You are on page 1of 11

Time-averaging and temporal-filtering in

wall-modeled large eddy simulation


Cite as: Phys. Fluids 33, 035108 (2021); https://doi.org/10.1063/5.0039651
Submitted: 08 December 2020 . Accepted: 26 January 2021 . Published Online: 03 March 2021

H. Hosseinzade, and D. J. Bergstrom

ARTICLES YOU MAY BE INTERESTED IN

Grid-point and time-step requirements for direct numerical simulation and large-eddy
simulation
Physics of Fluids 33, 015108 (2021); https://doi.org/10.1063/5.0036515

On the H-type transition to turbulence—Laboratory experiments and reduced-order modeling


Physics of Fluids 33, 024105 (2021); https://doi.org/10.1063/5.0036082

Taylor–Couette flows undergoing orthogonal rotation subject to thermal stratification


Physics of Fluids 33, 035107 (2021); https://doi.org/10.1063/5.0035546

Phys. Fluids 33, 035108 (2021); https://doi.org/10.1063/5.0039651 33, 035108

© 2021 Author(s).
Physics of Fluids ARTICLE scitation.org/journal/phf

Time-averaging and temporal-filtering


in wall-modeled large eddy simulation
Cite as: Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651
Submitted: 8 December 2020 . Accepted: 26 January 2021 .
Published Online: 3 March 2021

H. Hosseinzadea) and D. J. Bergstrom

AFFILIATIONS
Department of Mechanical Engineering, University of Saskatchewan, Saskatoon, Saskatchewan S7N 5A9, Canada

a)
Author to whom correspondence should be addressed: hadi.hosseinzade@usask.ca

ABSTRACT
A turbulent channel flow at a Reynolds number of Res ¼ 2000 is solved based on the spatially filtered Navier–Stokes equations using large
eddy simulation and an in-house code. A nonequilibrium wall model is implemented to predict the flow in the wall layer based on the
Reynolds-averaged approach. To mitigate the log-layer mismatch, which is often encountered in wall modeling, two temporal schemes are intro-
duced to average the wall layer solution and to filter the flow information input to the wall layer. It is found that the time periods used for the
time-averaging and temporal-filtering schemes affect the performance of the wall model. The results show that shorter time periods enable the
wall model to respond to the flow structures in the outer layer and correctly predict the friction velocity. However, the prediction of the friction
velocity also depends on the location of the matching point. Locating the matching point further from the wall results in better performance
due to the compatibility of the subgrid scale model with the grid resolution further from the wall. The temporal-filtering scheme is used to
remove nonessential high-frequency wavelengths that can disturb the functionality of wall modeling. Various combinations of the time-
averaging and temporal-filtering time periods are investigated for different locations of the matching point. Overall, it is concluded that using a
shorter period for time-averaging and a temporal-filtering period comparable to the turbulent diffusion timescale leads to improved results.
Published under license by AIP Publishing. https://doi.org/10.1063/5.0039651

I. INTRODUCTION shear stress to the velocity in the core. Assuming the plane-averaged
Wall-bounded flows are still a challenging research topic in the velocity at the first grid point to be located in the logarithmic law
simulation of high-Reynolds number turbulent flows.1–4 Large eddy region, he proposed that the local wall shear stress fluctuates around
simulation (LES) as an innovative approach to resolve unsteady tur- the mean value proportional to the local instantaneous velocity. This
bulent flows suffers from the excessive computational cost of resolv- simple algebraic correlation was adequate in the absence of pressure
ing flow regions close to solid surfaces, which are characterized by gradient effects.9 However, the advection and pressure gradient effects
very small length scales. The most energetic motions in these near- on high-Reynolds numbers and separated flows are significant and
wall regions are typically scaled with a spatial filtering scheme as small should be included in the model for the wall region.10 Balaras et al.11
as the viscous length scale, which decreases with increasing Reynolds suggested solving the full boundary layer partial differential equations
number.5 This means that the required grid resolution in the near- to estimate the wall shear stress. Subsequently, various wall models
wall regions increases as a function of Reynolds number. Chapman6 have been developed as categorized in Fig. 1. All these wall models,
indicated that the number of grid points varies as Ntotal / Re2:4x . The
except for the wall function approach, implement a form of the
high computational cost associated with resolving the flow on a fine momentum and continuity equations on a separate grid to resolve the
grid in the wall region at high-Reynolds numbers has led to the devel- wall layer. The additional grid can be embedded inside the LES
opment of alternative strategies near solid walls. In one of these domain or be employed as a zonal layer between the LES domain and
approaches, referred to as wall modeling, the flow in the wall layer is the wall. In this respect, the wall models can be divided into two
modeled by replacing the no-slip boundary conditions with the groups: hybrid RANS/LES (Reynolds-averaged Navier–Stokes) and
instantaneous wall shear stress. wall stress models. In the hybrid RANS/LES models, two distinct
The idea of wall layer modeling was introduced by Deadroff7 and regions, i.e., the near-wall region and the outer region, are indepen-
later further developed by Schumann.8 Schumann used an approxi- dently resolved using the RANS and LES approaches, respectively. The
mate wall boundary condition for a channel flow to relate the wall flow information is transferred only at the interface, which increases

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-1


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

Another topic that should be addressed in wall modeling is the


period of time used for averaging the flow field. Wall functions and
equilibrium wall models implement a time-averaging scheme over the
entire simulation period.8,23 The full RANS equations are applied to
resolve the wall layer in some nonequilibrium wall models.21 Taking
the average over too long a time period makes the wall model unre-
sponsive to the essential stress-carrying flow structures in the outer
layer. In contrast, other wall models apply a time-averaging scheme on
the thin boundary layer equations (TBLEs) to resolve the flow in the
near-wall region.1,23 The time-averaged approach implemented in this
study uses a predefined time period. Flow properties in the wall layer
are averaged over a certain number of (previous) time steps to provide
a more realistic instantaneous wall shear stress to the outer flow.
However, choosing the number of time steps for time averaging is a
FIG. 1. Categories of wall models for high-Reynolds number turbulent flows. challenge.
The wall layer also needs to respond to flow variations in the
LES. The flow information from the LES domain is used as input to
the wall layer. The input should correspond to the timescale of the
wall layer since the wall model is implemented in an unsteady RANS
the challenge of coupling the two layers, especially on unstructured
(URANS) framework. Otherwise, the wall model could be adversely
grids.12 Heinz13 reviewed the performance of hybrid methods on sepa-
affected by large velocity fluctuations. Thus, a temporal filtering is
rated flows and presented promising models for this category of wall
required to filter the input information and provide the wall layer with
modeling. However, the main drawbacks of the hybrid wall model are
the low-frequency fluctuations in the LES as a top boundary condi-
the artificial buffer layer and unrealistic physical structures, which can
tion.19,24 The temporal filtering is applied to the flow data transferred
contaminate the LES domain at some distance away from the match-
from the LES domain to the wall layer at the interface. This paper
ing layer.14 In contrast, the wall stress models return the wall shear
presents the results of wall-modeled LES in a high-Reynolds number
stress as a new boundary condition to the LES, and the wall-modeling
channel flow at Res ¼ 2000. The effects of the timescales used in time
grid is embedded inside the LES domain. There are different
averaging and temporal filtering on the wall-model performance are
approaches for the wall stress models as shown in Fig. 1. For a more
systematically studied.
detailed insight into each group, see Piomelli and Balaras,15 Larsson
et al.,14 and Bose and Park.9 The wall modeling allows the use of a
coarse grid for the LES in the near-wall regions where there are small II. GOVERNING EQUATIONS
flow structures. The effects of these small-scale motions must be The finite-volume method is used to discretize the incompress-
expressed in a time-averaged framework.15 Since the flow structures in ible spatially filtered Navier-Stokes equations with a second-order
the wall layer experience several life cycles during a single time step of accurate central differencing scheme on a structured collocated grid. A
the outer layer, the instantaneous governing equation is unable to two-step fractional step method and a second-order Adam–Bashforth
resolve them. scheme temporally resolve the governing equations. An alternating
Chapman6 estimated that the number of the grid points required direction implicit solver based on the tri-diagonal matrix algorithm is
to resolve a wall-modeled flow is significantly reduced (Ntotal / Re0:4 x ). applied for the pressure correction. Applying the spatial-filtering oper-
Although wall models can alleviate the computational costs in terms ation to the continuity and Navier–Stokes equations in LES results in
of memory requirements and run-times, they often result in poor pre- the following governing equations:
dictions for the skin friction that is related to the so-called “mismatch”
of the mean velocity profile in the overlap region.16–19 LES inherently e
u i;j ¼ 0; (1)
solves for the instantaneous velocity field, while wall models are based e
p ;i
on time-averaged approaches. Therefore, the wall layer experiences u_ i þ ðg
e ui uj Þ;j ¼  þ eu i;jj  sij;j ; (2)
q
large temporal fluctuations in velocity at the top boundary, whereas
overall it varies smoothly in time. This issue leads to the mismatch where e indicates a filtered parameter and sij are the SGS stresses.
between the flow properties in the wall layer and LES domain at their A dynamic nonlinear subgrid scale (SGS) model (DNM) is
interface. Regardless of what the wall-modeling approach is employed implemented to model the SGS stresses based on both the sym-
from the categories shown in Fig. 1, the wall modeling potentially suf- metric (strain rate) and antisymmetric (rotation rate) parts of the
fers from the mismatch.3,4,14,19,20 This study attempts to address the filtered velocity gradient.25 The DNM estimates the SGS stresses
mismatch issue using a sophisticated wall-modeling scheme that has from a nonlinear constitutive relation to close the momentum
demonstrated acceptable performance in different flow geometries. equation,
The so-called dynamic wall models have been developed to accommo-
date the unsteadiness at the top boundary and reduce the mismatch.21 sij ¼ CS bij  CW cij  CN gij ; (3)
The performance of the dynamic wall model proposed by Park and e 2 jSjS
g , c ¼ 4D e 2 ðe e2
where bij ¼ 2D S ik X e ike
e kj  X S kj Þ, and gij ¼ 4D
Moin22 in the simulation of a high-Reynolds number channel flow is ij ij
adopted. ðe
S ike
S kj  1 e
3 S mne
S nm dij Þ. D is the filter size, dij is the Kronecker delta,

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-2


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

FIG. 2. Interaction between the wall layer


and LES domain in a side view of the
channel flow: (a) the extent of the wall
layer on the vertical symmetry plane
showing streamwise velocity contours and
(b) a schematic of wall-modeling grid
embedded in the LES domain.

and Sij and Xij represent the strain rate and rotation rate, respectively. Since the wall layer is resolved using a form of the URANS equa-
Hereafter, the flow parameters in the LES domain are shown without tions, a time-averaging operator is required to average the flow param-
the tilde for simplicity. The local values of the coefficients CS ; CW , and eters. An exponentially weighted-averaging function ensures that the
CN are determined by applying the Germano identity to the constitu- flow parameters in recent time are more significant than older values.
tive relation evaluated on the grid scale and a coarser test grid.26 It is defined as follows:28
The wall layer is embedded in the LES domain; it extends ð
from the wall to meet the LES domain at the matching point as  1 t
/ðtÞ ¼ /ðt 0 ÞeDt=Tav dt 0 ; (6)
shown in Fig. 2. At the matching plane, the velocities and pressure Tav 1
are set as Dirichlet conditions for the wall layer, and the wall shear
where t and t 0 are the current and previous times, respectively, Tav is
stress is returned to the LES as a boundary condition. A dynamic
the characteristic averaging timescale, and Dt is the computational
nonequilibrium wall model is used to model the region near the
wall.22 This dynamic wall model is implemented in a set of time step. Using a first-order approximation and Leibniz’s rule for dif-
URANS equations based on the TBLE formulation proposed by ferentiation under an integral, Eq. (6) can be simplified into a linear
Balaras et al.,11 i.e., correlation.29 As such, a combination of weighted terms based on the
  previous time-averaged value and the new value is implemented to
@ui @ @ 1 @ p @ ui
@ estimate the time-averaged parameter in the current time and reduce
þ ð
u1uiÞ þ
 ð
unuiÞ ¼ 
 þ ð þ  t Þ ;
@t @x1 @xn q @xi @xn @xn the memory load,30
(4)  nþ1 ¼  /nþ1 þ ð1  Þ /
n ;  ¼ Dt=Tav
/ ; (7)
where n is the wall-normal direction and i ¼ 1 and 3 represent the 1 þ Dt=Tav
streamwise and spanwise directions, respectively. The overbar indi-
n
 is the time-averaged parameter from the previous time step
cates that the parameter is time-averaged. The value of the wall- where /
u n ) is estimated by solving the continuity equation to
normal velocity ( and /nþ1 is the new value. The characteristic averaging timescale is
conserve mass in the wall layer. The pressure gradient in the wall- comparable to the convective timescale taken as a ratio of the channel
normal direction is assumed to be negligible, and the instantaneous height to the velocity at the center of the channel. The time-averaging
pressure specified at the matching point is based on the LES field. An scheme using this strategy ensures that a wall model that can respond
eddy viscosity model is used to predict the residual stresses; it is to the instantaneous flow structures in the LES.
obtained from the relation Recall that the wall layer is coupled to the instantaneous velocity
 ij S ij and pressure fields in the LES at its upper boundary as shown in
R
 t ¼  t þ   ; (5) Fig. 2. These flow parameters can fluctuate at high frequencies and
2S ij S ij with short wavelengths. The input of high-frequency fluctuations to
where the modeled eddy viscosity,  t , is obtained from Prandtl’s the wall layer can lead to an overprediction or underprediction of the
mixing-length hypothesis ( t ¼ ðjyÞ2 jSjD), where the von Karman skin friction that becomes manifest in the mean velocity field as the
constant is j ¼ 0:41. The van Driest wall-damping function is log-layer mismatch.14 One solution to mitigate this mismatch is to
included as D ¼ ½1  exp ðyþ =Aþ Þ2 , with Aþ ¼ 26, and Rij is the filter the input flow data at the matching point over a particular time
resolved Reynolds stress in the wall layer. period before inputting it to the wall layer.19 Eliminating the high-
To begin the simulation, the initial velocities were estimated frequency unsteadiness reduces the log-layer mismatch. Hence, the
based on Spalding’s mean velocity profile27 and introducing random top boundary of the wall layer becomes more compatible with the wall
fluctuations on the order of 5% of the local mean velocity. The simula- model’s averaging timescale. A similar exponentially weighted opera-
tion was run for a sufficient time period to transition from the initial tor is employed to filter the velocity and pressure parameters in the
random fluctuations to real turbulence. LES at the matching point,24

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-3


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

Dt friction velocity (us ) is estimated based on the flow data in the wall
h/ðtÞi ¼ e /ðtÞ þ ð1  eÞ h/ðt  DtÞi; e¼ ; (8)
Tfl layer and used to nondimensionalize the mean velocities and Reynolds
stresses in both the LES and wall-modeling domains. The time-
where h:i represents a temporal-filtered parameter. In contrast to the averaged velocity at the first grid point of the wall layer is used to esti-
time averaging, the characteristic timescale Tfl used in temporal filter- mate the wall shear stress (sw ¼ l du dy ) and subsequently the friction
ing is defined based on the turbulent diffusion timescale across the velocity.
wall layer. As shown in Fig. 3, the mean velocity profile obtained from the
Studies have found that locating the matching point between the LES is in good agreement with both the DNS results of Hoyas and
top of the viscous layer (yþ > 50) and the outer layer (y=d < 0:2) Jimenez32 and the wall-modeled LES of Park and Moin.22 The velocity
accurately predicts the effects of the solid surface.31 In high-Reynolds profiles in both Fig. 3(a) using inner coordinates and Fig. 3(b) based
number flows, the efficacy of the SGS model deteriorates significantly on the outer coordinates match the reference profiles, although there
in the coarse mesh near the wall.4 This issue motivates locating the is an overprediction in the region close to the wall. This deviation is
matching point further into the LES domain where the grid resolution expected based on previous studies and relates to the fact that the SGS
is compatible with the requirements of the SGS model. The effects of model is unable to properly model the subgrid scale flow structures for
the matching point’s location on the time-averaging and temporal- the very coarse grid near the wall. The results of wall modeling have
filtering methods are systematically investigated in the present study. the potential to be used as an alternative for the prediction of the
III. NUMERICAL SETUP mean velocity below the matching point. This issue will be discussed
The computational domain for the turbulent channel flow at further in following paragraphs. The results of the WMLES overpre-
Res ¼ 2000 was 2pd  2d  43 pd in the streamwise, wall-normal, and dict the mean velocity profile in comparison to the results of Park and
spanwise directions, respectively, where d is the half of the channel Moin.22 This might be due to the domain size implemented by Park
height. A uniform grid was used in all directions with 96  64  64 and Moin,22 which is much longer and wider than the present flow
(Nx ; Ny ; Nz ) control volumes giving a resolution of (Dxþ ; Dyþ ; Dz þ ) domain. The effects of domain size in the streamwise direction have
 (130, 63, 130). The matching point is located at the fifth node off been studied for turbulent channel flows using DNS by Abe et al.,33
the wall where yþ  272, and a nonuniform grid with 30 grid points Lozano-Duran and Jimenez,34 and Lee and Moser.35 The results indi-
is implemented to resolve the wall layer with y1þ  0:8. The grid reso- cate that the large-scale flow structures diminish when the size of the
lution in wall-parallel planes is similar to that for the LES. The wall flow domain decreases for the same Reynolds number and grid resolu-
layer statistics are averaged over the convective timescale given by tion. Lozano-Duran and Jimenez34 showed that an overprediction can
Tav ¼ d=uc , where uc is the mean velocity at the center of the channel. be observed at the center of a channel using a very short domain. Note
The filtering timescale is determined as a fraction of Tfl ¼ hwm =jus , that the grid resolution at the center of a channel flow using the DNS
where hwm is the height of the wall layer and us is the friction velocity. is low and comparable to the grid resolution in the wall region using
The CFL (Courant–Friedrichs–Lewy) number is 0.3 in the LES region, the WMLES. Also, the velocity variance increases, and statistical noise
whereas the wall model uses CFL ¼ 15. A no-slip boundary condition is observed when a shorter domain is implemented.
is implemented at the upper and lower walls, and the flow is periodic The results are promising in the outer layer and center of the
in the streamwise and spanwise directions. The mean time step nondi- channel. They substantiate the performance of the self-calibrating
mensionalized by the viscous timescale (tv ¼ =u2s ) is dt þ ¼ 1:2. The mechanisms used in the DNM to estimate the forward and backward
simulation is initially run for 3000 time steps to eliminate any initial rates of turbulence kinetic energy transfer between the filtered and sub-
transients (25d=uc ). Thereafter, the flow statistics are sampled over a grid scales compared to the DSM.25 Consistent with the deficiency of
period of 75d=uc . the SGS model in the near-wall region, the Reynolds stresses shown in
Fig. 3(c) show a poor prediction for y  0:2d but improve toward the
IV. RESULTS center of the channel in comparison to the results of Park and Moin.22
The first set of results pertains to a fully developed channel flow The prediction of u0rms þ matches the profile of Park and Moin,22
for a Reynolds number of Res ¼ 2000 similar to the study of Park and although both underpredict u0rms þ compared with the DNS. On the
Moin22 but in a much shorter computational domain and with the other hand, v0rms þ and w0rms þ present better predictions above the
dynamic nonlinear SGS model rather than the dynamic Smagorinsky matching point and smoothly reduce to match the results of Park and
þ
model (DSM). The wall modeling uses the time-averaged TBLE Moin22 at the center of the channel. The turbulent shear stress (u0 v0 )
(URANS) instead of the full-RANS equations, and a temporal-filtering depicts a very good agreement with the reference data above the match-
scheme is introduced, which was recently investigated for a wall func- ing point. Consideration of Fig. 3(c) shows that only the v0rms þ profile
tion by Yang et al.19 The performance of the wall-modeled LES is ini- decreases to zero below the matching point, while both the u0rms þ and
tially evaluated by time averaging over a dimensionless time period of w0rms þ profiles increase in the wall region. Note that in the WMLES,
TW ¼ 1:0 as used by Park and Moin22 for averaging the eddy viscosity only the wall-normal velocity (v) takes the zero-velocity boundary con-
(TW is the averaging time period, Tav, nondimensionalized by d=uc ). dition at the wall. Overall, the results in Fig. 3 confirm the potential of
The initial dimensionless filtering time period is TF ¼ 1:0, which is the wall model. The next paragraphs will evaluate the effects of different
the ratio of the filtering time period (Tfl) to the turbulent diffusion time-averaging and temporal-filtering schemes on its performance.
timescale (hwm =jus ). This time period was recommended to reduce The effects of the temporal-filtering scheme will now be investi-
the strength of the log-layer mismatch.19 Figure 3 demonstrates the gated. Two scenarios are introduced, Scenario A and Scenario B, based
results of using the time averaging on the TBLE in the wall-modeling on the location of the matching point. In Scenario A, the matching
strategy and compares them with results taken from the literature. The point is located at the first node off the wall, while Scenario B moves

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-4


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

þ þ
FIG. 3. (a) Mean velocity profile in wall coordinates, (b) mean velocity profile in outer coordinates, and (c) Reynolds stresses (w0rms and u0rms profiles are shifted by þ0.5 and
þ1.0, respectively) at Res ¼ 2000 with TW ¼ 1:0; TF ¼ 1:0.

the matching point to the fifth grid point in the LES domain. Figure 4 small. In comparison, there is no significant difference between the
shows the mean velocity profiles for both scenarios with TW ¼ 1:0 two cases in Scenario B, which suggests that temporal filtering is
and TF ¼ 1:0 using inner coordinates. The case with no filtering is ineffective and perhaps unnecessary when the matching point is
also shown, which means that the instantaneous flow field from the located sufficiently far from the wall. On the other hand, it appears
LES is directly applied to the top boundary of the wall layer. Yang that temporal filtering is essential when the matching point is
et al.19 studied temporal filtering and no filtering for a WMLES chan- located at the first node off the wall as concluded by Yang et al.19
nel flow using a wall function based on the logarithmic law. The effect Note that the results of wall modeling below the matching point
of temporal filtering was clearly evident in their study since the match- (dashed-line) match very well with the DNS results of Hoyas and
ing point was located at the first node. The nonequilibrium wall model Jimenez.32 As noted above, the prediction for the mean velocity
implemented in the present study is able to respond to the inputs profile obtained from the wall modeling can serve as an alternative
based on the filtering time period being considered. The matching to the results of LES in the wall region. Note that Fig. 4 shows the
points are highlighted by vertical dotted lines, and the mean velocity results with TW ¼ 1:0 and TF ¼ 1:0 for the temporal-filtered cases.
profiles in Scenario B are shifted upward for clarity. Scenario A indi- The magnitude of the time-averaging and temporal-filtering time
cates that the temporal filtering reduces the mismatch although the periods should also be investigated to reveal their effects on the
difference between the results with and without temporal filtering is mean velocity profile.

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-5


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

Figure 5 presents the simulation results where the values the results are shown in Fig. 5(b). Compared with Fig. 5(a), it can be
TW ¼ 0:1 and TF ¼ 0:1 are introduced to evaluate the effects of concluded that the results are almost independent of the temporal-
shorter time periods for time averaging and temporal filtering. Recall filtering scheme. For both temporal-filtering cases, using a smaller
that a smaller value for TF implies that the wall layer at the top bound- time-averaging period (TW ¼ 0:1) better predicts the flow field. The
ary is now responsive to the higher frequency flow structures. The results using TF ¼ 1:0 are slightly closer to the DNS than the results
effects of temporal filtering are evaluated for TF ¼ 1:0 and 0.1. For with TF ¼ 0:1. The results of additional test cases that do not include
each of these values, two characteristic timescales for time averaging any temporal filtering (not shown) similarly confirm that shorter time
are investigated, i.e., TW ¼ 1:0 and 0.1. In Fig. 5(a), the upper set of periods for time averaging, i.e., TW ¼ 0:1, improve the prediction of
mean velocity profiles for Scenario A indicates that the wall layer with the mean velocity profile.
a temporal filtering using TF ¼ 0:1 underpredicts the mean velocity The results of Fig. 5 indicate that the performance of time averag-
profile (the consequence of the overpredicted friction velocity) regard- ing and temporal filtering depends on the location of the matching
less of the time-averaging period. In general, the results for TF ¼ 1:0 point. Scenario A underpredicts the mean velocity profile in all cases.
show a better prediction in comparison to the results for TF ¼ 0:1. It As discussed earlier, the use of the SGS model on a coarse grid in the
confirms again the role of temporal filtering in mitigating the mis- near-wall region is problematic and results in the overprediction of the
match when the matching point is located at the first node. For the friction velocity. It seems that higher values of TF and TW work better
cases with TF ¼ 1:0, the velocity profile at TW ¼ 1:0 matches better in Scenario A. On the other hand, the effect of the averaging schemes
with the DNS. A similar analysis was conducted for Scenario B, and is minimal in Scenario B. Based on the small changes observed, it
seems that a combination of a lower value of TW and a higher value of
TF improves the performance of the averaging schemes for Scenario B.
Note that a higher location of the matching point requires more grid
points to resolve the wall layer, which, then, increases the computa-
tional costs of the wall modeling. For example, Scenario A resolves the
wall layer in the wall-normal direction with 20 nodes compared to the
30 nodes used in Scenario B. The grid spacing of the last node of the
wall layer (matching point) in Scenario B is Dyþ  16, whereas it is
Dyþ  0:5 in Scenario A.
To further examine the interaction between the LES and wall
model, two probes were placed in the spanwise mid-plane of the chan-
nel close to the matching point (x ¼ pd; y ¼ hwm ; z ¼ 23 pd). One
probe tracked the instantaneous and temporal-filtered streamwise
velocities inside the LES domain. The second probe was located at the
last node in the wall layer close to the matching point to monitor the
time-averaged velocity inside the wall layer. Recall that the temporal-
filtered velocity is input to the wall layer as the top boundary condi-
tion. The velocity fluctuations from these probes over the time
FIG. 4. Temporal filtering vs no filtering with TW ¼ 1:0 and TF ¼ 1:0. advancement are shown in Fig. 6 (T is normalized by the convective
timescale). In Figs. 6(a) and 6(b), the profiles corresponding to

FIG. 5. The effects of various filtering and averaging time periods: (a) Scenario A with the matching point at the first node and (b) Scenario B with the matching point at the fifth
node.

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-6


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

FIG. 6. Tracking the streamwise velocity at the matching point (fifth node) for the instantaneous and temporal-filtered LES and time-averaged wall layer with the dimensionless
filtering time period: (a) TF ¼ 1:0 and (b) TF ¼ 0:1.

TW ¼ 1:0 are shifted a half unit upward for clarity, while the lower in determining the slope, since it varies too quickly to specify a local
part corresponds to TW ¼ 0:1. For both cases shown in Fig. 6(a), peak. The wall layer responds to the fluctuations at the matching point
where TF ¼ 1:0, the wall layer velocity and the temporal-filtered over a long time span (36 times the convective timescale) when
velocity at the matching point effectively do not respond to the strong TW ¼ 1:0. For comparison, TW was decreased to 0.1, and a different
fluctuations in the LES field. Figure 6(b) presents the results for a value trend was observed as shown in Fig. 7(b). In this case, the peaks occur
of TF ¼ 0:1. The velocity profiles in the upper region of Fig. 6(b) dem- at T ¼ 80; 82; 85, and 88 for the input, highest, middle, and lowest
onstrate that the temporal-filtered velocities now begin to respond to probes, respectively. This indicates that the entire wall layer adapts to
the LES field. However, the stronger temporal fluctuations are not evi- the new input in less than 500 time steps (8 times the convective time-
dent in the wall-model velocity. For the profiles in the lower region of scale). The slope of the fitted-blue line increases to 1/100 and for a sec-
Fig. 6(b), which are obtained with TW ¼ 0:1 and TF ¼ 0:1, a rapid ond peak reaches 1/30, which indicates a wall layer more responsive to
response to the LES is observed in the temporal-filtered velocity and the oscillations in the flow field at the top boundary. Figure 7 implies
the wall layer closely tracks the temporal-filtered profile due to the that an averaging time period comparable to the convective timescale
shorter time period used for time averaging. However, using TF ¼ 0:1 (TW ¼ 1.0) as used by Park and Moin22 makes the wall layer relatively
can lead to introducing nonessential high-frequency oscillations into slow to respond to the fluctuations in the LES region.
the wall layer and as such making the wall layer less responsive to the In the following analysis, the wall shear stress, which is the main
critical stress-carrying flow structures. This issue was observed in Fig. feedback to the LES from the wall layer, is tracked over time as shown
5(a) as an underprediction in the mean velocity profiles. Based on the in Fig. 8. The wall shear stress is representative of the velocity at the
physical trends shown in Fig. 6, the effects of different values of TW first node in the wall layer since it is calculated from sw ¼ l dudy . Figure
and TF on the mean velocity field in the wall layer will be evaluated in 8(a) presents sw fluctuations for four different cases in terms of TW
a more systematic manner below. and TF as previously discussed in Fig. 6. The fluctuations in the wall
The magnitude of the characteristic timescale for the time averag- shear stress and velocity in the LES are normalized by the correspond-
ing affects the response time of the wall layer. To illustrate this, three ing mean values. For both cases with TW ¼ 1:0, the wall shear stress
probes were used to track the streamwise velocity inside the wall layer profiles vary slowly with time. On the other hand, fairly rapid varia-
in the log-law region, the buffer region, and the first node off the wall. tions are observed in the two cases with TW ¼ 0:1, which implies that
The channel flow in Scenario B with TF ¼ 1:0 was used to perform a lower value for time averaging may be a better choice. The case with
the numerical experiment. The temporal-filtered velocity at the match- TF ¼ 0:1 includes high-frequency signals that are damped in the case
ing point was also monitored, which drives the flow inside the wall with TF ¼ 1:0. Figure 8(b) is provided to visualize the wall shear stress
layer. Figure 7 shows the dimensionless velocity fluctuations for each fluctuations in a wall-resolved LES (WRLES) channel flow. However,
probe over time. The temporal-filtered profile (input) approaches a due to the excessive resolution required for a wall-resolved channel
local peak at T ¼ 76, and the probe furthest from the wall at yþ ¼ 112 flow at Res ¼ 2000, a channel flow at Res ¼ 395 was considered. In
reaches a peak at T ¼ 81 as shown in Fig. 7(a). The next two probes Fig. 8(b), sw fluctuations for both the wall-resolved and wall-modeled
show the peak in their profile at T ¼ 96 and T ¼ 117. The last probe low-Reynolds number channel flows are presented together with the
located at yþ ¼ 0:87 is representative of the fluctuations in the wall results of the WMLES channel flow at Res ¼ 2000 with TF ¼ 1:0 and
shear stress. The blue line connects the peak points for all probes in TW ¼ 0:1. The amplitude of the fluctuations in the WMLES at
the wall layer and has a slope of 1/5200. Although a plausible peak in Res ¼ 395 is higher than what was observed for the WMLES at
the temporal-filtering profile is shown in the figure, it was not included Res ¼ 2000. The wall modeling was strongly recommended for high-

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-7


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

FIG. 7. Response of wall modeling to dif-


ferent inputs with the dimensionless aver-
aging time period: (a) TW ¼ 1:0 and (b)
TW ¼ 0:1.

FIG. 8. Wall shear stress fluctuations (a)


for various TF and TW values for the
WMLES channel flow at Res ¼ 2000
and (b) comparison of WMLES (at
Res ¼ 2000 and Res ¼ 395 with TF ¼
1:0 and TW ¼ 0:1) with WRLES at
Res ¼ 395 with/without temporal filtering
with TF ¼ 1:0.

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-8


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

Reynolds number flows.15 The course grid near the wall provides the
time-averaged effects of small flow structures in the wall region. Note
that when using the wall modeling in low-Reynolds number flows, a
part of the effects of these small flow structures is resolved by a higher
grid resolution compared with the grid resolution used in high-
Reynolds number flows. The sw fluctuations in the WRLES at
Res ¼ 395 are also presented to make a comparison between the
results of the WRLES and WMLES channel flows. The temporal-
filtered sw is also depicted since the wall model is fed with the
temporal-filtered input. A similar temporal-filtering scheme is imple-
mented for sw in the WRLES channel flow (TF ¼ 1:0). Both profiles
(WRLES and WMLES) are obtained based on the flow data at
yþ  0:8. It is obvious that the fluctuations obtained from WMLES
are less responsive compared to the temporal-filtered WRLES. This
difference is due to using a Reynolds-averaged approach in the
WMLES, whereas the wall region is instantaneously resolved using an
unsteady approach (LES) in the WRLES. However, the amplitude of
the fluctuations demonstrates that the WMLES varies with a wider
range compared to the WRLES at a lower Reynolds number.
Based on the previous results, we can conclude that TF ¼ 1:0
FIG. 9. Comparison of all scenarios with TF ¼ 1:0 and TW ¼ 0:1 using the
and TW ¼ 0:1 show a better prediction of the wall shear stress in the temporal-filtering (TF) and non-temporal-filtering (NTF) schemes. The profiles in the
WMLES. This corresponds to using a temporal-filtering period com- lower region pertain to Scenarios A and D, where the matching point is located at
parable to the turbulent diffusion timescale across the wall layer and the first node. The upper region presents the results for Scenarios B and C with the
using only a small fraction of the convective timescale for time averag- matching point located at the fifth and third nodes, respectively.
ing in the wall layer. At this point, two other scenarios are introduced
to study a wall-modeled LES with the matching point at the third schemes able to alleviate the mismatch but also the wall-normal loca-
node (Scenario C) and to evaluate the effects of changing the wall- tion of the first node in the wall layer should be considered, especially
normal grid distribution while keeping the matching point at the first for matching points located at the first node in LES.
node (Scenario D). In Scenario C, the number of grid points in the
wall-normal direction for the wall layer remains at 30, and the first V. CONCLUSIONS
grid point in the wall layer is held at y1þ  0:87 by changing the grid A wall-modeled LES channel flow was studied at Res ¼ 2000.
distribution. Scenario D has the matching point located at the first The instantaneous flow structures were solved by the filtered Navier-
grid point in the LES, and the first grid point in the wall layer decreases Stokes equations, and a dynamic nonlinear model was applied to cal-
to y1þ  0:17, while the number of grid points increases from 20 to 30. culate the residual stresses. The effects of time averaging in the wall
Figure 9 shows the mean velocity profiles for Scenarios A to D with layer and filtering the input flow data from LES to the wall layer at the
TF ¼ 1:0 and TW ¼ 0:1. Scenarios B and C are presented in the upper matching point, as well as the location of the matching point, were
region of the figure. In general, the temporal-filtered (TF) cases yield investigated. The performance of the SGS models in the near-wall
better predictions, while the cases without temporal filtering (NTF) region is inadequate due to the use of a very coarse grid, which, on the
slightly underpredict the mean velocity profiles. This is different from other hand, expedites the simulation run-time. This issue is observed
what was observed in Fig. 4, where the TF and NTF cases showed sim- in the predictions for both the mean flow velocity and Reynolds
ilar results. In that case, the time averaging occurred with TW ¼ 1:0, stresses in the region below the matching point. Time-averaging analy-
while a much shorter time period is implemented for time averaging sis shows that the use of too large a time period for averaging the wall
in Fig. 9. It is observed that moving the matching point from the 5th layer makes the wall model unresponsive to the flow motions at the
to the 3rd node has minimal effects on the final results. Note that both interface. Hence, neither the temporal-filtered nor unfiltered flow data
matching points are located above the buffer layer and in the log-law from the LES domain lead to a realistic wall shear stress from the wall
region, i.e., yþ ¼ 272 for Scenario B and yþ ¼ 155 for Scenario C. model. A smaller time-averaging period results in a more appropriate
The results for Scenarios A and D are shown in the lower part of boundary condition to the LES in terms of the wall shear stress.
Fig. 9. The temporal-filtered cases yield better results regardless of the Applying temporal filtering to the input to the wall layer is more effec-
location of the first node in the wall layer, while the cases without tem- tive for a matching point located at the first node off the wall. A
poral filtering typically show a larger mismatch. Interestingly, the cases temporal-filtering period comparable to the turbulence diffusion time-
in Scenario D give a better prediction for the mean velocity profiles in scale is recommended for all scenarios with different locations of the
comparison to the cases in Scenario A. This implies that moving the matching point. However, using the temporal filtering with TF ¼ 1:0
first node in the wall layer closer to the wall (e.g., y1þ  0:17) helps to and TW ¼ 0:1 for the time averaging makes a robust and responsive
mitigate the mismatch in the velocity profiles. The reason may be that wall layer for scenarios with matching points located at a node further
the wall layer becomes more dependent on the viscous sublayer than away from the wall. On the other hand, for scenarios with a matching
the flow at the top boundary to predict the friction velocity. It can be point at the first node (yþ < 100), a longer time period (i.e.,
concluded that not only are the temporal-filtering and time-averaging TW ¼ 1:0) for time averaging yields a better prediction. The

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-9


Published under license by AIP Publishing
Physics of Fluids ARTICLE scitation.org/journal/phf

14
application of these results to more complex flows, i.e., flows without J. Larsson, S. Kawai, J. Bodart, and I. Bermejo-Moreno, “Large eddy simulation
homogeneity and using more complex grids, will present additional with modeled wall-stress: Recent progress and future directions,” Mech. Eng.
Rev. 3, 15-00418 (2016).
challenges. For example, the appropriate location of the matching 15
U. Piomelli and E. Balaras, “Wall layer models for large eddy simulations,”
point to mitigate the mismatch is also affected by local flow character- Annu. Rev. Fluid Mech. 34, 349–374 (2002).
istics, which can vary significantly in complex flows. However, the 16
U. Piomelli, E. Balaras, H. Pasinato, K. D. Squires, and P. R. Spalart, “The
insight provided by the present study will be useful in such flows. inner-outer layer interface in large-eddy simulations with wall-layer models,”
Overall, this study demonstrates how the time-averaging and Int. J. Heat Fluid Flow 24, 538–550 (2003).
17
temporal-filtering schemes effectively reduce the mismatch of the J. Lee, M. Cho, and H. Choi, “Large eddy simulations of turbulent channel and
boundary layer flows at high Reynolds number with mean wall shear stress
mean velocity profile in the wall-modeled LES turbulent flow. boundary condition,” Phys. Fluids 25, 110808 (2013).
18
S. T. Bose and P. Moin, “A dynamic slip boundary condition for wall-modeled
ACKNOWLEDGMENTS large-eddy simulation,” Phys. Fluids 26, 015104 (2014).
19
X. I. Yang, G. I. Park, and P. Moin, “Log-layer mismatch and modeling of the
The authors are grateful to the Natural Sciences and fluctuating wall stress in wall-modeled large-eddy simulations,” Phys. Rev.
Engineering Research Council of Canada (NSERC) (grant no. Fluids 2, 104601 (2017).
20
05103-2016) for providing financial support for this project. S. Chen, Z. Xia, S. Pei, J. Wang, Y. Yang, Z. Xiao, and Y. Shi, “Reynolds-stress-
constrained large-eddy simulation of wall-bounded turbulent flows,” J. Fluid
DATA AVAILABILITY Mech. 703, 1–28 (2012).
21
S. Kawai and J. Larsson, “Dynamic non-equilibrium wall-modeling for large
The summary simulation data presented in this study are avail- eddy simulation at high Reynolds numbers,” Phys. Fluids 25, 015105 (2013).
able from the corresponding author upon request and with appropri- 22
G. I. Park and P. Moin, “An improved dynamic non-equilibrium wall-model
ate limitations. for large eddy simulation,” Phys. Fluids 26, 015108 (2014).
23
W. H. Cabot and P. Moin, “Approximate wall boundary conditions in the
large-eddy simulation of high Reynolds number flow,” Flow, Turbul. Combust.
REFERENCES
63, 269–291 (2000).
1 24
M. Wang and P. Moin, “Dynamic wall modeling for large-eddy simulation of X. I. A. Yang, J. Sadique, R. Mittal, and C. Meneveau, “Integral wall model for
complex turbulent flows,” Phys. Fluids 14, 2043–2051 (2002). large eddy simulations of wall-bounded turbulent flows,” Phys. Fluids 27,
2
T. J. Craft, S. E. Gant, H. Iacovides, and B. E. Launder, “A new wall function 025112 (2015).
strategy for complex turbulent flows,” Numer. Heat Transfer, Part B 45, 25
B. C. Wang and D. J. Bergstrom, “A dynamic nonlinear subgrid-scale stress
301–318 (2004). model,” Phys. Fluids 17, 035109 (2005).
3 26
A. Frère, C. C. de Wiart, K. Hillewaert, P. Chatelain, and G. Winckelmans, M. Germano, U. Piomelli, P. Moin, and W. H. Cabot, “A dynamic subgrid-
“Application of wall-models to discontinuous Galerkin LES,” Phys. Fluids 29, scale eddy viscosity model,” Phys. Fluids A 3, 1760–1765 (1991).
085111 (2017). 27
D. B. Spalding, “A single formula for the ‘law of the wall,’ ” J. Appl. Mech. 28,
4
H. J. Bae, A. Lozano-Duran, S. T. Bose, and P. Moin, “Dynamic slip wall model 455 (1961).
for large-eddy simulation,” J. Fluid Mech. 859, 400–432 (2019). 28
H. Xiao and P. Jenny, “A consistent dual-mesh framework for hybrid LES/
5
S. B. Pope, Turbulent Flows (Cambridge University Press, Cambridge, 2000). RANS modeling,” J. Comput. Phys. 231, 1848–1865 (2012).
6 29
D. R. Chapman, “Computational aerodynamics development and outlook,” R. Tunstall, D. Laurence, R. Prosser, and A. Skillen, “Towards a generalised
AIAA J. 17, 1293–1313 (1979). dual-mesh hybrid LES/RANS framework with improved consistency,”
7
J. W. Deadroff, “A numerical study of three dimensional turbulent channel Comput. Fluids 157, 73–83 (2017).
30
flow at large Reynolds numbers,” J. Fluid Mech. 41, 453 (1970). C. Meneveau, T. S. Lund, and W. H. Cabot, “A Lagrangian dynamic subgrid-
8
U. Schumann, “Subgrid scale model for finite difference simulations of scale model of turbulence,” J. Fluid Mech. 319, 353–385 (1996).
31
turbulent flows in plane channels and annuli,” J. Comput. Phys. 18, 376–404 S. Kawai and J. Larsson, “Wall-modeling in large eddy simulation: Length
(1975). scales, grid resolution, and accuracy,” Phys. Fluids 24, 015105 (2012).
9 32
S. T. Bose and G. I. Park, “Wall-modeled large-eddy simulation for complex S. Hoyas and J. Jimenez, “Scaling of the velocity fluctuations in turbulent chan-
turbulent flows,” Annu. Rev. Fluid Mech. 50, 535–561 (2018). nels up to Res ¼ 2003,” Phys. Fluids 18, 011702 (2006).
10 33
G. I. Park, “Wall-modeled large-eddy simulation of a high Reynolds number H. Abe, H. Kawamura, S. Toh, and T. Itano, “Effects of the streamwise compu-
separating and reattaching flow,” AIAA J. 55, 3709–3721 (2017). tational domain size on DNS of a turbulent channel flow at high Reynolds
11
E. Balaras, C. Benocci, and U. Piomelli, “Two-layer approximate boundary con- number,” in Advances in Turbulence XI—Proceedings of the 11th
ditions for large-eddy simulations,” AIAA J. 34, 1111–1119 (1996). EUROMECH European Turbulence Conference, Porto (2007), pp. 233–235.
12 34
G. I. Park and P. Moin, “Numerical aspects and implementation of a two-layer A. Lozano-Duran and J. Jimenez, “Effect of the computational domain on
zonal wall model for LES of compressible turbulent flows on unstructured direct simulations of turbulent channels up to Res ¼ 4200,” Phys. Fluids 26,
meshes,” J. Comput. Phys. 305, 589–603 (2016). 011702 (2014).
13 35
S. Heinz, “A review of hybrid RANS-LES methods for turbulent flows: M. Lee and R. D. Moser, “Direct numerical simulation of turbulent channel
Concepts and applications,” Prog. Aerosp. Sci. 114, 100597 (2020). flow up to Res ¼ 5200,” J. Fluid Mech. 774, 395–415 (2015).

Phys. Fluids 33, 035108 (2021); doi: 10.1063/5.0039651 33, 035108-10


Published under license by AIP Publishing

You might also like