fluids

Article
Analysis of Transition for a Flow in a Channel via
Reduced Basis Methods
Gaetano Pascarella *,† , Ioannis Kokkinakis and Marco Fossati
Aerospace Centre of Excellence, Department of Mechanical and Aerospace Engineering,
University of Strathclyde, 16 Richmond St, Glasgow G1 1XQ, UK; [email protected] (I.K.);
[email protected] (M.F.)
* Correspondence: [email protected]
† Current address: 75 Montrose street, Glasgow G1 1XJ, UK.




Received: 12 October 2019; Accepted: 29 November 2019; Published: 5 December 2019


Abstract: The study of the flow mechanisms leading to transition in a planar channel flow is
investigated by means of a reduced basis method known as Dynamic Mode Decomposition (DMD).
The problem of identification of the most relevant DMD modes is addressed in terms of the ability
to (i) provide a fairly accurate reconstruction of the flow field, and (ii) match the most relevant flow
structures at the beginning of the transition region. A comparative study between a natural method of
selection based on the energetic content of the modes and a new one based on the temporal dynamics
of the modes is here presented.

Keywords: reduced basis methods; flow feature extraction; transition; channel flow

1. Introduction
The onset of turbulence in fluids is to-date an unresolved puzzle, remaining at the focus of research
in the past century partly due to its significance in engineering applications. Mathematically, a complete
description of the fundamental physical problem has not yet been developed, while from a
philosophical stand-point, turbulence remains a problem of determinism and chaos, a transition
from laminar to turbulent flows.
The transitional two-dimensional structure that typically forms in a Blasius (laminar) boundary
layer, or any other similar flow, was first observed by Ludwig Prandtl and further studied
by Tollmien [1] and Schlichting [2] (TS). The TS instability is the initial part of the process of transition to
turbulence in viscous boundary layers for which the essential mathematics was developed in the 1950
by Schlichting [3] and inner mechanisms uncovered more recently by Baines et al. [4]. Kachanov [5]
noted that the TS instability formation process represents the receptivity and response of a boundary
layer to either external or internal perturbations. As a consequence, low-amplitude, two-dimensional
and unsteady TS waves are formed that grow exponentially and, when their rotational speed reaches
in magnitude about 1% or 2% of the free-stream velocity, breakdown causing the flow to become
heavily three dimensional and eventually reach a fully developed turbulent state further downstream.
The term fully developed turbulent regime or fully turbulent will be used here to indicate a regime
when turbulence quantities no longer vary in the streamwise direction.
It is well established that the process of transition begins from very low levels of background noise.
The mean flow, above a critical Reynolds number, becomes inherently unstable to small-amplitude
disturbances, and the development to turbulence ensues. The transition process can be accurately
described by either linear stability theory such as the Orr-Sommerfeld equation developed at the
beginning of the 20th century or Floquet theory. Use of the Orr-Sommerfeld equation provides for
the linear two-dimensional modes of disturbance to a viscous (NSE) parallel flow. The TS instability

Fluids 2019, 4, 202; doi:10.3390/fluids4040202 www.mdpi.com/journal/fluids

Fluids 2019, 4, 202 2 of 19

tends to become unstable under the conditions given by the Orr-Sommerfeld equation, initiating the
transition process and eventually leading to turbulence.
It has previously been shown in various publications that temporal transition in channel flows
can be simulated quite accurately by numerical means. Using direct numerical simulation (DNS),
Sandham and Kleiser [6] and Härtel and Kleiser [7] obtained a very good prediction of the transitional
flow structures present, such as the formation of Λ-vortices, roll-up of shear layers and appearance
of hairpin vortices, while Baines et al. [4] probed the mechanisms behind the Tollmien-Schlichting
instability. Further, large eddy simulation (LES) studies conducted on grids much coarser than those
required to conduct equivalent fully resolved DNS also exhibited similar transitional phenomena.
See Germano et al. [8], Schlatter et al. [9], Schlatter et al. [10] and references therein for further details.
More recent studies by Wu and Moin [11] and Sayadi et al. [12], triggered transition of a
weakly-compressible Blasius layer at a Mach number of 0.2, by two different instability mechanisms of
TS-type transition into developed turbulence. Most importantly, it was demonstrated how boundary
layers develop into a statistically self-similar and universal state independent of the transition
mechanism.
More relevant to the case of bypass transition examined herein, Schlatter et al. [13] clarified the
receptivity of the boundary-layer streaks behind the turbulent breakdown of a boundary layer subject
to free-stream turbulence. The rise of turbulent spots in the boundary layer were attributed to a streak
secondary-instability process. Initially forming as a weak wave packet located in the low-speed streak,
it begins to grow in strength while dispersing in the streamwise direction. During the breakdown to
turbulence, quasi-streamwise vortex structures were identified on the flanks of the low-speed region
and arranged in a staggered pattern.
A criteria for bypass transition is defined by Wu et al. [14] as the superposition of a laminar Blasius
boundary layer with a freestream of an initial turbulence intensity level, Tu, of approximately 1% to
4%. Their analysis elucidated how bypass transition proceeds through a sequence of characteristic
structure formations: (i) formation of Lambda vortices, (ii) followed by hairpin packets, (iii) leading
to infant turbulent spots, and finally (iv) hairpin forests. This sequence of events closely resembles
that reported by Sayadi et al. [15] with regards to the secondary instability and breakdown process
encountered in boundary layer natural transition. Wu et al. [14] go on to reason that Λ-vortices owe
their formation to three-dimensional velocity perturbations interacting with the near-wall spanwise
vorticity, and are thus not a result of oblique-wave excited long streaks. Though the latter can still be
found, it is argued that they may not play as a dynamically important role in the bypass transition
sequence as the infant turbulent spots which form earlier upstream.
Reduced basis Methods are a promising tool to analyse very complex unsteady phenomena in
fluid dynamics [16], such as transition to turbulence . They are able to extract few flow features,
which might be only a mathematical abstraction that allows to express the final solution as a simple
linear combination of few basis functions, but they can also have a physical meaning, as they could
represent coherent structures which are responsible for fundamental behaviours [17] (e.g., structures
which cause instabilities or are associated to fundamental time frequencies present in the flow field).
The present work aims at exploiting this last aspect of reduced basis methods in order to identify
the most important flow features, alias modes, which are responsible for bypass type transition in a
developing zero-pressure gradient boundary layer flow. In particular, Dynamic Mode Decomposition
(DMD), introduced by Schmid [18], is used as feature extraction method, since it is able to characterize
the extracted flow features in terms of pure frequencies and growth/decay rates. This property makes
DMD the first competitor of the Proper Orthogonal Decomposition (POD) [19], which is able to extract
a set of orthogonal basis functions that can be ordered in terms of their energetic importance, but do not
carry any meaningful information about frequencies and can be misleading in identifying fundamental
dynamic properties [20].
DMD has been widely used in literature for fluid analysis [21–24]. A crucial discussion in recent
literature has been how to properly select the most important dynamic modes from the entire set
Fluids 2019, 4, 202 3 of 19

available, in order to have a more physical understanding of the phenomenon. The selection of DMD
modes is indeed not straightforward as with other reduced basis techniques, since referring only on
modes amplitudes would not take into account frequencies and growth/decay rates of the modes,
which also can play an important role in the selection process. Different DMD variants have been
introduced which try to overcome this limit, either trying to select the most important modes a-priori,
such as optimized-DMD [25] and sparsity promoting DMD [26], or introducing a ranking once all the
DMD modes are extracted [27,28].
The present work aims at identifying a strategy of modes selection that will respond to two
important questions: (1) can we identify DMD modes that are associated to physical flow structures
that characterize the onset of transition and (2) can we identify those few DMD modes that better
allow reconstructing the flow field in the region leading to the transition phase? The assessment
and evaluation of DMD modes with respect to the two above criteria will be carried out by looking
at the similarity between the reconstructed flow field with a few selected modes and the actual
flow field in the time window preceding the actual transition zone, where turbulent structures will
already show a chaotic behaviour. In this study, a first attempt is made to identify the most important
modes responsible for bypass type transition in a developing zero-pressure gradient boundary layer
flow using DMD equipped with a new time local selection of the most dominant modes. Since a
transient phenomenon is considered, it is worth noticing here that another crucial aspect of DMD
recently discussed in literature is how much it is capable to provide consistent results when transient
phenomena are considered [29]. Specifically Page and Kerswell [30] have shown, for a simple Couette
flow, how DMD can fail when trying to describe the dynamics of a fluid system moving between two
equilibria along an heteroclinic orbit. Nevertheless, in the same work it was also shown how the DMD
looks accurate if applied to very short time windows not containing cross-over points of the dynamical
system. This is why the DMD extraction method applied in the present work is focusing on a small
window at the early stage of transition. The whole transition process is indeed an heteroclinic orbit
going from the unstable equilibrium of the laminar flow to the stable limit cycle of the fully-turbulent
flow. Therefore, applying DMD to the entire time window is very likely to give results which are
not consistent. The paper is organized as follows. Section 2 presents the DNS-like analysis of the
transition on a planar channel flow, Section 3 reports the basics of Dynamic Mode Decomposition
and its characteristics in terms of modes identification, and finally Section 4 reports the comparative
analysis and assessment of the proposed methods for modes selection in the characterization of the
onset of transition.

2. Transition in a Planar Channel Flow

The numerical simulation of transition in a plane channel flow follows on from previous numerical
results obtained by Kokkinakis and Drikakis [31], where emphasis was placed solely on the final,
statistically steady-state, fully turbulent regime. The purpose of this study on the contrary is to shed
light into the transition process, starting from the seemingly laminar initial condition up to the fully
developed turbulent state, and identify the dominant mechanisms regulating it.
The 9th-order WENO scheme and 1283 mesh resolution are employed, for which the accuracy of
the results were previously shown to be DNS–like [31]. The Reynolds number based on the friction
velocity (uτ ) and the channel half-height (Lh ) is equal to Reτ ≈ 395, whereas that based on the bulk
velocity (ub ) is Reb ≈ 6887. The y+ value of the first point from the wall is equal to 0.74, resulting in a
resolved mean friction Reynolds number of just over 393, within less than 0.5% of the target value.
The employed grid and numerical scheme were previously shown to be capable of accurately resolving
the turbulence properties of the test-case and set-up considered in this study [31].
In order to obtain a sufficient number of flow snapshots of the bypass transition, the initial velocity
field is perturbed with white noise in all directions by 5% of the local streamwise velocity (instead of
the 10% used in Ref. [31]), which in turn is obtained by a laminar Poiseuille parabolic profile. For an
Fluids 2019, 4, 202 4 of 19

initial random perturbation of 1%, very weak streamwise streaks were still observed but remained
inherently smooth for as long as it was computationally affordable/reasonable.
Furthermore, it was observed that a simulation’s numerical dissipation properties significantly
influenced the duration of the transition process, though the underlining physical mechanism remained
similar in all cases. Low-order schemes, such as the 2nd-order MC limiter, delayed the formation and
appearance of the streamwise streaks while substantially decreasing their growth rate. The increased
numerical dissipation of lower-order schemes severely dampens the initial perturbations, thus delaying
the onset of the streamwise streaks. On the contrary, the higher-order schemes better resolve the
initial random velocity perturbations and retain a much higher turbulent intensity. In either case,
the disturbances resolved near the midstream are carried towards the near-wall region as momentum
is transferred between the two regions during the evolution of the laminar boundary layer (BL),
leading eventually to the formation of the streamwise streaks. Figure 1 reports snapshots of the
DNS-like solution.

Figure 1. Q-criterion isosurfaces coloured by the velocity magnitude at four instants of time and
depicting the bypass transition process. From top-left to bottom right: t = 1450 Lh u− 1 −4
b (Q = 10 ),
laminar; t = 1480 Lh u− 1 −4 −1 −4 −1
b (Q = 10 ) and t = 1530 L h ub (Q = 5 × 10 ), transition; t = 1700 L h ub
(Q = 0.05), fully developed.

2.1. Streamwise Streaks

Helicity is a scalar quantity defined as the inner product between the velocity and vorticity vectors.
By definition, helicity vanishes in two-dimension. The physical meaning of helicity becomes a bit more
clearer when it is normalized, defined as:

u · Ω/ (|u| |Ω|) (1)

where u is the velocity vector, Ω is the vorticity vector and | · | represents their magnitude. The value of
normalized helicity ranges from –1 to 1. As normalized helicity physically represents the value of the
cosine angle between the velocity vector and vorticity vector, the extreme values of −1 and 1 represent
regions of a flow that is highly three-dimensional, whereas the centre value of zero corresponds to
a two-dimensional flowfield where the vorticity is normal to velocity, as shown in Figure 2 where
vorticity is streamwise.
Fluids 2019, 4, 202 5 of 19

Normalized helicity (Degani et al. [32]) is a useful indicator of how the velocity vector field is
oriented with respect to the vorticity vector-field for a given flowfield. For instance, at the centre of
streamwise vortices and streaks, such as TS waves or streamwise streaks, the velocity and vorticity
vector tend to align themselves parallel to each other and as a result the normalized helicity will attain
it’s maximum absolute value of one. This fact is utilized to locate the core of unidirectional vortices
such as Görtler or TS waves, as well as streamwise streaks. Figure 2 depicts the normalized helicity
prior to transition of the flowfield resolved in the Implicit Large Eddy Simulation (ILES) results.

Figure 2. Normalized helicity iso-surfaces depicting streamwise streaks as precursor to transition.

One possible transition process is widely known as bypass transition. Under a sufficiently
high level of free-stream turbulence intensity, generally >1%, streamwise elongated disturbances are
induced in the near-wall zone of an attached laminar BL, termed streamwise streaks or Klebanoff
distortions, such as those visualized in the top-left image of Figure 1. Under sufficient background
noise, the streaks can form in a laminar BL, and are characterized by high and low velocity that
alternate in the spanwise direction (wobble) at some distinguished periodicity and with a wavelength
in the order of the BL thickness O(δ). For the channel flow at a Reτ = 395 considered here, the streaks
are positioned at y/δ > 0.35.
Landahl [33] presumed the presence of a lift-up effect to explain the presence of streaks in
transitional flows. It is argued that coherent streamwise streaks can efficiently extract energy from the
mean flow via the lift-up effect, gradually strengthening and growing. This mechanism is reported
to occur for spanwise scales ranging from those of the near-wall streaks (∼100y+ ) to those of the
large-scale motions. The lift-up effect alone however does not determine the streak patters that
emerge during transition. Using numerical results, Chernyshenko and Baig [34] showed that the
combine action of the lift-up of the mean profile, mean shear, and viscous diffusion own different
streak pattern-forming properties and bear much greater influence than the pattern of the wall-normal
motions. For further details concerning the lift-up effect, a recent and thorough review is given by
Brandt [35].
Typically, in a nominally zero-pressure gradient boundary layer subject to high levels of
free stream turbulence, the mechanism of Tollmien–Schlichting waves transition is bypassed.
Nonetheless, during the transition window, it is evident that a mean streamwise momentum transfer
from the outer- towards the inner-boundary layer region still occurs. This is attributed to the streamwise
streaks contained within y/δ > 0.35 of the boundary layer.
The wavelength of the initial streak disturbance is largely determined by low-frequency
disturbances; high-frequency disturbances are mostly damped by the laminar shear layer. Once the
laminar BL is distorted by the streaks, it becomes susceptible to instabilities. Although the streaks
are of a large wavelength, the instabilities that form are of a short wavelength (top-right image in
Fluids 2019, 4, 202 6 of 19

Figure 1), and suggests that they are excited/triggered by high-frequency perturbations. Two of the
most noteworthy instability modes are the (i) sinuous mode, and (ii) varicose mode.
The Klebanoff distortions grow downstream both in length and amplitude and finally cause
breakdown–of the seemingly laminar BL to this point–with the formation of turbulent spots.
These turbulent spots coalesce farther downstream, finally leading to the onset of turbulence,
as indicated by the left and right bottom images in Figure 1.
Note that besides the magnitude of the initial condition velocity perturbation, the streamwise
streaks peaks and growth-rate are also sensitive to the shape of the imposed energy spectrum E(k )
and its integral length scale; for the same initial velocity perturbation, or turbulence intensity Tu,
but different different E(k ), a different transition process may be obtained.

2.2. Flowfield Structure

Figure 3d reports the streamwise velocity in the y-z plan at different instants of time:
t1 = t0 + 40 h/ub , t2 = t1 + 19 h/ub , t3 = t2 + 16 h/ub , t4 = t3 + 16 h/ub and t5 = t4 + 9 h/ub ,
while wall-normal velocity is presented in Figure 4. Two different 2-dimensional phenomena can
be observed to have formed at the exact time the streamwise streaks begin to breakdown. One is a
Rayleigh–Taylor instability (RTI) visible near the upper wall, whereas a Kelvin–Helmholtz instability
(KHI) like vortex can be observed at the lower half of the channel. These typically form between two
streamwise streaks that have become sufficiently strong to interact and alter one another. Asymmetries
in the structure between the two streamwise streaks begin to amplify any velocity perturbations, and
breakdown of the structures becomes iwhichmminent. This is also indicated by the bottom-left image
in Figure 1, along the streamwise direction.
Since the streamwise streaks are located close to each other and are located at a similar distance
from the wall, the spanwise velocity perturbations are the first to be amplified—not being restricted by
the presence of the wall or the proximity of the viscous sublayer. Simultaneously, the helicity structures
appear to move in a side-to-side snake-like fashion in the spanwise direction, which is unrestricted in
comparison to the wall-normal direction. Eventually, once the viscous shear stresses are overcome,
the spanwise collision between two streaks causes momentum to be ejected in a manner resembling
that of the mushroom shape typically encountered in RMI. It can therefore be deduced that prior to the
RT-instability formation, streamwise streaks breakdown due to asymmetries in their structure arising
from the superimposed random perturbations in the initial condition.
Once a streamwise streak disperses, the rest of the domain undergoes bypass transition to
turbulence. It should be noted that the process of the formation of the streamwise streaks occurs
while there is a gradual transfer of momentum from the upper midstream region of the boundary
layer, towards the lower near wall region as can be seen in Figure 5a, which has also been reported
in the DNS of Sandham and Kleiser [6], Härtel and Kleiser [7]. The streamwise streaks form while
the laminar boundary layer is developing; the viscous laminar boundary layer in conjunction with
perturbations are responsible for a momentum exchange between the midstream and near wall region
of the boundary layer.
Interest to develop methods capable of suppressing near-wall turbulence in boundary layers
in order to reduce friction drag has recently highlighted the importance of the lateral oscillations
exhibited by streamwise streaks prior to breakdown. In recent DNS studies conducted to develop
and examine such methods, spanwise oscillatory wall motion was introduced in a fully developed
turbulent channel flow (see Touber and Leschziner [36], Lardeau and Leschziner [37] and references
therein). It was found that the unsteady cross-flow straining causes major spanwise distortions to near
wall streak structures, leading to a pronounced reduction in wall-normal momentum exchange in the
viscous sublayer, successfully disrupting the turbulence contribution to the wall shear stress.
Fluids 2019, 4, 202 7 of 19

(a) t0 (b) t1 = t0 + 40 h/ub

(c) t2 = t1 + 19 h/ub (d) t3 = t2 + 16 h/ub

(e) t4 = t3 + 16 h/ub (f) t5 = t4 + 9 h/ub

Figure 3. Contours of streamwise velocity in the y-z (streamwise normal) plane.

(a) t0 (b) t1 = t0 + 40 h/ub

(c) t2 = t1 + 19 h/ub (d) t3 = t2 + 16 h/ub

(e) t4 = t3 + 16 h/ub (f) t5 = t4 + 9 h/ub

Figure 4. Contours of spanwise velocity in the x-y (spanwise normal) plane.

Fluids 2019, 4, 202 8 of 19

Closer examination of the data reveals that the point at which the RT-instability forms is an
inflection point in the streamwise velocity profile (see Figure 5d), which is verified by transition
theory. This creates two distinct regions in the laminar boundary layer that resemble a shear layer.
Hence, the turbulent spots that appear in the wall-normal velocity in Figure 4c,e are reminiscent of
transition vortex spots in a turbulent free shear layer. It can therefore be concluded: (i) viscosity causes
the formation of the laminar BL, (ii) within which the streamline streaks form given sufficient external
perturbation, leading to a gradual momentum transfer within the laminar BL, (iii) until the Euler, or
inviscid non-linear advective part of the full NSE, induces the instability breakdown mechanisms once
sufficient momentum is transferred to overcome the viscous shear stresses.

1.6 3.5

3 t=0
1.4
t=1
t=2
2.5 t=3
1.2
t=4
2 t=5
1

dU/dY-1
1.5
U

0.8
t=1 1
0.6 t=2
t=3 0.5
t=4
0.4 t=5
t=6 0

0.2 -0.5

0 -1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Y Y
(a) (b)

2 10
t=0
t=1
t=2
t=3 5
t=4
1.5 t=5
0
U dU/dY

-5
d U/dY

1
2

-10
t=0
t=1
-15 t=2
0.5
t=3
t=4
-20 t=5

0 -25
0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Y Y
(c) (d)

Figure 5. Wall normal profiles of various streamwise velocity functions during transition.
(a) Stream-wise velocity profile; (b) Wall normal derivative; (c) Wall normal derivative times streamwise
velocity; (d) Second order wall normal derivative.

3. Feature Extraction and Dynamic Mode Decomposition

Given the set of snapshots [u1 u2 . . . u Ns ] computed through the numerical simulation and ui ,
the variable to be processed at time ti (snapshot i), the DMD extraction technique is based on the
following assumption:
ui+1 = Tui , (2)

which defines the linear dynamics that best fits the initial set of snapshots. This assumption allows to
express the initial set of collected snapshots according to the linear dynamic theory:
Fluids 2019, 4, 202 9 of 19

Nm
u( x, t) ≈ û = ∑ αi φ̄i (x)eωi t , (3)
i =1

where φ̄i are the vectors of DMD modes, ωi are the DMD eigenvalues, αi are the modes amplitudes,
constant in time. For an in-depth description of the DMD algorithm used in the present work to extract
DMD modes the reader can refer to [38]. The main steps are briefly presented below. An eigenvalue
decomposition of the matrix T needs to be performed in order to extract DMD modes. Since for fluid
dynamics problems this matrix might be huge, its dimension needs to be reduced before performing
such a step. Therefore the following similarity transformation is applied to the matrix T:

T̃ = Φ∗ TΦ, (4)

where Φ represents the matrix of spatial functions needed to project the low-rank dynamics back to
the high-dimensional space and is computed from the singular value decomposition of the matrix of
snapshots U = [u1 u2 . . . u Ns −1 ]:
U = ΦΣA∗ . (5)

The last snapshot is not included in the matrix U in order to build the time shifted matrix
0
U = [u2 u3 . . . u Ns ] which allows to express Equation (2) in matrix form:
0
U = TU. (6)

The combination of Equations (4)-(6) provides an expression for the reduced matrix T̃:
0
T̃ = Φ∗ U AΣ−1 , (7)

which is usually many order of magnitude smaller than the matrix T. The singular value decomposition
of matrix U might allow to perform a further reduction considering only the first r columns of Φ,
if a singular value threshold can be computed and all the singular values below this threshold are
discarded [39]. Nevertheless, in the present work, no reduction is introduced at this stage and all the
singular vectors are considered, in order to prevent any possibility of excluding spatial functions which
have very low energetic content but still have an important contribution for the dynamic. The reduced
matrix T̃ is then used to compute the eigenvalue decomposition and the spatial DMD modes can
be recovered projecting back to the high-dimensional space the T̃ eigenvectors, using the following
transformation:
0
Φ = U AΣ−1 W, (8)

where W are the eigenvectors of the matrix T̃, also called DMD eigenvectors. The modes computed
with this relation are also known as exact DMD modes [38].
The eigenvalues λi of the matrix T̃ provide the growth/decay rate (real part) and the frequency
(imaginary part) of each dynamic mode, and therefore define the time dynamics.
They are related to the eigenvalues ωi in the Equation (3) as follows:

log(λi )
ωi = , (9)
∆t
where ∆t represents the physical time interval used to sample the snapshots collected in the matrix U.
Fluids 2019, 4, 202 10 of 19

The coefficients αi in the Equation (3) are computed solving the following optimization
problem [26]:
0
minr kU − ΦDα V k2 , (10)
α i ∈R

where Dα is the diagonal matrix of DMD coefficients αi , r is the number of DMD modes extracted, and
V is the Vandermonde matrix which contains the DMD eigenvalues and is defined as follows:

Ns −1
λ21,DMD
 
1 λ1,DMD ... λ1,DMD
Ns −1
1 λ22,DMD ...
 
λ2,DMD λ2,DMD 
V=
 .. .. .. .. ..  (11)
.

. . . . 
s −1
1 λ Ns −1,DMD λ2Ns −1,DMD ... λN
Ns −1,DMD

4. Onset of Transition and Dynamic Mode Decomposition (DMD) Characterization

The DMD method is applied to 1000 snapshots of the numerical simulation sampled in the time
interval highlighted in Figure 6, using an equi-spaced ∆t = 10−3 Lh u− 1
b . This is the maximum value
admissible to have time-resolved data, i.e., the maximum ∆t which still allows to follow the evolution
in time of flow structures. Each snapshot consists of all the three components of the velocity arranged
in a unique column vector, u = [u, v, w] T . The time interval selected considers the early stage of
transition, when the vortical laminar tubes that appear in the left half of the channel start showing
spatial oscillations. These oscillations are then amplified and eventually determine the interaction and
breakdown of the laminar tubes, which rearrange in more chaotic yet coherent turbulent structures.
The reader can refer to the first three time instants reported in Figure 1 to have a direct visualization
of the described process. Figure 7 reports the information coming from the DMD in terms of the
eigenvalues λi and the time dynamics αi eωi t of the dynamic modes extracted. Since the DMD modes
are normalized, the time dynamics is a direct indication of the importance of each mode with respect
to the others at all time instants.

(a) (b)

Figure 6. Shear stress evolution for the time interval used in Dynamic Mode Decomposition (DMD) (a).
Full time window from transition to a fully developed turbulent flow (b).
Fluids 2019, 4, 202 11 of 19

(a) DMD eigenvalues on the unit circle (b) time dynamics of DMD modes

Figure 7. DMD eigenvalues and time dynamics.

4.1. Modes Selection Approaches: Temporal Envelope vs. DMD Energy

The DMD algorithm reported in Section 3 extracts 999 dynamic modes out of the 1000 snapshots
collected in the chosen time window. The selection of the modes relevant to describe the onset of the
transition is not straightforward.
Each DMD mode is associated to a frequency and growth/decay rate (the real and imaginary
part of the DMD eigenvalues respectively) and picking up the dynamic modes only on the basis of
their frequencies might not be a good method of selection when the dominant frequencies of the fluid
dynamic phenomenon under consideration are not known a-priori. A natural way to rank the DMD
modes would be to consider their energetic content [27]. In this way, the most energetic ones will be
selected and this could penalize modes which are important for the dynamic but yet have a very low
energetic content. The present work compares the DMD mode selection based on an energy ranking of
the dynamic modes with a selection based on their time dynamics, which will be hereafter referred to
as t-envelope method.
The energy of each dynamic mode is computed according to the formula proposed by Tissot et al. [27]:

e2σi T − 1
Z T
1
Ei = kφi k2 e2σi t dt = kφi k2 , (12)
T 0 2Tσi

where T is the time window that covers the entire sampling, the σi are the real part of the DMD
eigenvalues ωi and the coefficients αi are considered instead of kφi k2 , since the dynamic modes have
been normalized. With an approach based on an energy ranking of the modes, once the energy has been
computed, only the most relevant modes on the basis of Ei are considered. The proposed t-envelope
method, instead, allows to perform a modes selection which is based on a criterion that is local in time
and does not rely on an integral quantity, such as the one reported in Equation (12). The algorithm is
straightforward and does not imply any computational step, besides the selection of local maxima.
It can be schematized with the following few steps

1 The time dynamics of all the DMD modes, namely the functions αi eωi t , are evaluated at the
sampling points and their amplitude is collected in a matrix Tdyn , where on each row there are the
time amplitudes of the corresponding DMD mode;
2 Since the DMD modes φi are normalized, these functions represent the actual contribution of each
mode to the resulting flow field; therefore the maxima for each column of the matrix Tdyn are
computed and the corresponding modes are selected;
3 Once these modes have been selected, the positions of the maxima in the matrix Tdyn are set to
zero, the new maxima are computed and the corresponding modes selected;
Fluids 2019, 4, 202 12 of 19

4 The procedure can be iterated for any levels of selection, depending on how much finer the user
wants to resolve coherent structures in the flow fields and how many information he means
to extract.

The criterion for modes selection at each level can be expressed with a mathematical formula
as follows:
Idm (t) = arg max kαi eωi t k with 0 < t < T. (13)
i =1,...,Nm

Once the Idm (t) is built at each level, all the modes are collected and used as inputs in the
Equation (3) to obtain an approximation of the dynamic of the transition. Naturally, this selection
process could give important information in terms of dominant frequencies and coherent structures,
looking at the selected modes and the corresponding eigenvalues.
Figure 8 reports, on the left, the t-enevelope selection of DMD modes using eight levels of selection
and, on the right, the energy spectrum of DMD modes computed according to Equation (12). On the
energy spectrum the modes selected with the t-envelope method, highlighted with black circles,
are also reported together with the same number of the most energetic ones, highlighted in red. Here it
can be noticed how some of the modes selected with the t-enevlope method are not the most energetic
ones but yet important for the dynamics. The total number of modes used, with a number of levels
fixed to eight for the t-envelope, is 32. Figure 9 shows the DMD spectrum where the red circles indicate
the growth/decay rate and frequencies of the DMD modes selected through the t-envelope and the
energy method. Similarly, Figure 10 shows the αi amplitudes versus the frequencies (imaginary part of
ωi eigenvalues) of the DMD modes selected with the two methods.
Figures 11 and 12 report the DMD modes selected with the t-envelope and the energy methods
respectively and used to compute the approximated flow field with the Equation (3). All the modes are
represented in terms of the Q-criterion computed on the three components of each mode. Since each
complex mode appears in the list of the selected modes with its complex conjugate, only half of the
modes is showed, excluding the one related to the mean, which does not vary over time (ω DMD = 0).
Therefore, being the number of selected modes 32, only the real part of 15 of them is reported. All the
selected modes are sorted according to the energy level defined in Equation (12). Comparing Figures 11
and 12 it can be observed how among the modes selected by the t-envelope method, modes exist that
are closely related to the stream tubes characterizing the flow field at the beginning of the time window
considered, which are not present in the energy selection process.

(a) (b)

Figure 8. DMD modes selection using t-envelope method (a) and energy spectrum of all the extracted
DMD modes (b): the red circles highlight the modes selected using the energy norm, the black circles
the ones selected using the t-envelope method with eight levels of selection.
Fluids 2019, 4, 202 13 of 19

(a) (b)

(c) (d)

Figure 9. DMD spectrum (ω): (a,c) whole spectrum and (b,d) detailed view on the selected modes. Red
circles highlight the DMD modes selected on the DMD spectrum using the t-envelope method (first
row) and the energy method (second row).

(a) (b)

Figure 10. Amplitude of α coefficients. Red circles highlight the DMD modes selected using the
t-envelope method (a) and the energy method (b).
Fluids 2019, 4, 202 14 of 19

Figure 11. Selected modes via the t-envelope approach coloured by φ magnitude. Modes are ordered
from left-to-right from top-to-bottom according to the corresponding energy content.
Fluids 2019, 4, 202 15 of 19

Figure 12. Selected modes via the energy approach coloured by φ magnitude. Modes are ordered from
left-to-right from top-to-bottom according to the corresponding energy content.
Fluids 2019, 4, 202 16 of 19

The difference among the selected modes becomes evident in Figure 13 where a more explicit
comparison of the onset of transition as obtained through the numerical simulation and as it is
reconstructed by the two set of modes is presented. The top row reports the DNS-like solutions at
three distinct instants of time in the time window of Figure 6. The left plot refers to a time instant at
the beginning of the window, the central plot to an intermediate one and the right plot to a time at
the end of the time window, right at the beginning of the actual transition phase are indicated by the
shear stress plot. The middle row is the reconstructed flow field using the t-envelope method and the
bottom row reports the reconstructed fields using the Energy method. In these Figures, the Q-criterion
of the reconstructed velocity field is shown.

Figure 13. Comparison of the onset of transition described through numerical simulation (first row),
DMD t-envelope (second row), DMD energy (third row); t = 1430Lh u− 1 −1
b (first column), t = 1480L h ub
−1
(first column), t = 1530Lh ub (third column).
Fluids 2019, 4, 202 17 of 19

It can be noticed how the t-envelope method is able to provide reconstructed flow fields that are
very closely related to the actual flow field. Indeed, the modes identified by the t-envelope clearly
isolate the instabilities of the laminar streaks as they grow, as time progresses and transition becomes
evident. Also, it is worth noticing how the reconstructed flow field by the t-envelope at the last instant
of time of the time window is visually very similar to the actual flow field. These elements support
the case for the adoption of the t-envelope method for the identification of structures leading to
transition and also for the identification of the few relevant modes for a fairly accurate reconstruction.
This is quite evident when a comparison with the energy-based reconstruction is realized, where the
identification of relevant flow structures and their evolution in time is not quite possible and also the
reconstruction at the last instant of time seems to be less accurate than in the case of the t-envelope.
Moreover, for the first two instants of time reconstructed with the Energy method, a different value for
the Q-criterion isosurfaces is used in order to be able to visualize the streamwise streaks, while for all
the other cases this value is kept constant. A quantitative analysis has been carried out as well to prove
that the t-envelope method is performing better than the energy selection criterion. Figure 14 shows the
error e computed as the L-2 norm over the entire domain of the difference of the two solution vectors:

ku DNS − u DMD k
e= p , (14)
Np

where Np is the number of grid points, u DNS is the solution vector coming from the simulation and
u DMD is the solution vector coming from the DMD reconstruction. The error is reported over time
for both the velocity magnitude and each component of the velocity vector. It can be noticed how the
t-envelope method always outperforms the energy method also in terms of e.

Figure 14. Relative error of the Energy and t-envelope method with respect to simulations for the
velocity vector: velocity magnitude and velocity components.

5. Conclusions
The present work has explored the problem of characterization of the onset of transition to
turbulence in channel flow by means of a reduced basis method called Dynamic Modes Decomposition.
The key contribution of the work consists in the formulation and validation of a method for the
selection of the DMD modes that are capable to identify structures relevant to the onset of the
transition. A natural method for the identification of structures based on the energy content has
proven not sufficient to identify and reconstruct the onset of transition and a new method based on
Fluids 2019, 4, 202 18 of 19

the modes dynamical content has been proposed that seems to provide better insight and consistency
with the mechanism of transition.

Author Contributions: Conceptualization, G.P., I.K. and M.F.; methodology, G.P.; software, G.P., I.K.; validation,
G.P. and I.K; formal analysis, G.P., I.K., M.F.; writing–original draft preparation, G.P., M.F.; writing–review and
editing, G.P., I.K. and M.F.; supervision, M.F.
Funding: This research received no external funding.
Acknowledgments: Results were obtained using the ARCHIE-WeSt High Performance Computer (www.archie-
west.ac.uk) based at the University of Strathclyde.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations
The following abbreviations are used in this manuscript:

POD Proper Orthogonal Decomposition

DMD Dynamic Mode Decomposition

References
1. Tollmien, W. Über die Entstehung der Turbulenz. 1. Mitteilung. Nachr. Ges. Wiss. Gött. Math.-Phys. Klasse
1929, 1929, 21–44.
2. Schlichting, H. Zur Enstehung der Turbulenz bei der Plattenströmung. Nachr. Ges. Wiss. Gött. Math.-Phys.
Klasse 1933, 1933, 181–208.
3. Schlichting, H. Boundary-Layer Theory; McGraw-Hill: New York, NY, USA, 1955.
4. Baines, P.; Majumdar, S.; Mitsudera, H. The mechanics of the Tollmien-Schlichting wave. J. Fluid Mech. 1996,
312, 107–124, doi:10.1017/S0022112096001930. [CrossRef]
5. Kachanov, Y.S. Physical Mechanisms of Laminar-Boundary-Layer Transition. Annu. Rev. Fluid Mech. 1994,
26, 411–482, doi:10.1146/annurev.fl.26.010194.002211. [CrossRef]
6. Sandham, N.D.; Kleiser, L. The late stages of transition to turbulence in channel flow. J. Fluid Mech. 1992,
245, 319–348, doi:10.1017/S002211209200048X. [CrossRef]
7. Härtel, C.; Kleiser, L. Subharmonic transition to turbulence in channel flow. Appl. Sci. Res. 1993, 51, 43–47,
doi:10.1007/BF01082512. [CrossRef]
8. Germano, M.; Piomelli, U.; Moin, P.; Cabot, W.H. A dynamic subgrid-scale eddy viscosity model. Phys.
Fluids A Fluid Dyn. 1991, 3, 1760–1765, doi:10.1063/1.857955. [CrossRef]
9. Schlatter, P.; Stolz, S.; Kleiser, L. LES of transitional flows using the approximate deconvolution model. Int. J.
Heat Fluid Flow 2004, 25, 549–558. doi:10.1016/j.ijheatfluidflow.2004.02.020. [CrossRef]
10. Schlatter, P.; Stolz, S.; Kleiser, L. Large-eddy simulation of spatial transition in plane channel flow. J.
Turbulence 2006, 7, N33, doi:10.1080/14685240600602929. [CrossRef]
11. Wu, X.; Moin, P. Forest of hairpins in a low-Reynolds-number zero-pressure-gradient flat-plate boundary
layer. Phys. Fluids 2009, 21, 091106, doi:10.1063/1.3205471. [CrossRef]
12. Sayadi, T.; Hamman, C.W.; Moin, P. Fundamental and subharmonic transition to turbulence in
zero-pressure-gradient flat-plate boundary layers. Phys. Fluids 2012, 24, 091104, doi:10.1063/1.4744967.
[CrossRef]
13. Schlatter, P.; Brandt, L.; de Lange, H.C.; Henningson, D.S. On streak breakdown in bypass transition. Phys.
Fluids 2008, 20, 101505, doi:10.1063/1.3005836. [CrossRef]
14. Wu, X.; Moin, P.; Hickey, J.P. Boundary layer bypass transition. Phys. Fluids 2014, 26, 091104,
doi:10.1063/1.4893454. [CrossRef]
15. Sayadi, T.; Hamman, C.W.; Moin, P. Direct numerical simulation of complete H-type and K-type transitions
with implications for the dynamics of turbulent boundary layers. J. Fluid Mech. 2013, 724, 480–509,
doi:10.1017/jfm.2013.142. [CrossRef]
16. Rowley, C.W.; Dawson, S.T. Model reduction for flow analysis and control. Annu. Rev. Fluid Mech. 2017,
49, 387–417. [CrossRef]
Fluids 2019, 4, 202 19 of 19

17. Taira, K.; Brunton, S.L.; Dawson, S.T.; Rowley, C.W.; Colonius, T.; McKeon, B.J.; Schmidt, O.T.; Gordeyev, S.;
Theofilis, V.; Ukeiley, L.S. Modal analysis of fluid flows: An overview. Aiaa J. 2017, 4013–4041. [CrossRef]
18. Schmid, P.J. Dynamic Mode Decomposition of numerical and experimental data. J. Fluid Mech. 2010,
656, 5–28. [CrossRef]
19. Berkooz, G.; Holmes, P.; Lumley, J.L. The proper orthogonal decomposition in the analysis of turbulent
flows. Annu. Rev. Fluid Mech. 1993, 25, 539–575. [CrossRef]
20. Rowley, C.W. Model reduction for fluids, using balanced proper orthogonal decomposition. Int. J. Bifurc.
Chaos 2005, 15, 997–1013. [CrossRef]
21. Schmid, P.J.; Li, L.; Juniper, M.; Pust, O. Applications of the dynamic mode decomposition. Theor. Comput.
Fluid Dyn. 2011, 25, 249–259. [CrossRef]
22. Seena, A.; Sung, H.J. Dynamic mode decomposition of turbulent cavity flows for self-sustained oscillations.
Int. J. Heat Fluid Flow 2011, 32, 1098–1110. [CrossRef]
23. Muld, T.W.; Efraimsson, G.; Henningson, D.S. Flow structures around a high-speed train extracted using
proper orthogonal decomposition and dynamic mode decomposition. Comput. Fluids 2012, 57, 87–97.
[CrossRef]
24. Pan, C.; Yu, D.; Wang, J. Dynamical mode decomposition of Gurney flap wake flow. Theor. Appl. Mech. Lett.
2011, 1, 012002. [CrossRef]
25. Chen, K.K.; Tu, J.H.; Rowley, C.W. Variants of dynamic mode decomposition: Boundary condition, Koopman,
and Fourier analyses. J. Nonlinear Sci. 2012, 22, 887–915. [CrossRef]
26. Jovanović, M.R.; Schmid, P.J.; Nichols, J.W. Sparsity-promoting dynamic mode decomposition. Phys. Fluids
2014, 26, 024103. [CrossRef]
27. Tissot, G.; Cordier, L.; Benard, N.; Noack, B.R. Model reduction using Dynamic Mode Decomposition. C. R.
Mécanique 2014, 342, 410–416. [CrossRef]
28. Kou, J.; Zhang, W. An improved criterion to select dominant modes from dynamic mode decomposition.
Eur. J. Mech.-B/Fluids 2017, 62, 109–129. [CrossRef]
29. Bagheri, S. Koopman-mode decomposition of the cylinder wake. J. Fluid Mech. 2013, 726, 596–623. [CrossRef]
30. Page, J.; Kerswell, R.R. Koopman mode expansions between simple invariant solutions. J. Fluid Mech.
2019, 879, 1–27. [CrossRef]
31. Kokkinakis, I.; Drikakis, D. Implicit Large Eddy Simulation of weakly-compressible turbulent channel flow.
Comput. Methods Appl. Mech. Eng. 2015, 287, 229–261, doi:10.1016/j.cma.2015.01.016. [CrossRef]
32. Degani, D.; Seginer, A.; Levy, Y. Graphical visualization of vortical flows by means of helicity. AIAA J.
1990, 28, 1347–1352, doi:10.2514/3.25224. [CrossRef]
33. Landahl, M.T. On sublayer streaks. J. Fluid Mech. 1990, 212, 593–614, doi:10.1017/S0022112090002105.
[CrossRef]
34. Chernyshenko, S.I.; Baig, M.F. The mechanism of streak formation in near-wall turbulence. J. Fluid Mech.
2005, 544, 99–131, doi:10.1017/S0022112005006506. [CrossRef]
35. Brandt, L. The lift-up effect: The linear mechanism behind transition and turbulence in shear flows. Eur. J.
Mech. B/Fluids 2014, 47, 80–96, doi:10.1016/j.euromechflu.2014.03.005. [CrossRef]
36. Touber, E.; Leschziner, M.A. Near-wall streak modification by spanwise oscillatory wall motion and
drag-reduction mechanisms. J. Fluid Mech. 2012, 693, 150–200, doi:10.1017/jfm.2011.507. [CrossRef]
37. Lardeau, S.; Leschziner, M.A. The streamwise drag-reduction response of a boundary layer subjected to a
sudden imposition of transverse oscillatory wall motion. Phys. Fluids 2013, 25, 075109, doi:10.1063/1.4816290.
[CrossRef]
38. Tu, J.H.; Rowley, C.W.; Luchtenburg, D.M.; Brunton, S.L.; Kutz, J.N. On dynamic mode decomposition:
Theory and applications. arXiv 2013, arXiv:1312.0041.
√
39. Gavish, M.; Donoho, D.L. The optimal hard threshold for singular values is 4/ 3. IEEE Trans. Inf. Theory
2014, 60, 5040–5053. [CrossRef]

c 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).