You are on page 1of 15

PHYSICS OF FLUIDS VOLUME 16, NUMBER 8 AUGUST 2004

ARTICLES

Analysis and low-order modeling of the inhomogeneous transitional


flow inside a T-mixer
Haysam Teliba)
Dipartimento di Ingegneria Aeronautica e Spaziale, Politecnico di Torino, Torino 10129, Italy
Michael Manhart
Fachgebiet Stromungsmechanik, Technische Universitat Munchen, D-85748 Garching, Germany
Angelo Iollo
Dipartimento di Ingegneria Aeronautica e Spaziale, Politecnico di Torino, Torino 10129, Italy
!Received 16 October 2003; accepted 19 March 2004; published online 8 June 2004"
A direct numerical simulation of the transitional flow !Re!300 to Re!700" inside a T-mixer
configuration has been carried out. Time records were collected and used to perform a proper
orthogonal decomposition !POD" of the flow. Changes of the flow characteristics in the frequency
spectra and extracted coherent spatial structures indicate flow transition across the investigated
Reynolds numbers. The POD modes were used to derive a low-order model of the flow. An a priori
test limits the possibilities of the modeling; for the periodic case it demonstrates that the flow can
be reduced to a system of a few degrees of freedom, while for the turbulent ones this results to be
extremely difficult because of the large number of degrees of freedom that are necessary to describe
the flow. 2004 American Institute of Physics. #DOI: 10.1063/1.1751204$

I. INTRODUCTION laminar-turbulent transition which takes place for a Reynolds


number of about 400.
Many active substances, such as drugs in pharmaceutics, The flow analysis relies on the KarhunenLoeve decom-
are obtained through continuous precipitation of nanopar- position. Also known as proper orthogonal decomposition
ticles in so-called T- or Y-mixers. These particles result from !POD", this approach allows the definition of an optimal
a chemical reaction of two counter-directed liquid solutions function space for the representation of the velocity field
that are mixed in a duct or a pipe. The product of the chemi- database. It was first introduced into the fluid flow commu-
cal reaction is lyophobic with respect to the mixture, and it nity by Lumley2 to identify, without any a priori knowledge,
precipitates as a solid or dispersed phase. These T- or any self-organized structures that might be present in the
Y-mixer configurations !Fig. 1" are applied in a wide range flow. Laboratory experiments show that such functions can,
of chemical reactors. The application of low-order models to in fact, be interpreted as actual three-dimensional structures
reproduce such complicated three-dimensional turbulent that characterize the dynamics of the flow !see e.g., Refs. 3
flows and the evaluation of the suitability of such models at and 4". The POD functions can, however, be thought of as
different flow regimes is the subject of the present investiga- defining a mere function space with desirable properties,
tion. which gives a synthetic description of the flow field. More-
It is obvious that the turbulence level, measured by the over, they represent an appropriate surrogate of Fourier
Reynolds number, plays a major role in the mixing process.1 analysis for inhomogeneous turbulent flows.5,6
Thus, by limiting our study to the analysis and modeling of
This technique is also used to devise low-dimensional
the nonreacting turbulent flow, we can still obtain important
models by projecting the NavierStokes equations onto the
information for process engineers. In this respect, the flow
POD functions. The main idea is based on the fact that sys-
field inside the mixer is obtained by direct numerical simu-
tems that have a large !infinite" number of degrees of free-
lation of the incompressible NavierStokes equations. All
dom often exhibit motion that is characterized by a small
relevant space and time scales are resolved, so that the solu-
phase space dimension. It has in fact been rigorously shown
tion database consists of the three velocity components of
that the dynamics for several partial differential equations
each space and time discretization point. The Reynolds num-
ber, based on the square section side, ranges from 300 to 700 !PDEs" is exponentially attracted onto a low-dimensional
in order to describe the flow characteristics across the subset of a phase space !an inertial manifold". For a complete
review of the method as a predictive tool !see Refs. 7 and 8".
In the considered flow, the possibility of representing the
a"
Electronic mail: haysam.telib@polito.it dynamics with a small special subset of the unknowns is

1070-6631/2004/16(8)/2717/15/$22.00 2717 2004 American Institute of Physics

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
2718 Phys. Fluids, Vol. 16, No. 8, August 2004 Telib, Manhart, and Iollo

not a priori clear, and part of the contribution of this work is II. DIRECT NUMERICAL SIMULATION
to clarify this issue. If we want to model an inhomogeneous
The flow was analyzed in the T-mixer geometry at five
flow using low-order models, we must at least make sure that
different Reynolds numbers, ranging from Re!300 to Re
the solution database is reasonably represented by the POD
!700, by direct numerical simulation !DNS". In this manner,
base. The classical test to clarify this issue consists in exam-
the flow characteristics were obtained at and beyond
ining the convergence rates of the POD series expansion. laminarturbulent transition, which takes place at about Re
Usually one considers as many POD modes as necessary to !400"50. Section II A describes the numerical method that
reach an arbitrary threshold !e.g., 99.9%" of the captured was used for the direct solution of the NavierStokes equa-
energy. This number scales9 with Re 9/4. In this work, we tions !1":
present an additional a priori test that quantifies the recon-
%ui
struction error of flow snapshots that were not included in !0,
the data pool used to build the POD expansion. The capabil- %xi
ity of the POD series to represent these snapshots is a preju- !1"
%ui %ui 1 %p % 2u i
dicial criterion for any low-order modeling, especially for #u j !$ #' 2 .
%t %x j & %xi %x j
long-term flow predictions. We can anticipate that the results
of such an a priori test indicate that as the Reynolds number
increases above the transition limit, only a third of the fluc- A. Numerical scheme
tuating kinetic energy is recovered by the POD expansion Our DNS method is based on a finite-volume method
with 400 modes, leaving little hope of low-dimensional rep- formulated on a nonequidistant staggered mesh. A Poisson
resentation. equation for the pressure correction is obtained using the
For the lowest Reynolds numbers !i.e., 300 and 400", the projection method according to Ref. 17. The flow variables,
previously mentioned a priori test indicates that the fluctu- velocity components, and pressure are defined on a nonequi-
ating energy is well resolved by the POD expansion. In such distant Cartesian mesh in a staggered arrangement. The ve-
cases, it makes sense to devise a low-order model. However, locity components are stored in the centers of the cell faces,
it may occur that the stable solution manifold is projected while the pressure is stored in the cell centers. The specific
onto an unstable subspace by the order reduction discrete formulations are derived by integrating the previous
procedure.10 This occurrence makes, for example, the mod- equations for a Newtonian fluid over the corresponding con-
trol cells that surround the definition points of the individual
eling of transitional flows, as shown in Ref. 11, questionable.
variables. The midpoint rule is used to approximate the
Another point to take into account is that models for the
fluxes. The required interpolations and the approximation of
interaction of the simulated scales with the unresolved scales the first derivatives are performed by linear interpolation
and with the boundary conditions are necessary. Dissipation !central" and central finite difference formulations, respec-
to small scales is almost invariably taken into account by the tively. This ensures second-order accuracy of the spatial
Heisenberg model.1215 This approach basically results in a discretization.18 The discrete solution is advanced in time by
larger viscosity term of the projected model. The influence of a leapfrog time step with time lagged diffusion term, given
boundary conditions and the influence of pressure gradients as
is either considered by imposing a mean flow or by using
u n#1 !u n$1 #2(t # C ! u n " #D ! u n$1 " $G ! p n#1 "$ , !2"
ad hoc models.
In the following, we instead take advantage not only of where C, D, and G denote the discrete convective, diffusive,
the flow snapshots, as in the usual POD approach, but also of and gradient operators, respectively. The pressure at the new
the flow dynamic behavior. This is done by tuning the low- time level p n#1 ! p n #(p n#1 is determined by solving the
order model to represent the available solution time series for Poisson equation
the empirical eigenfunctions. This approach is based on the 1
solution of an inverse problem for the linear terms in the Div# G ! (p n#1 "$ ! Div! u * " , !3"
2(t
low-order model. A full discussion of this method is pre-
sented in Ref. 16. Here, we present its application to a com- where u * is an intermediate velocity field obtained by solv-
ing Eq. !2" using pressure p n at the known time level. A
plicated real-life problem. This approach also prevents the
divergence-free field u n#1 is obtained after a velocity correc-
possibility of a low-order unstable projection and circum-
tion step, given as
vents the problem of modeling the dissipation to the small
unresolved scales. u n#1 !u * $2(tG ! (p n#1 " . !4"
The paper is set out as follows. The numerical simula- The Poisson equation !3" is solved by an iterative proce-
tion details are discussed in terms of validation of the ob- dure accelerated using a multigrid cycle. The smoother is
tained data. Subsequently, the POD is synthetically intro- based on the velocitypressure iteration presented in Ref. 19
duced. Next, in terms of POD functions, the flow regimes are with over-relaxation. This scheme gives the same conver-
characterized as the Reynolds number is increased. Finally, gence properties as a conventional GaussSeidel iteration
the modeling issue is discussed and results pertinent to a with successive over-relaxation. The advantage of the
model of the transitional flow are presented. present algorithm is the easy treatment of boundaries, at

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
Phys. Fluids, Vol. 16, No. 8, August 2004 Analysis and low-order modeling 2719

FIG. 1. The geometry of the configuration.

which only velocity boundary conditions have to be speci- B. Results


fied. The numerical scheme, its performance, and accuracy
The global behavior of the flow field is first analyzed at
have been documented in detail in Refs. 20 and 21.
different Reynolds numbers. The solution at a higher Rey-
The DNS has been carried out in a quadratic mixer ge-
ometry as shown in Fig. 1. The dimensions are nolds number was obtained using a previous run at a lower
(Lx,Ly,Lz)!(H,H,4H), H being the width of the mixer Reynolds number as initial solution. After a transition phase
and z the main flow direction. In what follows, all lengths are that takes some few channel through flows, the flow reaches
scaled by H. The feeding tubes are modeled by steady inflow statistical steadiness !see Fig. 2". We have taken care to start
boundary conditions at x!0.0 and x!1.0. They are assumed our flow analysis after the statistical steadiness has been
to be laminar pipe flows centered around (y,z)!(0.5,0.5) reached.
with a diameter of d!0.5. No-slip conditions are used at z We have looked at global values, such as the kinetic
!0.0 as well as at the walls in the y and z directions. At the energy integrated over the volume of the mixer or the wall
outflow boundary (z!4.0), the velocities are extrapolated by shear stress integrated over the top wall at y!1.0. The inte-
a zero-gradient condition. The computational domain is re- grated kinetic energy #Fig. 3!a"$ shows periodic behavior in
solved using (NX,NY ,NZ)!(112,160,256) grid points. This
time at Re!300. By increasing the Reynolds number to Re
resolution was chosen on the basis of estimations of the wall
!400, the behavior changes from periodic to a chaotic state.
shear stresses and the Kolmogorov microscale using previ-
ous coarser grid simulations. At higher Reynolds numbers, small-scale turbulence appears

FIG. 2. Time history of the kinetic en-


ergy integrated over the flow domain.
The time span of the POD data pool
T POD covers about 10 through flows of
a particle with average velocity. The
averages of the integrated kinetic en-
ergy are taken starting from t 0 .

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
2720 Phys. Fluids, Vol. 16, No. 8, August 2004 Telib, Manhart, and Iollo

FIG. 3. DNS integral values. !a" Kinetic energy inte-


grated over the flow domain. !b" Normalized integrated
shear stress at the top wall.

that causes an increase in the total flow energy in the domain.


The ratio of the energy content of the five flows normed to
the lowest Reynolds number is 1:1,07:1,12:1,16:1,19. While
at low Reynolds numbers we can observe high energy den-
sity at small wave numbers, the energy is shifted toward high
wave numbers at the turbulent regime. As the Reynolds num-
ber increases, turbulent movements develop perpendicular to
the mainstream direction, which transport high-speed flow
from the channel core to the wall boundary layers and vice
versa. Thus, much higher velocity gradients occur and the
shear stresses normalized by viscosity increase. As far as the
time history of shear stresses at the top wall of the T-mixer
domain is concerned #Fig. 3!b"$, Re!500 marks a remark-
able change of behavior: At Reynolds numbers below 500,
we obtain a symmetrical solution, in the sense that the inte- FIG. 4. Snapshot of the flow field: velocity contour plot in a plane near one
grated shear stresses on opposite walls are equal at every of the inlets (x!0.75) and instantaneous stream lines.

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
Phys. Fluids, Vol. 16, No. 8, August 2004 Analysis and low-order modeling 2721

FIG. 5. Visualization of instantaneous velocity field at Re!300. !a" Velocity vector plot at z!1.2. !b" Isolines of mainstream component u z of the velocity
vector at z!1.2. !c" Isosurface of velocity magnitude, !u!!0.6.

time step. Furthermore, the time signal of the wall stresses in the positive and negative y directions. The plane jet oscil-
for the lowest Reynolds number is periodic. At higher Rey- lates, as can be seen in Figs. 5!c", 6!c", 7!c", 8!c", and 9!c".
nolds numbers the behavior is clearly chaotic, so that the The intensity of these oscillations increases with higher Rey-
stresses on the opposite walls are uncorrelated. nolds numbers. At Reynolds numbers greater than Re!400,
The two incoming jets impact in the middle of the mixer small-scale structures overlay the large structure, that is
and explode radially !Fig. 4". The component in the positive formed by the plane jets and the vortices, as can be seen
z direction induces a velocity field that is very similar to flat from the isosurface plot in Figs. 7!c", 8!c", and 9!c".
jets. The negative z component impinges at the back wall and
is forced, by its high momentum, to change direction by 180 III. FLOW ANALYSIS USING PROPER ORTHOGONAL
to proceed toward the outlet of the domain. A highly chaotic, DECOMPOSITION
thus unsteady, flow field develops in this part of the domain.
We use proper orthogonal decomposition !POD" to ana-
As this flow again reaches the incoming jets, it is forced to
lyze and characterize the behavior of the flow in the T-mixer
elude them and thus is divided into four column-like streams.
with a variation of the Reynolds number. Before discussing
While these streams pass the jets, the outer part of the
the results, a synthetic description of the POD is given. For a
streams is turned by friction and exchange of momentum
complete derivation of the theory see, e.g., publications by
mechanisms by the high-speed jets. Thus, four characteristic
Sirovich22 or Holmes et al.7
vortices develop, one in each corner of the duct #Fig. 5!a"$.
Although the vortices are stable for the lowest Reynolds
A. The proper orthogonal decomposition
number, they do not remain constant in either the spatial or
in the time domain. The flat jet, that is formed by the de- If the characteristic eigenproblem is solved for the n
flected incoming jets, marks the separating plane between the %n matrix of the one-point time correlations of the velocity
right and the left two vortices #Fig. 5!b"$. The fluid of this jet fields, we obtain the n eigenvalues ) n and n eigenfunctions
directed outward feeds the vortices by reflection at the walls a n (t) of the problem.

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
2722 Phys. Fluids, Vol. 16, No. 8, August 2004 Telib, Manhart, and Iollo

FIG. 6. Visualization of instantaneous velocity field at Re!400. !a" Velocity vector plot at z!1.2. !b" Isolines of mainstream component u z of the velocity
vector at z!1.2. !c" Isosurface of velocity magnitude, !u!!0.6.

"
T
C ! t,t ! " a n ! t ! " dt ! !) n a n ! t " , !5"
mode * ni (x), which describes its spatial shape. It can easily
be shown that temporal as well as spatial orthogonality
holds:

"
with
1

"""
a m ! t " a n ! t " dt!) n , mn , !9"
1 T T
C ! t,t ! " ! u i ! x,t " u i ! x,t ! " dx. !6"

"" "
T V

*m n
i ! x " * i ! x " dx! , mn , !10"
By projecting the snapshots of the flow field onto the nth V
eigenfunction a n (t), we can calculate the nth POD mode of
the domain. and thus the energy decomposition is also orthogonal. By
construction, the POD modes maximize the represented en-

* ni ! x" ! " T
a n ! t " u i ! x,t " dt. !7"
ergy. In statistically stationary flows, the first mode, the most
parallel to all the solutions, represents approximately the
mean flow field. Thus, the first eigenvalue describes the en-
In doing so, we decompose the flow field into ergy content of the mean, and the sum of all the remaining
eigenvalues describes the energy of perturbation.
NT
The following results are obtained using 2500 snapshots
u i ! x,t " ! +
n!1
a n ! t " * ni ! x" . !8" u i (x,t) to build the time correlation matrix C(t,t ! ). Enlarg-
ing the basis does not lead to significant changes, as a com-
The eigenvalues represent the energy content of the modes. parison between the eigenfunctions a n (t) that are obtained
The nth eigenfunction a n (t) is the time evolution of the nth using 2500 and 5000 snapshots demonstrates. A filter was

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
Phys. Fluids, Vol. 16, No. 8, August 2004 Analysis and low-order modeling 2723

FIG. 7. Visualization of instantaneous velocity field at Re!500. !a" Velocity vector plot at z!1.2. !b" Isolines of mainstream component u z of the velocity
vector at z!1.2. !c" Isosurface of velocity magnitude !u!!0.6.

applied to reduce the spatial resolution of the snapshots from B. Results


4.8%106 points, used for the DNS, to 0.6%106 points for the
analysis. This was performed by averaging over two cells in 1. Eigenvalue spectra
each spatial direction. The loss of information is justified as
The double logarithmic plot of the eigenvalues #Fig.
structures much larger than the Kolmogorov scale are being 10!a"$ over the mode number shows three characteristic
dealt with. zones: The first zone is characterized by a steep decrease of
The results demonstrated in the Sec. III B are obtained the eigenvalues !mode number below 10". The slope of the
by finding the POD basis for energy. The series converges curves of the low Reynolds number cases is flatter in this
optimally, in terms of energy, but it is not guaranteed if the part. At Re!300, we can see #Fig. 10!a"$ a pairwise arrange-
energy is the best suited to describe the dynamics. One might ment of the eigenvalues for small mode numbers, which in-
speculate that, if instead, the instantaneous velocity fields, dicates the presence of traveling waves.
vorticity, or helicity is used, typical flow patterns may be In the second region !mode numbers between 10 and
represented better, and thus converge faster and provide a 200, approximately", we can see that, as the Reynolds num-
better basis. In order to clarify this question, a representative ber is increased above 500, the curves converge to a power
test is performed for the highest Reynolds number flow case. law indicating an inertial range. This inertial range covers
This test shows that the fastest basis to converge is the one about 200 modes at Re!700 and the eigenvalues decay at
found using velocity. However, the differences in conver- approximately n $2/3. Above the inertial range, a steeper gra-
gence are very small and the shape of the curve of the eigen- dient is obtained, corresponding to a dissipative range. Be-
values versus mode number of the different kernels shows low Re!500, the energy containing and dissipative range are
exactly the same characteristics as the one found for velocity. hardly to separate. Finally, above n!200 the eigenvalues go

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
2724 Phys. Fluids, Vol. 16, No. 8, August 2004 Telib, Manhart, and Iollo

FIG. 8. Visualization of instantaneous velocity field at Re!600. !a" Velocity vector plot at z!1.2. !b" Isolines of mainstream component u z of the velocity
vector at z!1.2. !c" Isosurface of velocity magnitude, !u!!0.6.

below numerical precision, which is at about the order of similar to a Reynolds decomposition, the fluctuating velocity
10$7 . field can be retrieved by subtracting a 1 (t)* 1i (x) from the
The percentage of the unsteady energy related to the instantaneous velocity field
second mode is highest for Re!400 #Fig. 10!b"$, although
unsteady energy increases with increasing Reynolds num- u !i ! x,t " !u i ! x,t " $a 1 ! t " * 1i ! x" . !11"
bers. This indicates the development of a dominant fluctuat-
ing structure, which seems to be unstable when the Reynolds In accordance, we observe that the average of every other
number is further increased. modethe fluctuationsis nearly zero.
For the lowest Reynolds numbers #Fig. 11!a"$, a clear
harmonic behavior can be identified for the first modes. At
2. Time domain
Re!400, the time development of the second modethe
The eigenfunctions a n (t) record the time evolution of most energetic fluctuationdeserves special attention. Its
the corresponding mode * ni (x). By construction #Eq. !10"$, characteristic frequency has changed totally with respect to
the L 2 -norm of * ni (x) is unity, thus, the L 2 -norm of the Re!300: it does not exhibit a more or less regular har-
corresponding coefficient a n (t) is equal to ) n , representing monic variation. Instead, it shows a flipping from a positive
the energy content of that mode #see Eq. !9"$. In Fig. 11, the to a negative value, which seems to occur at random times
time coefficients a n (t)/ !) n are depicted. They are scaled by on a very large time scale. The third mode again shows a
!) n so that their norm is unity. harmonic content with a characteristic frequency similar to
If a certain time coefficient is constant, then the corre- the frequency of the second eigenfunction of Re!300.
sponding spatial mode is the rescaled time average of the For the high Reynolds number cases, we can observe a
flow. In Fig. 11, the first time coefficient, the most energetic, more chaotic time history with some aspects of intermit-
remains nearly constant in time. Hence, in a way that is tency.

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
Phys. Fluids, Vol. 16, No. 8, August 2004 Analysis and low-order modeling 2725

FIG. 9. Visualization of instantaneous velocity field at Re!700. !a" Velocity vector plot at z!1.2. !b" Isolines of mainstream component u z of the velocity
vector at z!1.2. !c" Isosurface of velocity magnitude, !u!!0.6.

There is a significant change in the behavior of the functions a n (t). As the inlet boundary conditions are equal
second eigenfunction when the Reynolds number varies: for all the flow cases, the first mode, which describes the
the time signal of this POD mode is periodic at Re!300, mean downstream propagation of the mass, is almost identi-
it shows a very particular behavior at Re!400 and be- cal for all Reynolds numbers. The first mode of Re!700 is
comes a chaotic signal for the highest Reynolds numbers. plotted in Fig. 12!a". The steady characteristics of this flow
This change in behavior seems to mark the laminar can be seen; that is, the impacting incoming jets, the high-
turbulent transition. There are some indications that a speed flow in the vertical symmetry plane and the large vor-
subharmonic cascade is the dominant mechanism that leads
tices that develop at each corner of the channel. The second
to chaos for this flow.9,23 What leads us to state this is
POD mode describes the main unsteady structure in the flow
the occurrence of the doubled and the quadrupled period
of the characteristic period of Re!300 at Re!400. We ob- field, which is very similar for all Reynolds numbers. A su-
tained spectral information by applying a fast Fourier trans- perposition of this mode onto the first mode would result in
form to the first few functions a n (t): the third eigenmode the oscillation of the jet-like flow in the symmetry plane
obtained at Re!400 has the same characteristic frequency as #Fig. 12!b"$. The intensity of this structure is greatest for
the second eigenmode of Re!300, the sixth eigenmode has flow cases before transition, as the corresponding eigenvalue
the doubled and the second has the quadrupled period, which denotes #Fig. 10!a"$. On the contrary, for Re!600 and Re
is the minimum frequency we can resolve with this sample !700 higher mode numbers are more relevant, as the eigen-
length. values of these mode numbers indicate. The plot #Fig. 12!c"$
of the 301st mode for Re!700 shows the small scales that
3. Space domain occur in this flow due to the development of turbulence. At
The corresponding spatial POD modes * ni (x) are gath- first sight these structures seem to exhibit homogeneity, but
ered by projecting the velocity fields u i (x,t) onto the eigen- integral information is still present.

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
2726 Phys. Fluids, Vol. 16, No. 8, August 2004 Telib, Manhart, and Iollo

!13"

All of the coefficients that appear in the previous quadratic


model are computed using the available POD modes. Let us
concentrate on the two right-hand-side terms. If the first term
is rearranged using integration by parts, we can see that it
basically represents the effect of the pressure drop through
the mixer. The second term represents the damping due to
viscosity. The pressure drop is relatively unaffected by the
simulation of a limited number of scales. Dissipation occurs,
however, in the smallest scales and therefore for the simula-
tion of large scales the last term on the right-hand side can be
omitted, as is well known.
We are, however, left with the problem of modeling the
energy transfer to the unresolved scales and the energy input
due to the imposed boundary conditions.

B. An a priori test
An a priori test was performed to check the flow recon-
struction properties of the obtained POD eigenmodes for
flow fields not contained in the original snapshots. This test
FIG. 10. Eigenvalue spectra. Re!300, Re!400, Re!500, was performed for each Reynolds number. We used addi-
Re!600, Re!700. !a" Eigenvalues. !b" Partial sums of eigenvalues. tional velocity samples u i (x,t) that were not part of the data
pool used to calculate the POD basis. The check is per-
formed by projecting these snapshots onto the POD basis
* ni (x),

"""
IV. LOW-ORDER MODELING FOR FLOW PREDICTION
a n ! t " ! u i ! x,t " * ni ! x" dx, !14"
This section is developed to evaluate the feasibility of V

low-order modeling based on POD modes for moderately


reconstructing the flow field with these coefficients a n (t)
complex flows using a limited number of unknowns !less
than 100" and a quadratic model. The envisaged application NT

i ! x,t " ! + a ! t " * i ! x "


is a feedback flow control procedure in which the low-order u* n n
!15"
model is used as a short-term predicting tool. n!1

A. The Galerkin projection and calculating the error /:

The low-order model !LOM" is obtained by substituting 000 V # u i ! x,t " $u i* ! x,t "$ 2 dx
the expansion for u i (x,t) !8" in the NavierStokes equations / ! t,N T " ! , !16"
!1" and then projecting them onto the eigenmodes * ni (x) of 000 V # u i ! x,t " $a 1 ! t " * 1i ! x"$ 2 dx
the flow. In what follows, the brackets !," denote the L 2
which represents the relative error in recovering the fluctu-
definition of the inner product:
ating energy #Eq. !11"$ of a given snapshot with N T POD
eigenmodes. This error was ensemble averaged over 50 and

"""-
200 snapshots !not part of the POD data pool" in order to
! - i, . j "! i
! x" . j ! x" dx. !12" validate the results obtained.
V The results are shown in Fig. 13 and summarized in
Table I. They weakly depend on the number of snapshots
considered for the ensemble average. The convergence is not
As (!i ,! j )! , i j !10", the obtained equation is very encouraging for large Reynolds numbers. We see that a

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
Phys. Fluids, Vol. 16, No. 8, August 2004 Analysis and low-order modeling 2727

FIG. 11. Eigenfunctions: a n (t), a 1 (t), a 2 (t), a 3 (t). !a" Re!300. !b" Re!400. !c" Re!500. !d" Re!600. !e" Re!700.

great part of the turbulent energy of flow fields that were not underlines the possibility of reducing the complexity of a
contained in the POD data pool, cannot be captured for large system with the order of 107 degrees of freedom !DNS" to a
Reynolds numbers. system of the order of 105 degrees of freedom.
All the unsteady energy can be captured for Re!300. However, the previous test suggests that the empirical
With increasing flow complexity, the resolved energy de- eigenmodes should provide a good basis for at least the lami-
creases continuously until Re!700, where only a third of the nar and the transitional cases at Re!300, 400.
fluctuating energy can be captured with 400 empirical eigen-
modes. C. Results
If this check is performed with a member of snapshots
used to find the POD basis, the captured energy increases By integrating Eq. !13" using 20 modes without any
with the mode numbers according to the slope of the partial model for the right-hand side and for the influence of bound-
sums of the eigenvalues with increasing mode numbers. This ary conditions, we obtain the time history of all the 20
contradictory behavior seems to indicate that, for higher modes. Such time histories can be compared with the results
Reynolds numbers, the dimension of the attractor in the obtained by projecting DNS data onto the empirical eigen-
phase space is much higher than 2500, which is the number modes. Integration was performed using a fourth-order
of flow snapshots that has been used to generate the modes. RungeKutta scheme, in which one instantaneous solution
In this respect, the convergence of the POD basis mentioned of the time coefficients of the POD is used as the initial
in Sec. III should just be interpreted as a loss of sensitivity of condition for the a n (t). Although the behavior of the flow
the eigenmodes with respect to a further addition of snap- for Re!300 can be represented by a low-dimensional attrac-
shots. By a linear extrapolation, we can estimate the lower tor, the solution diverges immediately.
bound to capture 90% of the fluctuating energy for Re!700 This misfitting is a result of the truncation. It does not
at O(105 ) eigenmodes. allow the energy to be transfered to smaller scales and dissi-
Using such a number of modes is not consistent with pation is inhibited, and the boundary conditions are not taken
what we intend to denote as a low-order model, although this into account.

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
2728 Phys. Fluids, Vol. 16, No. 8, August 2004 Telib, Manhart, and Iollo

FIG. 13. Reconstruction error 1/2 vs mode number. !a" Average error using
50 snapshots. !b" Average error using 200 snapshots.

where the coefficients E i j , representing the dissipative, pres-


sure and truncation terms, are still unknown. Let us consider
FIG. 12. Various POD modes * ni (x). !a" First POD mode Re!700: isosur- one characteristic period of the flow for which the DNS so-
face of ! !1 ! , contour plot of ! !1 ! in a plane near one of the inlets (x lution exists. If these DNS snapshots are projected onto the
!0.75) and stream lines. !b" Second POD mode Re!500: isosurface eigenmodes, we can obtain the time evolution of the modes,
! !2 ! !0.4. !c" 301st POD mode Re!700: contour plot of ! !301! in a plane
near one of the inlets (x!0.75).
which we denote as a i,DNS . By building the difference be-
tween the time derivative a i,DNS and the right-hand side of
Eq. !17", we can obtain the following function of E i j :
In order to improve the behavior of the system, we
model all the effects that have been previously disregarded
by now using a linear term. In other words, matrix D i j in
L! Ei j "! "
T
! a j E i j $a j a k C i jk $a i,DNS" 2 dt. !18"

!13" is replaced by a new matrix E i j in which the dissipative By minimizing this function, we obtain
term of the Galerkin projection, the pressure term, as well as
the effect of truncation are taken into account. A full discus-
sion of this approach is given in Ref. 16.
The pressure difference between the inlet and outlet is TABLE I. Captured fluctuating energy with 400 eigenmodes.
the driving force. Its term in the Galerkin projection !13" Captured fluctuating energy
does not vanish at the inlet and outlet boundaries of the flow Flow case with 400 eigenmodes
domain. This pressure term is quadratic in a i , which can !Re" 50 snapshots 200 snapshots
easily be shown by introducing Poissons equation for the 300 99.9% 99.9%
pressure, but the linear approximation seems reasonable, as 400 91.5% 84.6%
the main effectdriving the flowis modeled by just the 500 60.0% 55.2%
first mode. Thus, the low-dimensional model is written 600 48.2% 44.4%
700 37.3% 38.6%
a i !a j E i j $a j a k C i jk , !17"

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
Phys. Fluids, Vol. 16, No. 8, August 2004 Analysis and low-order modeling 2729

FIG. 14. Comparison between DNS !solid" and LOM !dashed" results for Re!300 using nine modes. !a" Second eigenfunction a 2 (t). !b" Fourth eigenfunction
a 4 (t). !c" Sixth eigenfunction a 6 (t).

%L! Ei j" accordance with the data that is obtained by integrating the
!0, !19" NavierStokes equations, but a similar behavior has also
%Eij
been observed by Rempfer.24 The model fails to predict the
which represents a linear system for the coefficients E i j . dynamics for higher Reynolds numbers; that is, the turbulent
Solving this system, we retrieve the optimal coefficients E i j cases.
that minimize the model versus the DNS error. A full discus-
sion of this approach is given in Ref. 16.
V. CONCLUSIONS
If this system of ordinary differential equations !ODEs"
is integrated, we can obtain a good match for the short- as When analyzing this inhomogeneous, anisotropic
well as the long-term prediction for the periodic problem of T-mixer flow at different Reynolds numbers, a qualitative
Re!300 using only nine modes !Fig. 14". We can obtain a change of the solution of the flow domain can be observed. A
good qualitative result for Re!400 for the short-time predic- periodic behavior of the flow pattern at Re!300, due to the
tion with a system of 20 equations, also in consideration of oscillation of the high-speed flow in the symmetry plane, can
the fact that the system displays a chaotic behavior !Fig. 15". be seen. By increasing the Reynolds numbers, this oscillation
The subharmonic frequency of the second mode is developed grows until it breaks down between Re!400 and Re!500.
#Fig. 15!a"$. The harmonic behavior of the third and fourth Any further increasing of the Reynolds number leads to tur-
mode is also reasonably predicted #see Figs. 15!b" and bulent characteristics.
15!c"$. For the long-term runs, the model changes its behav- The proper orthogonal decomposition classifies and
ior toward a very regular periodic pattern. This is not in quantifies these observations by separating the spatial and

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
2730 Phys. Fluids, Vol. 16, No. 8, August 2004 Telib, Manhart, and Iollo

FIG. 15. Comparison between DNS !solid" and LOM !dashed" results for Re!400 using 20 modes. !a" Second eigenfunction a 2 (t). !b" Third eigenfunction
a 3 (t). !c" Fourth eigenfunction a 4 (t).

the time domain into a sum of empirical eigenmodes. These ACKNOWLEDGMENTS


eigenmodes, which are similar to Fourier modes, optimize
This paper is dedicated to Professor Luca Zannetti on the
the captured energy. The time coefficients of the POD modes
occasion of his 60th anniversary. The first author !H.T."
for Re!300 show a harmonic behavior. At Re!400, we can
would like to thank the German Academic Exchange Service
notice a subharmonic cascade, which indicates the transition
and the Richard Winter Foundation for partially supporting
to turbulence, as the characteristic frequency of Re!300, its
this work.
half and its fourth occur. The higher Reynolds number flows
exhibit turbulent features.
1
The obtained POD modes are used to derive a low-order J. Baldyga and J. R. Bourne, Turbulent Mixing and Chemical Reactions
!Wiley, Chichester, 1999".
model. An a priori test limits the possibilities of the model- 2
J. L. Lumley, The structure of inhomogeneous turbulence, in Atmo-
ing: for the periodic case of Re!300 and Re!400, the test spheric Turbulence and Wave Propagation, edited by A. M. Yaglom and
demonstrates that the flow can be reduced to a system of a V. L. Tatarski !Nauka, Moscow, 1967", pp. 166 178.
3
few degrees of freedom, and for the turbulent ones it shows J. H. Citriniti and W. K. George, Reconstruction of the global velocity
field in the axisymmetric mixing layer utilizing the proper orthogonal de-
that the number of modes needed to give a reasonable rep- composition, J. Fluid Mech. 418, 137 !2000".
resentation of the flow is O(105 ), leaving little hope of true 4
S. V. Gordeyev and F. O. Thomas, Coherent structures in the turbulent
low-dimensional modeling for fully developed turbulent planar jet. Part 1: Extraction of proper orthogonal decomposition eigen-
modes and their self-similarity, J. Fluid Mech. 414, 145 !2000".
flows. 5
B. Knight and L. Sirovich, Kolmogorov inertial range for inhomoge-
These results are confirmed by showing that when using neous turbulent flows, Phys. Rev. Lett. 65, 1356 !1990".
O!10" modes to describe the flow and a simple quadratic 6
R. D. Moser, Kolmogorov inertial range spectra for inhomogeneous tur-
model with an adequately computed linear term, it is pos- bulence, Phys. Fluids 6, 794 !1994".
7
P. Holmes, J. L. Lumley, and G. Berkooz, Turbulence, Coherent Struc-
sible to give a good dynamical modeling only of the transi- tures, Dynamical Systems and Symmetry !Cambridge University Press,
tional cases at Re!300 and Re!400, whereas for the high Cambridge, 1996".
Reynolds number cases such attempts fail. 8
G. Berkooz, P. Holmes, and J. L. Lumley, The proper orthogonal decom-

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
Phys. Fluids, Vol. 16, No. 8, August 2004 Analysis and low-order modeling 2731

17
position in the analysis of turbulent flows, Annu. Rev. Fluid Mech. 25, A. J. Chorin, Numerical solution of the NavierStokes equations, Math.
539 !1993". Comput. 22, 745 !1968".
9
L. D. Landau and E. M. Lifshitz, Fluid Mechanics, Number 6 in Course of 18
J. H. Ferziger and M. Peric, Computational Methods for Fluid Dynamics,
Theoretical Physics, Institute of Physical Problems !USSR Academy of 2nd ed. !Springer, Berlin, 1997".
Sciences, Moscow, 1987". 19
C. W. Hirt, B. D. Nichols, and N. C. Romero, Solaa numerical solution
10
D. Rempfer, On low-dimensional galerkin models for fluid flow, Theor.
algorithm for transient fluid flows, Los Alamos National Laboratory
Comput. Fluid Dyn. 14, 75 !2000".
11
D. Rempfer, Low-dimensional modelling and numerical simulation of Report LA-5852 !1975".
20
transition in simple shear flows, Annu. Rev. Fluid Mech. 35, 228 !2003". M. Manhart, F. Tremblay, and R. Friedrich, MGLET: a parallel code for
12
N. Aubry, P. Holmes, J. L. Lumley, and E. Stone, The dynamics of efficient DNS and LES of complex geometries, in Parallel Computa-
coherent structures in the wall region of a turbulent boundary layer, J. tional Fluid Dynamics 2000, edited by C. B. Jenssen, T. Kvamsdal, H. I.
Fluid Mech. 192, 115 !1988". Andersson, B. Pettersen, A. Ecer, J. Periaux, N. Satofuka, and P. Fox
13
B. Podvin, On the adequacy of the ten-dimensional model for the wall !Elsevier Science, Amsterdam, 2001".
layer, Phys. Fluids 13, 210 !2001". 21
M. Manhart, A zonal grid algorithm for DNS of turbulent boundary
14
B. Podvin and P. Le Quere, Low-order models for the flow in a differ- layers, Comput. Fluids 33, 435 !2004".
entially heated cavity, Phys. Fluids 13, 3204 !2001". 22
15 L. Sirovich, Turbulence and the dynamics of coherent structures, part I,
L. Ukeiley, L. Cordier, R. Manceau, J. Del Ville, M. Glauser, and J. P.
II, III, Q. Appl. Math. 45, 561 !1987".
Bonnet, Examination of large-scale structures in a turbulent plane mixing 23
layer, part 2: dynamical systems model, J. Fluid Mech. 441, 67 P. Berge, Y. Pomeau, and C. Vidal, Order Within ChaosTowards a De-
!2001". terministic Approach to Turbulence !Hermann, Paris, 1984".
24
16
B. Galletti, C. H. Bruneau, L. Zannetti, and A. Iollo, Low order modeling D. Rempfer, Koharente Strukturen und Chaos beim laminar-turbulenten
of laminar flow regimes past a confined square cylinder, J. Fluid Mech. Grenzschichtumschlag, Ph.D. thesis, Universitat Stuttgart, Stuttgart,
503, 161 !2004". 1991.

Downloaded 15 Feb 2005 to 147.210.16.211. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp

You might also like