See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/307620499

FFT integration of instantaneous 3D pressure gradient fields measured by

Lagrangian particle tracking in turbulent flows

Article in Experiments in Fluids · September 2016

DOI: 10.1007/s00348-016-2236-3

CITATIONS READS

33 328

4 authors:

Florian Huhn Daniel Schanz

Rapp OptoElectronic GmbH German Aerospace Center (DLR)
59 PUBLICATIONS 1,236 CITATIONS 251 PUBLICATIONS 2,282 CITATIONS

SEE PROFILE SEE PROFILE

Sebastian Gesemann Andreas Schröder

German Aerospace Center (DLR) German Aerospace Center (DLR)
90 PUBLICATIONS 1,421 CITATIONS 581 PUBLICATIONS 3,946 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Florian Huhn on 22 September 2017.

The user has requested enhancement of the downloaded file.

 Noname manuscript No.
(will be inserted by the editor)

1 FFT integration of instantaneous 3D pressure

2 gradient fields measured by Lagrangian particle
3 tracking in turbulent flows

4 F. Huhn · D. Schanz · S. Gesemann ·

5 A. Schröder

6
7 Received: date / Accepted: date

8 Abstract Pressure gradient fields in unsteady flows can be estimated through

9 flow measurements of the material acceleration in the fluid and the assumption
10 of the governing momentum equation. In order to derive pressure from its
11 gradient, almost exclusively two numerical methods have been used to spatially
12 integrate the pressure gradient until now: first, direct path integration in the
13 spatial domain, and second, the solution of the Poisson equation for pressure.
14 Instead, we propose an alternative third method that integrates the pres-
15 sure gradient field in Fourier space. Using a FFT function, the method is fast
16 and easy to implement in programming languages for scientific computing. We
17 demonstrate the accuracy of the integration scheme on a synthetic pressure
18 field and apply it to an experimental example based on time-resolved material
19 acceleration data from high resolution Lagrangian particle tracking with the
20 Shake-The-Box method.
21 Keywords Lagrangian particle tracking · Pressure reconstruction · Shake-
22 The-Box

23 1 Introduction

24 The pressure field in a turbulent flow reveals important coherent structures,

25 e.g., vortex cores, and it is most relevant to determine the loads exerted by the
26 fluid on a structure in the flow, e.g., a wing. Yet, spatially well resolved pres-
27 sure fields P (x, t) in three-dimensional unsteady flows are difficult to measure.
28 Optical non-intrusive flow measurement techniques, in particular the novel
29 Lagrangian particle tracking approach Shake-The-Box (Schanz et al., 2013a,
30 2016), provide accurate and dense fields of the material acceleration in the

F. Huhn
Department of Experimental Methods, German Aerospace Center (DLR), Institute of Aero-
dynamics and Flow Technology, Bunsenstrasse 10, D-37073 Göttingen, Germany
E-mail: [email protected]
 2 F. Huhn et al.

31 fluid. Since, for the particle tracking approach, the material acceleration data
32 is given at scattered particle positions, it is interpolated to a regular grid using
33 physical regularizations of the interpolant function, as e.g., zero curl for the
34 material acceleration field. The pressure gradient field ∇P (x, t) is then derived
35 from the Navier-Stokes equation for incompressible isothermal fluid and it is
36 integrated in space. For this spatial integration, almost exclusively two meth-
37 ods have been used in the aerodynamic and hydrodynamic flow measurement
38 community (van Oudheusden, 2013): first, the direct numerical integration of
39 ∇P along various paths in space with some spatial marching scheme based on
40 finite differences,
P (x + ∆x) = P (x) + ∇P ∆x, (1)
41 and second, the numerical solution of the Poisson equation,

∆P = ∇ · (∇P ) (2)

42 obtained by applying the divergence operator to the experimental pressure

43 gradient field. The necessary Dirichlet or Neumann boundary conditions for
44 (2) are determined from known flow conditions at the boundary or directly
45 from the measured data set. Examples in which the first method was used
46 include the experimental studies by Liu & Katz (2006); Dabiri et al. (2014);
47 Tronchin et al. (2015); the second method has been applied, e.g., by Ghaemi
48 et al. (2012); Ragni et al. (2012); de Kat & van Oudheusden (2012); Novara &
49 Scarano (2013); Pröbsting et al. (2013); Neeteson & Rival (2015). Charonko
50 et al. (2010) compare both methods for pressure reconstruction from 2D PIV
51 measurements.
52 Certainly, the problem of reconstructing a two- or three-dimensional scalar
53 field from its spatial gradient field, that is given as a discrete data set, is not
54 limited to pressure reconstruction from experimental flow data, but has also
55 been adressed in several other scientific fields. In computer vision, Frankot
56 & Chellappa (1988) were the first to propose a fast method to integrate the
57 surface gradient of a three-dimensional body, obtained from images with the
58 ’shape from shading’ visualization technique. Their method enforces integra-
59 bility and integrates the gradient field in Fourier space. Independently, Zhang
60 (1996) introduces the very same method to reconstruct water surface eleva-
61 tions from optical surface slope measurements. He additionally considers the
62 treatment of boundary effects in non-periodic domains. Both works consider
63 two-dimensional problems, but the extension of the method to three dimen-
64 sions is straightforward, as will be shown in the following.
65 This non-iterative integration technique in Fourier space is common in
66 numerical approaches (e.g. Laizet & Lamballais, 2009), but has surprisingly not
67 been considered for pressure reconstruction in the aerodynamics community,
68 so far. To the best knowledge of the authors, exclusively one group in medical
69 imaging applies this fast integration technique to fluid flows (Wang & Amini,
70 2005; Negahdar et al., 2013). They determine pressure gradient fields in blood
71 flows, experimentally measured with magnetic resonance imaging, in order
72 to obtain the pressure load on blood vessels. As a matter of low Reynolds
 Title Suppressed Due to Excessive Length 3

73 numbers, the spatial complexity of the investigated blood flow is rather low,
74 while we will consider multi-scale turbulent flows with a broad spectrum of
75 spatial scales here.
76 The aim of this paper is to present the efficient integration method for spa-
77 tial gradient fields and to show its applicability to 3D time-resolved pressure
78 fields in turbulent flows. The gradient field data we use is based on accu-
79 rate material acceleration fields that have become available recently through
80 the advent of the high-density Lagrangian particle tracking (LPT) algorithm
81 Shake-The-Box (Schanz et al., 2016) in combination with the interpolation
82 scheme FlowFit (Gesemann , 2016).
83 The paper is organized as follows. In Sec. 2, we recall the principles of
84 pressure reconstruction from flow measurements and present the integration
85 method of the pressure gradient. In Sec. 3, we show examples of pressure re-
86 construction for a synthetic pressure field to validate the integration method,
87 and in Sec. 4, we reconstruct the pressure field in an experimental turbulent
88 jet flow. Finally, Sec. 5 summarizes the results.

89 2 Data and Methods

90 2.1 Momentum equation

91 The momentum ρu of a fluid parcel in an isothermal viscid Newtonian fluid

92 evolves according to the Navier-Stokes equation
 
Du ∂u
ρ =ρ + (u · ∇)u = −∇P + µ∆u (3)
Dt ∂t
93 with constant density ρ, constant viscosity µ, and the material acceleration
94 Du/Dt, i.e., the acceleration of a fluid element along its trajectory. Away
95 from boundaries and strong shear flows, the pressure gradient force typically
96 dominates the momentum balance,
∇P  µ∆u, (4)
97 such that for flows with high Reynolds number, the viscous term is relatively
98 small and we get a good estimate of the pressure gradient field by
Du
∇P ≈ −ρ = −ρ a, (5)
Dt
99 where we denote the material (or Lagrangian) acceleration with a from here
100 on. In flows with low Reynolds number or in well resolved measurements of
101 boundary layers, the viscous term can be of the same order as the pressure gra-
102 dient and has to be taken into account in these cases, for example by a direct
103 computation of the viscous term from highly resolved velocity data. For tran-
104 sonic steady flows, van Oudheusden et al. (2007) describe a way to also include
105 compressibility effects into the pressure reconstruction. For an overview over
106 several formulations of the momentum equation for pressure reconstruction in
107 different flow conditions, see the review paper by van Oudheusden (2013).
 4 F. Huhn et al.

108 2.2 Measurement of Lagrangian acceleration

109 According to the term in brackets in (3), material acceleration could simply be
110 obtained from time-resolved velocity fields as measured with particle image ve-
111 locimetry (PIV). PIV is a robust method to obtain time-resolved velocity fields
112 from pairs or sequences of particle images. However, this indirect composed
113 measurement of material acceleration involves spatial and temporal deriva-
114 tives of the velocity field (3). The derivatives enhance measurement noise in
115 the velocity field, which can lead to noisy acceleration fields, and consequently
116 to noisy pressure gradient fields. Additionally, the spatial smoothing effect of
117 the correlation window in the PIV technique may lead to an underestimation
118 of the velocity gradient and to a bias of the derived material acceleration.
119 Lagrangian particle tracking (LPT) overcomes these problems by tracking
120 individual fluid tracers, such that entire time-resolved trajectories of single
121 particles xp (x0 , t0 , t) are known. While the well-known denotation Particle
122 Tracking Velocimetry (PTV) suggests the exclusive measurement of the fluid
123 velocity, with LPT, the acceleration can additionally be determined as a point
124 measurement at the position of individual tracers. Acceleration at particle
125 positions is obtained as a(xp (t)) = d2 xp (t)/dt2 . The main source of error
126 influencing the acceleration value is the uncertainty of the measured particle
127 position, while no spatial smoothing is involved in the image processing.
128 LPT has long been used in statistical turbulence research at relatively low
129 seeding densities (see e.g., Hoyer et al., 2005; Xu et al., 2007). Higher res-
130 olution has been reached by applying LPT to tomographic PIV data as a
131 post-processing step in order to increase the accuracy of acceleration mea-
132 surements (Schröder et al., 2011) and to reconstruct pressure fields (Novara
133 & Scarano, 2013). Only recently, the processable seeding densities in the par-
134 ticle images of direct LPT, measured in values of particles per pixel (ppp),
135 increased substantially to values as large as 0.1 ppp (Schanz et al., 2014). As-
136 suming a camera with a 1 Mpixel sensor, 0.1 ppp ideally corresponds to 105
137 simultaneously tracked particles, or an unstructured grid of roughly 50×50×50
138 positions with accurate acceleration data. This spatially resolved acceleration
139 field allows for a LPT-based pressure reconstruction in unsteady turbulent
140 flows.
141 Here, we use the LPT algorithm Shake-The-Box by Schanz et al. (2016).
142 The strength of this algorithm is to use the temporal information of already
143 established particle trajectories to predict the three-dimensional particle posi-
144 tion in the next time step and identify the corresponding intensity peak in the
145 camera images. The predicted particle position is then reprojected to the cam-
146 eras using an estimated optical transfer function (OTF) (Schanz et al., 2013b)
147 and the particle position is iteratively corrected to optimally fit the intensity
148 peaks in the camera images. Given the optimal particle position, the intensity
149 signature of the particle in the camera images is subtracted from the images
150 (iterative particle reconstruction, IPR (Wieneke, 2013)), in order to reduce
151 the complexity of the particle positioning problem. This procedure is crucial
152 to handle high seeding densities. The discrete particle positions of an identi-
 Title Suppressed Due to Excessive Length 5

153 fied trajectory are fit by an optimal one-dimensional cubic B-spline curve, that
154 takes the expected measurement noise into account. Temporal differentiation
155 yields velocity and acceleration along the trajectory.
156 In a subsequent step, the acceleration data is interpolated to a Cartesian
157 grid with the FlowFit algorithm (Gesemann , 2016). Smoothing B-splines are
158 fit iteratively to the scattered acceleration data. The B-splines are defined on a
159 fine auxiliary grid with a selectable mean resolution given in particles per cell
160 (ppc) with typical values of 0.1 ppc. The interpolant is smoothed by penalizing
161 high frequencies in cells containing particles and in empty cells. Additionally,
162 the curl of the acceleration field is penalized according to assumption (4) and
163 tends to zero which supports integrability. Finally, the interpolant is evalu-
164 ated on the output grid with a user defined spatial resolution. A high spatial
165 sampling resolution is beneficial for the Fourier transform in the pressure in-
166 tegration scheme below.

167 2.3 Integration scheme for pressure reconstruction

168 Following Frankot & Chellappa (1988) or Laizet & Lamballais (2009), we ob-
169 tain the pressure field P (x) by integrating the measured three-dimensional
170 pressure gradient fields
∂x P = −ρ ax (6)
∂y P = −ρ ay (7)
∂z P = −ρ az (8)
171 in Fourier space

k · ∇P
g kx ∂g
x P + ky ∂y P + kz ∂z P
g g
P̃ (k) = = , (9)
i|k|2 i(kx2 + ky2 + kz2 )
172 and transforming back to normal space
P 0 (x) = F T −1 P̃ (k). (10)
173 For a short derivation of Eq. (9), see Ref. (Rocholz, 2008). The tilde denotes a
174 Fourier transformed function, e.g., P̃x = F T (Px ), F T −1 is the inverse Fourier
175 transform, and kx , ky , kz are the components of the wave number vector k.
176 In (9), the separation of the curl-free longitudinal component of the vector
177 field corresponds to a projection of the pressure gradient onto the k-vector
178
g ]/|k|2 , and the integration in space corresponds to a division by ik.
k[k · ∇P
179 Equation (9) has a singularity at k = 0. In order to handle this, the amplitude
180 for the constant component is set to zero, P̃ (k = 0) = 0. By this operation,
181 the amplitudes of the constant component of the three pressure gradients is
182 lost, e.g., ∂g
x P (k = 0) = 0, but they can be reconstructed by adding linear
183 planes to the pressure field (Zhang, 1996)
P (x) = P 0 (x) + h∂x P ix + h∂y P iy + h∂z P iz, (11)
 6 F. Huhn et al.

184 where h·i is the spatial mean over the entire domain. Below, we will see that in
185 practice we set boundary conditions at the periodic domain that account for
186 the global linear pressure gradient. Due to small imperfections of the data in
187 Fourier space (truncation errors etc.), the back transform in Eq. (10) generates
188 a small imaginary part for the resulting pressure field. We only consider the
189 real part and neglect the imaginary part. Finally, in order to obtain absolute
190 pressure, the integration constant, a constant pressure offset P0 , obtained from
191 additional measurements or from theoretical considerations at the boundaries,
192 is added to the relative pressure field P (x).
193 By construction, the integration scheme (9) satisfies the integrability con-
194 straint (Frankot & Chellappa, 1988; Zhang, 1996). The longitudinal component
195 of a vector field is curl-free, i.e., partial derivatives commute, e.g.,

∂x ∂z P (x) = ∂z ∂x P (x). (12)

196 This is equivalent to the requirement that two different integration paths be-
197 tween two points in space yield the same result which imposes a smoothness
198 constraint on the integrated field. In the presence of measurement noise, inte-
199 grability is not necessarily fulfilled. For a direct path integration of the pressure
200 gradient field, an alternative integration method (van Oudheusden, 2013), the
201 average over many paths is usually computed (using the mean or the median
202 (Dabiri et al. , 2014)) in order to reduce the bias of not curl-free measurement
203 noise. Instead, when using the FFT integration or the Poisson equation, the
204 pressure gradient fields that are being integrated are curl-free by construction.
205 In order to put the integration scheme (9) into in a more general perspec-
206 tive, we can look at the complete Navier-Stokes equation (3) in terms of the
207 Helmholtz decomposition. According to the Helmholtz theorem, an arbitrary
208 vector field a0 decaying to zero at infinity can be decomposed into two com-
209 ponents, one of which is divergence-free and the other component is curl-free
210 (see, e.g., Griffiths , 1999),

a0 = −∇Φ + ∇ × A. (13)

211 The curl-free component can be written as the gradient of the scalar potential
212 Φ and the divergence-free component can be written as the curl of the vector
213 potential A. In the Navier-Stokes equation (3), the pressure gradient term is
214 curl-free and the viscous term is divergence-free due to incompressibility, i.e.,
215 we can decompose the measured field of inertial forces ρa into the pressure
216 force and the viscous force by Helmholtz decomposition,

a0 = ρa = −∇P + µ∆u = −∇Φ + ∇ × A (14)

217 with Φ = P and µ∆u = ∇ × A. While the scalar pressure field P is given by
218 (9), similarly, the vector potential is given by
" #
e0
−1 k × a
A = −F T (15)
i|k|2
 Title Suppressed Due to Excessive Length 7

Fig. 1 Sketch of the different modified boundary conditions for the non-periodic pressure
field P and pressure gradient field ∂x P (two periods of an arbitrary function are shown
here for demonstration purposes). (a) Two periods of discontinuous data with period L = 1.
(b) Continuous extension of ∂x P (blue) by mirroring, CE1, (Zhang, 1996). (c) Continuous
extension of P (black) by mirroring, CE2, (Wang & Amini, 2005). (d) Representation of the
discontinuity in pressure with an impulsive pressure gradient obtained from path integration,
PI.

219 where the cross-multiplication with the k-vector separates the divergence-free
220 (transversal) component of a0 . In flows where the viscous force is of the same
221 order as the pressure gradient force, this decomposition is a direct way to
222 seperate the viscous force and to determine the contribution of the different
223 forces. In more turbulent flows, the viscous force is typically small and the
224 divergence-free component of the material acceleration is dominated by mea-
225 surement noise. Therfore, we can use the concept of Helmholtz decomposition
226 (13) to construct synthetic pressure gradient fields with non-zero curl compo-
227 nent.

228 2.3.1 Boundary conditions for periodic domain

229 The Fourier transform assumes a periodic domain for the transformed fields,
230 a condition that is typically not met by measurement data. A continuous
 8 F. Huhn et al.

231 extension of the field by mirroring the data has been proposed to minimize
232 boundary artifacts due to non-periodicity of the data (Zhang, 1996; Wang &
233 Amini, 2005). In the following, we discuss this approach and propose a new
234 treatment of the boundaries that suppresses boundary artifacts while avoiding
235 the need for excessive memory. The different boundary conditions discussed
236 here are depicted in Fig. 1.

237 Continuous extension (CE) A periodic continuous extension avoids disconti-

238 nuities at the boundaries of the field subject to the FT by reflecting the field
239 about mirror planes along the boundaries. In (9) and (10), the Fourier trans-
240 form acts on the pressure gradient field and on the pressure field. A crucial
241 point is whether to construct continuous boundaries for ∇P or for P . Zhang
242 (1996) proposes to continuously extend ∇P by reflecting it about the axes of
243 the coordinate system such that

∇P (x, y) = ∇P (−x, y) = ∇P (x, −y) = ∇P (−x, −y). (16)

244 We denote this choice of boundary conditions with a continuous extension

245 of the gradient field as CE1. If this continuous gradient field is integrated
246 in space, however, the resulting pressure field is not necessarily continuous
247 at the boundaries (Fig. 1b). Wang & Amini (2005) enforce continuity of the
248 resulting pressure field P instead (Fig. 1c), by extending the gradient field in
249 a discontinuous way as

∂x P (x, y) = −∂x P (−x, y)

= +∂x P (x, −y) = −∂x P (−x, −y) (17)
∂y P (x, y) = +∂y P (−x, y)
= −∂y P (x, −y) = −∂y P (−x, −y), (18)

250 where the only difference to (16) is in the signs. These boundary conditions
251 are denoted as CE2. In numerical experiments with a synthetic multi-scale
252 pressure field (cf. Fig. 2), we find that the CE2 approach produces smaller
253 residuals for the reconstructed pressure field than the CE1 approach.

254 Boundary gradients from direct path integration (PI) A disadvantage of the
255 periodic continuous extension of the pressure field is the obvious increase of
256 the domain size and the corresponding increase of required memory space by
257 a factor of 8 in three dimensions due to the mirrored fields. Another way to
258 match the boundary of the non-periodic pressure field to the periodic FFT
259 computations is to consistently represent the high gradient of the pressure
260 discontinuity in the gradient field (Fig. 1d). Assume a periodically extended
261 pressure field with a discontinuity at the boundaries. The gradient of this
262 field must have (infinitely) high values at the boundary, which are missing
263 in the measured gradient data. The discontinuity in the pressure field can
264 be accounted for by introducing an impulse at the boundary of the gradient
265 field. The magnitude of the impulse is obtained by a direct path integration
 Title Suppressed Due to Excessive Length 9

266 from boundary to boundary in the gradient field. For example, along a one-
267 dimensional line in x-direction in the gradient field ∂x P with n sample points,
268 the value of the boundary impulse can be expressed as
n−1
1X
∂x P (1) = ∂x P (n) = − ∂x P (i) (19)
2 i=2
269 The two boundary values in the measured ∂x P data is then replaced by the
270 values from (19). ∂y P and ∂z P are corrected in the same way. This modification
271 of the boundaries is denoted as PI boundary conditions. After this correction
272 of the boundaries, the gradient field linewise satisfies
Z L
∂x P (x)dx = 0, (20)
0
273 also in y-direction and z-direction, which is equivalent with P being periodic
274 with period L. The continuous extension approach CE2 also satisfies (20),
275 while the approach CE1 does not.

276 2.3.2 Steps of the pressure reconstruction algorithm

277 Including the LPT measurement, we summarize the pressure reconstruction

278 with the following steps:
279 1. Reconstruct particle trajectories from time-resolved particle images of at
280 least three cameras using the STB algorithm (Schanz et al., 2016).
281 2. Fit an interpolating continuous function consisting of 1D cubic B-splines
282 to the trajectories and differentiate twice w.r.t. time to obtain the material
283 acceleration at particle positions (TrackFit, (Gesemann , 2016)).
284 3. Interpolate the material acceleration, given on an unstructured grid, to a
285 fine Cartesian grid, using 3D cubic B-splines and penalizing the curl of the
286 acceleration field (FlowFit, (Gesemann , 2016)).
287 4. Neglect the viscous term in (3), assume constant density ρ and compute
288 ∇P (5).
289 5. Modify ∇P by applying the CE2 or PI method for periodic spatial bound-
290 ary conditions from Sec. 2.3.1, in order to avoid boundary artifacts.
291 6. Compute the Fourier transforms of ∇P , evaluate (9), set P̃ (k = 0) = 0,
292 and transform back, using a Fast-Fourier-Transform (FFT) function.
293 7. Add a constant pressure offset P0 to obtain absolute pressure.
294 Regardless of the exact way to obtain the pressure gradient on a Cartesian
295 grid, be it a different measurement technique or the additional consideration
296 of theoretical models, e.g., a compressible flow with varying density, the inte-
297 gration method is anyhow applicable, starting from step 5.
298 Once we have the pressure field, ground truth data is not always available to
299 compare our experimental pressure results with. Yet, the random measurement
300 noise that is uncorrelated in time can be filtered out in the temporal domain
301 with a low-pass filter. Certainly, various sources of error are correlated in time
302 or even constant, such that a validation against other measurements or models
303 is in general highly desirable.
 10 F. Huhn et al.

Fig. 2 (a) Synthetic three-dimensional pressure field (central plane shown) constructed as
correlated noise. (b-d) Difference fields between reconstructed pressure field and ground
truth data for different boundary conditions. (a,b) and (c,d) share the same colorbar respec-
tively. (e,f) Relative error along a section (dashed line) for CE2 and PI boundary conditions.

304 3 Synthetic pressure field

305 3.1 Curl-free pressure gradient

306 In order to validate the accuracy of the integration method, we create a three-
307 dimensional synthetic pressure field on a cubic 257 × 257 × 257 domain as
308 ground truth data, compute the spatial gradient and reconstruct the pressure
309 field again. A direct comparison of the original and the reconstructed pressure
310 field reveals the accuracy of the integration. In order to generate the synthetic
311 pressure field, uniformly distributed noise in Fourier space is correlated with
 Title Suppressed Due to Excessive Length 11

312 a normalized correlation kernel of the form

f (k) = exp(−α|k|1/2 ). (21)

313 The slow decay of the kernel due to the exponent 1/2 generates a wide range
314 of scales, mimicking turbulent pressure fields with large scale gradients as well
315 as small scale structures related to vortices. The scale coefficient α controls
316 the spatial scale of the synthetic pressure field; by increasing α, the kernel de-
317 cays faster and small scales are damped. The gradient of the periodic pressure
318 field is computed in Fourier space by multiplying with ik, then, a non-periodic
319 100 × 100 × 100 subdomain is cropped in real space for the test. Figure 2a
320 shows the cropped pressure field at the central z-plane. Pressure is recon-
321 structed from the gradient fields, with and without accounting for periodic
322 boundary conditions. With unchanged boundaries, the reconstructed pressure
323 field significantly deviates from the original field (not shown). In the difference
324 field, small boundary artifacts and large scale gradients up to 100% relative
325 error appear. When the gradient field is continuously extended according to
326 CE1 (16), the deviation is still large. Figure 2b shows the difference between
327 original and reconstructed pressure field. In contrast, when the boundaries
328 are modified according to the CE2 method (17) or the PI method (19), the
329 reconstructed pressure field agrees with the original field within an error of
330 less than 1%. Figures 2c,d show the vanishing difference between original and
331 reconstructed pressure field. Profiles along a section in x-direction show the
332 small relative error (Fig. 2c,d). The CE2 boundary conditions have the advan-
333 tage of smaller boundary artifacts than the PI method, to the cost of a 8 times
334 larger mirrored volume that has to be integrated. On a 3.4 GHz standard PC,
335 constructing PI boundary conditions and integrating pressure takes 0.5 s on
336 1003 grid points, and 3.3 s for a 2003 grid.

337 3.2 Pressure gradient with added non-zero curl component

338 Spatial fields of the material acceleration that are obtained from experiments
339 typically have a non-zero curl which may be due to a significant contribution of
340 the viscous term or due to measurement noise. In the data processing scheme
341 presented in Sec. 2.3.2, we penalize the curl in the FlowFit interpolation scheme
342 and therefore the FFT integration starts with a gradient field with vanishing
343 curl. In general, however, it is interesting to quantify the sensitivity of the
344 integration scheme to gradient fields with non-zero curl. Therefore, a pressure
345 gradient field with a non-zero curl component is constructed as

anzc = ∇P (α) + β 0 [∇ × A(α)] . (22)

346 Both, the synthetic pressure field P (α) and the vector potential A(α), are con-
347 structed as correlated noise as described above. The size of the spatial scales
348 of the potential fields can be controlled via the parameter α in the correlation
349 kernel, see Eq. (21). The amplitude of the added curl-component is given by
 12 F. Huhn et al.

350 the parameter β 0 = β std(∇P )/std(∇ × A), such that for β = 0.2 the signal-
351 to-signal ratio between the added curl-component ∇ × A and the curl-free
352 component ∇P is 20%. Figure 3a illustrates the synthetic pressure gradient
353 field and Fig. 3b shows the added curl-component with amplitude β = 0.2.
354 Both field have the same spatial spectrum with α = 6.7. The pressure field
355 is reconstructed as described in Sec. 2.3.2 using CE2 and PI boundary condi-
356 tions and the difference to the original pressure field is computed. Figure 3c
357 and 3d show the relative error for the reconstruction. In principle, the curl-
358 component should have no effect on the reconstruction, since the pressure gra-
359 dient field anzc is projected onto the longitudinal direction in Fourier space,
360 such that only the curl-free component is integrated. However, the applied
361 periodic boundary conditions are adapted for a curl-free field and therefore
362 generate boundary artifacts when the curl is non-zero. In the interior of the
363 volume (dashed square), the reconstruction error is within a few percent for
364 both boundary conditions (CE2 and PI). In the boundary region, however,
365 artifacts become large with relative errors up to 20% for CE2 boundary con-
366 ditions (Fig. 3c). For PI boundary conditions, the error exceeds 20% (Fig. 3d).
367 The effect of non-zero curl on the pressure reconstruction with the proposed
368 FFT scheme depends on the amplitude of the curl-component (controlled with
369 parameter β) and on its spatial spectrum (controlled with parameter α). The
370 amplitude β is varied and the relative error is quantified in the interior and
371 exterior domain defined in Fig. 3 (dashed-line). Figure 4 shows the rms value
372 of the relative error for α = 6.7, i.e., the curl-component has the same spa-
373 tial spectrum as the pressure gradient field and varying amplitude β. The
374 CE2 boundary condition performs better than the PI boundary condition, as
375 already illustrated in Fig. 3c,d. In the interior of the domain, the reconstruc-
376 tion error of both methods stays below 8% for a maximal amplitude of the
377 curl-component of β = 20%.
378 The error induced by imperfect boundary conditions is largely reduced, if
379 the spectrum of the added curl-component is shifted towards smaller spatial
380 scales, as one would expect for spatially uncorrelated measurement noise. Fig-
381 ure 4b shows the rms relative error for α = 3.3. The error does not exceed 1%
382 in the interior, and 5% at the boundaries for the tested values of the ampli-
383 tude β. This result indicates that, in the tested range of values, the proposed
384 FFT integration scheme is accurate within few percent for pressure gradient
385 fields with a curl-component of moderate amplitude and with small dominant
386 spatial scales.

387 3.3 Pressure gradient with internal boundaries

388 In a third case, we test the integration scheme on a synthetic pressure field
389 around a body. This adds additional boundaries in the interior of the domain.
390 For the determination of loads on a mechanical structure in the flow, as a wing,
391 the flow field is typically measured in a volume around the structure, in order
392 to reconstruct the pressure field and integrate pressure over the surface of the
 Title Suppressed Due to Excessive Length 13

Fig. 3 Synthetic pressure gradient with non-zero curl. (a) Curl-free pressure gradient, (b)
added curl-component with amplitude β = 0.2. Both fields have the same spatial spectrum
with α = 6.7. (c) Relative error (Pr − P )/rms(P ) in the pressure reconstruction with CE2
boundary conditions and (d) PI boundary conditions. The dashed square splits the domain
in interior and exterior domain (boundary with 10% width).

393 structure (Tronchin et al., 2015; Ragni et al., 2012). Again, we generate a syn-
394 thetic periodic pressure field, compute the gradient in Fourier space, and crop
395 a non-periodic subset. In the center of the gradient field, an empty spherical
396 region is defined, simulating a sphere in the pressure field which adds internal
397 boundaries (Fig. 5a). Integration in Fourier space, only working on rectangular
398 domains, may seem unapt to integrate the pressure field bounded by curved
399 surfaces. For example, a pressure reconstruction that is directly computed on
400 a surface following unstructured grid defined by particle positions has recently
401 been proposed by Neeteson & Rival (2015). However, in our pressure gradi-
402 ent field, the empty region can be replenished with an interpolation by the
403 FlowFit routine, such that the FFT integration can still be applied. Using a
404 random subset of 5% of the grid points as input data for the FlowFit (Fig. 5b),
405 we interpolate the empty space with a smooth function (Fig. 5c). In the opti-
406 mization, the curl of the field is penalized in the cost function, such that the
407 result is nearly curl-free. In this way, no explicit boundary conditions have to
408 be given at the curved interior boundary. Strictly speaking, the exact micro-
409 scopic boundary conditions on the surface of the wall S are ∇P |S = 0, the
410 pressure gradient vanishes, since velocity and acceleration are identical zero at
411 the wall. Here, however, we assume that the measurement point closest to the
412 wall still is far from the wall in the microscopic sense and therefore experiences
413 a non-zero pressure gradient.
 14 F. Huhn et al.

Fig. 4 Relative error of pressure reconstruction in presence of non-zero curl. (a) Curl-
free component and curl-component of the synthetic pressure gradient field have the same
spatial spectrum with α = 6.7 (cf. Fig. 3a). (b) The added curl-component has smaller scales
(α = 3.3) than the curl-free component (α = 6.7).

414 Using PI boundary conditions at the exterior boundaries, the reconstructed

415 pressure field closely matches the original pressure field. In Fig. 6, the pressure
416 field is interpolated onto the surface of the sphere, and by visual inspection no
417 difference can be seen between the two fields. The relative error is in the order
418 of few percent. Figure 6c shows the relative error along an arbitrary circle on
419 the sphere. Note that, strictly speaking, the pressure values on the surface of
420 the sphere are extrapolations of the pressure field, since no gradient data is
421 available exactly on the surface. When the load on the sphere is computed by
422 integrating pressure over the entire surface of the sphere, the values between
423 original and reconstructed pressure field differ by 2.1% for the field shown.

424 4 Experimental pressure field

425 Our integration method for the pressure gradient field is now applied to an
426 experimental data set. In the experiment, a transitional water jet emanates
 Title Suppressed Due to Excessive Length 15

Fig. 5 (a) Gradient of synthetic pressure field ∂x P with internal boundaries around a
sphere, central xy-plane. (b) Random 5% subset of the data points shown in a slice of the
volume. (c) Curl-free interpolation in the spherical region and in an external buffer zone,
central xy-plane.

Fig. 6 (a) Original

p and (b) reconstructed pressure field on the spherical body. (c) Relative
error (Pr − P )/ hP 2 i along a circle on the sphere (dashed line).

427 in vertical upward direction from a circular nozzle with diameter d = 10 mm

428 at Re = 5,000. Close to the nozzle, vortex rings form, grow, destabilize and
429 decay further downstream. The LPT images are taken with three cameras at
430 a frequency of 1kHz. About 13,000 particles are tracked instantaneously by
431 the Shake-The-Box method in a volume of roughly 28 cm3 , corresponding to a
432 mean particle distance of 1.34 mm. The experiment is described in more detail
433 in Ref. (Schanz et al., 2013b) and references therein.
434 For the pressure reconstruction, we follow the steps outlined in Sec. 2.3.2.
435 The output of the Shake-The-Box LPT algorithm provides time-resolved ac-
436 celeration fields at instantaneous particle positions (Step 2). Figure 7a shows
437 the distribution of tracer particles in a vertical section of the field of view
438 for a fixed time frame. The particles are color-coded with the vertical compo-
439 nent of acceleration ay with a maximal magnitude of ∼ 20 m/s2 . Alternating
440 regions of positive and negative acceleration can be seen that are associated
441 with a chain of vortex rings. In step 3, the acceleration field is interpolated
442 onto a Cartesian grid with the FlowFit routine (Fig. 7b). We choose a grid
443 of size 111 × 201 × 111 with a spatial resolution of ∆x = 0.3 mm, which is
 16 F. Huhn et al.

Fig. 7 (a) Particles color-coded with vertical acceleration ay [m/s2 ] in a slice z ∈

[−3, 3] mm. (b) Smooth acceleration field ay on a Cartesian grid after interpolation with
FlowFit with surrounding buffer zone, z = 0 mm. (c) Resulting pressure field P [Pa] at
z = 0 mm.

444 roughly 1/4 of the mean inter-particle distance in the experiment, such that
445 all details of the acceleration field are sampled. The domain of the Cartesian
446 grid is chosen in a way that it entirely encloses the particle containing mea-
447 surement volume (enclosed by dashed line). The interpolation function in the
448 interior extends smoothly to the exterior buffer zone and it is curl-free in the
449 whole domain. Small contributions of the viscous term to acceleration can be
450 suppressed in this way, since the viscous term is divergence-free (curl-only)
451 (14). In step 4, we set the density to ρ = 1000 kg/m3 and compute the pres-
452 sure gradient ∇P . In step 5, the boundaries are modified according to the PI
453 method, in order to account for periodic boundary conditions of the pressure
454 field, cf. Eq. (19). In step 6, the relative pressure field is obtained after an
455 integration in Fourier space. A section of the pressure field at the center plane
456 of the jet is shown in Figure 7c with the exterior region in dark blue. Each
457 vortex ring in the jet can be seen in the pressure field as a pair of circular low
458 pressure regions. In upward (downstream) direction, the low pressure region
459 inside consecutive vortex rings intensifies, indicating a growing rotation, since
460 the pressure gradient around the pressure minimum mainly balances centrifu-
461 gal forces in the rotating vortex. At the upper part of the domain, the vortex
462 rings break up into less regular structures with an increasingly disordered pat-
463 tern in the pressure field. This transition to turbulence can even better be
464 seen in a three-dimensional representation of the pressure field. Figure 8 shows
465 iso-surfaces of positive (red) and negative (blue) pressure regions that reveal
466 the ring structure of the vortices and the increasing diameter of the growing
467 vortex rings.
468 So far, relative pressure fields have been presented. These pressure fields
469 lack the addition of a spatially constant pressure offset, the integration con-
 Title Suppressed Due to Excessive Length 17

Fig. 8 3D pressure field (iso-surfaces) and central slice of the velocity field (vectors) of a
transitional jet. Iso-surfaces are shown for P = −5 Pa (blue) and P = +5 Pa (red).

470 stant of the spatial integration. In the present experiment, no data of pressure
471 boundary conditions is available and the measurement volume barely reaches
472 into regions of stationary flow where a pressure reference could be obtained
473 from Bernoulli’s equation (van Oudheusden, 2013). Consequently, we cannot
474 experimentally determine the pressure offset.
475 Nevertheless, each individual snapshot of the relative pressure fields is nor-
476 malized by its zero spatial mean by construction, hP ix,y,z = 0, and we can
477 address the question, whether such relative pressure fields are already smooth
478 in time, or if a time-dependent pressure offset has to be added to reach the ex-
479 pected smooth temporal evolution of the pressure field. Figure 8 represents the
480 temporal evolution of 50 consecutive relative pressure fields with ∆t = 1 ms in
481 a space-time plot. The pressure profiles along a vertical line (y-direction) in-
482 tersecting the vortex rings are plotted against time with contour levels (black)
483 drawn at [−5, 0, 5] Pa. Alternating low and high pressure regions, correspond-
484 ing to the vortex ring pattern in Fig. 8, can be seen and the slope displays
485 the propagation velocity of the vortex rings. This diagram indicates an overall
486 smooth evolution of the relative pressure fields, since the contour lines of the
487 large scale vortex rings evolve in time with small oscillations. For a strong
 18 F. Huhn et al.

Fig. 9 Temporal evolution of pressure field P (y, t) [Pa]. Space-time plot of pressure along
a vertical line intersecting the vortex rings.

488 time-dependent pressure offset, all contour lines would oscillate in phase. This
489 points to a largely time-independent pressure offset.

490 5 Summary

491 In this paper, we propose a new method for the spatial integration of the pres-
492 sure gradient field in turbulent flows obtained from experimental Lagrangian
493 particle tracking at high seeding density (Shake-The-Box ). The pressure gra-
494 dient field is integrated in Fourier space, simply including a Fourier transform
495 of the pressure gradient, a multiplication and a back transform, resulting in a
496 relative pressure field with zero mean. A single integration constant, the pres-
497 sure offset, has to be added to the relative pressure fields in a final step. Before
498 integration, the boundaries of raw data have to be modified to adjust to the
499 periodic boundary conditions inherent to the Fourier transform. We present
500 two choices of periodic boundary conditions that result in numerical errors of
501 less than 1% in the test cases with a curl-free pressure gradient.
502 Prior to integration, the FlowFit interpolation scheme yields a representa-
503 tion of the scattered experimental acceleration data on a Cartesian grid. Due to
504 a penalization, the FlowFit interpolation results in material acceleration fields
505 with vanishing curl. In the case of significant curl in the acceleration field,
506 the FFT integration scheme also works, since the field is projected onto the
507 longitudinal (divergent) component in Fourier space, and only this component
 Title Suppressed Due to Excessive Length 19

508 is spatially integrated. Yet, using the periodic boundary conditions presented
509 here for non-curl-free acceleration fields leads to artifacts at the boundaries.
510 These artifacts are reduced, if the curl-component has small spatial scales
511 compared to the curl-free component, as it is typical for spatially uncorrelated
512 measurement noise. Depending on the amplitude of the curl-component, the
513 integrated field is accurate within few percent in the interior of the domain.
514 We also show that using the FlowFit interpolation scheme facilitates the
515 treatment of internal boundary conditions, especially for curved interior bound-
516 aries, as e.g., the wall of a body inside the domain. The empty domain in the
517 body can be filled with a smooth curl-free function, such that the integration,
518 not limited to the FFT scheme, can be applied on a cubic domain ignoring
519 internal boundaries. Similarly, external boundaries can be embedded inside
520 the cubic domain by extending the domain with a buffer zone. This provides
521 pressure data at the exact position of walls inside the measurement volume.

522 Acknowledgments

523 The authors would like to thank Roland Rocholz for drawing the authors’
524 attention to the presented integration scheme with his PhD thesis (Rocholz,
525 2008). The authors also thank Daniele Violato and Matteo Novara from TU
526 Delft for the experimental setup and for the collaboration on conducting the jet
527 experiment. Work including the experimental results has partly been funded
528 by the DFG-project Analyse turbulenter Grenzschichten mit Druckgradient bei
529 großen Reynoldszahlen mit hochauflösenden Vielkameramessverfahren (grant
530 KA 1808/14-1 and SCHR 1165/3-1).

531 References

532 Dabiri, J., Bose, S., Gemell, B.J., Colin, S.P., Costello, J.H.: An algorithm
533 to estimate unsteady and quasi-steady pressure fields from velocity field
534 measurements, J. Exp. Biol. 217, 331–336, 2014.
535 de Kat, R., van Oudheusden, B.W.: Instantaneous planar pressure determina-
536 tion from PIV in turbulent flow, Exp. Fluids 52:10891106, 2012.
537 Charonko, J.J., King C.V., Smith, B.L., Vlachos, P.P.: Assessment of pressure
538 field calculations from particle image velocimetry measurements, Meas. Sci.
539 Technol. 21, 105401, 2010.
540 Frankot, R.T., Chellappa, R.: A Method for Enforcing Integrability in Shape
541 from Shading Algorithms, IEEE Transactions on Pattern Analysis and Ma-
542 chine Inteligence 10:4, 1988.
543 Gesemann, S.: From particle tracks to velocity and acceleration fields using
544 B-splines and penalties, arXiv:1510.09034v1, 2016.
545 Ghaemi, S., Ragni, D., Scarano, F.: PIV-based pressure fluctuations in the
546 turbulent boundary layer, Exp. Fluids 53:18231840, 2012.
547 Griffiths, D.: Introduction to Electrodynamics, Prentice Hall, Upper Saddle
548 River, New Jersey, p. 555-557, 1999.
 20 F. Huhn et al.

549 Hoyer, K., Holzner, M., Lüthi, B., Guala, M., Liberzon, A., Kinzelbach, W.:
550 3D scanning particle tracking velocimetry, Exp. Fluids 39, 923–934, 2005.
551 Laizet, S., Lamballais, E.: High-order compact schemes for incompressible
552 flows: A simple and efficient method with quasi-spectral accuracy, J. Comp.
553 Phys. 228(16), 59896015, 2009.
554 Liu, X., Katz, J.: Instantaneous pressure and material acceleration measure-
555 ments using a four-exposure PIV system, Exp. Fluids 41, 227240, 2006.
556 Negahdar, M.J., Kadbi, Mo., Cha, J., Cebral, J., Amini, A.: Noninvasive 3D
557 Pressure Calculation From PC-MRI Via Non-Iterative Harmonics-based Or-
558 thogonal Projection: Constant Flow Experiment, 35th Ann. Int. Conf. IEEE
559 EMBS, Osaka, Japan, 2013.
560 Neeteson, N.J., Rival, D.E.: Pressure-field extraction on unstructured flow data
561 using a Voronoi-tesselation-based networking algorithm: a proof-of-principle
562 study, Exp. Fluids 56:44, 2015.
563 Novara, M., Scarano, F.: A particle-tracking approach for accurate material
564 derivative measurements with tomographic PIV, Exp. Fluids 54:1584, 2013.
565 van Oudheusden, B.W., Scarano, F., Roosenboom, E.W.M., Casimmiri,
566 E.W.F., Souverein, L.J.: Evaluation of integral forces and pressure fields
567 from planar velocimetry data for incompressible and compressible flows,
568 Exp. Fluids. 43, 153-162, 2007.
569 van Oudheusden, B.W.: PIV-based pressure measurement, Meas. Sci. Technol.
570 24, 032001, 2013.
571 Pröbsting, S., Scarano, F., Bernardini, M., Pirozolli, S.: On the estimation of
572 wall pressure coherence using time-resolved tomographic PIV, Exp. Fluids
573 54:1567, 2013.
574 Ragni, D., van Oudheusden, B.W., Scarano, F.: 3D pressure imaging of an air-
575 craft propeller blade-tip flow by phase-locked stereoscopic PIV, Exp. Fluids
576 52:463477, 2012.
577 Rocholz, R.: Spatio-Temporal Measurement of Short Wind-Driven Water
578 Waves, Dissertation, Ruperto-Carola University of Heidelberg, Germany,
579 2008.
580 Schanz, D., Schröder, A., Gesemann, S., Michaelis, D., Wieneke, B.: Shake
581 The Box: A highly efficient and accurate Tomographic Particle Tracking
582 Velocimetry (TOMO-PTV) method using prediction of particle positions,
583 10th Int. Symp. on Part. Imag. Vel., Delft, Netherlands, 2013a.
584 Schanz, D., Gesemann, S., Schröder, A., Wieneke, B., Novara, M.: Non-uniform
585 optical transfer functions in particle imaging: calibration and application to
586 tomographic reconstruction, Meas. Sci. Technol. 24, 024009, 2013b.
587 Schanz, D., Schröder, A., Gesemann, S.: Shake The Box - a 4D PTV algo-
588 rithm: Accurate and ghostless reconstruction of Lagrangian tracks in densely
589 seeded flows, 17th Int. Symp. Appl. Laser Tech. Fluid Mech., Lisbon, 2014.
590 Schanz, D., Gesemann, S., Schröder, A.: Shake-The-Box: Accurate Lagrangian
591 particle tracking at high particle densities, Exp. Fluids EXIF-D-15-00467,
592 accepted for publication, please complete, 2016.
593 Schröder, A., Geisler, R., Staack, K., Elsinga, G.E., Scarano, F., Wieneke, B,
594 Henning, A., Poelma, C., Westerweel, J.: Eulerian and Lagrangian views of
 Title Suppressed Due to Excessive Length 21

595 a turbulent boundary layer flow using time-resolved tomographic PIV, Exp.
596 Fluids 50, 1071-1091, 2011.
597 Tronchin, T., David, L., Farcy, A.: Loads and pressure evaluation of the flow
598 around a flapping wing from instantaneous 3D velocity measurements, Exp.
599 Fluids 56:7, 2015.
600 Wang, Y., Amini, A.: Integrable pressure gradients via harmonics-based or-
601 thogonal projection, in: Christensen, G.E., Sonka, M. (Eds.): Information
602 Processing in Medical Imaging 2005, LNCS 3565, pp. 431–442, Springer-
603 Verlag, Berlin Heidelberg, 2005.
604 Wieneke, B.: Iterative reconstruction of volumetric particle distribution, Meas.
605 Science Tech. 24, 024008, 2013.
606 Xu, H., Oullette, N., Vincenzi, D., Bodenschatz, E.: Acceleration Correlations
607 and Pressure Structure Functions in High-Reynolds Number Turbulence,
608 Phys. Rev. Lett. 99, 204501, 2007.
609 Zhang, X.: An Algorithm for calculating water surface elevations from surface
610 gradient data, Exp. Fluids 21, 43–48, 1996.

View publication stats