PHYSICAL REVIEW LETTERS week ending

PRL 100, 074301 (2008) 22 FEBRUARY 2008

Flapping States of a Flag in an Inviscid Fluid: Bistability and the Transition to Chaos
Silas Alben*
School of Mathematics, Georgia Institute of Technology, Atlanta, Georgia 30332-0160, USA

Michael J. Shelley
Applied Math Laboratory, Courant Institute, New York University, New York, New York 10012, USA
(Received 26 October 2007; published 21 February 2008)
We investigate the ‘‘flapping flag’’ instability through a model for an inextensible flexible sheet in an
inviscid 2D flow with a free vortex sheet. We solve the fully-nonlinear dynamics numerically and find a
transition from a power spectrum dominated by discrete frequencies to an apparently continuous spectrum
of frequencies. We compute the linear stability domain which agrees with previous approximate models in
scaling but differs by large multiplicative factors. We also find hysteresis, in agreement with previous
experiments.

DOI: 10.1103/PhysRevLett.100.074301 PACS numbers: 46.40.Jj, 46.40.Ff, 47.15.ki, 47.52.+j

A longstanding problem in the coupled motion of elastic
f (here, mass per unit area) that moves past the flag with
bodies and high-speed flows is the instability underlying free-stream speed U. Scaling space on L, and time on L=U,
the flapping of flags. One of the earliest attempted explan- the flag with position X
s; t (s is arclength; 0  s  1)
ations dates to Lord Rayleigh, who made an analogy to an evolves via Newton’s 2nd law as
instability of a fluid jet [1]. Since the 1930s, workers in
aeronautics and applied mechanics have addressed the R1 Xtt  @s
T s^   R2 @ss
n
^  pn;
^ (1)
stability problem using linearized and approximate flow where T is the tension that enforces inextensibility, [p] is
models [2 – 4], and empirical models [5]. Recent models the pressure jump across the flag, and  is the flag curva-
[6 –8] have recalled some elements of this earlier work ture. The tension has been scaled by
f U2 L, the pressure
while introducing new approximations. Realistic unsteady by
f U2 , and R1 and R2 are the two control parameters of
flow solvers allow for numerical simulations at moderate the dynamical system with R1 
s =
f L the dimension-
Reynolds numbers (O
102 ) [9,10] but accurate studies at less mass of the flag and R2  B=
f U2 L3 its dimension-
higher Reynolds numbers are difficult due to the necessity
less rigidity (also an inverse square velocity). The sheet is
to resolve the production and dynamics of thin free shear
held and clamped at s  0, its leading end, with zero
layers.
deflection. Free-end boundary conditions are assumed at
In this work we simulate the nonlinear dynamics of a
s  1 with T    s  0 there, and the tension can be
heavy elastic sheet which moves in a 2D inviscid fluid and
eliminated from Eq. (1) by integration of the s^ component
sheds a vortex sheet from its trailing edge. Most previous
from s  1.
flag models can be considered as approximations, with
As the flag is also a surface in the flow, by the kinematic
varying degrees of accuracy, to this model. We characterize
boundary condition it must move with normal velocity  of
the behavior of flapping flags at large amplitudes and over
the fluid, which is continuous across Cb although the
many flapping periods, and demonstrate a transition from
tangential velocity is not. Hence we can write
periodic to chaotic flapping as bending rigidity is de-
creased. We also determine the stability boundary of the X t  n^  ^s; (2)
flow-aligned state for the flag, in the two-dimensional Rs 0 0
parameter space of dimensionless flag inertia and bending where the choice   0 ds 
s 
s0  enforces the same
rigidity, and compare this with two recent models [7,8]. frame as is implicit in Eq. (1), i.e., s is independent of t.
This comparison indicates that greatly-simplified models The wake shed behind the flag is modeled as a (free) vortex
coupling flag rigidity, and flag and fluid inertia, can yield a sheet (labeled Cf in Fig. 1), or a -function distribution of
qualitative understanding of flag stability. Finally, we dem- vorticity along a curve in the plane. The flag can also be
onstrate that this model exhibits bistability over a range of considered a (bound) vortex sheet (labeled Cb in Fig. 1)
dimensionless rigidity, which is consistent with recent due to the tangential velocity jump across it. The velocity
experiments of Zhang et al. in soap-film flows [11] and induced by the singular vorticity distribution on C  Cb 
of Shelley et al. in water tunnels [8]. Cf is given by the Biot-Savart integral, and relative to the
The flag model.—We consider the 2D flag as an inex- background flow is:
tensible elastic sheet of length L, mass per unit length
s ,
and rigidity B, moving under the pressure forces of a 1 Z
x  X
s0 ?
u
x  x^  ds0 
s0  ; (3)
surrounding inviscid and incompressible fluid of density 2 C jx  X
s0 j2

0031-9007=08=100(7)=074301(4) 074301-1 © 2008 The American Physical Society

 PHYSICAL REVIEW LETTERS week ending
PRL 100, 074301 (2008) 22 FEBRUARY 2008

n 2 , as used by Krasny [14] and many others since. This

allows vorticity accumulations along the free sheet to form
s vortical spirals [14]. Here we set  equal to 0.2, small
Cb enough to capture well-defined vortical rolls in the wake.
Cf On the bound sheet (the flag) we retain the singular kernel,
as mollifying there makes the matrix underlying the
Newton iteration ill-conditioned.
FIG. 1 (color online). Flag and wake dynamics for R1  0:3
and R2  0:014 45. The time-averaged flow velocity field of the
The code shows convergence at the expected orders as
flag from t  0 to t  120 is shown (red arrows: average over the spatial and temporal grids are refined. We have repro-
60 flapping periods), with the background velocity removed. The duced known steady flows past rigid plates [15]. We have
flag position (Cb , in black, with normal n^ and tangent s^ labeled) also compared an impulsively started flat plate, for which a
and trailing vortical wake (the free-sheet Cf , blue-dashed line) is growth rate of circulation of
0 t1=3 is predicted [12,16].
shown at t  120. We find a temporal exponent of 0:34 0:02 between t  0
and 1, for   0:1 and 0.2, which is the same growth
where x? 
y; x,  is the vortex sheet strength (the reported by Jones for   0:2 [12].
jump in tangential velocity across the sheet), and u satisfies Results.—We first demonstrate that our flag model cap-
the condition of zero jump in normal velocity. The average tures phenomena previously observed in experiment (e.g.,
of velocities above and below C is [8,11]). Figure 1 (and supplementary materials [17]) shows
the result of a long-time simulation using R1  0:3 and
1
X
s  X
s0 ? R2  0:014 45, values for which only steady flapping with
W
s  x^  6 ds0 
s0  ; (4) a well-defined period is observed over a broad range of
2 C jX
s  X
s0 j2
initial perturbations. Here, the dynamics was induced by
where the Birkhoff-Rott integral is of principal value type. smoothly perturbing the leading-edge position [of ampli-
Hence,   W n. ^ The evolution of  on Cb is determined tude O
105 ], while smoothly ramping the background
by taking the tangential component of the difference of flow velocity up to unity. By t 10, exponential growth of
limits of the Euler equations above and below the flag the initial perturbation has saturated, and the flag has
(expressed in the frame where s is independent of t), entered a periodic flapping state, shown in Fig. 1 at t 
yielding 120. The trailing shed sheet has rolled up into spiral
vortices associated with local Kelvin-Helmholtz instabil-
t  @s 
    @s p; (5) ities [15], and has formed a ‘‘von Kármán’’ vortex street
where   W s^ . Using that ps1  0, Eq. (5) can be (see also the supplementary movies [17]). The time-
integrated across Cb to eliminate [p]. On Cf , p  0 and averaged velocity field with the background flow removed
the evolution of  is best considered in the average velocity is directed upstream towards the flag. The upstream fluid
frame where t  0, and hence is invariant. A convenient momentum thus generated is associated with a drag force
parametrization of Cf is then total circulation
0 to the free on the flag.
end of the flag with its initial value determined by the A sequence of snapshots for this case is shown in
Fig. 2(a), in which a simple periodic flapping dynamics
shedding condition from the flag end
_ 0  js1 (see
is seen. The flapping consists of backwards traveling waves
Jones [12]). We omit the small leading-edge vortex, which
of simple spatial form, with a wave envelope that is
gives a small perturbation to the transverse pressure forces
which drive the flag. We also neglect shear stresses in this roughly linear in distance from the flag support point.
simple model, which are dominated by pressure forces The temporal
R Fourier transform U^ ! of the bending energy
2
U  Cb  ds=2 is shown in Fig. 3(a). It is dominated by
when the flag undergoes O
1 deflections, but may none-
theless be important for determining the onset of flapping. peaks at two distinct frequencies, the higher of which is a
The numerical method.—We evolve Eqs. (1), (2), and harmonic of the lower, and also has a small long-time
(5), and boundary conditions, numerically by expanding component near frequency 0.4.
X, [p], and  on the flag as first-kind Chebyshev poly- For a range of larger R2 , these frequencies change
nomials. This representation is convenient for evaluating smoothly and reflect simple harmonic flapping. However,
the integral in Eq. (4) [13]. The system is discretized to Fig. 3(b) shows the temporal spectrum from a long-time
second-order implicitly in time (for stability purposes) simulation from the same initial conditions if R2 is very
with new segments of the free vortex sheet created at slightly decreased (to R2  0:014 36, 1% change). The
each time step [12]. The set of nonlinear discrete dominant frequencies shift abruptly higher and new fre-
Eqs. (1), (2), and (5) is solved by a Newton iteration. quency components appear. This is associated with a
At each time step, after convergence of the Newton change in flapping dynamics, illustrated by Figs. 2(b) and
iteration, Eq. (4) is used to update the position of the free 3(c), for which R2  0:0138[17]. In Fig. 2(b) the vortex
sheet. On the free sheet, the singular kernel x? =jxj2 is wake is compressed relative to that in Fig. 2(a), due to the
mollified and replaced by the smoothed kernel x? =
jxj2  higher dominant frequency over which vorticity changes
074301-2
 PHYSICAL REVIEW LETTERS week ending
PRL 100, 074301 (2008) 22 FEBRUARY 2008

a) 1.5
1
of chaotic dynamics. This is illustrated in Figs. 2(c) and
0.5 3(d) [17] for which R2  0:0025. The flag snapshots over
0 regular time intervals fill space within a roughly linear
-0.5 envelope, and the flag shape has become more irregular.
-1 The vortex wake seen in Fig. 2(c) retains predominantly
b) 1
6 7 8 single-signed vorticity above and below the center line, but
0.5 the structure of the vortices and the spacing between them
0 is irregular. The spectrum now shows no apparent domi-
d)
-0.5 nant frequencies but rather a wide range of excited fre-
-1 quencies. Here the time-averaged fluid drag is 0.427, still
c) 0.8 large but somewhat decreased from the previous case.
0.4 As a point of comparison, in Fig. 2(d) we show snap-
0
-0.4 shots from the model and from a recent experiment using a
-0.8 paper flag in air [18]. In both cases R1  0:37. For the
0 1 2 3 4 5 experiment, R2 is unavailable. For the model we take R2 
0:018, giving the longest wavelength mode; the snapshots
FIG. 2 (color online). Snapshots of the flag for fixed mass
vary little with R2 until the abrupt onset of a higher
(R1  0:3) and decreasing rigidities R2 . (a) The observed flap-
ping mode at R2  0:014 45, about a factor of 2 below the wavelength mode. The model shapes show larger curva-
critical R2  0:0262; (b) a higher energy flapping mode for tures but similar flapping amplitude (the model gives more
R2  0:0138; (c) a chaotic flapping mode at R2  0:0025. similar shapes using R1  0:25). Possible reasons for the
(d) Comparison of experimental flag snapshots in [18], differences are 3D effects and skin friction in the experi-
Fig. 12b, with model shapes. In both cases R1  0:37 (see text ment, and nonuniformity in the experimental flow.
for R2 ). Linear stability.—Here we study directly the linear
stability of the straight flag, and assess recent models of
the transition to instability. We infer stability directly by
sign. The fluid drag on the flag also increases dramatically imposing small initial perturbations and tracking their
with this transition from harmonic flapping, moving in growth (if any) in the full numerical system. By searching
time-average from 0.139 to 0.525 between panels a to b, over a large portion of the R1 -R2 space, we identified a
with the drag time-trace punctuated by intermittent high- boundary curve, plotted in Fig. 4, below which (in R2 )
drag ‘‘flag-snapping’’ events (as is somewhat evident from small perturbations grow exponentially, with a rate which
Fig. 2(b)]. In short, the sheet dynamics now has more increases with distance from the line, and above which
degrees of freedom in both time and space, the latter being
of larger amplitude and bending energy.
Further decreases in R2 introduce yet more spatial com- 0
10
plexity as well as the broad spectral content characteristic
-1
10
-3
R1~ R2
x10
a) -2
10
^
2
|Uω| 1 R2
0 -3
10
0 2 4 6 8 10 12 14
ω
x10-3 -4
b) 15 10
10
-3 -2 -1 0 1 2
5 10 10 10 10 10 10
0 R1
0 2 4 6 8 10 12 14
ω
x10-3
c) FIG. 4. Computed stability boundary in the R1 R2 plane. The
10 upper solid boundary gives the smallest R2 above which a small
5 2

0
leading-edge forcing (y
t  105
2t2 e
2t ) does not lead to
0 2 4 6 8 10 12 14 flapping. The lower solid boundary is the largest R2 below which
ω
x10-3 such forcing leads to exponential growth of elastic energy in
d)
3 time until the flag saturates with O
1 flapping, as shown in
2
1 Figs. 1 and 2. The solid line gives the scaling R1 R2 for
0 comparison at small flag masses. The black crosses mark the
0 2 4 6 8 10 12 14 cases shown in Fig. 2 [upper cross is (a) and (b), and lower is
ω
(c)]. The dashed line shows the stability boundary from [8], and
FIG. 3. Temporal power spectra of flag bending energy vs. the dash-dotted line the corresponding boundary plotted in Fig. 3
frequency, for R1  0:3 and R2 equal to 0.014 45 (a), 0.014 36 of [7] (showing only the portion R2 const, but having the same
(b), 0.0138 (c), and 0.0025 (d). asymptotic scalings as the model here and in [8]).
074301-3
 PHYSICAL REVIEW LETTERS week ending
PRL 100, 074301 (2008) 22 FEBRUARY 2008
0.014
a) A0= 0.4 b) 0.012
0.01
flapping, while below the perturbation decays to the
-2 straight flag case. This decay, and hence the bistability,
10 0.008

<U Elastic >

0.006 may be surprising given the conservative nature of the
0.004
A0 = 0.2 system. However, the flag is coupled to the fluid, and by
UElastic

0.002
0
2.6 2.65 2.7 2.75
-3
2.8 this can be convectively stabilized; if the flag perturbation
R2 x 10
-3 0.08 is sufficiently small, it is simply swept into the wake and
10 A0 = 0.1 c) moved downstream.
0.06

The hysteresis loop corresponding to this value of R1 

<UElastic>
0.04
A 0 = 0.06 0.02
0:3 is shown in Fig. 5(b) and that for R1  0:03 in Fig. 5(c).
-4
We find that the hysteresis loops extend over factors of
0
10 0 5 10 15 20 25
0.025 30 0.0258 0.0262 0.0266
0.0254 1.025 to 1.04 in R2 in the two cases, compared with a factor
R2
Time of 2 in the experiment of ([8], Fig. 4) where R1  0:05, and
a factor of 1.33 in [18] for 0:1 < R1 < 1. The larger
FIG. 5. Bistability and hysteresis in the onset of flapping with
changes of flag rigidity R2 . (a) For R1  0:03 and R2  2:74
hysteresis in the experiments may be due to the effects of
103 , four different values of transverse perturbations fluid viscosity and structural damping.
2
A0
2t2 e
2t are applied to the flag leading edge, for A0 in We thank M. Jones and J. Zhang for useful discussions.
the range 0.06 to 0.4, and bending energy is plotted in time. The We note that J. Zhang has recently made informal obser-
hysteresis loop for R1  0:03 is shown in (b), with the flat and vations of a transition in flag flapping behavior in soap-film
steady flapping states in (a) circled. The time-averaged bending flows similar to that seen between Figs. 2(a) and 2(b). S. A.
energy in the asymptotic steady state is plotted. (c) The hystere- acknowledges support from an NSF Mathematical
sis loop for mass R1  0:3, an order of magnitude larger than in Sciences Grant. M. J. S. acknowledges support from the
(b), showing a range of bistability which is also an order of DOE (No. DE-FG02-88ER25053).
magnitude larger.

perturbations in y do not increase. This stability boundary

shows a linear slope at small R1 , where a proportionately *[email protected]
[1] Lord Rayleigh, Proc. London Math. Soc. s1-10, 4 (1878).
small bending rigidity is needed for the instability to occur.
[2] T. Theodorsen, NACA, Technical Report No. 496, 1935.
At larger R1 the curve becomes nearly flat. A physical [3] T. von Kármán and J. M. Burgers, in Aerodynamic Theory,
interpretation for the transition to instability is that greater edited by W. Durand (Springer-Verlag, Berlin, 1935),
flag inertia allows small oscillations of the flag to increase Vol. 2, p. 346.
more readily against the damping of fluid inertia. Also [4] H. Glauert, The Elements of Airfoil and Airscrew Theory
plotted in Fig. 4 is the stability boundary (dashed line) (Cambridge University Press, Cambridge, England, 1947).
for a biinfinite flag from [8]. It has the same asymptotic [5] R. L. Bisplinghoff and H. Ashley, Principles of Aero-
scalings but differs by a large multiplicative factor. This elasticity (Dover, New York, 2002).
discrepancy is not surprising because the pressure distri- [6] A. D. Fitt and M. P. Pope, J. Eng. Math. 40, 227 (2001).
bution in the current model is dominated by vortex shed- [7] M. Argentina and L. Mahadevan, Proc. Natl. Acad. Sci.
ding, a term absent in the infinite flag model. The stability U.S.A. 102, 1829 (2005).
[8] M. Shelley, N. Vandenberghe, and J. Zhang, Phys. Rev.
boundary presented in Argentina & Mahadevan [7] for an
Lett. 94, 094302 (2005).
inviscid flow with vortex shedding has the same asymptotic [9] L. Zhu and C. S. Peskin, J. Comput. Phys. 179, 452 (2002).
scalings, though again with a different multiplicative factor [10] B. Connell and D. Yue, J. Fluid Mech. 581, 33 (2007).
(see Fig. 4). Their model approximates the flow around the [11] J. Zhang, S. Childress, A. Libchaber, and M. Shelley,
flexible flag as the steady potential flow past a hinged rigid Nature (London) 408, 835 (2000).
plate, scaled by a variable flag velocity. All three models [12] M. Jones, J. Fluid Mech. 496, 405 (2003).
have similar asymptotic scalings, presumably because the [13] M. A. Golberg, Numerical Solution of Integral Equations
dominant terms in the flag equation have the same scaling (Plenum, New York, 1990).
with physical parameters in the three models. [14] R. Krasny, J. Comput. Phys. 65, 292 (1986).
Bistability.—Our nonlinear flag model supports bista- [15] P. Saffman, Vortex Dynamics (Cambridge University
bility of the straight and flapping states, as has been ob- Press, Cambridge, England, 1992).
[16] D. I. Pullin, J. Fluid Mech. 88, 401 (1978).
served experimentally [8,11] and numerically [9]. To show
[17] See EPAPS Document No. E-PRLTAO-100-038805 for
this, the initial perturbation amplitude applied to the flag four supplementary movies, showing the transition from
was varied in a range of R2 near the stability boundary. single-frequency to chaotic flapping. For more informa-
Figure 5(a) shows the elastic energy in time for four differ- tion on EPAPS, see http://www.aip.org/pubservs/
ent initial vertical perturbations. A division between the epaps.html.
two states occurs for amplitude 0.1: Above this value, the [18] Y. Watanabe, S. Suzuki, M. Sugihara, and Y. Sueoka,
perturbed flag reaches a unique state of large-amplitude J. Fluids Struct. 16, 529 (2002).

074301-4