You are on page 1of 22

Experiments in Fluids (2019) 60:4

https://doi.org/10.1007/s00348-018-2648-3

RESEARCH ARTICLE

On the calculation of force from PIV data using the generalized


added‑mass and circulatory force decomposition
Eric Limacher1 · Chris Morton1 · David Wood1

Received: 6 February 2018 / Revised: 25 September 2018 / Accepted: 4 November 2018 / Published online: 26 November 2018
© Springer-Verlag GmbH Germany, part of Springer Nature 2018

Abstract
To understand the forces generated on an accelerating body in a fluid flow, it is useful to have the same framework for theory
and experiment. A candidate for this purpose is the decomposition of fluid-dynamic force into added-mass and circulatory
components. In its generalized form applicable to viscous incompressible flows, this formulation, referred to as the GAMC
formulation, is applied to planar particle image velocimetry data to calculate instantaneous forces in the present study. These
estimates are compared to direct force measurements and to an alternative force formulation from impulse theory, referred
to as the standard impulse formulation (SIF). The chosen test case is a nominally two-dimensional circular cylinder towed
through quiescent water under three acceleration profiles with peak Reynolds numbers between 5100 and 5150. For all three
motion profiles, the measured and filtered drag force is consistently greater than the calculated forces with a bias of 10–20%,
but the trends are in close agreement. Inspection of the presented equations reveals that the GAMC is less sensitive to near-
body vorticity data than the SIF, which has the following consequences. First, forces calculated using the GAMC formula-
tion are less sensitive to random error in the velocity than the SIF. This benefit comes at the cost of increased sensitivity to
errors in cylinder position, but the associated uncertainty is negligible in the present study. Second, the GAMC is much more
tolerant to the omission of near-body vorticity data, which is an attractive feature for PIV investigations. Finally, when no
data are omitted, the SIF is more sensitive to the force components induced by uncharacterized high-frequency vibrations.

1 Introduction Limacher et al. (2018), will be investigated. The GAMC for-


mulation is a decomposition of the fluid-dynamic force into
For many applications in which fluid-dynamic force data added-mass and circulatory force components, which makes
are required, direct force measurement is not practical. As it uniquely relatable to theoretical treatments of unsteady
a result, researchers have considered the use of velocity- forces. This formulation is valid for incompressible flows.
field information, acquired by means of particle image Exact force formulations in the literature are derived from
velocimetry (PIV), to calculate instantaneous forces on conservation of momentum and are equivalent to the follow-
submerged bodies, e.g. Noca et al. (1999), van Oudheusden ing standard momentum control volume analysis, expressed
et al. (2006), Mohebbian and Rival (2012) and DeVoria et al. for an impermeable body and a fixed control volume (Noca
(2014). A review of the various available formulations is 1997) as:
given by Rival and van Oudheusden (2017). In the present
𝜕t ∫V ∮S ∮S
𝜕
work, the experimental utility of the generalized added-mass 𝐅=− 𝜌𝐮dV − n̂ pdS + n̂ ⋅ {− 𝜌𝐮𝐮 + 𝐓}dS,
and circulatory (GAMC) force formulation, as derived in
(1)
where 𝐮 is the velocity in an absolute frame of reference,
Electronic supplementary material The online version of this
article (https​://doi.org/10.1007/s0034​8-018-2648-3) contains
p is the pressure, 𝜌 is the density, n̂ is the outward facing
supplementary material, which is available to authorized users. normal on S, which encloses the fluid volume V around
the body, and T is the viscous stress tensor. Despite their
* Eric Limacher equivalence, different force formulae arise in the literature
ejlimach@ucalgary.ca
from unique approaches to addressing the second integral of
1
Department of Mechanical and Manufacturing Engineering, Eq. (1). Some formulations explicitly solve for the pressure
University of Calgary, 2500 University Drive NW, Calgary, field (e.g. Kurtulus et al. 2007; Albrecht et al. 2013), while
AB T2N 1N4, Canada

13
Vol.:(0123456789)
4 Page 2 of 22 Experiments in Fluids (2019) 60:4

others remove it from the equation through substitution of The first integral on the right-hand side of Eq. (4) is referred
the Navier–Stokes equations and the application of various to as hydrodynamic impulse in Saffman (1992), as vortical
vector identities (e.g. Noca et al. 1999; DeVoria et al. 2014). impulse in Wu et al. (2015), and as shed-vorticity impulse in
These broad categories of approaches differ in their sen- Limacher et al. (2018). Given the prevalence of Eq. (4) in the
sitivity to experimental uncertainty, and it is not yet clear literature, either in this form or very similar, Eq. (4) will be
whether any one method is generally superior in precision referred to as the standard impulse formulation, abbreviated
or accuracy. as SIF. The first term on the right-hand side of Eq. (4) can
In selecting a force formulation, one must consider the be readily understood in terms of the circulation growth and
difficulty in resolving velocity vectors close to the body. convection of vortices in the wake. The second term, though it
This challenge is due to the scattered reflection of laser treats the acceleration of the body, is distinct from the classical
light near the body surface, which tends to saturate the added-mass force.
acquired images and prevent accurate PIV cross-correla- The absence of an added-mass force in the various force
tions. It is possible to express the force entirely in terms of formulations is inconsistent with modern low-order aerody-
surface integrals far from the body, as done in the works of namic models that usually include such a term, e.g. Wang
Mohebbian and Rival (2012) and Noca et al. (1999), to cir- and Eldredge (2013), Xia and Mohseni (2013) and Polet et al.
cumvent this experimental limitation. Such approaches are (2015). This discrepancy motivated the study of Limacher
suitable if one is interested only in bulk force estimates, et al. (2018). The resulting GAMC formulation contains the
but the various terms in these surface-integral formula- same two terms as those in Eq. (4) plus two more: one involv-
tions are difficult to interpret physically. ing the concept of image vorticity, and another dependent on
To aid in physical intuition, Wu et al. (2015) suggest the scalar potential describing the hypothetical irrotational
that vorticity-based formulae be used to permit the inter- flow around the body. The latter, when combined with the
pretation of instantaneous forces in terms of local mecha- last term of Eq. (4), gives rise to an explicit added-mass term,
nisms, i.e. to attribute components of force to the evolution which facilitates comparison between theory and experiment.
of vortical structures around the body. The formulation of The present study includes a comparison between the
Wu (1981), valid for an unbounded, incompressible fluid forces calculated from PIV data using the GAMC formu-
domain, can be employed for this purpose. For a single lation and the SIF, as well as a direct measurement with a
rigid body, his formulation reduces to force transducer. The test case selected for investigation is
[ ] a nominally two-dimensional circular cylinder that linearly
N − 1 ∫V
d 1 2
𝐅=𝜌 − 𝐱a × 𝝎dV + Vb 𝐮c − Vb 𝐱c × 𝜴 , accelerates from rest in a quiescent fluid. The choice of a
dt N−1
circular cylinder lends the advantage of a simple and exact
(2) expression for image vorticity, and the choice to accelerate
where N = {2, 3} is the dimension of the space and Vb is the from rest ensures that the full vorticity field remains within
body volume. 𝐱a is a position vector in an absolute frame of the observable field of view. The GAMC formulation is
reference, and 𝐱c is the location of the body centroid in that briefly summarized in Sect. 2, where it is simplified to a
frame. 𝐮c and 𝜴 are the linear and angular velocities of the form specific to the circular cylinder test case. The experi-
body, evaluated at its centroid. 𝝎 is the vorticity calculated mental setup and approach is presented in further detail in
from the absolute velocity field, i.e. 𝝎 = 𝛁 × 𝐮. The expres- Sect. 3.
sion in brackets on the right-hand side of Eq. (2) is invariant
to changes of origin due to the relationship between “total 2 Overview of the GAMC force formulation
vorticity” and 𝜴 (Wu (1981)):
In the GAMC formulation, the total force acting on a body
∫V (3)
𝝎dV = − 2𝜴Vb , submerged in an infinite, incompressible fluid domain is
expressed as
where the term “total vorticity” refers to the integral on the
left-hand side. As a result, we can select 𝐱c = 0 at every d𝐏
𝐅=−𝜌 , (5)
instant in time without loss of generality. Equation (2) thus dt
reduces to where 𝜌 is the fluid density, t is time and 𝐏 is the total
[ ] impulse, which can be expressed as the sum of four compo-
N − 1 ∫V
d 1
𝐅=𝜌 − 𝐱 × 𝝎dV + Vb 𝐮c , (4) nents Limacher et al. (2018):
dt
𝐏 = 𝐏v + 𝐏i + 𝐏𝜙 + 𝐏b . (6)
where 𝐱 denotes position in a coordinate system that is fixed
to the body centroid but does not rotate with it. The four terms on the right-hand side of Eq. (6) are
referred to, from left to right, as the shed-vorticity impulse,

13
Experiments in Fluids (2019) 60:4 Page 3 of 22 4

image-vorticity impulse, potential impulse and body-volume set of concentrated image vortices. A valid image-vorticity
impulse. These terms are given as distribution in Vb must induce a velocity everywhere on Sb
so as to exactly cancel the surface-normal velocity induced
N − 1 ∫V
1
𝐏v = 𝐱 × 𝝎dV, (7) by the vorticity in V (Saffman 1992). For a two-dimensional
circular cylinder, a point vortex at location 𝐱j gives rise to
an image vortex at the location 𝐱im with equal and opposite

N − 1 ∫Vb
1 circulation (Milne-Thompson 1960):
𝐏i = 𝐱 × 𝝎dV, (8)
( )
D2
𝐱im = 𝐱j . (13)
∮ Sb
n̂ 𝜙� dS, 4|𝐱j |2
𝐏𝜙 = − (9)
As h → 0, the set of discrete point vortices in V approaches
𝐏b = −𝐮c Vb , (10) a continuous distribution of vorticity, and the same is true of
where n̂ is an outward-facing unit normal on the body sur- the image vortex system in Vb. Since the image vortices will
face Sb . 𝜙′ is the scalar potential associated with the hypo- be at least as closely spaced within Vb as the approximated
thetical irrotational flow around the body as viewed in a point vortices in V, the error introduced by discretizing equa-
body-fixed frame of reference, i.e. 𝜙′ satisfies the boundary tion (8) is expected to be of the same order as the errors
conditions 𝛁𝜙� = 𝐔∞ − 𝐮c at infinity (where 𝐔∞ is the far- noted in Eq. (12).
field freestream velocity in an absolute frame of reference) The total vortical impulse, 𝐏v + 𝐏i , is discretized as
and 𝛁𝜙� = 0 on Sb (the impermeable boundary condition). [

( ) ]
In the present work, only two-dimensional problems are con- D2
𝐏v + 𝐏i = 1− 𝐱
4|𝐱j |2 j
× 𝚪j , (14)
sidered such that N = 2. For clarity, let it be repeated that 𝐮c j V
and 𝝎 are evaluated in an absolute frame of reference, while
𝐱 = [x, y]T denotes position relative to the body centroid in where the sum is taken over the domain V only.
a non-rotating coordinate system. The potential impulse and body-volume impulse terms
When applied to PIV data on a uniform square grid in the are calculated easily:
present study, the shed-vorticity impulse integral is evalu- ( 2)
𝜋D
ated using the midpoint method of numerical integration: 𝐏𝜙 = 𝐮c , (15)
2
[ ]
∑ ( 2)
2
𝐏v ≈ h 𝐱j × 𝝎j , (11) 𝜋D
j V 𝐏b = − 𝐮c . (16)
4
where h is the constant spacing of velocity vectors in both The classical added-mass force is recovered from these two
the x- and y-directions, and all data points in the domain non-circulatory terms by means of the relation
of integration are indexed by the subscript j. A two-dimen-
sional Taylor series expansion of the integrand in Eq. (7), d 𝜋D2
𝐅am = −𝜌 (𝐏𝜙 + 𝐏b ) = − 𝐚, (17)
as delineated in Appendix A, reveals that the approximation dt 4
in Eq. (11) leads to truncations errors, en , on the order of where 𝐚 = d𝐮c ∕dt is the cylinder acceleration. Equation (4)
( ) can also be expressed in terms of impulse, where the total
en ∼ 
1 4 2
h ∇ (𝐱 × 𝝎) . (12) impulse is calculated as
24
𝐏 = 𝐏v + 𝐏b (18)
In the present study, the dimensionless spacing of velocity
vectors, h∗ = h∕D , will be of order h∗ ∼ (10−2 ), such that rather than using Eq. (6). These formulations are equivalent
(h∗ )4 ∼ (10−8 ) and the truncation errors have negligible when the no-slip condition holds, which necessarily requires
effect on the calculated force coefficients for data points in that
the interior of the fluid. 𝐏i ≡ − 𝐏𝜙 . (19)
The approximation in Eq. (11) treats each fluid element
This relationship, referred to as the conservation of image-
as a concentrated point vortex of circulation 𝚪j = 𝝎j h2 at
vorticity impulse (Limacher et al. 2018), will be discussed
the centre of the fluid element. By further assuming that
in the present study. Although Eqs. (6) and (18) are theo-
this point-vortex approximation yields accurate estimates of
retically equivalent, they may yield different results when
induced velocity as per the Biot–Savart law, the image-vor-
applied to experimental data containing uncertainty. Here-
ticity distribution can be approximated as the corresponding
after, the formulation acronyms—GAMC and SIF—will be

13
4 Page 4 of 22 Experiments in Fluids (2019) 60:4

used as subscripts to distinguish quantities calculated from The prescribed cylinder motion is restricted to the
these respective equations. x-direction. Although cylinder vibrations contribute to the
The various quantities are normalized by cylin- raw y-impulse signals through the irrotational terms in Eqs.
der diameter and maximum cylinder velocity, Umax , i.e. (15) and (16), these contributions will be of zero mean. As
𝛤 ∗ = Γ∕(DUmax ), u∗c = uc ∕Umax , y∗ = y∕D , and 𝐱∗ = 𝐱∕D . will be shown, y-force coefficients (i.e. lift coefficients)
Normalized x-impulse, P∗x = Px ∕(D2 Umax ), is expressed in remain near zero after filtering during the entire period of
the GAMC formulation as observation. Measurement of y-acceleration was deemed
[ ( ) ] unnecessary to the objectives of this study, and lift coeffi-
∑ 1 𝜋 cients are reported only for posterity. y-impulse is calculated
P∗x,GAMC = 1− ∗ ∗
yj 𝛤j + u∗c . (20)
j
4|𝐱j |
∗ 2
V 4 simply as
[ ( ) ]
Upon time differentiation, the first term yields the circula- ∑ 1
∗ ∗ ∗
Py,GAMC = − 1− x 𝛤 , (26)
tory force, and the second yields the added-mass force. The j
4|𝐱j∗ |2 j j
V
x-component of normalized impulse is calculated in the SIF
as [ ]
[ ] ∑
∑ 𝜋 ∗ P∗y,SIF =− xj∗ 𝛤j∗ , (27)
P∗x,SIF = y∗j 𝛤j∗ − u . (21) j
j V 4 c V

and the lift coefficients are calculated as


The second term in Eq. (21) does not yield the added-mass
force upon differentiation; through Eq. (5), this term lends a dP∗y,GAMC
Cy,GAMC,raw = − 2 , (28)
force component in the same direction as the body accelera- dt∗
tion, not in opposition to it. This force component appears in
previous work with the same sign (Wu 1981; Eldredge 2007; dP∗y,SIF
Kriegseis and Rival 2014). Cy,SIF,raw = − 2 . (29)
dt∗
For either formulation, the drag coefficient, Cx , is defined
as Once again, these force coefficients are filtered to yield the
reported values of Cy,GAMC and Cy,SIF in Sect. 4.
Fx� Time derivatives in the preceding equations will be
Cx = 2
, (22) approximated with a second-order central differencing
1∕2𝜌DUmax
scheme:

dP∗x ( ∗ ∗ )
dMk Mk+1 − Mk−1
=−2 , (23) ≈− , (30)
dt∗ dt∗ 2𝛥t∗
where Fx′ is the force per unit length on the cylinder in
where M is the parameter undergoing differentiation, the
the x-direction and t∗ = tUmax ∕D . In the present study, an
subscripts k − 1, k and k + 1 are the chronological indices of
accelerometer is used to measure the acceleration in the
consecutive data points, and 𝛥t∗ is the dimensionless time
x-direction, permitting the calculation of drag coefficients
between them.
as follows:
[ ( ) ]
d ∑ 1 𝜋
yj 𝛤j − a∗x,raw , 3 Experimental methods
∗ ∗
Cx,GAMC,raw = − 2 ∗ 1−
dt j
4|𝐱 j
|
∗ 2 2
(24)
3.1 Experimental setup and PIV methodology
[ ]
d ∑ ∗ ∗ 𝜋 Sketches of the experimental setup in the water channel at
Cx,SIF,raw = − 2 ∗ yj 𝛤j + a∗x,raw , (25)
dt j
2 the University of Calgary are shown in Fig. 1a, b. Using this
apparatus, an acrylic hollow circular cylinder of diameter
2 is the dimensionless raw accel-
where a∗x,raw = ax,raw D∕Umax D = 25.4 mm was towed through quiescent water of depth
eration measurement. These force coefficients are filtered as 410 mm. A clearance of 2 mm, or 0.079D, existed between
described in Sect. 3.2 to yield the reported force coefficients, the bottom of the cylinder and the channel floor; the sub-
Cx,GAMC and Cx,SIF , in Sect. 4. merged aspect ratio of the cylinder was thus about 16.1.
The tunnel width is 385 mm, resulting in a blockage ratio

13
Experiments in Fluids (2019) 60:4 Page 5 of 22 4

Fig. 1  a Sketch of the experi-


mental setup for acquiring
particle images. b Side view
of the experimental setup with
key dimensions. Sketches not
to scale

Fig. 2  Magnitude of a
programmed dimensionless
velocity, u∗c , and b programmed
dimensionless acceleration, a∗,
versus dimensionless time, t∗,
for all three test cases

(a) (b)

of 6.6%. The cylinder was mounted to an ATI Mini40 six- Table 1  Reynolds number, Case a∗p Re
axis force transducer above the waterline, with the cylinder Re, and peak dimensionless
acceleration, a∗p, for the three 1 0.50 5150
piercing the free surface. An accelerometer (Analog Devices acceleration cases
ADXL337) was fixed to the mount between the cylinder and 2 0.75 5138
the transducer with one of its axes aligned with the direction 3 1.00 5101
of cylinder motion. The transducer was rigidly mounted to
a belt-driven traverse system (Parker HPLA 80) which sits
above the water channel. Extruded polystyrene foam sheets tance of 500 mm, or 19.7D, for all cases, although the flow
were attached to the traverse fore and aft the cylinder, mov- field could only be observed over a towed distance of about
ing with the cylinder and skimming the free surface. These 4D. The water was left to settle for 10 minutes between tri-
sheets prevent surface distortions that would unevenly reflect als. Four repeated trials for each case were completed.
scattered laser light into the cameras, thereby introducing Due to water temperature drift during the experimen-
error into the PIV correlations. tal campaign of about 0.4 ◦ C, Reynolds numbers for the
The traverse was programmed to accelerate the cylinder three cases varied between of 5100 and 5150, calculated as
to a maximum velocity of Umax = 0.198 m/s under three dif- Re = Umax D∕𝜈 , where 𝜈 is the kinematic viscosity. Reynolds
ferent symmetric triangular acceleration profiles. The pro- number estimates are reported in Table 1.
grammed dimensionless accelerations, defined as Since the force formulation includes an area integral over
a∗ = du∗c ∕dt∗ , are plotted versus dimensionless time in the full fluid domain, PIV data were acquired all around the
Fig. 2a, b, exhibiting peak dimensionless accelerations of cylinder. Two aligned laser sheets, measured at low laser
a∗p = {0.50, 0.75, 1.00}. The cylinder was towed for a dis- power to be about 4 mm in thickness across the region of

13
4 Page 6 of 22 Experiments in Fluids (2019) 60:4

interest, were directed at the cylinder midspan from either with DaVis 8.3 to yield particle image-pair correlations.
side of the water channel to eliminate shadows. A pulsed An iterative multi-grid cross-correlation algorithm was
Nd:YLF laser with a wavelength of 527 nm and an energy used (Scarano and Riethmuller 2000), with a final interro-
of 20 mJ per pulse (Photonics DM20-527) was used in con- gation window size of 16 × 16 pixels. A 50% overlap was
junction with a beam splitter and other required optics as applied between adjacent interrogation windows, resulting
depicted in Fig. 1a. A converging–diverging pair of spherical in a vector spacing of 0.0265D for the leeward camera, and
lenses (focal lengths of 150 mm and − 100 mm) was used to 0.0260D for the forward camera. The velocity data were
reduce the cross-section of the beam. On the longer of the post-processed using a universal outlier detection technique
two beam paths, an additional spherical lens, of focal length (Westerweel and Scarano 2005), and the application of a
1000 mm, was placed in the beam path to compensate for spatial denoising filter over a 5 × 5 vector window.
finite divergence of the beam to match the thickness of the The force measurements were synchronized with the cap-
two sheets in the region of interest. ture of particle images and the motion of the traverse using a
Two Phantom Miro LAB340 cameras, with 35 mm Nikon rising trigger signal. The trigger signal, the analog accelera-
AF Nikkor lenses, were mounted below the channel at loca- tion signal, and the six channels of data from the force trans-
tions ahead and behind the cylinder, permitting complete ducer were all recorded on the same data acquisition device.
optical access to the flow-field; in the results, the cameras There was an estimated 24 ms delay between the start of
will be referred to as the forward and leeward cameras. To image capture and the start of the commanded motion pro-
achieve the desired FoV overlap while maintaining fine spa- file. This was determined by analysing the dewarped parti-
tial resolution, the cameras were mounted at inward angles cle images, and determining the delay which gives the best
towards each other, and Scheimflug adapters were used to agreement between the observed and expected positions
align the focus planes horizontally. By angling the cameras, of the cylinder. Examples of dewarped particle images are
local spanwise velocities can introduce bias in the in-plane shown in Fig. 3. The force data acquisition is assumed to
velocity estimates, but this source of error should remain align with the start of image capture, as both devices were
small since the flow is expected to be predominantly two- triggered by the same signal over short cabling distances. In
dimensional. For subsequent analysis, the velocity-field data the results to follow, the time t∗ = 0 coincides with the start
for x∗ < 0 are taken from the forward camera, and those for of the commanded motion profile, i.e. 24 ms after the ris-
x∗ ≥ 0 are taken from the leeward camera. ing edge of the trigger signal. All force measurements were
The water was seeded with silver-coated hollow glass captured at a sampling rate of 2000 Hz.
spheres of mean diameter 11.7 μ m (Potters 110P8), which
were mixed evenly by running the water channel pump for 3.2 Filtering procedure
about ten minutes and letting the water settle for more than
an hour. Synchronized particle images were then acquired The measured forces and the forces calculated from PIV data
from both cameras at a rate of 500 Hz using LaVision’s must be filtered to remove information that is not meaningful
DaVis 8.4 software. The laser was fired in single-pulse to the present investigation (e.g. inertial forces due to vibra-
mode, synchronized with image capture using a LaVision tions, electrical interference, etc.). The filtering procedure
high-speed controller. Particle images were later processed is described in the present section.

Fig. 3  Dewarped particle


images at t∗ = 2.92 (200th
captured image) from the a
forward camera and b leeward
camera. The images are overlaid
with contours of cylinder cross-
section indicating the estimate
cylinder location. The solid red
line indicates the surface visible
to the camera and the dashed
black line indicates the contour
obscured from the camera by
the cylinder

13
Experiments in Fluids (2019) 60:4 Page 7 of 22 4

A discretely sampled signal 𝜓(t) is filtered in the fre- artifacts are possible for the modified Gaussian filter, but their
quency domain by manipulation of its discrete Fourier magnitude is expected to be small due to the low magnitude
transform, 𝛹 (f ). Unwanted frequencies are attenuated by of the negative extrema of g′.
multiplication of 𝛹 (f ) with a windowing function, G(f), and The rate of image capture for PIV is slower than that of the
the filtered time-domain signal is recovered using an inverse force and acceleration data acquisition (500 Hz rather than
Fourier transform. In the present study, the sampling time of 2000 Hz), and PIV data were only acquired over a compara-
these signals is about tsam = 4.1 s, yielding a frequency reso- tively short time window. As such, prior to filtering, the calcu-
lution in the Fourier domain of about 𝛥f = 1∕tsam = 0.24 Hz. lated force is cast onto the time vector associated with the force
For signals that contain sharp transients, as is the case with data acquisition. During the time of image capture, these data
the force signal at the onset of motion, a filtering artifact are linearly interpolated; before and after this time period, the
known as Gibb’s phenomenon, or colloquially as ringing, calculated drag coefficients are set to constant values of zero
must be considered. This effect is characterized by spurious and one, respectively, where the latter approximates the mean
oscillations in the vicinity of the discontinuities in the time steady-state drag.
domain. Ringing can be avoided altogether by choosing a
Gaussian window function, G(f; c): 3.3 Force data processing
[ ( )]
2
1 f Three steps are necessary in the processing of the measured
G(f ;c) = exp − , (31) forces: (1) the inertial force needs to be removed, (2) the bias
2 c
force given by the transducer under acceleration needs to be
where c is the standard deviation of the Gaussian curve. characterized and removed, and (3) the force needs to be fil-
In the present study, c cannot be chosen so as to achieve tered to remove high-amplitude oscillations due to vibrations
acceptably low attenuation in the passband while simultane- induced by the traverse system.
ously attenuating unwanted high-frequency information. The The mount and accelerometer, whose total mass is
following modified Gaussian function with a flat passband mm = 0.032 kg, were towed through the same motion profiles
region and an unambiguous stopband region has thus been shown in Fig. 2 with the cylinder removed. The bias force
selected: coefficient is determined as
Fx,bias = Fx,raw + mm ax,raw , (33)
f ≤ fpb

⎪1 � � �2 �
− 0.157 fpb < f ≤ fsb
⎪ f −fpb
G� (f ) = ⎨ 1.157exp − 2
fsb −fpb

⎪0 f > fsb

(32)

where fpb and fsb are the passband and stopband frequencies,
respectively. The region between fpb and fsb is a Gaussian
curve with c = c0 = 0.5(fsb − fpb ) , stretched vertically to
equal zero at f = fsb and one at f = fpb.
(a)
The lowpass filter to be used in the present study employs
a modified Gaussian window function with fpb = 5 Hz and
fsb = 14 Hz. This particular choice is a compromise between
the tendency to promote ringing, and the desire to attenuate
forces due to high-frequency mechanical vibrations that can-
not be adequately characterized, and that would be difficult
for an independent researcher to reproduce. Except where
otherwise stated, comparisons between the measured forces
and those calculated from PIV data will employ this filter.
The frequency-domain representation of this filter, G� (f ), is
plotted alongside a Gaussian window function, G(f ;2c0 ), in (b)
Fig. 4a.
Figure 4b shows the time-domain representations of the
Fig. 4  a Gaussian and modified Gaussian window functions, G(f ;2c0 )
Gaussian and modified Gaussian filters, g(t) and g� (t). The and G� (f ). b Time-domain representation of the Gaussian and modi-
presence of a zero crossing in g� (t) indicates that ringing fied Gaussian filters, g(t) and g� (t)

13
4 Page 8 of 22 Experiments in Fluids (2019) 60:4

where ax,raw is the raw measured acceleration, and Fx,raw is The estimated bias mass is used to remove the bias force
the raw measured force in the x-direction. Throughout the from subsequent measurements with the cylinder attached.
present study, measured forces will be non-dimensionalized Designating the raw drag force measurement prior to inertial
according to subtraction to be Fx,raw
′ , the hydrodynamic portion of the
measured drag on the cylinder, Fx,raw , is calculated as
F(⋅) �
C(⋅) = 2
, (34) Fx,raw = Fx,raw + (moh + mb )ax,raw , (37)
1∕2𝜌DHUmax
where moh is the overhung mass attached to the force trans-
where H is the submerged height of the cylinder, and ducer, measured to be 0.202 kg. The measured and filtered
(⋅) indicates the force of interest. For one trial of case 3, x-forces (i.e. drag) and force coefficients will then be desig-
Fig. 5 shows the measured dimensionless acceleration, nated Fx,meas and Cx,meas, respectively. Similarly, Fy,meas and
2 , and the bias x-force coefficient, C
a∗x,raw = ax,raw D∕Umax x,bias,
Cy,meas will represent measured and filtered forces and force
to be in phase with one another. The acceleration signal coefficients in the y-direction (i.e. lift).
is filtered according to the procedure in Sect. 3.2 to yield Of note in Fig. 5 is the acceleration peak, and correspond-
a∗x,meas. The filtered acceleration shows good agreement with ing force peak, just before t∗ = 0.5, preceded by near-zero
the programmed acceleration, a∗. force and acceleration. This feature appears in all cases. The
The ratio relating the bias force and the acceleration con- traverse experiences a delay of about 60 ms from the com-
stitutes the dimensionless bias mass internal to the trans- manded starting moment to the actual onset of movement,
ducer that must be accounted for. Expressing the relationship perhaps due to static friction. Quickly thereafter, the traverse
as recovers from this initial error and follows the programmed
trajectory. It is estimated that this initial behaviour invali-
Cx,bias = m∗b a∗x,raw , (35) dates the data in the time period t∗ < 0.75.
the dimensionless bias mass that minimizes the quantity The traverse motor is connected to the apparatus shown
∑ in Fig. 1b by a toothed belt, and vibration is generated at
(Cx,bias,k − m∗b a∗x,raw,k )2 (36) the frequency with which the teeth pass over the pulleys.
k
The frequency of vibrations is thus proportional to traverse
is m∗b = 0.767 ± 0.011, indicating an additional mass of speed, and at the chosen maximum speed, Umax, the traverse
mb = 0.101 ± 0.002 kg internal to the transducer which vibrates at 19.8 Hz. The natural frequency of the submerged
contributes to the measured inertial force (calculated as cylinder is 18.6 Hz, as determined by an impulse test on the
mb = 0.5m∗b 𝜌D2 H ). These values represent an average across stationary cylinder in still water. Despite the proximity of
nine datasets (three trials of each of the three acceleration the excitation and natural frequencies, the associated cylin-
cases), and the stated uncertainty ranges represent two stand- der displacements relative to the programmed trajectory are
ard deviations across these trials. small. In the particles images, the cylinder surface appears
to oscillate with an amplitude of several pixels, which is less
than the vector spacing of 8 pixels. Associated errors in the
calculation of impulse will be discussed.
Although the vibrations generate small cylinder displace-
ments, their associated accelerations contribute significantly
to the unfiltered drag force coefficient, Cx,raw, as shown for a
sample trial of the a∗p = 1.00 case in Fig. 6. In the present
study, the high frequency force oscillations due to vibrations
are filtered, and the force formulations are validated in their
ability to capture the low frequency trends associated with
vortex evolution. The chosen filter fully attenuates the vibra-
tions at the maximum speed, but is unable to fully attenuate
the lower frequency vibrations that occur during acceleration.
An artifact of the sudden motion onset therefore appears in the
filtered drag coefficients, Cx,meas, around t∗ = 0.5 in Fig. 6; this
artifact is also observed for all other trials and all other cases.
Fig. 5  Measured (unfiltered) bias force coefficient, Cx,bias, measured
Once the traverse reaches a programmed speed of about
(unfiltered) dimensionless acceleration, a∗x,raw, measured (filtered)
dimensionless acceleration, a∗x,meas, and programmed dimensionless 0.7Umax, the vibrations excited by the traverse reach the stop-
acceleration, a∗, versus dimensionless time, t∗, for one trial with a band frequency of fsb = 14Hz and subsequent vibrations are
dimensionless peak acceleration of a∗p = 1.00

13
Experiments in Fluids (2019) 60:4 Page 9 of 22 4

Fig. 7  Domain of integration in the calculation of impulse, divided


into subdomains for which data is taken from the forward and leeward
Fig. 6  Raw and filtered measured drag coefficient, Cx,raw and Cx,meas, cameras
for one trial of case 3 (a∗p = 1.00)

4 Results
fully attenuated. The time at which this speed is reached coin-
cides roughly with the time to flow separation for all cases 4.1 Vorticity contours and force histories
(within a dimensionless time of 0.2), such that hydrodynamic
oscillations due to vortex evolution after separation are not PIV data were acquired for four diameters of cylinder travel.
corrupted by vibrations. For each of the three cases, Fig. 8 shows sequences of nor-
malized vorticity contours at regular intervals of dimension-
3.4 Calculation of vorticity and impulse less distance travelled, s∗ = s∕D , where s is measured rela-
tive to the starting position. The vorticity contours represent
Dimensionless vorticity, 𝜔∗ = 𝜔D∕Umax, is calculated from averaged fields over four trials.
the normalized velocity data by means of a second-order After flow separation, the formation of symmetric vorti-
central differencing scheme. To calculate vorticity for data ces is observed, in agreement with previous observations
points adjacent to the cylinder, the normalized velocity field (Fackrell 2011; Laschka 1985). The dimensionless time of
within the cylinder was set equal to [u∗cx,raw , 0]T , and these separation, ts∗, is estimated for each case from inspection of
data were included in the curl operation. u∗cx,raw = u∗cx,raw (t∗ ) the averaged vorticity contours. These times and the corre-
is acquired by integrating the dimensionless raw acceleration sponding dimensionless distances travelled, s∗s , are listed in
measurement: Table 2. In all three acceleration cases, the subsequent evolu-
tion of the vorticity field involves an interaction between the
t∗
separated shear layers and the reversed boundary layers
∫0
u∗cx,raw (t∗ ) = a∗x,raw dt∗ . (38)
behind the separation points. For a period of time, the pri-
mary vortex cores are fed vorticity directly from the sepa-
The calculated vorticity field protrudes one vector into the rated shear layer; as the reversed boundary layers grow, the
cylinder, and these data within the cylinder are included in flux of vorticity emanating from the shear layers is inter-
the vortical impulse calculations. rupted. This is accompanied by the ejection of opposite-
The calculation of the vortical impulse terms is performed signed vorticity from the cylinder surface. This process is
using vorticity data within the rectangular domain bounded by periodic and results in the formation of secondary vortical
− 1 ≤ x∗ ≤ 2 and − 1 ≤ y∗ ≤ 1 (see Fig. 7). Data for x∗ ≥ 0 structures which are eventually incorporated into the pri-
are taken from the leeward camera and data from x∗ < 0 are mary vortex cores. One period of this process is portrayed
taken from the forward camera, and the vortical impulse is the in greater detail in Fig. 9, which shows vorticity contours for
sum of the impulse calculated in each of these subdomains.

Table 2  Dimensionless times a∗p ts∗ s∗s


and distances travelled at the
estimated moment of flow 0.50 2.28 0.493
separation (ts∗ and s∗s ) for each
0.75 1.66 0.421
acceleration case
1.00 1.42 0.456

13
4 Page 10 of 22 Experiments in Fluids (2019) 60:4

Fig. 8  Contour plots of normalized vorticity, 𝜔∗, for a∗p = 0.50 (left column), 0.75 (middle column), and 1.00 (right column). Dimensionless dis-
tance travelled, s∗, increases from top to bottom

13
Experiments in Fluids (2019) 60:4 Page 11 of 22 4

case 3 ( a∗p = 1.00 ) at twelve stations from s∗ = 1.8 to P∗𝜙x =


𝜋 ∗
u ,
2 cx,raw
[ ]
s∗ = 2.9. Throughout the period of observation, the primary ∑ (40)
1
vortex cores remain in close proximity to the cylinder. For P∗ix = − h∗2 y∗ 𝜔∗ .
4|𝐱j∗ |2 j j
further detail, refer to the videos in the supplementary j V

material.
Measured and filtered drag and lift coefficients for four In Fig. 12a, it is demonstrated that Pix and − P𝜙x follow the
repeated trials for each acceleration case are plotted in same trend, but exhibit distinct low-amplitude oscillations
Fig. 10 versus dimensionless distance travelled, s∗. For the which are consistent across repeated trials. The dimension-
two greater acceleration cases, the force histories after the less period of the oscillations is about 0.4, which corre-
end of cylinder acceleration are in remarkable agreement sponds to the frequency of traverse vibrations. The proximity
with one another. The local minima and maxima in the drag of the traverse vibration frequency to the natural frequency
after the end of acceleration can be attributed to the forma- of the cylinder leads to variations in the phase and magni-
tion of secondary vortical structures as described above. The tude of the vibrational oscillations along the span of the
dimensionless period of these drag oscillations is approxi- cylinder. Thus, the acceleration measurement above the
mately 0.9, which corresponds to a frequency of 9 Hz. waterline does not reliably capture the oscillations that occur
at the illuminated plane and affect the near-body velocity
field. Lower frequency components of the motion, however,
4.2 Comparison of measured and calculated forces
should be in phase all along the cylinder. After filtering
to yield Pix,filt and −P𝜙x,filt , we find remarkable agreement
A comparison of the measured and calculated forces is
across acceleration cases in Fig. 12b.
presented in Fig. 11. Four independent trials are plotted
together, demonstrating high repeatability. The same quali-
tative trends are captured by the measured and calculated 4.4 Force contributions from individual impulse
forces. For all three acceleration cases, the drag coefficients terms
obtained using both impulse formulations are in close agree-
ment with one another, but are consistently 10–20% less than The contributions of the vortical impulse term to the drag
the measured drag coefficients. Possible reasons for this dis- coefficient are calculated as
crepancy include end effects and underestimation of viscous [ ]
drag, as discussed further in Sect. 5.1. The filtered x-compo- ∑
∗2 d ∗ ∗
nent of the added-mass force coefficient is calculated from Cx,v,raw = − 2h y 𝜔 (41)
dt∗ j j j
the measured and filtered x-acceleration according to
𝜋 [ ( ) ]
Cam = − a∗x,meas . (39)
2 ∗2d ∑ 1 ∗ ∗
Cx,i,raw = 2h y 𝜔 . (42)
dt∗ j 4|𝐱j∗ |2 j j
Cam is plotted alongside the drag coefficients in Fig. 11. The
measured and calculated drag coefficients follow the added-
mass trend closely up until flow separation, at which point These force coefficients are then filtered to yield shed-vorti-
the trends diverge considerably. The filtered lift coefficients city and image-vorticity impulse contributions, Cx,v and Cx,i,
are also plotted, and remain near zero for all cases due to the respectively. The contributions of the irrotational impulse
symmetry of the flow field. terms are calculated using the measured, filtered and nor-
For all three cases, an artifact of the traverse startup malized cylinder acceleration signal according to
behaviour persists around t∗ = 0.5 despite filtering, appar- Cx,𝜙 = −𝜋a∗x,meas (43)
ent in both the measured force and measured acceleration
used to calculate Cam . The behaviour of the traverse at the 𝜋 ∗
Cx,b = a (44)
onset of motion is described in Sect. 3.3, and the data prior 2 x,meas
for t∗ < 0.75 should not be treated as valid. All four of these contributions are plotted for all three cases
in Fig. 13, alongside their sum, Cx,GAMC . Due to the con-
4.3 Conservation of image‑vorticity impulse servation of image-vorticity impulse, the contributions Cx,i
and Cx,𝜙 tend to cancel; after the cylinder reaches constant
The principle of image-vorticity impulse conservation pre- velocity, the contribution of both of these term tends to zero.
sented in Limacher et al. (2018) requires that 𝐏i ≡ − 𝐏𝜙 . During acceleration, Cx,v and Cx,i contain residual oscilla-
This principle is tested with the present dataset using the tions due to vibrations that are nearly perfectly antiphased,
x-components of these impulses, calculated as such that they cancel upon summing. The effect of vibrations

13
4 Page 12 of 22 Experiments in Fluids (2019) 60:4

Fig. 9  Dimensionless vorticity


contours, 𝜔∗ for dimensionless
distances from 1.8 to 2.9 for
case 3 (a∗p = 1.00). Second-
ary vortical structures form
due to interaction between the
separated shear layer and the
reversed boundary layer aft of
the separation point

13
Experiments in Fluids (2019) 60:4 Page 13 of 22 4

Fig. 10  Measured and filtered


(filter 1) drag and lift coef-
ficients, Cx,meas and Cy,meas,
versus dimensionless distance
travelled, s∗

on the near-body vorticity field, and thence on the force, is coefficient using all of the available data, Cx,GAMC , is also
captured by the potential impulse contribution in the GAMC plotted (in blue) for reference. Although these results are
formulation. This cancellation of the near-body vorticity specific to the present case, they offer a preliminary estimate
contribution is not present in the SIF, and the effect of vibra- of the required proximity to the cylinder at which data must
tions on the force is captured in part by the shed-vorticity be acquired to achieve a desired degree of accuracy. The
impulse contribution. For further discussion on vibrations, greatest deviations occur shortly after flow separation, where
see Sect. 5.3. a large concentric data omission misses all of the relevant
vorticity field evolution. For the smallest omission of data
4.5 Sensitivity of the impulse formulations (ro∗ = 0.04), the maximum deviation of the calculated drag
to the omission of vorticity data coefficient from Cx,GAMC is 0.04.

Two types of data omission are considered: (1) the concen-


tric removal of all data within a constant radial distance 5 Discussion
ro∗ of the cylinder surface, i.e. omission of all data within
|𝐱∗ | ≤ ro∗ + 1∕2 , and (2) the omission of all data from the 5.1 Underestimation of force using impulse
forward camera, i.e. omission of all data for x∗ < 0 . The formulations
two omission types reflect two important challenges to PIV
investigation, namely the resolution of data near the sur- Using both the SIF and the GAMC formulation, the calcu-
face and the ability to resolve data all around a body. The lated drag coefficients were consistently lower than the
effects of these omissions on the calculated drag coefficients measured drag coefficients, with deficiencies typically
are demonstrated in Fig. 14, where the concentric omission between 10 and 20% of the measured value. Errors of this
is performed for ro∗ = 0.04 . The force calculated using the magnitude could be attributed to three-dimensional effects
SIF becomes significantly inaccurate due to the concen- in the flow-field. A simple check on the effect of ground
tric omission. Although the forward-camera data omission clearance was undertaken. The original clearance of 2 mm
affects the SIF during acceleration, its drag estimates are not was reduced to 1 mm, and the resulting measured and fil-
greatly affected under steady translation, indicating that the tered force histories for the a∗p = 1.00 case are plotted in
boundary layer on the forward part of the cylinder becomes
Fig. 16. The reduction in ground clearance has negligible
steady. Comparatively, neither of these omissions yield sig-
effect until about t∗ = 2.5, thereafter exhibiting small dis-
nificant changes in the calculated force using the GAMC
crepancies of alternating sign. These differences are not
formulation.
negligible, but they do not indicate a unidirectional bias
The concentric data omission is further investigated for
caused by the ground plane.
the GAMC formulation for the a∗p = 1.00 case. Figure 15
At the free surface, the generation of waves could con-
presents the filtered drag coefficients calculated with the tribute additional drag, although a method to test this
GAMC formulation when data within |𝐱∗ | ≤ ro∗ + 1∕2 are hypothesis was not devised. The use of the extruded poly-
omitted for four different values of ro∗; the calculated drag styrene sheets to skim the free surface would not prevent

13
4 Page 14 of 22 Experiments in Fluids (2019) 60:4

(a)

(b)

Fig. 12  a Raw x-component of image-vorticity impulse and raw


x-component of potential impulse for four repeated trials of the
a∗p = 1.00 test case. b Filtered x-component of image-vorticity
impulse and filtered x-component of potential impulse for one trial of
each of the three acceleration cases. The expected agreement between
these impulse signals is referred to as the conservation of image-vor-
ticity impulse

removed were undertaken to eliminate this feature of the


setup as a source of error. The results in Fig. 17 show that
no bias was induced by the foam sheets.
Another possible explanation for the observed bias is
the method of vorticity calculation at the cylinder surface
and the under-resolution of the boundary layers. Prior to
discretization, the presented impulse formulations are
exact, and the vortical impulse terms account for both
pressure and viscous components of force. It is conceiv-
able that numerical effects at the boundary could cause the
impulse formulations to under predict the viscous com-
Fig. 11  Measured and calculated drag coefficients (using the GAMC
and SIF formulations) versus dimensionless time for all three cases
ponent of drag. For steady flow over a nominally two-
with peak dimensionless accelerations of a a∗p = 0.50, b a∗p = 0.75, dimensional cylinder, previous studies indicate a viscous
and c a∗p = 1.00. Four repeated trials are plotted for each case. Data in drag coefficient of around 0.05 at about Re = 3000 (Thom
the time period t∗ < 0.75, as indicated by the grey patch, are invalid 1933), and a percentage of total drag due to skin fric-
due to the startup behaviour of the traverse described in Sect. 3.3
tion at about 3% for Re = 10, 000 [given by Achenbach
(1968) with reference to Thom (1929)]. During accelera-
this source of drag, as the sheets were constructed with a tion from rest, the boundary layers would be thinner with
generous radial clearance of 15–20 mm around the cyl- greater normal velocity gradients than under steady trans-
inder to prevent any chance of contact with the cylinder. lation. The contribution of skin friction to total drag could
Repeated force measurements with the sheets in place and therefore exceed the steady-state estimates in the present

13
Experiments in Fluids (2019) 60:4 Page 15 of 22 4

(a)

(a)

(b)

(b)

(c)

Fig. 13  Filtered contributions of individual impulse terms to the drag


coefficient for all three cases with peak dimensionless accelerations
of a a∗p = 0.50, b a∗p = 0.75, and c a∗p = 1.00. The subscripts v, i, 𝜙,
and b denote the contributions from the shed-vorticity, image-vorti-
city, potential and body-volume impulse terms, respectively

(c)
unsteady investigation; that is, the observed bias is plausi-
bly of the same order as the expected viscous drag. Fig. 14  Drag coefficients from the GAMC formulation and the SIF
with all vorticity data included (blue and cyan solid lines) in the
calculation and with two different omissions of vorticity data. Force
5.2 Uncertainty in calculated impulse is recalculated including only data points at a radial distance of
r0∗ = 0.04 or greater from the cylinder surface (upward-facing triangle
Using standard error propagation techniques, the uncertainty markers), and then again with only data points in the region x∗ ≥ 0
in the calculated impulses are determined due to uncertainty (downward-facing triangle markers)
in the PIV velocity data, which is estimated for every vector
according to the method of Wieneke (2015) within LaVision’s
DaVis 8.3 software (n.b. the estimates furnished by the DaVis deviations of random scatter). An example contour plot of the
software have been multiplied by two to yield uncertainties normalized x-velocity uncertainty, 𝛿u∗ = 𝛿u∕Umax, is shown
with a 95% confidence integral, i.e. covering two standard in Fig. 18a for one trial of the a∗p = 1.00 case, alongside the

13
4 Page 16 of 22 Experiments in Fluids (2019) 60:4

Fig. 18  a Normalized x-velocity uncertainty, 𝛿u∗ for one trial at


a∗p = 1.00 at t∗ = 3.70, and b normalized vorticity contours, 𝜔∗, at the
same instant. Uncertainty is greatest in regions of high vorticity, espe-
cially near the cylinder surface

Fig. 15  Drag coefficients calculated with the GAMC formula- corresponding normalized vorticity contour plot at the same
tion with and without the concentric omission of data within instant in Fig. 18b. The regions of highest uncertainty tend to
|𝐱∗ | < ro∗ + 1∕2 for a single trial of the a∗p = 1.00 case be in regions of high vorticity, especially near the cylinder
surface where uncertainty in the velocity magnitude is around
10% of the maximum cylinder velocity.
The normalized x-impulse uncertainty due to random error
in the PIV correlations, 𝛿P∗x,GAMC,vel = 𝛿Px,GAMC,vel ∕(D2 Umax ),
is given by
[ {( ) }2 ]1∕2
∑ 1
𝛿P∗x,GAMC,vel = 1− ∗ ∗2 ∗
y h 𝛿𝜔k ,
k
4|𝐱k∗ |2 k
(45)
where h is the normalized vector spacing in both directions,

and 𝛿𝜔∗k is the calculated vorticity uncertainty. The sum in


Eq. (45) is taken over every data point in the fluid volume,
Fig. 16  Measured and filtered drag coefficient measurements at a V. For the SIF, the normalized impulse uncertainty is
ground clearance of 2 mm (blue lines) and 1 mm (red lines). Three
repeated trials of each ground clearance are shown for the a∗p = 1.00 [ { }2 ]1∕2
acceleration case ∑
𝛿P∗x,SIF,vel = y∗k h∗2 𝛿𝜔∗k . (46)
k

The derivation of Eqs. (45) and (46) is provided in Appendix


B. Since |𝐱j | > 0.5 for all locations outside the cylinder, the
inequality

1
1− <1 (47)
4|𝐱j∗ |2

holds for all data points in the domain V, and the SIF is more
sensitive to velocity uncertainty than the GAMC formulation
in the absence of other sources of error. The impulse uncer-
tainties calculated using Eqs. (45) and (46) are plotted in
Fig. 19 for four repeated trials of the a∗p = 1.00 case. Trends
Fig. 17  Measured and filtered drag coefficient measurements with
and without the extruded polystyrene sheets skimming the free sur- and magnitudes of uncertainty are similar for the other cases.
face. Three repeated trials of each arrangement are shown for the
𝛿P∗x,SIF,vel is consistently greater than 𝛿P∗x,GAMC,vel , having up
a∗p = 1.00 acceleration case
to twice the magnitude. The low level of uncertainty between

13
Experiments in Fluids (2019) 60:4 Page 17 of 22 4

D2 Umax, are plotted versus dimensionless time in Fig. 20 for


one trial of the ap = 1.00 case. They are an order of magni-
tude smaller than the random errors 𝛿P∗x,GAMC,vel and
𝛿P∗x,SIF,vel , and thus contribute negligibly to Eqs. (50) and
(51).
For positional errors of similar magnitude to the grid
spacing, the evaluation of vorticity for data points nearest
the body can also be affected, altering the values of the sums
in Eqs. (48) and (49). To definitively show that the calcu-
lated impulses and force histories are insensitive to cylinder
position error, Fig. 21 shows the filtered force histories for
Fig. 19  Normalized impulse uncertainty due to PIV uncertainty for
both formulations, 𝛿P∗x,GAMC,vel and 𝛿P∗x,SIF,vel, as calculated using Eqs.
one a∗p = 1.00 case with four different offsets applied to the
(45) and (46) versus t∗ for the four repeated trials of the a∗p = 1.00 p o s i t i o n d a t a : 𝛥𝐱∗ = [0.02, 0]T , 𝛥𝐱∗ = [−0.02, 0]T ,
case
𝛥𝐱∗ = [0, 0.02]T , and 𝛥𝐱∗ = [0, −0.02]T . The altered force
histories are denoted Cx,GAMC,𝛥, accompanied in the legend
of Fig. 21 by the associated offset 𝛥𝐱∗ . The four altered
t∗ = 2.5 and t∗ = 3.5 is likely a result of better illumination
curves are well aligned in Fig. 21 with the original force
during this period of time. Although any additional error in
estimate, Cx,GAMC . The largest differences between any of
the impulse signal will tend to be amplified by differentiation
these altered force estimates and Cx,GAMC is about 0.05,
in the calculation of force coefficients, it is of sufficiently
which occurs shortly after flow separation in the time inter-
high frequency in the present study that, after filtering, the
val from 1.5 < t∗ < 2.
variability in the force coefficients for both formulations is
The insensitivity of the force estimates to cylinder posi-
similar.
tion error recommends the chosen method of near-body vor-
In the present study, impulse uncertainty due to errors
ticity calculation as robust. By setting all velocity vectors
in the estimated cylinder position are of secondary impor-
within the cylinder to an estimated cylinder velocity prior
tance. Errors in the estimated cylinder position, denoted by
to calculating vorticity, errors in the cylinder position that
ep , yield an offset in the position variables used in the cal-
are of similar or lesser magnitude than the grid spacing will
culation of shed-vorticity and image-vorticity impulse. This
have little effect on the calculated force.
offset is constant across the domain of integration, and thus
cannot be treated as a collection of uncorrelated random
5.3 Capturing the effect of vibrations on force
errors as was done with the vorticity uncertainty. In Appen-
dix B, positional offset errors are propagated to errors in the
The unattenuated artifacts of vibrations prior to flow separa-
calculated impulses to yield
tion (see Fig. 11) give insight into how the two formulations
( ) capture the effect of vibrations on force. The direct effect
∑ 1
𝛿P∗x,GAMC,p = e∗p h∗2 1+ 𝜔∗j , (48) of vibrations is to introduce oscillations into the near-body
j
4|𝐱j∗ |2

[ ]

𝛿P∗x,SIF,p = e∗p h∗2 𝜔∗j . (49)
j

These positional errors are combined with the random errors


according to

𝛿P∗2
x,GAMC
= 𝛿P∗2
x,GAMC,p
+ 𝛿P∗2
x,GAMC,vel (50)

𝛿P∗2
x,SIF
= 𝛿P∗2
x,SIF,p
+ 𝛿P∗2
x,SIF,vel
. (51)

Conservatively, let us consider positional errors of


ep = 0.02D, which is slightly less than the vector grid spac- Fig. 20  Normalized impulse uncertainty due to cylinder position
ing. 𝛿P∗x,GAMC,p and 𝛿P∗x,SIF,p, which have been normalized by error for both formulations, 𝛿P∗x,GAMC,p and 𝛿P∗x,SIF,p, as calculated
using Eqs. (48) and (49) versus t∗ for the four repeated trials of the
a∗p = 1.00 case

13
4 Page 18 of 22 Experiments in Fluids (2019) 60:4

cylinder acceleration to calculate the irrotational contribu-


tions to force, and on the programmed cylinder velocity in
the calculation of vorticity at data points adjacent to the
cylinder. For four repeated trials of the a∗p = 0.50 case,
Fig. 22 shows the drag coefficients calculated by means of
this modified processing, which is denoted by the prime
superscripts in Cx,GAMC

, Cx,SIF

, and Cam
′ . The residual artifact

of the traverse startup behaviour, which persists after filter-


ing, is indeed better resolved by the SIF than the GAMC
formulation. However, this elevated sensitivity is of limited
utility. Although such low-frequency artifacts may be better
resolved with the SIF than when using the GAMC formula-
tion, the actual high-frequency vibrational contribution to
the raw forces cannot be determined without accurate char-
acterization of instantaneous acceleration. The accurate cal-
culation of Cx,b and the removal of bias and inertial compo-
nents of the raw force signal requires fully resolved
Fig. 21  Comparison of the measured and filtered drag coefficient, acceleration information.
Cx,meas, the calculated and filtered drag coefficient using the GAMC
formulation, Cx,GAMC, and four modified calculations in which the
position data for all grid points is shifted by some offset 𝛥𝐱∗. The
shown data is for one trial of the a∗p = 1.00 case 5.4 Calculating force with incomplete vorticity field
data

vorticity field. In Fig. 13, the early modulations in the Cx,v For the various reasons noted in the introduction, the experi-
and Cx,i signals, attributable to vibrations, were antiphased. mental determination of the complete vorticity field around
Insofar as the effect of vibrations is restricted to the near- a solid body is difficult, which poses an obstacle to the
body vorticity field, vibrations have negligible effect on the application of impulse-based force formulations. Efforts to
total vortical contribution to force in the GAMC formula- develop corrections to these formulation for missing vorti-
tion. By contrast, the modulations in Cx,v are not fully can- city data are therefore well motivated.
celled in the SIF, where Cx,SIF = Cx,v + Cx,b . In the absence The recent work of Graham et al. (2017) addressed this
of accurate characterization of vibrational motions, the SIF challenge for a two-dimensional flat plate. They delineate
may thus be expected to capture more (but not all) of the a method to determine a sheet-vorticity distribution on the
oscillatory force trends induced by vibrations. plate that would satisfy the no-slip condition in the pres-
Consider the case where an acceleration measurement is ence of incomplete vorticity field data. If the vorticity field
not available. We would then rely on the programmed were fully resolved, the no-slip condition would be directly

Fig. 22  Comparison of the


measured and filtered drag
coefficient, Cx,meas, with the
drag coefficients calculated
using the programmed traverse
motion rather than the measured
acceleration, Cx,GAMC

and Cx,SIF

.
In the absence of information
regarding the high-frequency
vibrations, the SIF captures
more of the residual modula-
tions in the force history that
persist after filtering

13
Experiments in Fluids (2019) 60:4 Page 19 of 22 4

observed and the strength of this vortex sheet would be image-vorticity concept is also valid in three dimensions,
found to be zero at all points. The vortex sheet represents a its implementation in practice is likely to be cumbersome.
correction for any missing data, and its associated impulse
is included in the impulse calculation. Their corrected for-
mulation yields good agreement with measured forces for a 6 Conclusions
translating plate test case, whereas the uncorrected SIF fails
due to the missing data. The decomposition of fluid-dynamic force into added-mass
The formulation of Graham et al. (2017) is presented as a and circulatory components is a common feature of unsteady
correction to the SIF, but it may also be viewed as a special aerodynamic theory, and was recently shown to be applica-
case of the GAMC formulation. They note that the vortex ble to viscous flows in Limacher et al. (2018). This general-
sheet can be decomposed into a portion owing to the vor- ized added-mass and circulatory (GAMC) formulation is
ticity in the wake, termed the circulatory component, and closely related to the typical representation of impulse the-
a portion owing to the motion of the plate. The impulse of ory in the literature, referred to as the standard impulse for-
the latter component, upon time differentiation, yields the mulation, or SIF. Both of these formulations are discretized
classical added-mass force. The derivation of the GAMC and applied to planar particle image velocimetry (PIV) data
formulation in Limacher et al. (2018) makes it clear that the to calculate force on a linearly accelerating cylinder under
inclusion of an added-mass term requires the concomitant three acceleration profiles. The resulting estimates capture
inclusion of an image-vorticity term. This condition is met the trend of the force well, as determined by comparison to
in the work of Graham et al. (2017), as their calculation of direct measurement with a force transducer. However, the
the vortex-sheet strength accounts for the effect of image calculated forces consistently underestimate the measured
vortices in the conformally mapped circle domain. forces by 10–20%. End effects are investigated, but could
The success of the method of Graham et al. (2017) can not be totally eliminated as a possible explanation for the
be attributed to the nearness of the missing vorticity data discrepancy. It is speculated that limitations on spatial reso-
to the body. Through the lens of the GAMC formulation, lution could cause the impulse formulations to inadequately
we can generalize this statement to arbitrary geometries. resolve the viscous component of drag, and this possibility
Due to the cancellation of the image-vorticity and shed- merits further study.
vorticity impulse terms for the near-body region (Limacher After flow separation, the symmetrical formation of a
et al. 2018), the exclusion of thin attached boundary layers vortex pair is observed in the wake. Subsequent oscillations
from the calculation of total impulse is not likely to have a in the drag coefficients, observed in the measured and cal-
detrimental effect on the accurate calculation of force, as culated force histories, are attributed to the formation of
observed in the present study. secondary vortical structures. These structures are generated
through the interaction of the separated shear layer and the
5.5 Evaluation of image‑vorticity impulse for other reversed boundary layer aft of the separation points. The
geometries secondary structures are incorporated into the primary vor-
tex cores, which remain in close proximity to the cylinder
The GAMC formulation in its general form is applicable to throughout the period of observation. Force histories show
any geometry in two or three dimensions, but its application qualitative agreement across acceleration cases when plotted
requires the evaluation of a valid image-vorticity distribution against dimensionless distance travelled.
and its associated impulse. This task is straight forward and A key difference between the SIF and the GAMC formula-
well known for two-dimensional circular cylinders, and can tion is the inclusion of an image-vorticity impulse term in the
be extended to other two-dimensional geometries by means latter. Cancellation of the image-vorticity and shed-vorticity
of conformal mapping. The image-vorticity distribution on impulse terms in the region near the body give rise to three
a thin flat plate, for example, is evaluated by conformally important consequences: (1) the GAMC formulation is less
mapping the domain around the plate to the circle domain sensitive to uncertainty in the velocity data, as determined by a
representing the cylinder cross section, and evaluating the standard uncertainty propagation analysis. This benefit comes
image vorticity as done herein. Conveniently, the total vor- at the cost of increased sensitivity to errors in the estimated
tical impulse, 𝐏v + 𝐏i , can then be evaluated in the circle cylinder position, but this source of error was found to be neg-
domain, so long as the conformal map has been chosen so ligible in the present study. In the present study, the difference
as not to distort the domain at infinity (see Limacher et al. in impulse uncertainty between these two formulations proved
(2018) for a proof of this statement). In the case of geom- to be inconsequential to the filtered drag estimates; the vari-
etries for which an analytical conformal map to the circle ability in drag coefficients across repeated trials is similar for
domain is not available, numerical methods are available both formulations. (2) The GAMC is less sensitive to the omis-
(Driscoll and Trefethen 2002). Unfortunately, although the sion of near-body vorticity data. The omission of vorticity data

13
4 Page 20 of 22 Experiments in Fluids (2019) 60:4

within a radial distance of 0.04D from the cylinder surface Appendix B: Propagating uncertainty
yielded changes in the calculated force that are less than 10% in velocity data to uncertainty in calculated
using the GAMC formulation, but on the order of 100% using force
the SIF. The GAMC similarly outperformed the SIF when
all data on the forward side of the cylinder (i.e. x∗ < 0) were If some quantity M is a function of the set of variables {Lj },
omitted from the calculation. (3) In the absence of precise the uncertainty in M can be calculated from the uncertainties
knowledge regarding the vibrational velocities of the cylinder in each Lj according to (Moffat 1988):
(superimposed on the commanded cylinder velocity), the SIF
[ ( )2 ]1∕2
is capable of resolving a greater portion of forces due to these ∑ 𝜕M
vibrations. However, neither formulation can fully resolve 𝛿M = 𝛿Lj . (56)
j
𝜕Lj
high-frequency vibrational forces without accurate charac-
terization of the instantaneous cylinder acceleration. This methodology can be applied to the calculation of the
vortical impulse terms, the x-components of which are cal-
Acknowledgements The authors thank the Natural Sciences and Engi-
neering Research Council (NSERC) for funding this research. culated for each formulation using

Pvx,SIF = h2 yj 𝜔j ,
Appendix A: Discretization errors (57)
j
in numerical integration of vortical impulse
( )
∑ D2
The shed-vorticity impulse, as defined in Eq. (7), is evaluated Pvx,GAMC = h2 1 − 2 yj 𝜔j . (58)
using the approximation in (11). For a single elemental area, j 4rj
its contribution to 𝐏v is calculated as
Errors in these calculated impulses result from two key
[ ]
sources: (1) uncertainty in the velocity vectors due to error in
𝐏v,j ≈ h2 𝐱j × 𝝎j .
V
the PIV correlations, and (2) errors in the estimated cylinder
position. Errors of the first kind can be treated as random,
This approximation is equivalent to a one term Taylor series and the standard error propagation approach in Eq. (56)
expansion. Setting can be applied. In this approach, the error in each vorticity
data point is treated as an independent random variable. In
𝐟 (x, y) = 𝐱 × 𝝎, (52) fact, neighbouring vorticity data points may have correlated
we can express errors, but the assumption that these errors are uncorrelated
permits tractable analysis. By contrast, uncertainty arising
𝜕𝐟 || 𝜕𝐟 |
𝐟(x, y) = 𝐟(x0 , y0 ) + (x − x0 )| + (y − y0 ) || due to errors in the position variable are not properly treated
𝜕x |(x0 ,y0 ) 𝜕y |(x0 ,y0 ) as uncorrelated across the domain of integration. Since the
1 2
𝜕 𝐟 | 1 𝜕 𝐟|
2
position variable is measured relative to the cylinder centre,
+ (x − x0 )2 2 || + (y − y0 )2 2 || +⋯
2 𝜕x |(x0 ,y0 ) 2 𝜕y |(x0 ,y0 ) we expect the errors in yj to be dominated by errors in the
(53) estimated cylinder location. Thus, at any instant in time,
Integrating over a square area of dimensions h × h centred the positional error is treated as a constant offset across the
on (x0 , y0 ), we obtain domain.
Let yj be the measured y-position in the cylinder-fixed
h∕2 h∕2 [ ] frame of reference and let Yj be the true position such that
h2 𝜕 2 𝐟 𝜕 2 𝐟
∫−h∕2 ∫−h∕2
2
𝐟(x, y)dxdy = h 𝐟(x0 , y0 ) + + + ⋯.
24 𝜕x2 𝜕y2 yj = Yj + ep , (59)
(54)
where ep is the error in the position estimate. The vortical
The first-order terms in the series do not contribute to the impulse is then calculated by means of the SIF as
integral, and the leading-order error, en, introduced by trun- ∑ ∑
cating the series to the zeroth order term is of order Pvx,SIF = h2 Yj 𝜔j + ep h2 𝜔j ,
( ) (60)
en ∼ 
j j
1 4 2
h ∇ (𝐱 × 𝝎) . (55) where h is the grid spacing in both directions. Vorticity is
24
calculated from velocity using a central differencing scheme:
1
𝜔i,j = (v − vi−1,j − ui,j+1 + ui,j−1 ), (61)
2h i+1,j

13
Experiments in Fluids (2019) 60:4 Page 21 of 22 4


where u and v are the velocities in the x- and y-directions, Pvx,GAMC = h2 y�j 𝜔j .
i and j are the x- and y-indices of the PIV dataset. The sec- (66)
j
ond term on the right-hand side of Eq. (60) represents the
impulse error due to cylinder position error, which we will Equation (66) has the same form as (57), and the uncertainty
denote as 𝛿Px,SIF,p. We expect the total circulation to be zero due to random error in the PIV correlations is given by

in the domain around a surging cylinder, but the sum j 𝜔j � �1∕2
may not be evaluated to be identically zero due to errors in �
𝛿Px,GAMC,vel = h 2
y�2 2
j 𝛿𝜔j
the vorticity.
j
Using Eqs. (56) and (61), the uncertainty in the calculated (67)
� �2 1∕2
vorticity can be expressed as ⎡� ⎤
2⎢ D2
√ =h 1− 2 y2j 𝛿𝜔2j ⎥ .
1 ⎢ rj ⎥
𝛿𝜔i,j = 𝛿v2i+1,j − 𝛿v2i−1,j − 𝛿u2i,j+1 + 𝛿u2i,j−1 , (62) ⎣ j ⎦
2h
where 𝛿u and 𝛿v are the velocity uncertainties. Within the
DaVis 8.3 software, estimates of 𝛿u and 𝛿v based on cor- Comparing Eqs. (67) and (63), it is clear that the GAMC
relation statistics are furnished by the method of Wieneke formulation is less sensitive to random error in the PIV cor-
(2015) (n.b. herein, the estimates furnished by the DaVis relations because
software are multiplied by two to yield uncertainties with a
95% confidence integral, i.e. covering two standard devia- D2
0< <1 (68)
tions of random scatter). Their method accounts for random 4rj2
errors which tend to blunt the correlation peak to reduce
the quality of particle displacement estimates. The resulting for all positions in the fluid domain outside the cylinder
component of impulse uncertainty, 𝛿Px,SIF,vel , is determined where rj > D∕2 . However, the cost paid for this reduced
using Eqs. (56) and (57) to yield sensitivity to random error is an increased sensitivity to
[ ]1∕2 error in the cylinder position estimate. As above, this error
∑ is treated by considering an offset to all location estimates
𝛿Px,SIF,vel = h2 y2j 𝛿𝜔2j , (63) in the evaluation of Eq. (67). For reasons that will become
j
apparent, let us consider the magnitude of the total vortical
impulse vector rather than solely the x-component:
where the sum over the index j includes every data point in
the two-dimensional vorticity field, i.e. 𝛿𝜔j in Eq. (63) is )(
∑ D2
corresponds to 𝜔i,j in Eq. (61). Let us assume that the errors |𝐏vt,GAMC | = h 2
1 − 2 rj 𝜔j
𝛿Px,SIF,p and 𝛿Px,SIF,vel are additive according to j 4rj
( ) (69)
∑ D2
𝛿P2vx,SIF = 𝛿P2x,SIF,p + 𝛿P2x,SIF,vel = h2 rj − 𝜔j .
4rj
[ ]2 [ ] j
∑ ∑ (64)
= e2p h4 𝜔j +h 4
y2j 𝛿𝜔2j .
j j Positional error will be considered radially; let rj be the
measured radial position, and let Rj be the true radial posi-
Equation (64) yields an estimate of uncertainty in the tion. These are related by the position error ep by
x-impulse values calculated by means of the SIF, account- rj = Rj + ep . (70)
ing for both random errors in the PIV correlations and
Equation (69) can be rewritten as
errors in the estimated position of the cylinder. Estimates of
y-impulse uncertainty would follow an identical pattern. Let ( )
∑ D2
us now repeat a similar analysis for the total vortical impulse |𝐏vt,GAMC | = h2 Rj + ep − 𝜔j . (71)
4(Rj + ep )
uncertainty for the GAMC formulation. For convenience, j

let us define
So long as the error ep is small, the error in impulse magni-
( 2
) tude due to radial position error ep can be estimated as
D
y�j = 1− yj , (65)
4rj2
𝜕|𝐏vt,GAMC |
𝛿|𝐏vt,GAMC |p ≈ ep , (72)
𝜕ep
where rj = |𝐱j |, and rj2 = xj2 + y2j , such that

13
4 Page 22 of 22 Experiments in Fluids (2019) 60:4

where Graham W, Pitt Ford C, Babinsky H (2017) An impulse-based approach


to estimating forces in unsteady flow. J. Fluid Mech. 815:60–76
( ) Kriegseis J, Rival DE (2014) Vortex force decomposition in the
𝜕|𝐏vt,GAMC | ∑ D2
= h2 1+ 𝜔j (73) tip region of impulsively-started flat plates. J. Fluid Mech.
𝜕ep j
4(Rj + ep )2 756:758–770
Kurtulus DF, Scarano F, David L (2007) Unsteady aerodynamic
forces estimation on a square cylinder by TR-PIV. Exp. Fluids
( ) 42(2):185–196
∑ D2
2 Laschka B (1985) Unsteady flows—fundamentals and application. In:
=h 1 + 2 𝜔j . (74)
j 4rj AGARD conference Proceedings No. 386, North Atlantic Treaty
Organization
Limacher E, Morton C, Wood D (2018) Generalized derivation of the
Conservatively, let it be assumed that this magnitude added-mass and circulatory forces for viscous flows. Physical
error contributes entirely to the x-direction of impulse, i.e. Review Fluids 03:014701
𝛿Px,GAMC,p ≈ 𝛿|𝐏vt,GAMC |p . Finally, we arrive at Milne-Thompson LM (1960) Theoretical hydrodynamics, 4th edn.
MacMillan, New York
( )
∑ Moffat RJ (1988) Describing the uncertainties in experimental results.
2 D2 Exp Therm Fluid Sci 1:3–17
𝛿Px,GAMC,p ≈ ep h 1 + 2 𝜔j . (75)
j 4rj Mohebbian A, Rival DE (2012) Assessment of the derivative-moment
transformation method for unsteady-load estimation. Exp Fluids
53(2):319–330
As before, we take the position error and random error to be Noca F (1997) On the evaluation of time-dependent fluid-dynamic
additive according to forces on bluff bodies. Ph.D. thesis, California Institute of
Technology
� � � �2 Noca F, Shiels D, Jeon D (1999) A comparison of methods for evalu-
� D2 ating time-dependent fluid dynamic forces on bodies, using only
𝛿P2x,GAMC =e2p h4 1 + 2 𝜔j
4rj velocity fields and their derivatives. J Fluids Struct 13:551–578
j
Polet DT, Rival DE, Weymouth GD (2015) Unsteady dynamics of rapid
� �2 (76) perching manoeuvres. J Fluid Mech 767:323–341
⎡� 2 ⎤
D Rival DE, van Oudheusden B (2017) Load-estimation techniques for
+ h4 ⎢ 1 − 2 y2j 𝛿𝜔2j ⎥.
⎢ j 4rj ⎥ unsteady incompressible flows. Exp Fluids 58(3):20
⎣ ⎦ Saffman PG (1992) Vortex dynamics. Cambridge University Press,
Cambridge
Scarano F, Riethmuller ML (2000) Advances in iterative multigrid PIV
Comparison of Eqs. (64) and (76) affirms the earlier asser- image processing. Exp Fluids 29(Suppl 1):S051–S060
Thom A (1929) An investigation of fluid flow in two dimensions. HM
tion that the GAMC formulation is less sensitive to random Stationery Office, Richmond
error in the PIV data but more sensitive to cylinder position Thom A (1933) The flow past circular cylinders at low speeds.
error, with both conclusions resulting from the inequalities Proc R Soc Lond Ser A Contain Pap Math Phys Character
in (68). 141(845):651–669
Van Oudheusden BW, Scarano F, Casimiri EWF (2006) Non-intru-
sive load characterization of an airfoil using PIV. Exp Fluids
40(6):988–992. https​://doi.org/10.1007/s0034​8-006-0149-2
References Wang C, Eldredge JD (2013) Low-order phenomenological modeling
of leading-edge vortex formation. Theor Comput Fluid Dyn
Achenbach E (1968) Distribution of local pressure and skin friction 27:577–598
around a circular cylinder in cross-flow up to Re = 5 × 106 . J Westerweel J, Scarano F (2005) Universal outlier detection for PIV
Fluid Mech 34(4):625–639 data. Exp Fluids 39(6):1096–1100
Albrecht T, del Campo V, Weier T, Metzkes H, Stiller J (2013) Deriv- Wieneke B (2015) PIV uncertainty quantification from correlation sta-
ing forces from 2D velocity field measurements. Eur Phys J Spec tistics. Meas Sci Technol 26:074002
Top 220(1):91–100 Wu JC (1981) Theory for aerodynamic force and moment in viscous
DeVoria AC, Carr ZR, Ringuette MJ (2014) On calculating forces from flows. AIAA J 19(4):432–441
the flow field with application to experimental volume data. J Wu JZ, Ma HY, Zhou MD (2015) Vortical flows. Springer, Berlin
Fluid Mech 749:297–319 Xia X, Mohseni K (2013) Lift evaluation of a two-dimensional pitching
Driscoll TA, Trefethen LN (2002) Schwarz–Christoffel mapping, vol flat plate. Phys Fluids 25(9):091901
8. Cambridge University Press, Cambridge
Eldredge JD (2007) Numerical simulation of the fluid dynamics of Publisher’s Note Springer Nature remains neutral with regard to
2D rigid body motion with the vortex particle method. J Comput jurisdictional claims in published maps and institutional affiliations.
Phys 221(2):626–648
Fackrell S (2011) Study of the added mass of cylinders and spheres.
Ph.D. thesis, University of Windsor

13

You might also like