Professional Documents
Culture Documents
DOI 10.1007/s00348-010-0911-3
RESEARCH ARTICLE
Received: 7 January 2010 / Revised: 14 May 2010 / Accepted: 5 June 2010 / Published online: 2 February 2011
Ó Springer-Verlag 2011
Abstract The dynamic mode decomposition (DMD) is a Experiments and numerical computations are the two pil-
data-decomposition technique that allows the extraction of lars upon which this transfer relies. In the numerical field,
dynamically relevant flow features from time-resolved ever-increasing computer resources have recently enabled
experimental (or numerical) data. It is based on a sequence the analysis of rather complex flows as to their global
of snapshots from measurements that are subsequently stability, receptivity and controllability (Edwards et al.
processed by an iterative Krylov technique. The eigen- 1994; Lehoucq and Scott 1997; Theofilis 2000). Modern
values and eigenvectors of a low-dimensional representa- techniques, mainly adapted from fast iterative algorithms
tion of an approximate inter-snapshot map then produce for linear algebra problems, are becoming commonplace in
flow information that describes the dynamic processes many large-scale and multi-physics calculations. On the
contained in the data sequence. This decomposition tech- experimental side, remarkable progress has been made in
nique applies equally to particle-image velocimetry data the measurement of flow quantities. In particular, particle-
and image-based flow visualizations and is demonstrated image velocimetry (PIV) techniques are becoming more
on data from a numerical simulation of a flame based on a and more powerful, and the degree of spatial and temporal
variable-density jet and on experimental data from a lam- resolution often rivals equivalent numerical simulations. In
inar axisymmetric water jet. In both cases, the dominant addition, the use of high-speed cameras and image-based
frequencies are detected and the associated spatial struc- visualization techniques has provided unprecedented
tures are identified. insight into fast fluid dynamic processes. In contrast to
numerical simulations, however, the quantitative evalua-
tion of these experimental data lags behind the possibilities
1 Introduction for computational results. The reason for this gap lies in the
above-mentioned algorithms. Nearly all the algorithms in
An objective and quantitative description of flow behavior current use in quantitative fluid dynamics rely on a model
is the underlying principle of many scientific inquiries in equation, commonly the (linearized) Navier–Stokes equa-
computational and experimental fluid dynamics. A pro- tions and their variants. This model appears in the algo-
found understanding of all relevant processes also builds rithms as a matrix-vector multiplication, which is pivotal in
the foundation for the transfer of fundamental fluid effects maintaining orthogonality or in adding robustness to the
into technological applications and fluid-based devices. respective method (Lehoucq and Scott 1997). This neces-
sary model information is easily furnished by a numerical
code, but is impossible to access in experimental mea-
surements. In physical experiments, only the measured data
are available. To achieve the same level of quantitative
description of fluid processes, the common algorithms have
P. J. Schmid (&)
to be modified to eliminate their reliance on a model and to
Laboratoire d’Hydrodynamique (LadHyX),
Ecole Polytechnique, Palaiseau 91128, France restrict their input to data only. The dynamic mode
e-mail: peter@ladhyx.polytechnique.fr decomposition (DMD) (Schmid and Sesterhenn 2008;
123
1124 Exp Fluids (2011) 50:1123–1130
Schmid 2010) achieves this goal by extracting the relevant that, for experimental data, the snapshots vj are produced
dynamics from a sequence of data; similar to the Arnoldi by a nonlinear process. Invoking a linear-tangent
algorithm, a high-degree matrix polynomial is fitted to the approximation, a linear mapping A from one snapshot to
data sequence. An inter-snapshot linear map is then iden- the next is assumed; this mapping is furthermore taken as
tified that acts as a low-dimensional approximation of the constant over the data sequence. We can thus write
system dynamics. This linear map is computed by pro- vjþ1 ¼ Avj : ð2Þ
cessing the data sequence, generated by a nonlinear pro-
cess, and represents the optimal linear operator (in a least- Following the idea underlying Krylov techniques and, in
squares sense) that describes the evolution of the flow over particular, the Arnoldi method (Greenbaum 1997;
a small time interval. The eigenvalues and eigenvectors of Trefethen and Bau 1997), we express and approximate
this map then capture the principal dynamics contained in the linear map A using an (N - 1)-dimensional snapshot
the snapshot basis. basis VN-1
1 . This step can be expressed as
The representation of a nonlinear process by a linear
sample-to-sample map is closely linked to the concept of a fv2 ; v2 ; . . .; vN g ¼ Afv1 ; v2 ; . . .; vN1 g
Koopman operator, an analysis tool for dynamical systems. fv1 ; v2 ; . . .; vN1 gS ð3Þ
This type of spectral analysis of nonlinear processes pro- or
vides the mathematical foundation of the dynamic mode
decomposition and has recently been applied to complex VN2 ¼ AVN1 VN1 S ð4Þ
1 1
fluid flows (Rowley et al. 2009).
Snapshot-based analysis of fluid flows has previously with S as a companion matrix that simply shifts the
been accomplished by the proper orthogonal decomposi- snapshots 1 through N - 1 (via the subdiagonal matrix
tion (POD) (Lumley 1970; Sirovich 1987; Berkooz et al. entries) and approximates the last snapshot N by a linear
1993), which gained widespread popularity by extracting combination of the previous N - 1 snapshots (Ruhe 1984).
coherent structures from experimental or numerical data. It is thus necessary to best express the last measurement by
This statistical technique provides a hierarchy of spatial a linear combination of the previous measurements. This
structures that are mutually orthogonal and cumulatively procedure will result in the low-dimensional system matrix
capture the energy content of the flow. However, the S. It is then known that the eigenvalues of S, also referred
extraction of the temporal dynamics is not immediate, and to as the Ritz values, approximate some of the eigenvalues
the interpretation of each structure’s contribution to the of the full system matrix A. The associated eigenvectors of
overall dynamic process is not trivial. Nevertheless, POD S provide the coefficients of the linear combination that is
structures are widely used in model reduction and have necessary to express the modal structure within the
shown significant success in feedback control applications snapshot basis. We obtain
(Noack et al. 2003).
min
vN VN1 1 s
ð5Þ
s
123
Exp Fluids (2011) 50:1123–1130 1125
123
1126 Exp Fluids (2011) 50:1123–1130
expressing the last column v80 by a linear combination of the second mode (DM2) is characterized by large structures
the previous 79 snapshot vectors according to (5). This spanning the mean shear layer; its imaginary part (not
matrix then represents an optimal linear snapshot-to-snap- shown) is shifted by 90°. Both real and imaginary part,
shot map that can further be analyzed. The complex when combined with the temporal behavior of the mode
eigenvalues of S describe the temporal dynamics contained exp(x t) with x & 5.97i, represent a large-scale structure
in the data sequence. In accordance with standard con- advected in the downstream direction. This structure has
vention, the eigenvalues of S are logarithmically mapped been identified as the most relevant within the processed
onto the complex plane, i.e., x ¼ logðeigðSÞÞ=Dt with data sequence. Subsequent dynamic modes, DM4,6,8,10
Dt ¼ 0:053; such that their real part xr represents expo- show increasingly smaller spatial features (reminiscent of
nential growth or decay (depending on the sign) and their higher harmonics), but also increasingly higher frequencies
imaginary part xi contains the temporal frequency (see x4,6,8,10 &i{11.94, 17.91, 23.88, 29.84}.
Fig. 2a). Since real data have been processed, complex In order to further demonstrate the decomposition into
eigenvalues appear as conjugate pairs resulting in a sym- dynamic modes, a superposition of the most relevant
metric spectrum with respect to the axis xi = 0. modes is attempted. This is accomplished by computing the
The spectrum (see Fig. 2a) displays an eigenvalue near amplitude of each mode contained in the original data
the origin, which signifies a corresponding structure that is sequence. The temporal evolution of, say, the k-th mode is
steady in time. The associated dynamic mode (DM1) shows then given by the dynamic mode DMk, properly scaled by
a structure reminiscent of the mean flow. The real part of its amplitude and multiplied by its temporal dependence
30
20
10
ω
i
0
−10
−20
−30
ωr
−40
−0.04 −0.02 0 0.02
Fig. 2 Results from a dynamic mode decomposition of a data sequence (with N = 80) illustrated in Fig. 1. a Dynamic mode spectrum,
logarithmically mapped to the complex plane. b Selected dynamic modes, visualized by the real part of the passive tracer
123
Exp Fluids (2011) 50:1123–1130 1127
Fig. 3 Animation of a
superposition of dynamic
modes. (top) The three most
dominant dynamic modes have
been used; (bottom) eleven
modes, based on all modes
shown in Fig. 2, have been
used. The time instances
correspond to the ones chosen
for the original sequence,
displayed in Fig. 1
exp(xk t). In this manner, each structure can be animated structures—or for model reduction—by including only the
individually resulting in a hierarchy of dynamic processes. most relevant structures of interest. The above demonstra-
Alternatively, a superposition of structures gradually reconsti- tion has been intended for instructive purposes; the pro-
tutes the full dynamic of the original data sequence. Figure 3 cessing of experimental data is more challenging due to the
shows the result of two animations of such superpositions, with presence of noise and uncertainties. The decomposition of an
three and eleven dynamic modes, respectively. The time image-sequence taken from experiments in a water channel
instances correspond to the ones depicted in Fig. 1 for the will be dealt with in the next section.
original data sequence. In the reduced model based on only
three dynamic modes (mean mode DM1, DM2 and its complex
conjugate DM3) the overall large-scale structures are, as 4 Application to an axisymmetric water jet
expected, captured both in space and in time; small-scale details
and sharper features, however, are still missing with this severe The dynamic mode decomposition will be applied to an
a truncation. The superposition based on eleven dynamic image-based flow visualization of a laminar axisymmetric
modes (all modes shown in Fig. 2: mean mode DM1 and jet. In this configuration, a low Reynolds number axisym-
DM2,4,6,8,10 with their complex conjugates) recovers the origi- metric water jet enters, without swirl, into a quiescent pool
nal process to an astonishing degree. of water (Fig. 4). An axisymmetric vortex sheet develops,
This demonstration shows that the dynamic mode which subsequently rolls up according to a Kelvin–Helm-
decomposition (DMD) is effective in detecting and extract- holtz-type instability (Gallaire and Chomaz 2003). Due to
ing the relevant dynamic features of the flow. No assumption the slow speed of the flow, no three-dimensional effects
of a energy-based ranking nor the use of second-order sta- have been observed, and the flow is consequently taken as
tistics, as required for a proper orthogonal decomposition, axisymmetric throughout the window of observation.
has been necessary. The DMD simultaneously provides Snapshots from a sequence of images are then extracted
temporal information (expressed in the spectrum) and spatial by a video camera at a rate of 25 Hertz. These snapshots
details (contained in the modes) of the dynamic process. A are subsequently processed into a pixelized colormap rep-
superposition of the modes, weighted by their amplitude and resenting the evolution of a passive dye tracer. A localized
temporal behavior, gives a low-dimensional representation 3 px 9 3 px Laplacian filter is applied to the raw images to
of the full dynamics that can equally be used for filtering—by eliminate noise, mostly in the freestream. The resulting
neglecting modes with low amplitudes or noisy spatial images used in this experiment are shown in Fig. 4. They
123
1128 Exp Fluids (2011) 50:1123–1130
show the typical roll-up of the outer jet shear layer due to axial direction is observable, in addition to a more pro-
Kelvin–Helmholtz instabilities. Even in this short sequence nounced slanting of the coherent structures. The axial
of images, the combined advective and diffusive nature of extent of this mode covers nearly the entire observation
the instability is clearly visible, and the scale and speed of window and shows structure even close to the nozzle exit.
the underlying instability can be estimated by inspection Again, a typical staggering of the real and imaginary
alone. components of this mode appears. The eigenvalue corre-
Nevertheless, the images by themselves do not represent sponding to this dynamic mode has a larger temporal fre-
an objective and quantitative means to gain insight into the quency; this tendency supports the fact that the dynamic
prevalent perturbation dynamics. Moreover, only the most mode decomposition has successfully extracted a disper-
dominant features can be observed, whereas more subtle sion relation of the flow field from the furnished data
and smaller-scale instabilities may be missed entirely. For sequence.
this reason, we apply the dynamic mode decomposition The next higher mode, shown in Fig. 5 4r,4i, has even
(DMD) outlined above. Each image is converted into a smaller coherent features, again concentrated in regions of
column-vector, and a rectangular matrix of these vectors maximum mean shear. Significant amounts of structure can
(from 50 images taken) is formed, which represents the be observed upstream near the nozzle. The spatial structure
matrix VN1 with N = 50 from the derivation above. The reveals scales of small to moderate size further downstream
system matrix S is subsequently extracted, and its eigen- of the nozzle, which substantiates the observation that
values, again logarithmically mapped, give a quantitative small-scale features appear further downstream in the axial
picture of the underlying temporal disturbance dynamics. direction. The corresponding eigenvalue shows an expec-
The spatial features corresponding to the identified (and ted larger temporal frequency.
numbered) eigenvalues are shown in Fig. 5. The fifth and final dynamic mode is displayed in Fig. 5
As before, an eigenvalue appears at the origin. This 5r,5i. It shows a higher frequency and a nearly neutral
eigenvalue (which indicates neither exponential growth/ stability behavior. It is thus expected to be readily obser-
decay nor oscillatory behavior) accounts for a steady state vable over the processed time interval.
contained in the data sequence. As such, it represents the The moderate fluid velocities in this jet experiment
mean flow or the time-invariant component of the flow. make the decomposition into dynamic modes via DMD and
This steady component is shown in Fig. 5 (top right). A the interpretation of the results a relatively straightforward
characteristic spreading of the shear layer is observed in the exercise. For high-speed flows, this may become increas-
downstream (top-to-bottom) direction, but no spatial length ingly difficult due to the presence of very small-scale
scale in the axial jet direction can be detected. incoherent fluid elements. Nevertheless, the DMD tech-
More interesting flow features are contained in the nique is able to optimally identify and extract the dynam-
higher modes. These DMD modes now represent the per- ically most relevant structures. The relevance of structures
turbation (rather than the mean flow) dynamics. The second among the many identified modes can be further quantified
DMD mode (Fig. 5 2r,2i) shows a characteristic pattern by computing the amplitude of each structure contained in
located at the shear layer, which represents the roll-up of the original data sequence. In this manner, it is possible to
the axisymmetric vortex sheet and the resulting entrain- discern dynamically irrelevant structures and focus on the
ment process of quiescent ambient fluid into the jet. The more coherent fluid elements and their associated temporal
real and imaginary part of the modal patterns are staggered dynamics.
due to a phase difference of 90o between them. A clear It should be mentioned that in our case, the extracted
spatial scale in the downstream direction is visible, which spectrum is symmetric with respect to the temporal fre-
readily translates into a corresponding (local) wavenumber quency. This is a consequence of processing real-valued
in the axial coordinate. data that yields eigenvalues that are either purely real (non-
The third mode, depicted in Fig. 5 3r,3i, shows even oscillatory) or come in complex conjugate pairs. The
more sophisticated patterns. A slightly smaller scale in the dynamic mode decomposition, however, can also be
123
Exp Fluids (2011) 50:1123–1130 1129
sequence of an axisymmetric
water jet. (top left) DMD 30
mean flow
spectrum, logarithmically 5
mapped to the complex plane. 20
(top right) Dynamic mode DM1, 4
representing the mean flow. 3
10
Dynamic modes (2 through 5)
2
corresponding to the labeled 1
eigenvalues; middle row real 0
ωi
part, bottom row imaginary part
−10
−20
−30
−40
−4 −2 0
ωr
(2r) ( 3r) ( 4r) ( 5r)
applied to snapshots that have been preprocessed by, e.g., a 5 Summary and conclusions
Fourier transform in a homogeneous coordinate direction
and are thus complex valued. In this case, the DMD A novel decomposition technique, referred to as dynamic
spectrum will be non-symmetric with respect to the real mode decomposition (DMD), has been introduced and
axis. demonstrated on image-based flow visualizations. It relies
123
1130 Exp Fluids (2011) 50:1123–1130
123