Physica D 202 (2005) 218–237
Stability and accuracy of periodic ﬂow solutions obtained
by a POD-penalty method
S. Sirisup, G.E. Karniadakis∗
Division of Applied Mathematics, Brown University, 182 George Street, Providence, RI 02912, USA
Received 14 May 2004; accepted 11 February 2005
Communicated by I. Mezic
We develop a new penalty method to derive low-dimensional Galerkin models for ﬂuid ﬂows with time-dependent boundary
conditions. We then outline a procedure based on bifurcation analysis in selecting the proper values of the penalty parameter(s)
that yield asymptotically stable periodic solutions of the highest possible accuracy. We illustrate this new approach by studying
ﬂow past a circular cylinder using direct numerical simulation (DNS) data, and a wave-structure interaction problem using
particle image velocimetry (PIV) data.
© 2005 Elsevier B.V. All rights reserved.
MSC: 37E99; 65P99
Keywords: Penalty methods; Low-dimensional; Dynamical systems; Galerkin projections
Low-dimensional systems for unsteady ﬂuid ﬂows, based on the proper orthogonal decomposition (POD), have
had mixed success in predicting the correct dynamics even at exactly the same set of parameters for which the POD
modes were obtained. Speciﬁcally, an erroneous state may be obtained after long-time integration even if the correct
periodic state is set to initialize the simulation—the numerical solution eventually drifts to a new erroneous state.
This has also been observed in other systems, for example in the Kuramoto–Sivashinsky equations . Empirical
ﬁxes based on artiﬁcial dissipation, e.g. , can only correct the dynamics in short-term integration, and more
rigorous procedures need to be followed to guarantee asymptotic stability, e.g. see .
∗Corresponding author. Tel.: +1 401 863 1217; fax: +1 401 863 3369.
E-mail address: email@example.com (G.E. Karniadakis).
0167-2789/$ – see front matter © 2005 Elsevier B.V. All rights reserved.
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 219
A distinction, however, should be made between autonomous low-dimensional dynamical ﬂow systems and
non-autonomous ones. We have observed, for example in many POD studies with diverse ﬂow systems [4,5], that
autonomous systems are more susceptible to this “drifting” while non-autonomous systems may reach asymptot-
ically stable and accurate states without incorporating any special treatment. An example of a non-autonomous
system is oscillatory ﬂow past a circular cylinder. We have observed through accurate numerical integration of the
corresponding POD system (for millions of time steps) that asymptotic stable and accurate states can be reached,
at least for an external frequency close to the natural frequency of the system, i.e. the vortex shedding frequency in
The implementation of complicated boundary conditions in Galerkin systems has historically been a matter
of some controversy ; see also [7–9]. An in-depth study of boundary conditions for Galerkin POD systems
was performed in . Herein we introduce a penalty method, similar in spirit with the “tau” method in spec-
tral methods , but more ﬂexible in many aspects as we will see in this study. In particular, we will study
two different systems based on data obtained from direct numerical simulations (DNS) and experimental results
using particle image velocimetry (PIV) . A new aspect of the current work is the use of the penalty param-
eter(s) as bifurcation parameter(s) in order to perform stability analysis using a standard package, e.g. AUTO
Penalty methods have been used in the past successfully in implementing boundary conditions for different
types of numerical discretizations. For example, ﬁnite difference schemes on complex geometries have been de-
veloped in  using a penalty method to impose Dirichlet boundary conditions. Also, a penalty method was
developed in  to enforce boundary conditions for shock-free compressible Navier–Stokes simulations. A
similar penalty method was used in imposing boundary vorticity constraints in . In general, the penalty ap-
proach enforces the boundary conditions but also accounts for the governing equation at the boundary in a con-
tinuous manner, thus relaxing some of the numerical stiffness associated with very steep gradients at Dirichlet
From the fundamental point of view, we pose the following question:
•Is there a range in the penalty parameter τfor which the periodic solutions of the ﬂow system are asymptotically
stable, and are there any particular values of τfor which the solution is most accurate.
For full numerical discretizations of any type, the accuracy of the solution scales as the inverse of the penalty
parameter in well-resolved simulations. However, efﬁciency considerations require that a ﬁnite value for τmust
be used. As τ→∞we have a strong imposition of the boundary condition while for small values of τwe have a
weak imposition of the boundary conditions. The question then is how do low-dimensional discrete systems differ
from their “full-blown” counterparts in this respect? More importantly, should we impose the boundary conditions
for low-dimensional systems in a strong form or in a weak form. Intuitively, we expect to have stability above a
threshold value in τbut that does not imply good accuracy in predicting the ﬂow dynamics.
The above are some of the issues we address in the current work. The outline of the paper is as follows: in the
next section, we describe the data sets based on which the POD-penalty system is derived. Then, the POD-penalty
formulation is given in some detail. Subsequently, the results of the bifurcation study using the penalty parameter are
presented for the two cases corresponding to DNS and experimental data. Finally, a summary and a brief discussion
is included in the last section.
2. Data gathering
We construct POD modes based on data obtained both from DNS as well as from experiments. We report here
on two prototype cases we have studied with the penalty method.
220 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
Fig. 1. Computational domain for DNS.
2.1. Direct numerical simulation
We consider ﬂow past a circular cylinder for which both two- and three-dimensional POD models have been
time-dependentinﬂowpastacircular cylinder at Reynolds number Re =100and500. The inﬂow velocity is uniform
but oscillatory in time, and is given by
The amplitude of the forcing term A, is kept the same at A=0.1 for both Reynolds numbers. The forcing frequency
is chosen so that we have a lock-in (resonant) state; this is done by choosing the frequency to be close to the Strouhal
number which is 0.1667 for Re =100 and 0.22 for Re =500.
The computational domain is shown in Fig. 1. A time-dependent boundary condition is imposed at the inﬂow
boundary 1; periodicity is imposed on 3and 4while on 2the zero Neumann condition on velocity is im-
posed and the pressure is set to be zero. On the cylinder surface 5the no-slip boundary condition is prescribed.
Converged solutions were obtained using the spectral/hp element method . The domain is discretized into
412 triangular elements while seventh-order Jacobi polynomial basis are used to obtain resolution independent
2.2. Particle imaging velocimetry (PIV) experiment
Here we consider wave interaction with a vertical, surface-piercing cylinder, see Fig. 2. This ﬂow gives rise to
complex forms of wake structure due to the orbital particle trajectories of the incident wave, and the sweeping of
previously generated vortices past the cylinder due to the oscillatory nature of the wave. The cylinder is ﬁxed and the
ﬂow motion is sustained by the wave action. We will employ the experimental results obtained using particle image
velocimetry (PIV) by Yang and Rockwell . For this ﬂow there are two important non-dimensional parameters
that need to be speciﬁed: First, the Keulegan–Carpenter number deﬁned by
in which Aois either the displacement amplitude of the cylinder motion or the amplitude of the oscillatory ﬂow and
Dis the cylinder diameter. Secondly the Stokes number
in which fis the frequency of the motion and νis the kinematic viscosity.
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 221
Fig. 2. Experimental set up .
Quantitative images were obtained using a technique of high-image-density particle velocimetry (PIV). There
are 13 phased-averaged snapshots available for a time period T=0.89. These images will be used to extract the
POD modes. Details of the experiments and the imaging approach are described in detail in . A brief summary
is given next.
The vertical, rigidly suspended cylinder, which is shown in the schematic of Fig. 2, was maintained stationary
during all experiments. It had a diameter of D=12.7mm and a length of L=876mm. The submerged length
of the cylinder was 700mm. The value of the Keulegan–Carpenter number were KC =2πAo/D =13.9atthe
depth of the laser sheet, which is indicated in Fig. 2 as 51mm beneath the quiescent free-surface. The amplitude
Aocorresponds to the radius of the particle orbit of the wave, which was also determined at the depth of 51mm.
Furthermore, the value of the Stokes number was β=fD2/ν =164 for this experiment. The corresponding value
of Reynolds number is Re =KC ×β=2280.
3. POD-penalty systems
In order to employ time-dependent boundary conditions in low-dimensional models, we formulate a new method
to construct Galerkin systems. In particular, we incorporate the boundary conditions directly into the Navier–Stokes
equations as constraints, enforced via suitable penalty parameters. In the next section we will demonstrate how to
select the penalty parameters through bifurcation analysis in order to achieve asymptotically stable and accurate
222 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
Herewe employ the hierarchical POD modes as a trial basis to represent the velocity ﬁeld. In addition, we employ
a Galerkin projection of the Navier–Stokes equations onto these modes to derive dynamical systems to simulate the
ﬂow. Let us decompose the total ﬂow ﬁeld Vas
where U0is the time-averaged ﬁeld. We express uas the linear combination of the POD modes as written in the
u(x, y, t)=φu
j(x, y)aj(t),v(x, y, t)=φv
where aj(t) are the unknown coefﬁcients and φ=(φu,φ
v) deﬁnes the vector of the POD modal basis.
In the following we derive separately the low-dimensional system for the DNS data and the experimental
3.1. DNS: POD-penalty system
The Galerkin projection of the Navier–Stokes equations with penalty terms included onto the jth POD mode is
where τ1is the penalty parameter and U∞is the imposed velocity at the inﬂow boundary 1(see Fig. 1). The
function ϒ(x) is deﬁned as
We note here that on boundary 1we do not impose any boundary conditions as it is now treated as part of the
interior domain. The treatment of the pressure term is of particular importance, so we analyze the corresponding
Galerkin projection by using the Gauss’s theorem to obtain
Obtaining the POD modes from DNS of an incompressible ﬂow ﬁeld leads to divergence-free eigenmodes, and
thus the pressure term inside the domain is eliminated (ﬁrst term in the above equation). On the side bound-
aries 3and 4we assume periodicity and hence the pressure boundary terms cancel each other. On the outﬂow
boundary 2the pressure is set to zero in the corresponding DNS. The inﬂow boundary 1should not be in-
cluded in the computation of the second term of Eq. (4) since we have already included it in the Navier–Stokes
equations. Therefore, there is no contribution from the pressure on this boundary in the integration by parts pro-
cedure. Finally, on the cylinder boundary the test function is zero and thus there is no pressure contribution there
In summary, the Galerkin projection leads to the dynamical system:
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 223
M], where Mis the number of POD modes. The term fj(a) includes the convective and
viscous terms and has the form:
Also, Gj(a) is the boundary penalty term, which is written as follows (in summation convention):
where φm(y|n) means the function of yobtained by evaluating φmon n.
Since U∞(t) is time-dependent, we obtain a non-autonomous system. More details on the derivation of the above
formulation for the Navier–Stokes equation are presented in .
3.2. Experiment: POD-penalty system
The POD-penalty system for the experimental data is derived similarly. By referring to Fig. 3, we now employ
twopenalty terms τ1andτ2for the boundaries 1and 2, respectively.The Galerkin projection of the Navier–Stokes
equations with penalty terms onto the jth POD mode is now
where the projection vector φjis deﬁned as previously, and U1
∞are the velocity vectors at the boundaries 1
and 2, respectively (see Fig. 3). The function ϒi(x) is deﬁned as
Unlike the earlier DNS study where the ﬂow is two-dimensional, in this case the true ﬂow is three-dimensional
but only a two-dimensional slice is visualized via PIV. Correspondingly, imposing the divergence-free condition
on the two-dimensional POD modes is not appropriate. To this end, we will employ the divergent POD modes and
let the penalty terms “counteract” the divergent contributions (ﬁrst term of Eq. (4)); how accurate is this procedure
will be tested by the results presented in the next section. Intuitively, it can be justiﬁed as the penalty term controls
effectively the boundary mass ﬂuxes (on 1and 2), and thus by adjusting the value of τ1and τ2, respectively,
we can counteract any mass sources or sinks due to the pressure contributions in the domain interior. The pressure
contributions from the boundaries vanish due to periodicity and Dirichlet boundary conditions, similarly to DNS
case. Regarding the representation of the time-dependent velocity boundary condition at 1and 2, we have found
that it is accurate to use a Fourier series with 16 Fourier modes to represent the time-periodic forcing at those
boundaries. A systematic investigation of this has been presented in .
The Galerkin projection of the two-dimensional governing equations leads to the dynamical system:
224 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
Fig. 3. Computational and PIV domains for KC =13.9.
M], where Mis the number of POD modes. The term fj(a) has the same form as in Eq. (6).
The boundary penalty term G∗
j(a) is modiﬁed as follows:
3.3. Transformation to an autonomous system
In the next section we will show how to track periodic branches of the dynamical systems described by Eqs.
(5) and (8). However, in order to effectively use the AUTO dynamical system package  to track the periodic
branch, these non-autonomous systems need to be transformed to autonomous systems. To this end, we introduce
the nonlinear oscillators
This particular system has an asymptotically stable solution given by
βt) and q(t)=cos(ˆ
We then incorporate the nonlinear oscillator to the POD-penalty system in order to obtain an equivalent autonomous
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 225
3.3.1. DNS: POD-penalty autonomous system
M], fj(a) is given by Eq. (6) and Gj(a,p) is now deﬁned as
((1.0+Ap, 0) −U0(y|1)) ·φj(y|1)dy.
Therefore, we have replaced the term sin(ωt)inEq.(1) with p(t).
3.3.2. Experiment: POD-penalty autonomous system
The transformation of the non-autonomous system to an equivalent autonomous one for the case of POD-penalty
system derived from experimental data is somewhat more complicated. For this POD-penalty system, we have the
representation of the velocity vectors at the boundaries 1and 2in the form of Fourier series as
where Tis the period, Nthe number of Fourier modes (for this case N=16), yja grid boundary point, and i=1
or 2 for velocity vectors at the boundaries 1and 2, respectively.
We then can transform Eq. (8) into an equivalent autonomous system as follows:
where with a=[a1,a
M], fj(a) is the same as Eq. (6),p=[p1,p
n=1...N. Note that here ˆ
β=nπ/ T .
j(a,p,q) is deﬁned as
Here nyis the number of grid points on 1and 2, and wkis the weight for the trapezoid integration.
226 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
4.1. DNS: Galerkin POD-penalty system
The Galerkin POD-penalty systems for Reynolds number Re =100 and 500 are derived by employing 100
snapshots per period for both cases. We ﬁrst present representative results of the stability of these solutions and
subsequently we investigate their accuracy.
4.1.1. Stability of periodic solutions
Here, we study stability of the solutions of the Galerkin POD-penalty model through bifurcation analysis. We
choose the bifurcation parameter to be the penalty constant τ1. In order to use the AUTO bifurcation tracking
package, for this case, the asymptotically stable periodic solution must be provided—this of course is not known a
priori. To this end, we assume a constant (typically large) value of τ1and obtain the corresponding solution of the
non-autonomous system. However, it is not certain that this solution will have the same period as the forcing period
orevenbeingperiodic [21–23]. Toovercomethis,we will studythe stability of theparticular solution for thatspeciﬁc
value of τ1using Poincar´
e maps, following the work of [24–26]. Speciﬁcally, we have obtained the stability of the
periodic solution using Poincar´
e maps, which we used to ﬁnd the return times of the periodic solution, following
procedures outlined in [24,25]. We then employed the algorithm in  in order to ﬁnd the Floquet multipliers of
the periodic solution.
Let us examine a speciﬁc case to illustrate this approach. We consider the low Reynolds number Re =100 case
with the number of POD modes M=6 and integrate Eq. (5) for a few values of the penalty parameter, say in
the range of τ1∈[2000,3000]. We found, through the Poincar´
e map, that in this range an asymptotically stable
periodic solution does indeed exist. Hence, we can choose any value of τ1∈[2000,3000] to apply AUTO in order
to study stability of periodic solutions more systematically. We also found that the period of the asymptotic state
is T=5.9988 while the corresponding period from the full DNS is identical to this value. As we will show in the
next section, agreement in the period does not imply agreement in the ﬂow ﬁeld dynamics between the DNS and
the low-dimensional system.
Repeating this procedure with higher truncations at M=10 and 20 at the same Reynolds number, we found
similarly asymptotically stable periodic states, which can be used as starting points for the AUTO bifurcation
analysis. A similar study was performed for the POD-penalty system at Reynolds number Re =500 for M=6,
10 and 20 POD modes. With the penalty parameter τ1=3000, the system posses an asymptotically stable solution
with period T=4.5454, which is in agreement with the results from the full DNS. Other large values of the penalty
parameter yield similar results. After obtaining the asymptotically stable periodic solution, we use it as a starting
point for AUTO, and track the stability of the periodic solution by decreasing τ1until loss of stability is detected.
This produces the lowest value of the penalty parameter that guarantees stability. Speciﬁcally, for all the systems
examined here loss of stability shows bifurcation into a torus. The results of this analysis for both Reynolds numbers
are listed in Table 1, where in the third column the minimum values of τ1for stability are presented.
Values of penalty parameter for stability and accuracy requirements
Re Modes Lowest penalty parameter for stability Best penalty parameter for accuracy
Re =100 6 5.24127 ≈5200
10 5.23467 ≈6000
20 3.20842 ≈6500
Re =500 6 1.39258 ≈9000
10 1.26819 ≥2.5×105
20 1.27168 ≥3.0×106
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 227
Fig. 4. Simulation using M=10 POD modes for Re =100, τ1=5×105. DNS: ; POD-penalty simulation: solid line.
4.1.2. Accuracy of periodic solutions
We now turn our attention to the accuracy of the ﬂow dynamics predicted by the low-dimensional system at
different values of the τ1parameter. In Fig. 4 we plot the phase portraits predicted by the POD-penalty system for
τ1=500,000 against the DNS corresponding results for the system with truncation at M=10. At this value of
τ1an asymptotically stable state is obtained with the correct time period but as can be observed in this plot the
accuracy in the ﬂow dynamics is poor. In order to improve this accuracy we ﬁrst deﬁne a relative error for each
penalty parameter by
iare the maximum of the predicted POD modal coefﬁcients corresponding to the low-dimensional
system and DNS, respectively.
In Fig. 5 we plot the results of the bifurcation analysis for Re =100 and M=10 for the ﬁrst four POD modes.
We see that for the higher values of the penalty parameter τ1the accuracy of a∗
icompared to Q∗
iis worse than for
228 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
Fig. 5. Bifurcation diagram using M=10 POD modes for Re =100. a∗
icorresponds to solid line and denotes the maximum of the POD-penalty
i, denoted by dash line, is the corresponding maximum coefﬁcient of the DNS predictions.
the lower values of τ1. Similar results are also observed for the case of Re =100 with M=6 and 20, and also at
Re =500 with M=6. However, this trend is not universal. For example, for the truncation M=20 at Re =500
(see results in Fig. 6), the agreement in the ﬂow dynamics is good for large values of τ1but there is a lower bound
of the penalty parameter τ1below which this agreement is lost. Similar results were obtained at Re =500 with
M=10. Therefore, it is the combination of the penalty parameter and truncation parameter for certain complexity
in the ﬂow dynamics (here governed by the Reynolds number) that determines the quality of the prediction in the
InFig.7,therelativeerror, deﬁned in Eq. (10), is plottedagainstthepenaltyparameterinordertodeterminethebest
penalty parameter for the case Re =100 with M=10. The best penalty parameter is found to be approximately
6000. A qualitatively different result is provided in Fig. 8, where the relative error for the case Re =500 with
Fig. 6. Bifurcation diagram using M=20 POD modes for Re =500. The legend is as in Fig. 5.
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 229
Fig. 7. Left: relative error for the POD-penalty model with respect to DNS with M=10 POD modes for Re =100, and a close-up on the right.
M=20 is plotted against the penalty parameter. Here we ﬁnd that the error decreases monotonically with the
penalty parameter but above some value, much greater than the stability bound, the accuracy saturates.
A summary of our studies to determine the best values of the penalty parameter τ1for the best possible accuracy
is presented in the last column of Table 1. Following the results of the bifurcation analysis, we then performed
integration of the POD-penalty system using the parameters in Table 1. The corresponding results for two typical
cases are shown in Fig. 9 for Re =100 and Fig. 10 for Re =500. For the latter case we observed the following: for
the simulation of the POD-penalty system with M=10, the system can predict correctly up to the ﬁfth mode while
the prediction of higher modes is erroneous. When we increase the number of modes in the system to M=20, the
accuracy of prediction of the dynamics is much better, i.e. good accuracy is now obtained up to 15th mode, see
Fig. 11. We recall that these two cases correspond to a lower bound in the penalty parameter for best accuracy as
determined by the bifurcation analysis. This ﬁnding could possibly suggest that for cases with a lower bound for
the most effective penalty parameter a higher truncation is required to achieve better accuracy; it is not clear if this
result will be true in other ﬂow problems, however. We also note that integrating the POD-penalty Galerkin system
Fig. 8. Relative error for the POD-penalty model with respect to DNS with M=20 POD modes for Re =500, and a close up on the right.
230 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
Fig. 9. Simulation using M=10 POD modes for Re =100, τ1=6000. DNS: ; POD-penalty simulation: solid line.
requires approximately 10 periods to reach an asymptotically stable state even if the “exact” DNS conditions are
used in the initialization process.
In summary, the lesson learned from the DNS study is that above a certain threshold in the value of penalty
parameter, stability of the periodic solution is obtained. However, the best accuracy may be obtained for speciﬁc
values of the penalty parameter that seem to depend strongly on the ﬂow dynamics and the truncation in the
POD-penalty low-dimensional system.
4.2. Experiment: Galerkin POD-penalty system
In this section, we will present the results for the POD-penalty for the experimental data. We have used two
truncations in the number of modes here, M=6 and 12. The number of Fourier modes that represents the periodic
time-dependent boundary condition is set to N=16. For this POD-penalty model, there are two penalty parameters
τ1andτ2that need to be speciﬁed. However,wewill adopt here a procedure where we trackthese penalty parameters
by ﬁxing one of them at a point that the asymptotically stable periodic solution for the POD-penalty system is
obtained. As in the previous study with DNS data, this asymptotically stable periodic solution might not be accurate
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 231
Fig. 10. Simulation using M=20 POD modes for Re =500, τ1=4×106. DNS: ; POD-penalty simulation: solid line.
but it will be used as a starting point for AUTO to track the minimum value of either τ1and τ2for asymptotic
stability of the periodic solutions.
4.2.1. Stability of periodic solutions
From preliminary numerical experiments for both M=6 and 12 we have determined that at τ1=1000 and
τ2=1000 the corresponding POD-penalty systems possess an asymptotically stable solution; see Fig. 12 for these
speciﬁc parameters. As in the previous case with the DNS data, the periodic solution is then studied through the
e map to examine its stability. We have found that the periodic solution for these speciﬁc penalty parameters
is indeed asymptotically stable with period of T=0.89, which is in agreement with the data from PIV.
Havingdeterminedthe starting pointfor AUTO,thetracking of stabilityof the periodicsolution is then performed
by ﬁxing τ1=1000 and decreasing τ2until loss of stability is detected. We then switch the role of τ1and τ2and
perform an analogous analysis. This study produces the lowest values of the penalty parameter for stability, and the
corresponding results are presented in Table 2. We have found that for both values of M=6 and 12 the periodic
solution loses its stability when one of the Floquet multipliers crosses the unit circle at −1. This was also observed
232 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
Fig. 11. Higher modes from the simulation using M=20 POD modes for Re =500, τ1=4×106. DNS: ; POD-penalty simulation: solid
for M=6 with τ2=1000 ﬁxed while varying τ1. However, in the case of the truncation with M=12 the periodic
solution loses its stability when one of the Floquet multipliers crosses the unit circle at 1.
4.2.2. Accuracy of periodic solutions
In order to ﬁnd the most effective values of the penalty parameters that produces the most accurate asymptotically
stable periodic solution compared to the data from PIV, we also track the periodic branch by increasing τ1or
Values of penalty parameters for stability limit and accuracy requirement for the POD-penalty system corresponding to experiment with
Modes Lowest penalty parameter for stability Best penalty parameters for accuracy
τ1=1000 τ2=1000 τ1=1000 τ2=1000
M=6 201.237 27.935 ≈2×105,∗≈7.5×104
M=12 142.652 601.109 ≥2×107≥7.5×107
∗Corresponding asymptotically stable periodic solution has a relative error greater than 5%.
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 233
Fig. 12. Simulation using M=12 POD modes for PIV data, τ1=1000, τ2=1000. PIV: ; POD-penalty simulation: solid line.
τ2while ﬁxing the other one at 1000. Then, we compute the relative error in order to ﬁnd the best value of
the penalty parameter for accuracy. In Fig. 13, the relative error is presented for the case of M=12 with ﬁxed
τ2=1000. A summary of best values of the penalty parameter for all cases is presented in the last column of Table
2. Using the values from the table we then integrate in time the POD-penalty system to reach the asymptotic stable
periodic states. The results of such simulations are presented in Figs. 14 and 15 as phase portraits. There is very
good agreement with the corresponding experimental data with the higher truncation giving higher accuracy, as
We have developed a Galerkin POD-penalty method to construct low-dimensional dynamical systems for un-
steady ﬂuid ﬂows with time-dependent boundary conditions. Penalized boundaries are incorporated directly in the
Galerkin statement of the Navier–Stokes equations, and thus information about the pressure ﬁeld on such bound-
aries is not required. The resulting dynamical system is non-autonomous, so we couple it to an equivalent nonlinear
oscillator in order to study its stability using standard bifurcation analysis.
234 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
Fig. 13. Relative error for the POD-penalty model from PIV data. Here, M=12 with ﬁxed τ2, and a close up on the right.
Fig. 14. Simulation using M=6 POD modes for PIV data, τ1=7.5×104,τ2=1000. PIV: ; POD-penalty simulation: solid line.
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 235
Fig. 15. Simulation using M=12 POD modes for PIV data, τ1=8×107,τ2=1000. PIV: ; POD-penalty simulation: solid line.
We study the stability and accuracy of periodic solutions using the penalty parameter(s) as bifurcation param-
eter(s). We consider two prototype ﬂows based on results from direct numerical simulations (ﬂow past a circular
cylinder), and from experiments (wave–structure interaction). The results from both studies are qualitatively similar.
We ﬁnd that there is a threshold value of the penalty parameter above which asymptotic stability of the periodic
solution is guaranteed. This is an expected result, similar to what is known for numerical discretizations of Navier–
Stokes equations. The surprising, however, ﬁnding is that the accuracy of the solution predicted by the Galerkin
POD-penalty system does not improve as the penalty parameter increases, as it is the case for full numerical dis-
cretizations. Instead, there is a speciﬁc range within which the solution is accurate. In particular, depending on the
number of modes (i.e. truncation) and the ﬂow complexity (i.e. Reynolds number) the best solution may correspond
to a speciﬁc value of the penalty parameter or a range well above the threshold value for stability.
In numerical discretizations that employ the penalty approach to impose Dirichlet or other type of boundary
conditions,asthe penalty parameter approachesa very large number (e.g. inverseofmachineprecision)theboundary
conditions are imposed exactly, i.e. in a strong form. Correspondingly, the error in the solution scales inversely
proportional to the penalty parameter. Our ﬁndings here suggest that for low-dimensional systems, imposing the
boundary conditions in a strong form may lead to an erroneous solution. Similar trends have been observed in
spectral penalty methods for simulations of high Reynolds number turbulence at relatively low resolution . This
236 S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237
can have great consequences in constructing effective low-dimensional dynamical systems as well as in formulating
proper boundary conditions in large-eddy simulations. However, generalization of this conclusion to other ﬂow
problems has to be tested very carefully.
The ﬁrst author gratefully acknowledges the Development and Promotion of Science and Technology Talents
(DPST) project from Thailand for providing his scholarship during his graduate studies at Brown University. The
authors would like to thank Dr. Y. Yang and Prof. D. Rockwell for providing PIV data. This work was supported
by ONR and NSF, and computations were performed at the facilities of NCSA (University of Illinois at Urbana-
Champaign) and at DoDs NAVO MSRC.
 N. Aubry, W.Y. Lian, E.S. Titi, Preserving symmetries in the proper orthogonal decomposition, SIAM J. Sci. Comput. 14 (2) (1993)
 W. Cazemier, R.W. Verstappen, A.E. Veldman, Proper orthogonal decomposition and low-dimensional models for driven cavity ﬂows,
Phys. Fluids 10 (7) (1998) 1685–1699.
 S. Sirisup, G.E. Karniadakis, A spectral viscosity method for correcting the long-term behavior of POD models, J. Comp. Phys. 194 (1)
 X. Ma, G.E. Karniadakis, H. Park, M. Gharib, DPIV-driven ﬂow simulation: a new computational paradigm, Proc. R. Soc. London A 459
(2031) (2003) 547–565.
 S. Sirisup, G.E. Karniadakis, Y. Yang, D. Rockwell, Wave-structure interaction: simulation driven by quantitative imaging, Proc. R. Soc.
London A 460 (2043) (2004) 729–755.
 X. Zhou, L. Sirovich, Coherence and chaos in a model of turbulent boundary layer, Phys. Fluid A 4 (12) (1992) 2855–2874.
 G. Berkooz, P. Holmes, J.L. Lumley, N. Aubry, E. Stone, Observations regarding ‘Coherence and chaos in a model of turbulent boundary
layer’ by X. Zhou and L. Sirovich [Phys. Fluids A 4 (1992) 2855], Phys. Fluid 6 (4) (1994) 1574–1578.
 L. Sirovich, X. Zhou, Reply to ‘Observations regarding ‘Coherence and chaos in a model of turbulent boundary layer’ by X. Zhou and L.
Sirovich, [Phys. Fluids A 4 (1992) 2855], Phys. Fluid 6 (4) (1994) 1579–1582.
 P. Holmes, J.L. Lumley, G. Berkooz, Turbulence, Coherence Structures, Dynamical Systems and Symmetry, Cambridge University Press,
 J.F. Gibson, Dynamical Systems Models of Wall-Bounded, Shear-Flow Turbulence, PhD thesis, Cornell University, 2002.
 D. Gottlieb, S.A. Orszag, Numerical Analysis of Spectral Methods: Theory and Applications, SIAM-CMBS, Philadelphia,
 Y. Yang, D. Rockwell, Wave interaction with a vertical cylinder: spanwise ﬂow patterns and loading, J. Fluid Mech. 460 (2002) 93–
 E.J. Doedel, R.C. Paffenroth, A.R. Champneys, T.F. Fairgrieve, Yu.A. Kuznetsov, B. Sandstede, X. Wang, Auto 2000. Continuation and
bifurcation software for ordinary differential equations (with homcont). Technical Report, Caltech, 2001.
 S. Abarbanel, A. Ditkowski, Asymptotically stable fourth-order accurate schemes for the diffusion equation on complex shapes, J. Comp.
Phys. 133 (2) (1996) 279–288.
 J.S. Hesthaven, D. Gottlieb, A stable penalty method for the compressible Navier–Stokes equations. I. Open boundary conditions, SIAM
J. Sci. Comput. 17 (3) (1996) 579–612.
 J. Trujillo, G.E. Karniadakis, A penalty method for the vorticity–velocity formulation, J. Comp. Phys. 149 (1999) 32–58.
 A.E. Deane, I.G. Kevrekidis, G.E. Karniadakis, S.A. Orszag, Low-dimensional models for complex geometry ﬂows: application to grooved
channels and circular cylinders, Phys. Fluids A 3 (10) (1991) 2337–2354.
 X. Ma, G.E. Karniadakis, A low-dimensional model for simulating 3D cylinder ﬂow, J. Fluid Mech. 458 (2002) 181–190.
 G.E. Karniadakis, S.J. Sherwin, Spectral/hp Element Methods for CFD, Oxford University Press, 1999.
 S. Sirisup, Issues in low-dimensional modeling of unsteady ﬂows: convergence, asymptotic stability and reconstruction procedures, PhD
thesis, Division of Applied Mathematics, Brown University, 2005.
 T.A. Burton, Stability and Periodic Solutions of Ordinary and Functional Differential Equations, Academic Press, 1985.
 T. Yoshizawa, Stability Theory and the Existence of Periodic Solutions and Almost Periodic Solutions, Springer-Verlag, 1975.
S. Sirisup, G.E. Karniadakis / Physica D 202 (2005) 218–237 237
 G.Makay,On some possibleextensionsofMassera’s theorem, ProceedingsoftheSixthColloquium on the QualitativeTheoryofDifferential
Equations (In EJ Qualitative Theory of Differential Equations), vol. 16, 2000, pp. 1–8.
 M. Hannon, On the numerical computation of Poincar´
e maps, Physica D 5 (1982) 412–414.
 W. Tucker, Computing accurate Poincar´
e maps, Physica D 171 (3) (2002) 127–137.
 R. Seydel, Practical Bifurcation and Stability Analysis: From Equilibrium to Chaos, Springer-Verlag, 1994.
 P.J. Diamessis, J.A. Domaradzki, J.S. Hesthaven, A spectral multidomain penalty method model for the simulation of high Reynolds
number localized incompressible stratiﬁed turbulence, J. Comput. Phys. 202 (1) (2005) 298–322.