbulent flow capability of COSA, now featur-
the authors of this paper using the compressible research solver ing both the K  x turbulence model of Wilcox [38] and the K  x
COSA with a low-speed preconditioner optimized for time-depen- Shear Stress Transport (SST) model of Menter [28]. As explained in
dent flows [9]. A prototype of the oscillating wing for extracting the next section, both models are two-equation eddy viscosity tur-
energy from an oncoming water stream has been recently tested bulence models, which require the solution of two additional
in water at Lac-Beauport near Quebec City. The measured data transport equations, one for the turbulent kinetic energy, and
have confirmed fairly high values of the energy conversion effi- one for a second turbulent variable called specific turbulence dissi-
ciency of this device. Based on the data reported in [20], the Rey- pation rate. The K  x SST model is an extension of the original
nolds number of the considered operating condition was K  x model of Wilcox, developed to (1) greatly reduce the sensi-
0.48 million. In this condition, the flow regime is predominantly tivity of the original K  x model to the somewhat arbitrary value
turbulent, and CFD analyses aiming to provide accurate and quan- of the specific dissipation rate enforced at the farfield boundaries
titative data for investigating the fluid dynamics of the oscillating of the computational domain, and (2) enhance the solution accu-
wing and optimizing its design ought to include the effects of tur- racy of turbulent flows by improving the capability of the K  x
bulence. In a recent article of Kinsey and Dumas [19], the hydrody- model to predict the onset and amount of separation in adverse
namics of the oscillating wing for power production at a Reynolds pressure flows. The numerical results of [28] and later comparative
number of 0.5 million is investigated by means of two-dimensional analyses performed for internal [21] and external [11] turbulent
(2D) and three-dimensional turbulent incompressible FLUENT flow flows highlight that the SST model achieves both objectives. Other
simulations using the Spalart–Allmaras turbulence model [34]. extensions of the original K  x model aiming to achieve the same
Although the article [19] does not report a comparative study of objectives have also been developed by Wilcox [39]. One-equation
the hydrodynamics of the oscillating wing in the laminar and tur- eddy viscosity turbulence models [34,3] require the solution of
bulent regimes, cross-comparison of the laminar and turbulent only one transport equation. Historically, the development of
flow simulations reported in the articles [18,19] points to the fact one-equation models has followed that of two-equation models,
that the efficiency of the energy conversion appears to increase sig- and its main motivation has been to reduce the computational cost
nificantly as the Reynolds number increases from low values, typ- associated with two-equation models while limiting the accuracy
ical of laminar regimes, to fairly high values, at which a loss with respect to flow simulations based on two-equation mod-
predominantly turbulent regime is expected. els. Several comparative analyses of realistic turbulent flow prob-
To date, there appears to exist more experimental and proto- lems using both one- and two-equation turbulence modeling
type-based studies regarding the use of the oscillating wing device highlight that, though the results of modern two-equation models
to extract energy from oncoming water rather than air streams. often appear to be closer to experimental data, the solution differ-
Nevertheless, overall feasibility studies and detailed aerodynamic ences between one- and two-equation models are indeed often
analyses aiming at developing oscillating wing devices to extract small. A wider review of turbulence modeling is beyond the scope
energy from the wind are ongoing. In 2008 Platzer et al. [31] have of this paper, and the interested reader is referred to the review
proposed the use of the oscillating wing technology for the devel- article [33] for a wider overview of the present state, challenges
opment of ‘flying flapping-wing power generators for the purpose and needs of turbulence modeling for engineering applications,
of tapping into the abundant energy available in the global jet and long term projections for the progress in this area.
streams’. They have also provided preliminary multidisciplinary The turbulent COSA code solves the two systems of algebraic
assessments of the technical viability of this concept, which has equations resulting from the time- and space-discretization of
been patented by Bradley and Platzer in 2009 [6]. The interest in the Reynolds-averaged Navier–Stokes (RANS) (or mean flow) equa-
the use of the oscillating wing device to extract energy from air tions and the turbulence model equations by means of an explicit
streams is also highlighted by the increasing number of computa- multigrid algorithm based on a four-stage Runge–Kutta smoother.
tional studies in this area [2,30,36]. The two systems are solved using a strongly coupled approach
One of the primary objectives of this report is to thoroughly [25,23,12], whereby the mean flow and the turbulence equations
investigate the effects of flow turbulence on the detailed aerody- are solved simultaneously in the iterative process. This integration
namic features accounting for the energy extraction of the oscil- approach has been shown to lead to significantly faster conver-
lating wing by performing a comparative aerodynamic analysis gence rates than the loosely coupled method [24,22], whereby the
of the device at the laminar flow condition with a Reynolds num- mean flow and turbulence equations are solved separately and of-
ber of 1100 considered in [18,9], and a turbulent regime with a ten with different methods. It is also possible to use a ’hybrid’ inte-
Reynolds number of 1.5 million. The presented analyses will high- gration approach, whereby multigrid is applied simultaneously to
light that the different aerodynamics of the laminar and turbulent the two systems, and time-marching on each grid level is decou-
regimes result in significantly different levels of power conversion pled [37], but this approach has not been adopted in the COSA
efficiency. Although the turbulent flow analysis of the oscillating code, which instead features the standard strongly coupled ap-
wing operating at a Reynolds number of 0.5 million has been re- proach: the multigrid solver is applied simultaneously to the mean
ported in [19], it is instructive to perform comparative turbulent/ flow and turbulence equations, and the two systems are time-
laminar CFD analyses of this device because the Reynolds number marched simultaneously on each grid level. The turbulent COSA
of real installations is likely to vary both due to site-dependent code adopts the strongly coupled integration method also for com-
design specifications and off-design operation. Hence, detailed puting time-dependent problems, whereby the explicit multigrid
knowledge of how the fluid dynamics of oscillating wing devices integration is used to solve the unsteady RANS (URANS) equations
varies on a wide range of the Reynolds number for given geomet- coupled to the K  x SST turbulence model. For such time-depen-
ric and kinematic characteristics is needed to accurately assess dent problems, the turbulent multigrid solver also features a point-
the energy yield of this device. From this perspective, the present implicit treatment of certain terms arising from the discretization
report is complementary to the other published studies in this of the physical time-derivatives. This approach is an extension of
area. the stabilization process reported by Melson et al. [27], and it
138 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

enables the use of fairly high Courant Friedrichs Lewis (CFL) num- tion rate of turbulent energy. The definition of the total energy is
bers, thus significantly reducing the number of multigrid cycles re- E = e + jvj2/2 + K, where e denotes the internal energy per unit mass.
quired to achieve a user-given reduction of the flow residuals. Introducing the assumption h of perfect gas, the
i equation of state can
A further relevant contribution of this report to the analysis of be written as p ¼ ðc  1Þ qE  12 q j v j2  qK , where p denotes the
steady and time-dependent turbulent flows based on the RANS static pressure. The generalized convective flux vector Uc is:
model is the systematic investigation of the uncertainty affecting
Uc ¼ Ec i þ Fc j  vb U ð2Þ
computed results due to the choice of different solid wall boundary
conditions (BCs) for the specific dissipation rate x, the second tur- where Ec and Fc are respectively the x- and y-components of Uc, and
bulent variable of the K  x turbulence models. Presented results are given by:
allow cross-comparisons of the variations of the computed solu- 0
tions caused by different choices of the turbulent wall boundary Ec ¼ ½qu qu2 þ p quv quH quK qux
0
conditions, the variations associated with the use of different tur- Fc ¼ ½qv quv qv 2 þ p qv H qv K qv x
bulence models, and the variations associated with the use of dif-
ferent levels of spatial and temporal resolution to be made. The vector vb is the velocity of the boundary S, and the flux term
The paper first presents the governing equations and the algo- vbU is its contribution to the overall flux balance, which is nonzero
rithmic features of the recently developed parallel structured mul- only in the case of unsteady problems with moving boundaries. The
ti-block turbulent COSA solver. It then provides a description of the symbol H denotes the total enthalpy per unit mass, the expression
oscillating wing device and its main kinematic parameters, fol- of which is H = E + p/q. The generalized diffusive flux vector Ud is:
lowed by a section on the validation of the turbulent solver, and Ud ¼ E d i þ F d j ð3Þ
numerical analyses of the impact of the turbulent wall boundary
condition on the computed flow solution. A detailed comparative where Ed and Fd are respectively the x- and y-components of Ud,
study of the power-extracting oscillating wing in laminar and tur- and are given by:
bulent flow conditions is then reported, which highlights the dif- 2 3
0
ferences of unsteady aerodynamic characteristics observed in the 6 7
two regimes, and explains the reasons for the different perfor- 6 sxx 7
6 7
mance achieved in the two regimes. A summary of the presented 6 s xy 7
Ed ¼ 6 7
6 usxx þ v sxy þ ðl þ rK l Þ @K  q 7
material and further prospective investigations are provided in 6 T @x x7
the closing section. 6 7
4 ðl þ rK lT Þ @K
@x
5
@x
ðl þ rx lT Þ @x
2. Governing equations 2 3
0
6 sxy 7
Internal and external viscous flows can be computed by solv- 6 7
6 7
ing the Navier–Stokes (NS) equations, which are a system of Npde 6
6
syy 7
7
nonlinear partial differential equations (PDEs) obtained by impos- Fd ¼ 6 us þ v s þ ðl þ r l Þ @K  q 7
6 xy yy K T @y y7
ing the conservation of mass, momentum and energy over a con- 6 7
6 ðl þ rK lT Þ @K 7
4 @y 5
trol volume. For 2D laminar flows Npde = 4 because the
@x
momentum equation has only two scalar components. In the case ðl þ rx lT Þ @y
of turbulent flows, the effects of turbulence are often taken into
The scalars sxx, sxy and syy are the Cartesian components of the vis-
account by averaging the NS equations on the time-scales of tur-
cous stress tensor s. Its expression is s = sM + sR, where sM is the
bulence. This process leads to the so-called Reynolds-averaged
molecular viscous stress tensor, and sR is the Reynolds stress tensor.
Navier–Stokes equations and the appearance in these equations
The tensor sM depends on the divergence of the flow velocity vector
of the Reynolds stress tensor. Making use of the Boussinesq
v, and the strain rate tensor s given by
approximation, this tensor depends mainly on the product of
the strain rate tensor and a turbulent or eddy viscosity. In the s ¼ ðrv þ rvT Þ=2 ð4Þ
CFD solver used for this study, the latter variable is computed M
by means of the two-equation K  x SST turbulence model [28]. For a Newtonian fluid, one has s = 2l[s  1/3(r  v)I], where l is
Thus, the simulation of two-dimensional turbulent flows requires the dynamic viscosity. Making use of the Boussinesq approxima-
the solution of a system of Npde = 6 PDEs, namely four RANS PDEs, tion, the tensor sR is:
 
one PDE describing the convection, diffusion, creation and 1 2
destruction of the turbulent kinetic energy, and one describing sR ¼ 2lT s  r  v  qKI ð5Þ
3 3
the evolution of the specific dissipation rate, the second turbulent
variable of the SST model. The turbulent viscosity lT is given by
Given a control volume C with boundary S, the Arbitrary lT ¼ a1 qK= maxða1 x; F 2 jXjÞ ð6Þ
Lagrangian–Eulerian integral form of the system of the 2D time-
dependent RANS equations and the SST model equations is: in which a1 = 0.31, X is the flow vorticity, and F2 is a function of
! K, x, the molecular kinematic viscosity m and the distance from
Z I Z
@ the wall d. The expression of F2 can be found in reference [28].
U dC þ ðUc  Ud Þ  dS  S dC ¼ 0 ð1Þ
@t CðtÞ SðtÞ CðtÞ The scalars qx and qy are the Cartesian components of the heat
flux vector q = jrT, where j is the thermal conductivity and T
The array U of conservative flow variables is defined as: is the static temperature. For turbulent flows, the thermal conduc-
tivity is the sum of a molecular and turbulent component, which
U ¼ ½q qu qv qE qK qx0
can be related to known values of the molecular Prandtl number
where the superscript 0 denotes the transpose operator, and q, u, v, Pr and the turbulent Prandtl number PrT using the equation
E, K and x are, respectively, the flow density, the x- and y-compo- j = lcp/Pr + lTcp/PrT, where cp is the constant-pressure specific
nent of the flow velocity vector v, the total energy per unit mass, heat.
the turbulent kinetic energy per unit mass and the specific dissipa- The source term S appearing in Eq. (1) is
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 139

S ¼ ½0 0 0 0 SK Sx 0 Cartesian velocity derivatives needed to calculate the stress tensors

sM and sR are determined by considering the local generalized cur-
where
vilinear coordinates (n, g) associated with the grid lines, using the
2 chain rule to relate the Cartesian derivatives of the velocity compo-
SK ¼ sR rv  DK ¼ lT Pd  ðr  vÞqK  DK ð7Þ
3 nents to their derivatives in the (n, g) system and the grid metrics,
and using second order finite-differences to approximate all re-
c R 2 cqK quired derivatives. The Cartesian derivatives of the flow velocity
Sx ¼ s rv  Dx þ CDx ¼ cqPd  ðr  vÞ  Dx þ CDx ð8Þ
mT 3 mT at the cell centers required to evaluate the source terms of K and
and x are determined using the same approach.
  For steady problems the time-derivative appearing in Eq. (1)
1 vanishes, and space-discretizing all remaining terms on a compu-
Pd ¼ 2 s  r  v rv ð9Þ
3 tational grid consisting of Ncell finite volumes leads to a system of
DK ¼ b qK x Dx ¼ bqx2 ð10Þ nonlinear algebraic equations of the form:
1 RU ðQ Þ ¼ 0 ð12Þ
CDx ¼ 2ð1  F 1 Þqrx2 rK  rx ð11Þ
x
The entries of the array Q are the unknown flow variables at the Ncell
where mT is the turbulent kinematic viscosity, and the variables rK, cell centers discretizing the physical domain. The array Q can be
rx, c, b⁄ and b are weighted averages of the constants of the stan- viewed as made up of Ncell subarrays, each of which stores the Npde
dard K  x model [38] and the constants of the standard K   flow unknowns at a particular cell center. The length of Q is there-
model [17] with weights F1 and (1  F1), respectively. The symbol fore (Npde  Ncell). The array RU stores the cell residuals, and its
rx2 is a constant of the standard K   model, and the function F1 structure is the same as that of Q. For each cell, the Npde residuals
(its expression is provided in Ref. [28]) depends on the local values are obtained by adding the convective fluxes Uc;f and the diffusive
of K, x, m, q, d, rK and rx. It can be shown that the production fluxes Ud;f through all its faces, and, for the K and x residuals, by
term Pd is always positive. Thus the source term SK of the K-equa- also adding the associated source terms evaluated at the cell center,
tion has a term which is always positive (production term propor- and given respectively by Eqs. (7) and (8).
tional to Pd), a term which is always negative (destruction term
DK) and a term which is positive or negative depending on the sign 3.2. Numerical integration of steady problems
of r  v. Similarly to SK, the source term Sx of the x-equation also
has a term which is always positive (production term proportional The RANS and the SST turbulence model equations are solved
to Pd), a term which is always negative (destruction term Dx), and a with a time-marching algorithm using the so-called strongly cou-
term which is positive or negative depending on the sign of r  v. pled approach [25,23,12] whereby the two sets of equations are
The source term Sx, however, features an additional cross-diffusion time-marched simultaneously. In COSA, the unknown flow vector
term CDx which can be positive or negative. As highlighted in past Q is computed by solving iteratively Eq. (12). A fictitious time-
studies on the numerical integration of the K  x turbulence model derivative (dQ/ds) premultiplied by the cell volumes is added to
equations [25,23], the identification of positive and negative source this system, and this derivative is then discretized with a four-stage
terms is of crucial importance when using an explicit multigrid Runge–Kutta (RK) scheme. The numerical solution is thus marched
integration of the equations of turbulence, due to the different in pseudo-time until the steady state is achieved. The convergence
numerical treatment to be adopted for positive and negative source rate is greatly enhanced by means of local time-stepping (LTS), var-
terms. iable-coefficient central implicit residual smoothing (IRS) and a
full-approximation scheme (FAS) multigrid (MG) algorithm. Denot-
3. CFD solver ing by Ds the local pseudo-time-step, l the RK cycle counter, k the
RK stage index, and ak the kth RK coefficient, the RK smoother is:
3.1. Space discretization
W0 ¼ Q l
The structured multi-block finite volume cell-centered parallel Wk ¼ W0  ak DsV 1 LIRS ½RU ðWk1 Þ þ f MG  ð13Þ
CFD code COSA discretizes all components of the generalized con- Q lþ1 ¼ W NS
vective flux vector Uc defined by Eq. (2) making use of a second or-
der upwind scheme. The discretization of the convective fluxes is where LIRS denotes the IRS operator, and fMG is the MG forcing func-
based on Van Leer’s MUSCL extrapolations and Roe’s flux-difference tion, which is nonzero when the smoother (13) is used on a coarse
splitting. Denoting by n the outward normal of the face of a grid level after a restriction step. The diagonal matrix V stores the vol-
cell, and by dS the area of such a face, the continuous convective umes of the grid cells. It can be viewed as a block-diagonal matrix
flux component through the face is Uc,f = (Uc  n)dS. Let Uc;f denote of size (Ncell  Ncell) with each block being the identity matrix of size
its numerical approximation, the expression of which is reported in (Npde  Npde) multiplied by the volume of the cell the block refers to.
[9]. Here the superscript ⁄ and the subscript f denote numerical The smoother (13) has a very poor convergence rate, due to the
approximation and face value, respectively. The turbulent solver stiffness of the iterative operator caused by the large negative
can compute the turbulent convective flux components using source terms of the turbulence model, such as DK and Dx. To
either a first order or a second order upwind scheme. The authors alleviate this problem, a semi-implicit integration strategy can be
of this paper have observed that the choice of either approach adopted: the negative source terms of the turbulence equations
yields negligible differences between the two solutions for sub- are evaluated at the RK stage k rather than at the stage k  1. Liu
sonic flows (such differences are non-negligible for transonic and Zheng [25] adopted this approach to develop an efficient
flows). All the results reported herein have been obtained using strongly coupled MG iteration for the compressible RANS equa-
the second order discretization for all convective flux components tions coupled with the standard K  x model. Those authors treat
and making use of Van Albada’s flux limiter. implicitly also the negative source term proportional to r  v
The discretization of the diffusive fluxes is based on second or- when the velocity divergence is positive. The turbulent MG inte-
der finite-differencing. More precisely, at the midpoint of each cell gration implemented by COSA uses a similar approach, but also
face the components of the vectors rT, rK and rx and the presents important differences (highlighted later in this
140 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

subsection) with respect to the scheme of Liu and Zheng, due pri- In general, when using the SST turbulence model, one would
marily to the modeling differences of the standard K  x and the adopt Eq. (17) rather than Eq. (18). Numerical experiments per-
K  x SST models. The implicit treatment of the negative source formed by the authors of this paper, however, reveal that the re-
terms of the K and x equations in the RK smoother leads to the fol- sults computed with either approach present fairly small
lowing two iterations to update qK and qx, respectively: differences for low-speed flows, such as those analyzed in this
n paper. For this reason and due to lower computational cost asso-
ð1 þ ak DsDþ ÞðqKÞk ¼ ðqKÞ0 þ ak Ds Dþ ðqKÞk1  b ½ðqK xÞk ciated with the use of Eq. (18), all analyses presented herein are
o based on the use of this equation. It should be noted that the
ðqK xÞk1   V 1 RK ðWk1 Þ ð14Þ
computational cost reduction achieved by using Eq. (18) rather
 þ than Eq. (17) is even higher when using the SST turbulence mod-
cD k k1 el with a harmonic balance (HB) RANS solver [8], a technology
ðqxÞk ¼ ðqxÞ0  ak Ds ½ðqKÞk  ðqKÞk1  þ b½ðqx2 Þ  ðqx2 Þ 
mT applicable to the fast simulation of strongly non-periodic flows

such as those associated with the oscillating wing device. This
þV 1 Rx ðWk1 Þ ð15Þ is because, using the stabilization process of the multigrid HB
NS iteration described in [8] and used for the COSA HB simula-
To simplify the notation, the IRS operator and the MG forcing term tions reported in [10], the size of the matrix inversions at each
have been temporarily omitted in these two equations. The symbols grid cell required to update all NH complex harmonic components
RK and Rx denote the complete cell residual arrays of qK and qx of qK and qx is [2(2NH + 1)  2(2NH + 1)] using Eq. (17) and
respectively, and Dþ ¼ maxð0; 23 r  vÞ. Eqs. (14) and (15) form a sys- [(2NH + 1)  (2NH + 1)] using Eq. (18). For highly nonlinear peri-
tem of two quadratic equations in the unknowns (qK)k and (qx)k. odic flows, the value of NH needed for a satisfactory time-resolu-
Linearizing the quadratic terms, one obtains a system of two linear tion often exceeds 5. Since these matrices are dense and
equations in the unknowns (qK)k and (qx)k, and the RK smoother unstructured, Gaussian elimination is used for their inversion,
for updating both the RANS and the SST variables incorporating and the computational cost of such inversions is proportional to
IRS and MG can be written as: the third power of the system size. Therefore, the use of Eq.
W0 ¼ Ql (18) rather than Eq. (17) for updating the harmonics of the SST
k
turbulence variables yields a reduction of the computational cost
ðI þ ak DsAÞW ¼ W0 þ ak DsAWk1 of nearly one order of magnitude. It is expected that the errors
ð16Þ
 ak DsV 1 LIRS ½RU ðWk1 Þ þ f MG  associated with the choice of Eq. (18) when using the SST model
may be significant for high-speed problems, due to the higher
Q lþ1 ¼ WNS
values of r  v caused by compressibility effects. Further investi-
The matrix A is block-diagonal and has size (Ncell  Ncell). The only gations will be carried out to verify this assumption. The primary
nonzero elements of each (Npde  Npde) block on the diagonal of A cause of the likely inadequacy of Eq. (18) for updating the SST
are those of the bottom right (2  2) partition, and this occurrence turbulence variables for high-speed flows, however, is not the
results in the coupling of the update process of the turbulent vari- non-negligible magnitude of r  v, but rather the SST expression
ables. The abovesaid partition is: of lT, which in general prevents decoupling the semi-implicit up-
" # date of qK and qx. Other two-equation turbulence models,
ðDþ þ b xÞ b K including the baseline model of Menter [28], feature expressions
Að5 : 6; 5 : 6; i; jÞ ¼ ASST ¼ cDþ ð17Þ
mT 2bx of lT structurally similar or identical to that of Wilcox’s K  x
model, and therefore enable an exact computationally cheaper
in which all variables are evaluated at the RK stage k  1, and the update of the turbulence variables for both low- and high-speed
indices i and j identify the grid cell the matrix block refers to. The flows.
cross-diffusion term CDx can also be positive or negative depending In the update process performed by the RK smoother, the
on the local flow conditions, and therefore, when negative, it could new estimate of x is prevented from assuming unphysically
be treated like D+ in the semi-implicit integration. However, this low values by limiting it with a minimum threshold based on
approach would make the implementation substantially more com- the production term Pd of Eq. (9), following the guidelines of
plex and less efficient because the term CDx depends on rK and [25]. When using MG the residuals of the x equation are also
rx. The evaluation of these gradients at stage k would couple the limited before being restricted to a coarser grid, as proposed in
update process of several cells, thus requiring the inversion of sig- [25]. The turbulent viscosity is computed on each grid level,
nificantly larger systems. For this reason, it has been preferred to but the production terms Pd and D+ are computed only on the
treat the term CDx explicitly regardless of its sign. It should be finest grid level and reported to the coarser levels by using the
noted that this term is absent in the standard K  x model. Another restriction operator. As the flow solution is also updated by
difference between the semi-implicit integration of the standard the prolongation operator, the K and x prolonged corrections
K  x model reported in references [25,23] and that of the SST are limited as proposed by Eliasson and Wallin [12]. The high-
model is that, in the former case, qx can be updated independently order restriction operator reported in [12] has also been imple-
of qK. This is however not possible in the SST case, since the ele- mented, and found to greatly enhance the robustness of the
ment (2, 1) of ASST is not zero. Due to this feature, a (2  2) matrix strongly coupled MG iteration.
inversion is required at each grid cell to update qK and qx. The dif- The integration of time-dependent problems is accomplished
ferent turbulent variables update of the K  x and SST models oc- by using the strongly coupled explicit integration approach re-
curs because the expression of the turbulent viscosity of the ported above for the solution of the URANS and SST model equa-
former model is obtained by setting F2 = 0 in Eq. (6). This operation tions. A point-implicit Runge–Kutta (PIRK) smoother [27] is also
results in the relationship K/mT = x, which can be used to remove adopted (the algorithmic details are provided in Appendix A).
the dependence of Eq. (15) on K. By performing this substitution, The PIRK smoother enables the use of higher CFL numbers than
the bottom right partition of each block of A becomes: the conventional non-stabilized fully explicit Runge–Kutta (FERK)
" #
ðDþ þ b xÞ b K smoother, thus significantly reducing the number of multigrid cy-
Að5 : 6; 5 : 6; i; jÞ ¼ AKx ¼ ð18Þ cles required to achieve a user-given reduction of the flow
0 cDþ þ 2bx residuals.
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 141

Table 1
y Flat plate drag coefficient computed with K  x and SST turbulence models using
different wall boundary conditions for x.
h0 t / =0 t / =1
T θ0 T SST Kx
t wlc1w 3.13  103 3.35  103
t / =0.25 wlc0w 3.29  103 3.52  103
T t / =0.75 mentw 3.09  103 3.31  103
t / =0.5 T
T

D
yr t / =0.25 L
T t / =0 imposed. The subsonic outflow conditions are determined by
T R
xr
R enforcing a user-given value of the static pressure and extrapolat-
α Y ing the remaining variables from the interior of the domain with
X suitable procedures, whereas the supersonic outflow conditions
t / =0.5
T are obtained by extrapolating all flow variables from the interior.
Several variants of the farfield BCs for internal problems, including
Fig. 1. Top plot: prescribed motion of oscillating wing. Bottom plot: airfoil motion one which preserves the nominal second order of the numerical
in reference system moving with freestream velocity for power-extracting mode.
scheme for problems with strong flow gradients on the farfield
boundaries, are reported in the article [7]. For both external and
3.3. Boundary conditions internal problems, the values of K and x are set to user-given val-
ues or extrapolated from the interior depending on whether an in-
In the case of external flow problems, the farfield BCs for the flow or outflow condition occurs.
mean flow equations are based on suitable combinations of one- At solid wall boundaries, the static pressure and the density are
dimensional (1D) Riemann invariants and user-given freestream extrapolated from the interior, and both velocity components are
data, namely sound speed, entropy and velocity components. The set equal to the wall velocity, which is zero only in the case of
way these data are combined depends on whether the fluid stream problems with motionless grids. The turbulent kinetic energy is
enters or leaves the computational domain [14]. In the case of set to zero, whereas two options are available for the calculation
internal flow problems, the subsonic inflow BCs for the mean flow of xw, the value of the specific dissipation rate at the wall. One ap-
equations are constructed by extrapolating the outgoing 1D char- proach is that proposed by Wilcox [38], which also allows the wall
acteristic variable and enforcing user-given values of total temper- roughness to be taken into account. The value of xw is:
ature, total pressure and flow direction. For supersonic inflow u2s
conditions, all thermodynamic and kinematic variables are xw ¼ SR ð19Þ
mw

30 5
theory
SST, wlc0w
25 K-ω, wlc0w
SST, mentw
K-ω, mentw
20 4
3
cf ⋅10

15
+
u

10 Spalding 3
SST, wlc0w
5 K-ω, wlc0w
SST, mentw
K-ω, mentw
0 2
0 2 4 6 0 0.2 0.4 0.6 0.8 1
30 5
theory
SST, wlc1w
25 K-ω, wlc1w
SST, mentw
K-ω, mentw
20 4
3
cf ⋅10

15
u+

10 Spalding 3
SST, wlc1w
5 K-ω, wlc1w
SST, mentw
K-ω, mentw
0 2
0 2 4 6 0 0.2 0.4 0.6 0.8 1
+
log10 (y ) x

Fig. 2. Turbulent flat plate analysis. Top left plot: comparison of Spalding’s velocity profile and velocity profiles at x = 0.5 computed with K  x and SST models using Wilcox’s
(SR = 100) and Menter’s wall BC. Top right plot: comparison of theoretical skin friction coefficient (cf) and cf computed with K  x and SST models using Wilcox’s (SR = 100)
and Menter’s wall BC. Bottom left plot: comparison of Spalding’s velocity profile and velocity profiles at x = 0.5 computed with K  x and SST models using Wilcox’s
(SR = 2500) and Menter’s wall BC. Bottom right plot: Comparison of theoretical cf and cf computed with K  x and SST models using Wilcox’s (SR = 2500) and Menter’s wall BC.
142 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

1 ρ 1
0 ρu 0
ρv
-1 ρE -1
ρK
-2 ρω -2
-3 -3

Δlr -4 -4
-5 -5
-6 -6
-7 -7
-8 Mentw -8 Wilcw
-9 -9
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
MG cycle MG cycle

Fig. 3. Convergence histories of turbulent SST analyses of flat plate boundary layer. Left plot: simulation using Menter’s BC. Right plot: simulation using Wilcox’s BC
(SR = 2500).

where us is the friction velocity, mw is the kinematic viscosity at the originates from the sign of the forces that the flow generates on
wall, and SR is a constant depending on the wall roughness. For a the oscillating airfoil. Based on the imposed motion and the up-
smooth wall, Wilcox’s original model foresees SR = 100. More recent stream flow conditions, the airfoil experiences an effective angle
studies, however, propose SR = 2500 for a smooth wall [13]. The im- of attack (AoA) a and an effective velocity ve given respectively by:
pact of the choice of SR on the computed viscous drag is analyzed in  
the section on the code validation. The other method for the calcu- aðtÞ ¼ arctan v y ðtÞ=u1  hðtÞ ð23Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
lation of xw is that proposed by Menter [28], according to whom:
v e ðtÞ ¼ u21 þ v y ðtÞ2 ð24Þ
60mw
xw ¼ ð20Þ
bD2w The maximum values of a and ve have a major impact on the ampli-
tude of the peak forces in the cycle, and also on the occurrence of
where Dw is the distance to the next grid point away from the wall. dynamic stall. The maximum effective AoA reached in the cycle is
As highlighted in the section on the code validation, Eq. (20) approximated by the modulus of its quarter-period value, that is
appears to improve the numerical stability of the MG-based inte- amax  ja(T/4)j. As explained in [18], the power-extracting regime
gration of the governing equations. However, it may prevent a (in a mean sense, over one cycle) occurs when a(T/4) < 0. This con-
grid-independent solution being attained since the value of xw de- dition is represented in the bottom sketch of Fig. 1, which provides
pends on the grid refinement at the wall boundary. To prevent such a time-sequence viewed in a reference frame moving with the far-
solution grid-dependence, some authors (e.g. [29]) adopt grid- and field flow at u1, so that the effective AoA a(t) is made visible from
flow-dependent bounds of the local values of xw provided by Eq. the apparent trajectory of the airfoil. In this sketch, the resultant
(20). Conversely, the use of Eq. (19) yields values of xw which, force R is first constructed from typical lift and drag forces (right-
above a minimum level of grid refinement, are grid-independent hand side) and then decomposed into X and Y components (left-
without requiring additional constraints. Moreover, its general form hand side). One sees that the vertical force component Y is in phase
[38] has the additional advantage of incorporating surface rough- with the vertical velocity component vy of the airfoil over the entire
ness effects [13]. cycle. This implies that the wing extracts energy from the fluid as
long as no energy transfer associated with the component X of
4. Oscillating wing device the aerodynamic force takes place. This is clearly the case since
the airfoil does not move horizontally. The aerodynamic analyses
Here an oscillating wing is defined as an airfoil experiencing of [18] and also those presented in this paper highlight that the
simultaneous pitching h(t) and heaving h(t) motions. The following aerodynamic phenomena taking place during the wing oscillation
mathematical representation of the imposed motion is that are substantially more complex than the quasi-steady model dis-
adopted in [18]. Taking a pitching axis located on the chord line cussed above. More specifically, the extent and efficiency of the en-
at position xp from the leading edge (LE), the airfoil motion is ergy extraction are heavily influenced by the occurrence of
expressed as: unsteady leading edge vortex shedding (LEVS) associated with dy-
namic stall and the LEVS timing with respect to the airfoil motion.
hðtÞ ¼ h0 sinðxtÞ ! XðtÞ ¼ h0 x cosðxtÞ ð21Þ Taking a wing span of one unit length, the instantaneous power
hðtÞ ¼ h0 sinðxt þ /Þ ! v y ðtÞ ¼ h0 x cosðxt þ /Þ ð22Þ extracted from the flow is the sum of a heaving contribution.
Py(t) = Y(t)vy(t) and a pitching contribution Ph(t) = M(t)X(t), where
where h0 and h0 are respectively the pitching and heaving ampli- M is the resulting torque about the pitching center xp. Denoting
tudes, X is the pitching velocity, vy is the heaving velocity, x is by c the airfoil chord, and C P  P=ð12 q1 u31 cÞ a power coefficient,
the angular frequency and / is the phase between heaving and the nondimensional power extracted over one cycle is given by:
pitching. In this study, / is set to 90o, and the NACA0015 airfoil is
Z  
selected. The freestream velocity is denoted by u1 and the angular 1 T
v y ðtÞ XðtÞc
C P ¼ C Py þ C Ph ¼ C Y ðtÞ þ C M ðtÞ dt ð25Þ
frequency x is linked to the oscillation frequency f by the relation- T 0 u1 u1
ship x = 2pf. The prescribed oscillating motion is depicted in the
top sketch of Fig. 1. where C Y ðtÞ ¼ YðtÞ=ð12 q1 u21 cÞ and C M ðtÞ ¼ MðtÞ=ð12 q1 u21 c2 Þ. The
An oscillating symmetric airfoil can operate in two different re- efficiency g of the power extraction is defined as the ratio of the
gimes: propulsive or power-extracting mode. This distinction extracted mean total power (P) and the total available power (Pa)
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 143

6.0 6.0

x/c=0.620 x/c=0.731
5.0 5.0

4.0 wlc1w 4.0

wlc0w

y/c⋅102
3.0 mentw 3.0
exp
2.0 2.0

1.0 1.0

0.0 0.0
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4

9.0 9.0

x/c=0.786 x/c=0.842
7.5 7.5

6.0 6.0
y/c⋅102

4.5 4.5

3.0 3.0

1.5 1.5

0.0 0.0
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4

12.0 12.0

x/c=0.897 x/c=0.953
10.0 10.0

8.0 8.0
y/c⋅102

6.0 6.0

4.0 4.0

2.0 2.0

0.0 0.0
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
u/u∞ u/u∞

Fig. 4. Turbulent flow past NACA4412 airfoil: comparison of measured boundary layer velocity profiles (profiles labeled ‘exp’) and velocity profiles computed with SST
turbulence model using Wilcox’s wall BC with SR = 2500 (profiles labeled ‘wlc1w’), Wilcox’s wall BC with SR = 100 (profiles labeled ‘wlc0w’), and Menter’s wall BC (profiles
labeled ‘mentw’) at six chordwise positions.

Table 2 efficiency g defined by Eq. (26) corresponds to the classical power

Turbulent flow past NACA4412 airfoil: lift, drag and pitching moment coefficients coefficient obtained by means of Betz’s analysis [5]. Such a theory
computed with SST turbulence model using different wall boundary conditions for x. is based on the 1D mass and linear momentum balance applied to
CL CD CM a stationary inviscid streamtube surrounding a generic power-
wlc1w 1.696 0.03743 0.06348 extracting device, and it shows that the upper limit of g is 16/
wlc0w 1.643 0.04219 0.06088 27  100  59.3%. Therefore, Eq. (26) provides the relationship be-
mentw 1.680 0.03884 0.06257 tween the mean power coefficient C P defined by Eq. (25) and Betz’s
theory power coefficient (g).

in the oncoming flow passing through the swept area (the flow
5. Validation
window):

P Py þ Ph c The predicting capabilities of the steady Euler solver and the

g  ¼ ¼ CP ð26Þ second order accuracy of the convective flux discretization have
Pa 12 q1 u31 d d
been demonstrated by comparing a sequence of numerical solu-
where d is the overall vertical extent of the airfoil motion. This dis- tions of an inviscid steady 2D internal flow problem computed
tance depends on the heaving and pitching motion parameters h0, with progressively more refined grids and the associated analytical
h0 and /, and, for the kinematic conditions considered in this report, solution [7]. The predicting capabilities of the time-domain
takes values slightly larger than 2h0. The power extraction Euler solver and the second order accuracy of the time- and
144 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

1 1
ρ
0 ρu
ρv 0
-1 ρE
ρK -1
ρω
-2 -2
-3 -3
Δlr
-4 -4
-5 -5
-6 -6
-7 Mentw -7 Wilcw
-8 -8
500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
MG cycle MG cycle

Fig. 5. Convergence histories of turbulent SST analyses of flow field past NACA4412 airfoil. Left plot: simulation using Menter’s BC. Right plot: simulation using Wilcox’s BC
(SR = 2500).

Table 3 4.5
Main geometric parameters of three grids adopted for assessing impact of spatial
refinement on computed oscillating wing turbulent flow field. 3.5

Grid nA nC nN Dw Ncell 2.5

Cx 1.5
6
Coarse 440 100 160 3.2  10 102,400
Medium 432 360 320 3.2  106 368,640
Fine 864 720 640 1.6  106 1,474,560 0.5

-0.5
0 0.2 0.4 0.6 0.8 1
3
space-discretization have been demonstrated by considering a
time-dependent problem resulting from the superposition of a uni- 2
form freestream and a steady vortex. The verification has been per- 1
CY

formed by comparing a sequence of numerical solutions of this 0

problem computed with progressively more time- and space-re- -1
fined grids and the associated analytical solution [9]. The assess- -2
ment of the predicting capabilities and the second order accuracy
-3
of the time-dependent laminar solver has been based on the anal- 0 0.2 0.4 0.6 0.8 1
ysis of the vortex shedding behind a circular cylinder and the use of
1
Richardson’s extrapolations, whereas the laminar viscous predict-
ing capabilities have also been verified by comparing a computed 0.5
laminar flat plate boundary layer and the analytical solution of Bla-
CM

sius [9]. The predicting capability of the code for inviscid problems 0
with moving grids is verified in [8], which compares the time-
-0.5
dependent pressure difference across a pitching flat plate in a uni-
form freestream to Theodorsen’s analytical solution. Here, the pre- -1
dicting capabilities of the turbulent solver are assessed by 0 0.2 0.4 0.6 0.8 1
comparing (a) a computed turbulent flat plate boundary layer t/T
and available theoretical results, and (b) a computed turbulent set-up 1: coarse, wlc1w, 128
separated airfoil boundary layer and available experimental data. set up 2: coarse, wlc1w, 256
set-up 3: coarse, wlc1w, 512
5.1. Flat plate set-up 4: coarse, wlc1w, 1024

Fig. 6. Comparison of force coefficients of oscillating wing (operating condition A),

The turbulent flow over a flat plate leading to the formation of a varying time-refinement. Top plot: horizontal force coefficient. Middle plot: vertical
turbulent boundary layer is considered. The computational domain force coefficient. Bottom plot: pitching moment coefficient.
is rectangular and the flat plate lies on the lower horizontal bound-
ary. The LE of the flat plate is in the origin of the Cartesian system,
and its trailing edge (TE) is at x = 1, where the (vertical) outlet refinement analyses conducted for each of the modeling set-ups
boundary is positioned. The inlet boundary is at x =  1/3, and discussed below have revealed that the solution computed with
the upper horizontal side is a farfield boundary positioned at the grid defined above presents negligible differences from the
y = 1. The adopted Cartesian grid has 384 mesh intervals along y, solution computed by using grids with substantially higher spatial
and the size of these intervals increases from the lower horizontal refinement. This turbulent flow field has been computed using
boundary to the upper horizontal boundary starting from a mini- both the K  x and the SST turbulence models, and, for each turbu-
mum value of 2.5  107. The grid has 256 equal mesh intervals lence model, three analyses have been performed, one computing
along x; 192 are on the flat plate and 64 in the space between xw with Wilcox’s condition (Eq. (19)) and SR = 100, one computing
the LE and the inlet boundary. The freestream Reynolds number xw with Menter’s condition (Eq. (20)) and one computing xw with
Re is 6  106, and the freestream Mach number is M1 = 0.2. Mesh Eq. (19) and SR = 2500, as proposed in [13]. The four profiles of the
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 145

4.5 Eq. (20). The skin friction coefficient cf along the flat plate obtained
with these four simulations is reported along the y-axis of the top
3.5
right subplot of Fig. 2. The theoretical estimate for this problem
2.5 ((cf)theory = 0.025 ⁄ (Re  x)1/7) is also reported for reference. One
Cx

1.5 notices that the differences between the cf profiles obtained by

using the two turbulence models with the same wall BC are of
0.5
the same order of magnitude as those observed when using one
-0.5 turbulence model with two different wall BCs, and this highlights
0 0.2 0.4 0.6 0.8 1
the impact of the adopted condition for xw on the computed
3
solution.
2 The velocity profiles obtained with the K  x and SST simula-
1 tions using either Eq. (19) with SR = 2500 or Eq. (20) are presented
CY

0 in the bottom left subplot of Fig. 2, where use of the former condi-
-1 tion is denoted by the label ‘wlc1w’. The cf profiles obtained with
-2 these analyses are instead reported in the bottom right subplot
of Fig. 2. The bottom left subplot highlights that the four velocity
-3
0 0.2 0.4 0.6 0.8 1 profiles are substantially closer to each other than in the top sub-
plot and this occurs because the rescaled Wilcox’s wall BC yields
1
wall values of x closer to those of Menter’s condition. The same
0.5 conclusion holds for the cf-profiles of the bottom right subplot.
The differences between the two solutions obtained using a given
CM

0 turbulence model with either Eq. (20) or Eq. (19) with SR = 2500 are
significantly smaller than those between the former solution and
-0.5
that obtained with Eq. (20) with SR = 100. The remaining significant
-1 differences between the computed cf profiles for a given wall BC
0 0.2 0.4 0.6 0.8 1 and use of either turbulence model are likely to be caused by struc-
t/T tural differences between the turbulence models, such as the lower
set-up 2: coarse, wlc1w, 256 sensitivity of the SST model to the freestream value of x. The the-
set up 5: medium, wlc1w, 256 oretical value of the drag coefficient CD for the considered configu-
set-up 6: fine, wlc1w, 256 ration is 3.14  103, whereas the values of CD obtained with the
set-up 7: medium, wlc0w, 256 six presented simulations are reported in Table 1, and these data
emphasize again the impact of the wall BC for x on the computed
Fig. 7. Comparison of force coefficients of oscillating wing (operating condition A), viscous drag.
varying space-refinement and xw BC. Top plot: horizontal force coefficient. Middle All simulations have been performed using the so-called im-
plot: vertical force coefficient. Bottom plot: pitching moment coefficient.
proved auxiliary state farfield BCs for internal flows [7] on the ver-
tical left and right boundaries of the computational domain, and a
standard external-flow farfield condition on the top horizontal
Table 4
Overall power coefficient and energy extraction efficiency of oscillating wing
boundary [14]. Symmetry conditions are imposed on the portion
(operating condition A), varying space-refinement, time-refinement and xw BC. of the lower horizontal boundary between the inlet boundary
and the LE of the flat plate, and a no-slip condition is applied on
Set-up 1 Set-up 2 Set-up 3 Set-up 4 Set-up 5 Set-up 6 Set-up 7
the flat plate. All simulations have been run for 2500 MG cycles
CP 1.001 1.013 1.017 1.017 1.014 1.020 1.019 with three grid levels and CFL = 4. The convergence histories of
g (%) 39.06 39.55 39.70 39.70 39.57 39.81 39.76 the four RANS PDEs and the two turbulence model PDEs of the
SST analysis using Eq. (20) are reported in the left subplot of
Fig. 3, whereas the six convergence histories of the SST analysis
using Eq. (19) with SR = 2500 are shown in the right subplot of
nondimensionalized velocity component parallel to the flat plate the same figure. The variable on the x-axis is the number of MG cy-
on a line orthogonal to the flat plate itself at x = 0.5, computed with cles, and the variable Dlr on the y-axis is the logarithm in base 10 of
the K  x and SST analyses using either Eq. (19) with SR = 100 or the RMS of the cell residuals of the considered conservation equa-
Eq. (20), are reported in the top left subplot of Fig. 2. The variable tion normalized by the RMS of the cell residuals at the first MG
on the x-axis is the logarithm in base 10 of y+, the nondimension- iteration. The convergence histories of the simulation based on
alized wall distance, and its expression is y+ = (usd)/mw. The variable Eq. (19) are fairly oscillatory, and the overall reduction of the resid-
on the y-axis is u+, the nondimensionalized velocity component uk uals is smaller than that achieved by the simulation based on Eq.
parallel to the wall, which, in this case, is the x-component of the (20). This is due to the difference in the estimates of the strong
velocity vector. Its expression is u+ = uk/us. The subplot also reports streamwise gradient of xw at the LE on the fine and coarser grids,
Spalding’s profile, which is a power-series interpolation of experi- caused by insufficient spatial resolution in the streamwise direc-
mental data joining the linear sublayer to the logarithmic region of tion of the coarser grids. This circumstance results in oscillations
the turbulent boundary layer occurring on a flat plate in the ab- of the flow residuals strictly localized to the LE of the flat plate.
sence of a streamwise pressure gradient. It is observed that the As expected, this problem does not occur without MG. This phe-
velocity profiles computed with either turbulence model using nomenon is also absent when using Eq. (20) with or without MG,
Eq. (20) (profiles labeled ‘SST, mentw’ and ‘K  x, mentw’) are very as this condition does not introduce any significant streamwsie
close to each other, and are both fairly close to Spalding’s profile. gradient of x at the wall on any grid level. For flow problems fea-
The velocity profiles computed with either turbulence model using turing rounded LEs, the oscillatory character of the convergence
Eq. (19) with SR = 100 (profiles labeled ‘SST, wlc0w’ and ‘K  x, histories of the turbulent MG solver based on the boundary condi-
wlc0w’) are also very close to each other, but are farther away from tion (19) is less pronounced than in the right subplot of Fig. 3. This
the Spalding’s estimate with respect to the solutions obtained with is because of the smaller streamwise gradients of xw associated
146 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

0.7 0.1

0.5
0.05
0.3

0.1
0
-0.1

-0.3
-0.05
-0.5

-0.7 -0.1
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 -0.1 -0.05 0 0.05 0.1

0.15 0.04

0.13
0.02

0.11
0
0.09

-0.02
0.07

0.05 -0.04
0.45 0.47 0.49 0.51 0.53 0.55 0.96 0.98 1 1.02 1.04

Fig. 8. Medium-refinement grid for oscillating wing analyses. Top left plot: near-airfoil area. Top right plot: leading edge area. Bottom left plot: airfoil upper side area at mid-
chord. Bottom right plot: trailing edge area.

with the use of Eq. (19) with respect to the flat plate problem, Fig. 4 present the comparison between measured and computed
where the sharp LE essentially leads to a flow singularity. It is also velocity profiles at the six chordwise positions indicated by the
expected that increasing the streamwise grid refinement in the value of the x/c variable reported in each subplot. For both
wall proximity reduces the residual oscillations under analysis, the experimental and computed results, the variable y/c along
and preliminary numerical tests conducted by the authors appear the y-axis denotes the distance from the airfoil, measured along
to confirm this hypothesis. Therefore, in view of its previously dis- a line orthogonal to the chord and intersecting the chord at the
cussed advantages, the wall BC (19) with SR = 2500 has been used indicated value of x/c. The variable u/u1 along the x-axis denotes
for all other simulations presented in the remainder of this study, the velocity component parallel to the chord taken along the
unless otherwise stated. abovesaid line orthogonal to the chord itself and nondimensional-
ized by the freestream velocity u1. The computed profiles labeled
5.2. NACA4412 airfoil ‘wlc1w’ are those obtained using Wilcox’s wall BC with SR = 2500,
the computed profiles labeled ‘wlc0w’ are those obtained using
In this subsection, the turbulent flow field past the NACA4412 Wilcox’s wall BC with SR = 100, and the computed profiles labeled
airfoil corresponding to the condition of maximum lift is consid- ‘mentw’ are those obtained using Menter’s wall BC. Examination
ered. The freestream Mach number is 0.2, and the AoA is 13.87°. of the velocity profiles presented in Fig. 4 confirms that the re-
The Reynolds number based on the airfoil chord and the free- sults obtained using different wall BC for x can differ signifi-
stream velocity is 1.52  106. This operating regime is character- cantly, and also that the solution obtained using Wilcox’s wall
ized by a flow reversal in the rear part of the airfoil suction side. BC with SR = 2500 is significantly closer to that obtained with
Detailed hot-wire boundary layer measurements have been per- Menter’s wall BC. The present test case has also been used by
formed at NASA Aimes and reported in [35]. The C-grid adopted Menter for validating the predictive capabilities of the SST model,
for the flow simulations reported below is that available on the and the level of agreement between the COSA SST solution using
web site of the NASA CFD code CFL3D [32]. This grid has 177 Menter’s wall BC and the experimental results is indeed the same
points along the airfoil, 41 points in the C-cut, and 81 points in as that between Menter’s own computed solution and the exper-
the normal-like direction, giving 20,480 cells. The farfield bound- imental data [28]. The agreement between the solutions obtained
ary is at about 20 chords from the airfoil, and the distance of the with Menter’s and Wilcox’s wall BCs can be further improved by
first grid points off the airfoil surface from the airfoil surface is increasing the value of the constant SR appearing in Eq. (19) be-
4.0  103% of the chord. The turbulent flow field under investi- yond 2500, but this has not been done because the recalibration
gation has been computed using only the SST turbulence model. of this parameter is beyond the scope of this paper. The values
In order to further assess the impact of the wall BC for x on of the lift coefficient CL, the drag coefficient CD, and the pitching
the accuracy of the computed solution, three analyses have been moment coefficient CM obtained with the three COSA SST simula-
performed, one computing xw with Wilcox’s condition (Eq. (19)) tions are reported in Table 2. The significant variations of the
with SR = 2500, one using Eq. (19) with SR = 100, and one comput- force coefficients with the wall BC for x highlight once again
ing xw with Menter’s condition (Eq. (20)). The six subplots of the impact of this modeling choice on the computed forces.
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 147

3 using Wilcox’s wall BC are expected to improve further by increas-

1 2 3 4 5 6 h 80°
ing the spatial refinement of the computational grid.
2 α 60°
θ
40°
1 6. Results
20°

α,θ
0 0°
h

This section focuses on the 2D turbulent analysis of an oscillat-

-20°
-1 ing wing simultaneously heaving and pitching, and is subdivided
-40°
into two subsections. The first subsection presents the physical
-2 -60° and computational set-up of the flow analyses, along with the
KINEMATICS
-80° assessment of the temporal and spatial refinement of the grids
-3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 used in this study. The second subsection investigates in detail
t/T the unsteady flow mechanisms enabling the efficiency of the en-
ergy extraction to be controlled and possibly maximized, and it
3
1 2 3 4 5 6 analyzes the qualitative and quantitative differences between a
CPa
turbulent flow regime with Reynolds number of 1.5  106 and
2 the laminar regime analyzed in [18,9].

1 6.1. Physical and numerical problem set-up

The wing section selected for this study is the NACA0015 airfoil.
0 Here two operating conditions are considered: one characterized
by a high efficiency of the energy extraction in the laminar flow re-
CP gime considered in [9] (case A), and the other characterized by a
-1
Cy
lower efficiency in the same laminar regime (case B). In both case
vy/u∞
CM A and B, the heaving amplitude h0 equals one chord and the pitch-
-2 ing center is at xp = 1/3. Case A is characterized by a pitching ampli-
Ωc/u∞
TURBULENT tude h0 of 76.33° and a nondimensionalized frequency f⁄ = f c/u1 of
-3 0.14, where f is the frequency in Hertz. In case B, h0 = 60.00° and
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 f⁄ = 0.18. In both case A and B, M1 = 0.1. In the turbulent simula-
t/T tions of both operating conditions, the value of the Reynolds num-
ber based on the freestream velocity and the airfoil chord is
3 Re = 1.5  106; for the laminar simulations of both operating condi-
1 2 3 4 5 6
CPa
tions reported in [9] and further analyzed in the following subsec-
2 tion, Re = 1100.
The time-dependent turbulent flow fields past the oscillating
wing have been computed using multi-block moving grids. In all
1 simulations, the whole computational grid is animated simulta-
neously by a heaving and a pitching motion component defined
0 by Eqs. (21) and (22) respectively. The grid does not deform, and
it undergoes a rigid-body motion corresponding to the prescribed
CP motion of the wing. In order to assess the sensitivity of the turbu-
-1 lent solutions to the level of spatial refinement, the operating con-
Cy
vy/u∞ dition A has been simulated using three C-grids with different
CM spatial resolution, namely a mesh with 102,400 cells (coarse), a
-2
Ωc/u∞
mesh with 368,640 cells (medium) and one with 1,474,560 cells
LAMINAR (fine). In all cases, the grid coordinates are nondimensionalized
-3 by the chord of the airfoil, and the farfield boundary is at about
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
50 chords from the airfoil. The number of mesh intervals on the air-
t/T
foil (nA), the number of intervals on the C-cut (nC), the number of
Fig. 9. Analysis of oscillating wing with h0 = 76.33o and f⁄ = 0.14 (case A). Top plot: intervals in the normal-like direction (nN), and the distance of
kinematic parameters. Middle plot: turbulent regime dynamics. Bottom plot: the first grid points off the airfoil from the airfoil surface (Dw) for
laminar regime dynamics. the three grids are reported in Table 3. In order to assess the sen-
sitivity of the turbulent solutions to the level of temporal refine-
ment, the operating condition A has also been simulated with the
All simulations have been run for 2500 MG cycles with three coarse grid using 128, 256, 512 and 1024 time-intervals per period.
grid levels and CFL = 1. The convergence histories of the four RANS All turbulent simulations have been run until the maximum differ-
PDEs and the two turbulence model PDEs of the SST analysis using ence between CY over the last two oscillation cycles became less
Eq. (20) are reported in the left subplot of Fig. 5, whereas the six than 0.7% of the maximum CY over the last cycle. The number of
convergence histories of the SST analysis using Eq. (19) with oscillation cycles typically required to fulfill this requirement has
SR = 2500 are shown in the right subplot of the same figure. The varied between two and ten depending on the spatial and temporal
convergence histories of the simulation based on Eq. (20) are refinement, and also on whether the simulation has been started
smoother than those of the simulation based on Eq. (19). In both from a freestream condition or restarted from the solution of a
cases, however, all residuals decrease by at least four orders of simulation using the same grid but different temporal refinement.
magnitude. Additionally, for the same reasons discussed in the pre- The only exception is the fine grid simulation using 256 time-inter-
ceding subsection, the convergence histories of the simulation vals per period, which has required the simulation of 12 oscillation
148 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

Fig. 10. Streamlines and vorticity contours for wing positions labeled 1–4 in Fig. 9. Left plots: turbulent regime. Right plots: laminar regime.

cycles starting from a freestream condition to achieve the afore- error of the other force coefficients at a given period is similar to
mentioned periodicity error. It has been chosen to monitor the that of CY. The threshold of 0.7 has been chosen because it gives
periodicity error of CY because the vertical force component gives the best trade-off between accuracy and computational cost of
the highest contribution to the extracted power. The periodicity the analyses. Reducing the CY periodicity error below this value
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 149

Fig. 11. Vorticity contours for wing positions labeled 3–6 in Fig. 9. Left plots: turbulent regime. Right plots: laminar regime.

yields insignificant variations of the periodic solutions with respect odic flow field. The periodic profiles of the horizontal force coeffi-
to those reported in the manuscript. For all turbulent analyses of cient CX, the vertical force coefficient CY and the pitching moment
the oscillating wing presented in this report, y+ has been found coefficient CM computed with the coarse grid using the abovesaid
to be smaller than one at all grid points and all times of the peri- four levels of time-refinement (set-ups 1 to 4) are reported in
150 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

3 iod. These profiles highlight that the computed solution is largely

h 80°
independent of the number of intervals per period provided that
2 α 60°
θ at least 512 time-intervals per period are used. Some small varia-
40° tions between the solutions using 256 and 512 intervals per oscil-
1
20° lation cycle are instead observed. The periodic profiles of the three

α,θ
0 0° force coefficients computed using 256 time-intervals per period
h

-20° with the coarse, medium and fine grids (set-ups 2, 5 and 6 respec-
-1
-40° tively) are reported in Fig. 7. It is seen that some differences exist
-2 -60° among all three solutions, indicating that the solution computed
KINEMATICS with the medium grid is not completely grid-independent and
-80°
-3 the solution computed with the fine grid may also not be fully
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
grid-independent. It should be noted that the spatial resolution
t/T of the coarse and medium-refinement grids past the airfoil is very
3 similar. The medium-refinement grid has many more cells than the
CPa coarse grid in the area behind the airfoil, and this choice has been
2 made to better resolve the vortex propagation behind the wing.
The medium-refinement and the fine grids, conversely, are topo-
logically similar over the entire domain, which means that the
1 medium-refinement grid has been obtained by removing every
second grid line of the fine grid. The fact that the differences be-
0 tween the solutions of the coarse and medium-refinement grids
are smaller than those observed using the medium-refinement
and fine grids may indicate that the resolution of the vorticity field
-1 CP behind the wing does not have a strong effect on the value of the
Cy
vy/u∞ forces acting on the wing. A similar occurrence has been reported
-2 CM in computational aerodynamics studies of stalled horizontal axis
Ωc/u∞ wind turbine blades [15]. Fig. 7 also reports the force coefficients
TURBULENT determined with the medium-refinement grid using 256 intervals
-3 per period and Wilcox’s wall BC (Eq. (19)) with SR = 100 (set-up 7).
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Comparing these profiles with all other results of Figs. 7 and 6 re-
t/T
veals that the variations resulting from using different xw BCs with
3 given temporal and spatial resolutions are larger than those result-
ing from varying the spatial and temporal resolutions with a given
CPa
turbulent wall BC.
2
The values of the overall power coefficient averaged over the
considered cycle C P and the power extraction efficiency g ob-
1 tained with these seven analyses are reported in the first and sec-
ond row of Table 4 respectively. These data indicate that (a) the
difference of C P , one of the main integral functionals of engineer-
0
ing interest, obtained with the coarse grid using 256 time-inter-
vals per cycle (set-up 2) and 512 time-intervals per cycle
CP (set-up 3) is only about 0.4%, and (b) the difference of the same
-1
Cy
functional obtained using 256 time-intervals per cycle with the
vy/u∞
CM medium-refinement grid (set-up 5) and the fine grid (set-up 6)
-2
Ωc/u∞ is only about 0.6%. For these reasons, and also to keep computa-
LAMINAR tional costs within affordable limits, set-up 5 has been used for all
turbulent analyses presented in the remainder of this study. This
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 has been done also because, as shown below, the differences be-
t/T tween the examined turbulent and laminar regimes are substan-
tially larger than the solution variations highlighted in this
Fig. 12. Analysis of oscillating wing with h0 = 60o and f⁄ = 0.18 (case B). Top plot:
kinematic parameters. Middle plot: turbulent regime dynamics. Bottom plot:
subsection. Consequently, the use of a non-perfectly temporal
laminar regime dynamics. and spatial grid-independent computational set-up for the turbu-
lent analyses is not believed to affect significantly the conclusions
of the following study. Four local views of the medium-refine-
Table 5
ment grid adopted for the analyses reported in the following sub-
Power coefficient and power extraction efficiency in turbulent and laminar regimes section are provided in Fig. 8.
for operating conditions A and B. The CFL number of all turbulent simulations has been set to 3,
Turbulent Laminar
2500 MG iterations per physical time-step have been performed,
and CFL ramping has been used. All calculations have been per-
CP C Py C Ph g (%) CP C Py C Ph g (%)
formed using the PIRK MG iteration, since the maximum CFL num-
A 1.014 1.166 0.152 39.6 0.862 0.824 0.038 33.6 ber that could be used retaining the numerical stability of the FERK
B 0.450 0.707 0.257 18.8 0.274 0.397 0.123 11.4 integration has been found to be 1.
As reported in [9], the laminar analyses of both operating con-
ditions have been instead performed with 128 time-intervals per
Fig. 6. The y-axis of each subplot reports the force coefficient and period and running the simulations for eight cycles of oscillation,
the x-axis reports a time variable nondimensionalized by the per- starting from a freestream condition. This has resulted in the
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 151

0 0
ρ
ρu
-1 -1
ρv
ρE
-2 ρk -2
ρω
-3 -3

Δlr
-4 -4

-5 -5

-6 -6

-7 -7
1 2
-8 -8
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500

0 0

-1 -1

-2 -2

-3 -3
Δlr

-4 -4

-5 -5

-6 -6

-7 -7
3 4
-8 -8
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500

0 0

-1 -1

-2 -2

-3 -3
Δlr

-4 -4

-5 -5

-6 -6

-7 -7
5 6
-8 -8
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
MG cycle MG cycle

Fig. 13. Convergence histories of turbulent simulations of oscillating wing for wing positions labeled 1–6 in Fig. 9.

maximum difference between CY over the last two oscillation cy- other at 43.0% of the period, whereas the turbulent CP curve has
cles being about 0.1% of the maximum CY over the last cycle. only one pronounced peak at 23.4% of the period. The analyses re-
ported in [18,9] show that the high values of the laminar CP after
6.2. Aerodynamic analysis the first peak and the following secondary peak (that at 43.0% of
the period) are due to the optimal synchronization of the wing mo-
The evolution of the main kinematic parameters, and the force tion and the LEVS associated with dynamic stall. This results in the
and power coefficients of the oscillating wing for the operating heaving force and the heaving velocity being in phase over most of
condition A are analyzed in Fig. 9, the top subplot of which reports the cycle. As shown later in this section, however, in the turbulent
the time-dependent values of the vertical position h of the wing, its regime the generation of the LE vortex starts later than in the lam-
angular position h and the effective AoA a computed with Eq. (23). inar regime; the vortex travels close to the airfoil towards its TE
One notes that the maximum AoA achieved in case A is about 35°. with the same speed it has in the laminar regime, and therefore
The middle and bottom subplots of Fig. 9 provide the main power it leaves the TE region later than it does in the laminar regime.
and force coefficients over one period of oscillations in the turbu- As a consequence, the optimal synchronization between wing mo-
lent and laminar regime respectively. There are two important dif- tion and LEVS seen in the laminar regime is lost in the turbulent
ferences between the curve of the overall power coefficient CP in regime if the same kinematic conditions are used in both cases.
the turbulent and laminar regimes. One is that the laminar CP curve The other difference between the turbulent and laminar CP curves
over one semi-period has two positive peaks, one at 20.3% and the is that the peak value of the former (2.43) is significantly higher
152 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

than the peak value of the latter (1.48). As shown later in this sec- the turbulent case, the LEVS starts at position 4 and the main vor-
tion, this is due to the fact that, before the LEVS starts, the turbu- tex leaves the TE region at position 6, Given that the temporal dis-
lent boundary layer does not experience significant flow tance between positions 3 and 5, and that between positions 4 and
reversals, unlike what seen in the laminar regime, in which separa- 6 is the same, the streamwise velocity of the main vortex is the
tion starts at the TE and travels to the LE until the laminar LEVS same in the laminar and turbulent regimes. These observations
commences. As a consequence, in the phase preceding the turbu- highlight the loss of optimal timing between LEVS and wing mo-
lent LEVS, the turbulent lift is higher than its laminar counterpart, tion occurring in the turbulent regime caused by the delayed onset
and the turbulent drag is lower than its laminar counterpart for a of LEVS. On the other hand, comparing the vorticity levels of the
given wing position. This phenomenon explains why the turbulent main turbulent vortex (position 5) and those of the main laminar
heaving force coefficient CY is significantly higher that its laminar vortex (position 4), it is noted that the former vortex is stronger,
counterpart until about 30% of the period, when the turbulent flow and this would have a beneficial effect on the heaving power pro-
separation at the LE starts. At this point, due to the delayed start of duction if a better LEVS/wing motion synchronization were
the turbulent LEVS, the heaving force decreases much more rapidly achieved.
than in the laminar case, leading to a reduction of the extracted The evolution of the main kinematic parameters, the force and
heaving power. The middle and bottom subplots of Fig. 9 also re- power coefficients of the oscillating wing for the operating condi-
port the nondimensionalized heaving velocity vy/u1, and the com- tion B are analyzed in Fig. 12. The time-dependent values of h, h
parison of these two subplots shows that, due to the and a are reported in the top subplot of Fig. 12, which highlights
aforementioned lack of synchronization between wing motion that the maximum AoA achieved in case B is about 11.5°. The anal-
and turbulent LEVS, the laminar heaving power production re- yses reported in [18,9] show that no LEVS exists in the laminar re-
mains positive over a longer portion of the period with respect to gime for the operating condition B. The turbulent analyses reported
the turbulent case. The sign of the power contribution of the pitch- in the present study reveal that also in the turbulent regime there
ing moment to the overall power at each point of the cycle can be is no LEVS for this operating condition. The middle and bottom
determined by comparing the sign of the nondimensionalized subplots of Fig. 12 provide the main power and force coefficients
angular velocity Xc/u1 reported in the last two subplots of Fig. 9 over one period of oscillations in the turbulent and laminar regime
and that of the pitching moment coefficient CM. The general shape respectively. The shape of the turbulent and laminar CP curves is
of the CM curve in the turbulent and laminar regimes is similar, but very similar, pointing to the fact that in this operating regime the
the turbulent CM curve is shifted to the right with respect to its main aerodynamic features are the same in both flow regimes. This
laminar counterpart, and the amplitude of the turbulent pitching is confirmed by the fact that the same observation is also made for
moment is slightly larger than that of its laminar counterpart. As the heaving force and pitching moment coefficients. As a conse-
a result, the turbulent mean value of the pitching power C Ph turns quence, the phase between the heaving velocity and force, and that
out to be negative, whereas this parameter is positive in the lami- between the angular velocity and pitching moment are also very
nar regime. The value of the power coefficient C Pa associated with close in both flow regimes. The main difference between the turbu-
the available power in case A is also reported in the last two sub- lent and the laminar regimes of the operating condition B is that
plots of Fig. 9 for reference. the amplitude of all turbulent force and power coefficients is larger
To examine in greater detail the aforementioned flow phenom- than the amplitude of the laminar curves. This difference is due to
ena, the flow snapshots at the six positions denoted by the circled the reduction of the heaving force and pitching moment caused, in
labels 1–6 in Fig. 9 are considered. These labels denote the wing the laminar regime only, by thicker boundary layers and flow sep-
configuration at 12.5%, 18.7%, 25.0%, 34.4%, 45.3% and 54.7% of aration in the portion of the cycle where the AoA increases.
the cycle respectively. The four left subplots of Fig. 10 depict the The mean values of the overall power coefficient C P , the heaving
contours of the flow vorticity Xf and the streamlines when the power coefficient C Py , and the pitching power coefficient C Ph , and
wing is at positions 1–4 in the turbulent regime, whereas the four the efficiency g of the power extraction process computed by
right subplots present the same analysis for the laminar regime. At means of Eq. (26) for cases A and B in the turbulent and laminar re-
position 1, the TE separation on the lower airfoil side starts in the gimes are reported in Table 5. In case A, the efficiency in the turbu-
laminar regime, whereas no TE separation is visible in the turbu- lent condition is nearly 40%, about 6 percentage points more than
lent regime. Additionally, the vorticity contours close to the lower the laminar analysis predicts. This difference is caused by the sig-
side of the airfoil highlight that the laminar boundary layer is sub- nificantly higher values of the turbulent C Py . As discussed above,
stantially thicker than the turbulent boundary layer, due to Rey- this occurrence is due to the fact that the increase of the turbulent
nolds number effects. At position 2, the laminar separation has CY due to a thinner and more stable boundary layer outweighs the
already passed the mid-chord position of the airfoil, whereas only reduction of heaving power caused by the loss of synchronization
a small TE separation has now appeared in the turbulent regime. At between heaving motion and LEVS. The turbulent C Py could be fur-
position 3, the laminar separation has reached the LE and the LEVS ther increased by re-tuning the kinematic parameters in order to
begins, whereas the turbulent TE separation has slightly increased improve the abovesaid synchronization. Note also that the turbu-
and the turbulent boundary layer on the lower side has thickened. lent C Ph is negative, whereas the laminar C Ph is positive. This is a
At position 4, the strong laminar vortex has already achieved the consequence of the loss of synchronization between pitching mo-
mid-chord position, whereas the turbulent LEVS has just started. tion and LEVS. Also this power loss could be removed by optimiz-
These four snapshots confirm that the turbulent boundary layer ing the wing kinematics for the turbulent regime. In the operating
of the high-Reynolds number regime is substantially thinner and regime B, the turbulent efficiency is nearly 19%, about 7 percentage
less prone to separation than its laminar counterpart in the low- points more than the laminar analysis predicts. No LEVS is present
Reynolds number regime, and this explains the higher levels of ex- in either flow regime for this operating condition, and this effi-
tracted power before the turbulent LEVS starts. The evolution of ciency difference is entirely due to the thinner turbulent boundary
the turbulent and laminar LEVS is compared in Fig. 11, where the layer and the absence of flow separation in the turbulent regime.
four left subplots report the Xf contours when the wing is at posi- As for the convergence of the turbulent simulations, the choice
tions 3 to 6 in the turbulent regime, whereas the four right sub- of the numerical control parameters reported in the preceding sub-
plots present the same variable for the laminar regime. It is section has resulted in all force coefficients achieving full conver-
noted that, in the laminar case, the LEVS starts at position 3 and gence at all physical times of the wing motion, and the RMS of
the main vortex leaves the TE region at position 5, whereas, in the residuals of the mean flow and the x equations dropping by
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 153

at least four orders of magnitude. The residual of the K equation have been reported, along with a novel point-implicit Runge–Kutta
has instead dropped by between 1.5 and 2 orders in all cases. This integration approach for URANS analyses featuring this type of tur-
is highlighted in the convergence histories reported in the six sub- bulence model. Such an approach allows the use of significantly
plots of Fig. 13, which refer to the six wing positions labeled 1–6 in higher CFL numbers (up to 3 times those usable with the conven-
Fig. 9. Similar residual drops have also been observed solving test tional fully-explicit RK integration), yielding significant reductions
cases of the same type of the oscillating wing using a commercial of the overall computational time.
package with numerical methods and modeling features similar
to those of COSA. Acknowledgments
Due to the complex turbulent aerodynamics of the oscillating
wing, the use of Large Eddy Simulation (LES) may yield more accu- This work made use of the facilities of N8 HPC provided and
rate predictions of the flow field. However, it is hard to foresee the funded by the N8 consortium and EPSRC (Grant No. EP/K000225/
level of improvement in the prediction of the timedependent force 1). The Centre is co-ordinated by the Universities of Leeds and
and power components. Furthermore, LES simulations for the con- Manchester.
sidered high Reynolds number would require extremely large com-
putational resources. The turbulent analyses presented above
provide a significant contribution to the aerodynamic knowledge Appendix A. Numerical integration of time-dependent
base of this device based on the relatively cheaper URANS technol- problems
ogy. These URANS data will be a valuable source for cross compar-
isons with detailed high-Reynolds number measurements, which The physical time-derivative of system (1) is discretized with a
are likely to be available before LES data. second order backward finite-difference. The set of nonlinear alge-
braic equations resulting from the space- and time-discretization
of system (1) is then solved using the same explicit integration
7. Conclusions procedure used for steady problems. The discretization of the
physical time-derivative of the unknown flow state by means of
A comparative analysis of the unsteady aerodynamics of the en- a second order backward finite difference and the introduction of
ergy-extracting oscillating wing device in two limiting laminar and the fictitious term V(dQ/ds)n+1 yield the equation:
turbulent regimes has been performed by means of a compressible nþ1
NS research code using the K  x SST turbulence model for the tur- dQ
V þ Rg ðQ nþ1 Þ ¼ 0 ðA:1Þ
bulence closure. ds
The comparative aerodynamic analysis of the oscillating wing
where
device operating in a laminar flow regime at Re = 1100 and a fully
turbulent regime at Re = 1.5  106 has been performed for two 3Q nþ1  4Q n þ Q n1
different kinematic operating conditions: a high-laminar effi- Rg ðQ nþ1 Þ ¼ V þ RU ðQ nþ1 Þ ðA:2Þ
2 Dt
ciency condition (case A), characterized by a nondimensionalized
motion frequency of 0.14 and a pitching amplitude of 76.33°, and The symbol Rg denotes a residual vector which also includes the
a low-laminar efficiency condition (case B), characterized by a source terms associated with the discretization of the physical
nondimensionalized motion frequency of 0.18 and a pitching time-derivative @U/@t contained in Eq. (1). Note that, also for
amplitude of 60°. The comparative turbulent/laminar analysis of time-dependent problems with moving bodies, the matrix V is inde-
case A shows that the optimal synchronization between wing mo- pendent of the physical time-level (denoted by the superscripts
tion and LEVS observed in the laminar regime is lost when oper- n + 1, n and n  1) because in this report only rigid-body grid mo-
ating in the turbulent regime. Despite this, and thanks to the tion is considered. The symbol Dt indicates the user-given physical
higher stability and lower thickness of the turbulent boundary time-step. Eq. (A.1) can thus be viewed as a system of (Npde  Ncell)
layer, the power extraction efficiency in the turbulent regime is ordinary differential equations in which the unknown is Qn+1, the
nearly 40%, whereas that in the laminar regime is about 34%. flow state at time-level n + 1. The calculation of Qn+1 is performed
The dependence of the phase between wing motion and LEVS iteratively by discretizing the fictitious time-derivative (dQ/ds)n+1
on the Reynolds number points to the importance of incorporat- of Eq. (A.1) with a four-stage RK scheme, and marching the equa-
ing turbulent flow effects in the optimization of the kinematic tions in pseudo-time until a steady state is achieved. Such steady
parameters of this device when its operating regime is fully tur- state is the flow solution for the physical time being considered.
bulent. By doing so, an efficiency even higher than that of 40% re- Similarly to the case of steady flows, the convergence rate is en-
ported herein for the considered turbulent Reynolds number is hanced by means of LTS, variable-coefficient central IRS and FAS
likely to be achieved. The operating condition B is characterized MG. Once the flow solution at the physical time-level n + 1 has been
by the absence of LEVS. Also in this case, the turbulent estimate calculated, the array Qn is moved to Qn1, the array Qn+1 is moved to
of the efficiency is higher than its laminar counterpart, but this Qn, and the calculation of a new time-level is started. This procedure
difference appears to be caused by the different characteristics corresponds to Jameson’s dual-time-stepping approach to the inte-
of the turbulent and laminar boundary layers at the considered gration of time-dependent problems.
Reynolds numbers. This solution procedure may become unstable when the physi-
The study has also highlighted the strong impact of the turbu- cal time-step Dt is significantly smaller than the pseudo-time-step
lent wall boundary condition for the specific dissipation rate x Ds. This instability was reported in [1], and thoroughly investi-
on the computed steady and time-dependent solutions. Systematic gated by Melson et al. [27]. The latter study, considering the simu-
analyses highlight that different choices of such a BC lead to vari- lation of turbulent flow problems by means of the thin-layer
ations of the solution significantly larger than those observed by Navier–Stokes and the algebraic Baldwin–Lomax turbulence model
using different turbulence models with the same wall BC, and also [4], elegantly solved the stability problem by treating implicitly the
larger than those observed by varying the temporal and spatial res- Qn+1 term of the physical time-derivative within the RK integration
olutions of the simulation for given turbulence model and turbu- process. In the COSA solver, this strategy has been generalized and
lent wall BC. The mathematical models underlying the developed for use with the differential K  x SST model. The resid-
implementation of the turbulent steady and time-domain solver ual Rg is split into the contribution depending on the Qn+1 term of
154 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155

the physical time- derivative, and a term Rd equal to the difference References
of Rg and the aforesaid Qn+1 term:
  [1] Arnone A, Liou M-S, Povinelli LA. Multigrid time-accurate integration of
V 3 nþ1 Navier–Stokes equations. Technical memorandum NASA TM 106373 ICOMP-
Rg ðQ nþ1 Þ ¼ Q þ gðQ n ; Q n1 Þ þ RU ðQ nþ1 Þ 93-37. Cleveland (OH, USA): Lewis Research Center; 1993.
Dt 2 [2] Ashraf MA, Young J, Lai JCS, Platzer MF. Numerical analysis of an oscillating-
wing wind and hydropower generator. AIAA J 2011;49(7):1374–86.
where g(Qn, Qn1) =  2Qn + 0.5Qn1. This equation can also be [3] Baldwin BS, Barth TJ. A one-equation turbulence transport model for high
written as: Reynolds number wall-bounded flows. Technical memorandum NASA TM
102847. Hampton (VA, USA): Lewis Research Center; 1990.
3V nþ1 [4] Baldwin BS, Lomax H. Thin-layer approximation and algebraic model for
Rg ðQ nþ1 Þ ¼ Rd ðQ nþ1 Þ þ Q ðA:3Þ separated turbulent flows. AIAA paper 78-257, January 1978. 16th AIAA
2 Dt
aerospace sciences meeting, Huntsville, Alabama.
Discretizing the fictitious time-derivative of Eq. (A.1) with a multi- [5] Betz A. Das Maximum der Teoretisch Möglichen Ausnützung des Windes
Durch Windmotoren. Z Gesamte Turbinenwesen 1920;20:307–9.
stage RK scheme, introducing the decomposition of Rg provided by [6] Bradley RA, Platzer MF. Oscillating-wing power generator with flow-induced
Eq. (A.3), and considering the Qn+1 term at stage k rather than at pitch-plunge phasing. US patent US20090121490 2009;A1.
stage (k  1) yields the following modified RK algorithm: [7] Campobasso MS, Baba-Ahmadi MH. Ad-hoc boundary conditions for CFD
analyses of turbomachinery problems with strong flow gradients at Farfield
boundaries. J Turbomach 2011;133(4).
W0 ¼ Ql [8] Campobasso MS, Baba-Ahmadi MH. Analysis of unsteady flows past horizontal
ðI þ ak ðbTD I þ DsAÞÞWk ¼ W0 þ ak DsAWk1  axis wind turbine airfoils based on harmonic balance compressible Navier–
ðA:4Þ Stokes equations with low-speed preconditioning. J Turbomach 2012;134(6).
ak DsV 1 Rd ðWk1 Þ [9] Campobasso MS, Drofelnik J. Compressible Navier–Stokes analysis of an
oscillating wing in a power-extraction regime using efficient low-speed
Q lþ1 ¼ WNS preconditioning. Comput Fluids 2012;67:26–40.
[10] Da Ronch A, McCracken AJ, Badcock KJ, Widhalm M, Campobasso MS. Linear
where bTD = 1.5Ds/Dt, l is the RK cycle counter, and Ql is shorthand frequency domain and harmonic balance predictions of dynamic derivatives. J
Aircraft 2013;50(3):694–707.
for Q nþ1
l . The stability analysis of [27] shows that the stability of [11] Ekaterinaris JA, Menter FR. Computation of oscillating airfoil flows with one-
algorithm (A.4) no longer depends on the ratio Ds/Dt. However this and two-equation turbulence models. AIAA J 1994;32(12):2359–65.
formulation is still unsuitable when IRS and MG are also used, be- [12] Eliasson P, Wallin S. A positive multigrid scheme for computations with two-
equation turbulence models. In: European congress on computational methods
cause both acceleration techniques have to be applied to a residual in applied sciences and engineering, Barcellona, Spain; September 2000.
term that vanishes at convergence, and this is not the case of Rd. The [13] Ferrer E, Munduate X. CFD predictions of transition and distributed roughness
solution is to introduce the residual Rg which does vanish at conver- over a wind turbine airfoil. AIAA paper 2009-269, January 2009. 47th AIAA
aerospace sciences meeting including the new horizons forum and aerospace
gence. Given that: exposition, Orlando, Florida.
[14] Jameson A, Baker TJ. Solution of the Euler equations for complex
DsRd ðWÞ ¼ bTD VW þ DsRg ðWÞ configurations. AIAA paper 83-1929, July 1983. 6th AIAA computational fluid
dynamics conference, Danvers, Massachusetts.
the IRS-MG-tailored counterpart of algorithm (A.4) is: [15] Johansen J, Sorensen NN, Michelsen JA, Schreck S. Detached-eddy simulation of
flow around the NREL phase VI blade. Wind Energy 2002;5:185–97.
W0 ¼ Ql [16] Jones KD, Lindsey K, Platzer MF. An investigation of the fluid-structure
k
interaction in an oscillating-wing micro-hydropower generator. In:
ðI þ ak ðbTD I þ DsAÞÞW ¼ W0 þ ak ðbTD I þ DsAÞWk1  Chakrabarti, Brebbia, Almozza, Gonzalez-Palma, editors. Fluid Structure
ðA:5Þ Interaction 2. Southampton, United Kingdom: WIT Press; 2003. p. 73–82.
ak DsV 1 LIRS ðRg ðWk1 Þ þ f MG Þ [17] Jones WP, Launder BE. The calculation of low-Reynolds-number phenomena
with a two-equation model for turbulence. Int J Heat Mass Transfer
Q lþ1 ¼ WNS 1973;16:1119–30.
[18] Kinsey T, Dumas G. Parametric study of an oscillating airfoil in a power-
Note that the matrix multiplying Wk at the second line of algorithm extraction regime. AIAA J 2008;46(6):1318–30.
(A.5) is block-diagonal with Ncell blocks. In each block the top left [19] Kinsey T, Dumas G. Computational fluid dynamics analysis of a hydrokinetic
turbine based on oscillating hydrofoils. J Fluids Eng
(4  4) partition is proportional to the identity matrix through the
2012;134:021104.1–021104.16.
coefficient (1 + akbTD), the bottom right (2  2) partition is given [20] Kinsey T, Dumas G, Lalande G, Ruel J, Mehut A, Viarogue P, et al. Prototype
by the sum of the (2  2) identity matrix multiplied by (1 + akbTD) testing of a hydrokinetic turbine based on oscillating hydrofoils. Renew Energy
and a non-diagonal (2  2) given by Eq. (17) or Eq. (18), depending 2011;36:1710–8.
[21] Koubogiannis DG, Athanasiadis AN, Giannakoglou KC. One- and two-equation
on whether the exact or approximate update of (qx) is used, and all turbulence models for the prediction of complex cascade flows using
other entries are zero. Similarly to the case of the integration of the unstructured grids. Comput Fluids 2003;32:403–30.
steady equations, this structure of the matrix premultiplying Wk re- [22] Lee S, Choi DW. On coupling the Reynolds-averaged Navier–Stokes equations
with two-equation turbulence model equations. Int J Numer Methods Fluids
sults in the coupling of the update process of the turbulent vari- 2006;50(2):165–97.
ables, whereas it still enables the four mean flow variables to be [23] Lin FB, Sotiropoulos F. Strongly-coupled multigrid method for 3-D
updated without any actual matrix inversion. Due to the fact that incompressible flows using near-wall turbulence closures. J Fluids Eng
1997;119:314–24.
the Qn+1 term arising from the backward finite-difference of the [24] Liu F, Zheng X. Staggered finite volume scheme for solving cascade flows with
physical time-derivative is evaluated at stage k, algorithm (A.5) is a K  x turbulence model. AIAA J 1994;32(8):1589–97.
said to be based a point-implicit Runge–Kutta (PIRK) integration [25] Liu F, Zheng X. A strongly coupled time-marching method for solving the
Navier–Stokes and K  x turbulence model equations with multigrid. J
of the time-dependent mean flow and turbulence equations. The Comput Phys 1996;128(2):289–300.
standard fully explicit Runge–Kutta (FERK) integration method is [26] McKinney W, DeLaurier J. The Wingmill: an oscillating-wing windmill. J
retrieved by setting bTD = 0 in this algorithm. The integration Energy 1981;5(1):109–15.
[27] Melson ND, Sanetrik D, Atkins HL. Time-accurate Navier–Stokes calculations
scheme of the steady equations is instead obtained by also replacing
with multigrid acceleration. In: Proc. 6th copper mountain conference on
Rg with RU in algorithm (A.5). Several numerical tests, including the multigrid methods; 1993. p. II423–37.
analyses of the oscillating wing presented in this paper, have high- [28] Menter FR. Two-equation turbulence-models for engineering applications.
lighted that the turbulent PIRK integration significantly improves AIAA J 1994;32(8):1598–605.
[29] Park SH, Kwon JH. Implementation of K  x turbulence models in an implicit
the stability of the fully-coupled integration, enabling stable pseu- multigrid method. AIAA J 2004;42(7):1348–57.
do-time-marching with larger CFL numbers than with the standard [30] Platzer MF, Ashraf MA, Young J, Lai JCS. Development of a new oscillating-wing
FERK integration. This yields significant reductions of runtimes, due wind and hydropower generator. AIAA paper 2009-1211, January 2009. 47th
AIAA aerospace sciences meeting including the new horizons forum and
to the reduction of the overall number of MG cycles required to aerospace exposition, Orlando, Florida.
achieve a user-given reduction of the flow residuals.
 M. Sergio Campobasso et al. / Computers & Fluids 88 (2013) 136–155 155

[31] Platzer MF, Young J, Lai JCS. Flapping-wing technology: its potential for air [36] Wang J, Liang C, Miesch M. High-order accurate CFD simulation of an
vehicle propulsion and airborne power generation. ICAS paper 2008-1.5.2, oscillating-wing wind power generator. AIAA paper 2013-0990, January
September 2008. 26th Congress of the international council of the aeronautical 2013. 51st AIAA Aerospace sciences meeting including the new horizons
sciences, Anchorage, Alaska. forum and aerospace exposition, Grapevine (Dallas/Ft. Worth Region), Texas.
[32] Rumsey CL. CFL3D Version 6.6, June 2011. http://cfl3d.larc.nasa.gov/+. [37] Wasserman M, Mor-Yossef Y, Yavneh I, Greenberg JB. A robust implicit
[33] Spalart PR. Strategies for turbulence modelling and simulations. Int J Heat multigrid method for RANS equations with two-equation turbulence models. J
Fluid Flow 2000;21(3):252–63. Comput Phys 2010;229:5820–42.
[34] Spalart PR, Allmaras SR. A one-equation turbulence model for aerodynamic [38] Wilcox DC. Reassessment of the scale-determining equation for advanced
flows. La Rech Aerospatiale 1994;1:5–21. turbulence models. AIAA J 1988;26(11):1299–310.
[35] Wadcock AJ. Structure of the turbulent separated flow around a stalled airfoil. [39] Wilcox DC. Formulation of the k  x turbulence model revisited. AIAA J
Contract report NASA CR-152263, NASA Aimes Research Center, Moffett Field, 2008;46(11):2823–38.
CA, USA; February 1979.