Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Chemical Engineering Science 279 (2023) 118881

Contents lists available at ScienceDirect

Chemical Engineering Science


journal homepage: www.elsevier.com/locate/ces

Decomposition of power number in a stirred tank and real time


reconstruction of 3D large-scale flow structures from sparse pressure
measurements
Kirill Mikhaylov a , Stelios Rigopoulos b , George Papadakis a,∗
a
Department of Aeronautics, Imperial College London, London, SW7 2AZ, UK
b
Department of Mechanical Engineering, Imperial College London, London, SW7 2AZ, UK

A R T I C L E I N F O A B S T R A C T

MSC: We consider the flow inside an unbaffled stirred tank at Re = 600 and decompose the pressure field into
76B47 components that are constant, linear and quadratic with respect to the temporal coefficients of velocity POD
93B30 (proper orthogonal decomposition) modes. The pressure components are used to analyse the power number of
62L12
individual blades and blade combinations. We also employ an identification algorithm to derive a linear dynamic
estimator that allows the reconstruction (in real time) of the instantaneous 3D velocity field from sparse pressure
Keywords:
measurements at the impeller blades. The first pair of POD modes are reconstructed with reasonable accuracy
Flow reconstruction from pressure
measurements using the pressure signal from a single sensor. The results improve with 6 or 12 sensors. Application of the
Reduced order modelling estimator to Re=500 and 700 shows that it is robust to changes in operating conditions. The work opens the
System identification possibility for real time control of mixing with targeted feeding location and rate.

1. Introduction surface can then be used to reconstruct the velocity field (Taylor and
Glauser, 2004; Hosseini et al., 2015). The second method is to use the
The flow inside stirred tanks has been extensively studied using both velocity POD modes to decompose the pressure field through the Pois-
computational fluid dynamics (CFD) and experimental methods, such son equation. This allows the pressure field to be decomposed as a sum
as particle image velocimetry (PIV). Both methods have revealed the
of pressure components, that are constant, linear and quadratic with re-
formation of large scale flow structures that determine momentum and
spect to the temporal coefficients of the velocity POD modes (Noack et
scalar transfer.
al., 2005, 2003). This approach retains the inherent physical link be-
Proper orthogonal decomposition (POD) has recently been used to
tween the velocity and pressure decompositions.
decompose the 3D velocity field and identify these large scale flow
In this paper, we present for the first time the shape of pressure
structures in unbaffled (Janiga, 2019; Mikhaylov et al., 2021) and baf-
modes that correspond to the large scale structures in the stirred tank.
fled (Mayorga et al., 2022) stirred tanks. POD can represent the most
important flow features (in terms of energy) with a small number of The pressure field on the impeller blade surfaces is inherently linked
degrees of freedom. In stirred tanks POD has been applied to the de- to the power number, a non-dimensional parameter that quantifies the
composition of the velocity field only, however it can be also used to power required to rotate the impeller at a given speed. The contribu-
decompose the pressure field. There are two methods to achieve this. tions of the time-average flow and the large scale structures to the mean
The first is to apply POD directly to the pressure field and hence ob- power number and its fluctuations can then be quantified through the
tain a linear decomposition (Arndt et al., 1997; Picard and Delville, obtained pressure components.
2000; Tamura et al., 1999; Borée, 2003; Hosseini et al., 2015; Tay- In our earlier study (Mikhaylov et al., 2021), we extracted the large
lor and Glauser, 2004). If a velocity decomposition is also performed scale structures through POD and used velocity data collected at a lim-
then the pressure and velocity decompositions can be linked by using ited number of sensor locations to reconstruct the temporal evolution
Extended Proper Orthogonal decomposition (Borée, 2003; Taylor and of these structures. The N4SID algorithm was used to identify a lin-
Glauser, 2004; Hosseini et al., 2015). Pressure measurements on a wall ear, multiple-input, multiple-output estimator directly from the data.

* Corresponding author.
E-mail address: g.papadakis@imperial.ac.uk (G. Papadakis).

https://doi.org/10.1016/j.ces.2023.118881
Received 29 December 2022; Received in revised form 27 April 2023; Accepted 12 May 2023
Available online 25 May 2023
0009-2509/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 1. Sketch of (a) the stirred tank and (b) the impeller geometry. The axes of the Cartesian coordinate system are shown in blue colour.

This approach had also been successfully used previously in 2D laminar Table 1
flows (Guzmán-Inigo et al., 2019). While our earlier study (Mikhaylov Basic dimensions of the stirred tank and impeller, normalised by the tank diam-
et al., 2021) was successful in predicting the evolution of the structures, eter T.
it relied on placing velocity probes within the flow. In the present study, D S H C L W d t
we demonstrate a less intrusive method in which only pressure data on
0.500 0.060 1.000 0.333 0.125 0.100 0.010 0.015
the impeller surface (and in some cases also on the vessel wall) are used.
Pressure profiles on the blade surfaces have previously been collected
experimentally (Lane et al., 2001; Steiros et al., 2017), which shows the The impeller is located at a distance H/3 from the tank bottom and
feasibility of our approach. Furthermore, the absence of velocity sensors has diameter equal to T/2. A schematic is shown in Fig. 1 and charac-
allows for much easier tank cleaning and prevents the accumulation of teristic dimensions, normalised by T, are provided in Table 1. The top
precipitated solids or biological matter in industrial applications. 0.02T of the impeller shaft is considered stationary as an approxima-
Once the model has been trained, experimental pressure readings tion to a stirrer guide and seal. This is the same geometry investigated
can be exploited for 3D flow reconstruction. This opens the possibility by the authors in Mikhaylov et al. (2021) and similar to that studied
for real time reaction control with targeted feeding location and rate in Tamburini et al. (2018) and Scargiali et al. (2017) using DNS and
(based on knowledge of the instantaneous characteristics of the flow experiments respectively.
structures) that could lead to a reduction of mixing time and waste The Reynolds number, defined in the usual way as 𝑅𝑒 = 𝜌𝑁𝐷2 ∕𝜇
product. For example, in Ducci and Yianneskis (2007) it was observed (where 𝜌 = 1000 kg∕m3 is the density, 𝑁 is the impeller speed and 𝐷
experimentally that the insertion of a passive scalar in the vortex core is the impeller diameter) is set to 600. Simulations at two additional
of the macro-instability leads to a reduction in mixing time by 20-30% values, Re=500 and 700, were also performed. For all cases, the com-
compared with injection outside the vortex core. Other studies (Assirelli putational domain was discretised with ∼ 7 × 106 cells. The reader may
et al., 2002, 2005, 2008) have investigated the effect on mixing of consult Mikhaylov et al. (2021) for code validation and grid indepen-
placing feed tubes for iodine-iodate reaction in the path of the trail- dence tests.
ing vortices compared with the standard location at the tank surface As the tank is unbaffled, the geometry has rotational symmetry and
and other locations in the flow. It was observed that the amount of the equations can be formulated and solved in a rotating reference
waste product was reduced to 5% when placing static feed tubes in the frame. The flow is assumed incompressible and the governing equations
path of the trailing vortices compared with 20% when placing the feed are given by,
tubes at the tank surface. In Assirelli et al. (2005, 2008) the effect of
𝜕𝐮 1
placing feed tubes in the path of the trailing vortices that rotate with + (𝐮 ⋅ ∇)𝐮 = − ∇𝑝+𝜈∇2 𝐮 − [𝝎 × (𝝎 × 𝐱) + 2𝝎 × 𝐮], (1a)
𝜕𝑡 𝜌
the impeller was investigated. A 7-fold decrease in the amount of waste
product was observed compared with a feed tube at the tank surface ∇ ⋅ 𝐮 = 0, (1b)
case and a reduction of waste product to 80% of that of the static feed where the last two terms in the momentum equation represent the cen-
tube in the impeller wake. trifugal and Coriolis forces. A Dirichlet boundary condition of 𝒖𝒘 =
To compare the reconstruction performance using pressure instead −𝝎 × 𝒙, where 𝝎 is the angular velocity vector and 𝒙 = (𝑥, 𝑦, 𝑧) is the
of velocity data, we choose the stirred tank geometry and flow condi- position vector, is imposed on the top, bottom and peripheral walls, as
tions to match our earlier study (Mikhaylov et al., 2021). We consider well as the top 0.02T of the shaft to account for the stirrer shaft guide.
an unbaffled stirred tank, with a base case Reynolds number 600. We
have performed however additional simulations to collect data for 210 3. Velocity POD modes
rotations (compared with 120 in Mikhaylov et al., 2021) in order to
elucidate the low frequency features of the power number. We also con-
To extract the large scale flow structures, we apply the POD ap-
sider two off-design conditions to test the performance of the obtained
proach, see Holmes et al. (2012) for details. The fluctuating components
estimator at Reynolds numbers 500 and 700.
of the velocity field are decomposed as,

2. Tank geometry and operating conditions 𝑖=∞


∑ 𝑖=∞
∑ 𝑖=𝑛

̄
𝒖(𝒙, 𝑡) = 𝒖(𝒙) + 𝒖′ (𝒙, 𝑡) = 𝒖̄ + 𝑎𝑖 (𝑡)𝜙𝑖 (𝒙) = 𝑎𝑖 (𝑡)𝜙𝑖 (𝒙) ≈ 𝑎𝑖 (𝑡)𝜙𝑖 (𝒙),
A covered unbaffled tank agitated by a 6-blade Rushton impeller 𝑖=1 𝑖=0 𝑖=0

is considered. The height of the tank (H) is equal to its diameter (T). (2)

2
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

where 𝒖̄ is the time-averaged velocity, 𝒖′ is the fluctuating velocity, of POD modes, and the inherent noise of any measuring device. This is
and 𝑎𝑖 (𝑡) and 𝜙𝑖 (𝒙) are the temporal coefficient and spatial distribution a linear relationship between the pressure measurements and the POD
of the 𝑖-th POD mode, respectively. In (2) we assume that 0-th mode coefficients 𝑎𝑖 (𝑡), it has exactly the same form as for the case of velocity
̄
corresponds to the mean velocity, i.e. 𝑎0 (𝑡) = 1 and 𝜙0 (𝒙) = 𝒖(𝒙). An ap- measurements, and thus the derivation of the flow estimator proceeds
proximation to the full flow field is obtained by truncating the series, in an identical fashion to that described in Mikhaylov et al. (2021). The
see last term in the above equation, where 𝑛 is the order of the trunca- resulting system has the form,
tion. The modes are ordered by descending energy content, hence 𝑛 is
selected so that a given percentage of the total energy is captured in the 𝐗(𝑡𝑗+1 ) = 𝐀′ 𝐗(𝑡𝑗 ) + 𝐁′ 𝐬(𝑡𝑗 ) + 𝐰(𝑡𝑗 ), (5a)
truncated expansion.
𝐚(𝑡𝑗 ) = 𝐇 𝐗(𝑡𝑗 ) + 𝐯(𝑡𝑗 ),

(5b)
The POD modes are computed by first assembling a snapshot matrix,
which is formed by storing, column by column, the three-dimensional where 𝐗(𝑡𝑗 ) is the estimator state at time instant 𝑡𝑗 , and 𝐰(𝑡𝑗 ) and 𝐯(𝑡𝑗 )
velocity field at different time instants. The snapshot matrix is then represent white Gaussian noise. The input 𝐬(𝑡𝑗 ) is taken to be pressure
weighted by a diagonal matrix which contains the cell volumes, and readings at selected sensor points on the blades and the output 𝐚(𝑡𝑗 ) are
finally Singular Value Decomposition is performed on the weighted ma- the temporal coefficients of the first few POD modes corresponding to
trix; for details the reader may refer to Mikhaylov et al. (2021). the large-scale structures. The matrices 𝐀′ , 𝐁′ and 𝐇′ are unknown and
In the present study, for all Reynolds numbers, the simulation is are computed using the N4SID algorithm (Van Overschee and De Moor,
run for approximately 40 rotations to allow the flow to reach a statis- 1994).
tically steady state (probe points in the flow field were used to verify The overall procedure consists of three steps. In the first step, the
this). For the 𝑅𝑒 = 600 case, the simulation is run for approximately 210 DNS simulations are performed to generate the input data 𝒔(𝑡) and the
additional rotations (compared with 120 in Mikhaylov et al. (2021)). output data 𝒂(𝑡) (the latter are obtained from the SVD of the weighted
Approximately 100 snapshots were recorded per rotation (one snapshot velocity snapshot matrix, see section 3). In the second step, the matri-
every 57 time steps) to provide a dataset of approximately 21000 snap- ces of the linear, time-invariant system (5) are computed by maximising
shots. The first 10000 snapshots (approximately 100 rotations) were the fit with the DNS data from the training data set. In the third step,
used for training the estimator and the remaining 110 rotations were the model performance is evaluated using unseen data from the valida-
used for its validation, see also section 4. For the other two Reynolds tion data set at design and off-design conditions. The training dataset is
numbers, after the initial 40 rotations, data for 20 rotations are col- approximately 100 rotations long and the validation dataset at the base
lected to serve as validation datasets. case is approximately 110 rotations long, while at off-design conditions
In Mikhaylov et al. (2021), it was found that the first two modes is approximately 20 rotations long.
were paired with a contribution of 14% each to the turbulence kinetic
energy, with a normalised frequency of 1.5𝐹𝑁 , where 𝐹𝑁 is the impeller 5. Analysis of power number
rotation frequency. The third and fourth modes were also found to be
paired, with smaller contributions of 2.6% each to the turbulence kinetic 5.1. Power number signal and spectra
energy and a frequency of 3.0𝐹𝑁 . The two pairs of modes were found to
correspond to rotating vortices and were primarily concentrated in the 𝑃𝑜
The power number in a stirred tank is defined as 𝑁 𝑝 = ,
areas close to the impeller. Identical results were found in the present 𝜌𝑁 3 𝐷5
work. It should be noted that due to the 60◦ rotational symmetry, the where the overbar denotes the time average and 𝑃 𝑜 is the mean power
modes obtained can be rotated by 60◦ (or multiples thereof) relative provided to the flow by the rotating impeller. Since the system is closed,
to those obtained with datasets of different length; apart from that the 𝑁 𝑃 is also equal to the total energy dissipation rate integrated over the
modes are identical to those shown in Mikhaylov et al. (2021). tank volume.
The time signal of instantaneous power number 𝑁𝑝 (𝑡) computed by
4. Flow estimation from sparse pressure measurements integration of pressure over the surface of the blades, as well as by the
volume integral of dissipation rate, is shown in Fig. 2a. The agreement
We now aim to construct a linear estimator that extracts the tempo- between the two calculation methods is very good, with 𝑁̄ 𝑝 ≈ 2.1 for
ral coefficients, 𝑎𝑖 (𝑡), and thus the whole fluctuating velocity field, from both, and the same general trends in the two signals. The power spec-
sparse pressure measurements. It can be shown (see Appendix A for the trum of the 𝑁𝑝 (𝑡) is shown in Fig. 2b. We split the signal into several
derivation) that the pressure fluctuation 𝑃 ′ (𝐱, 𝑡) can be written in terms windows with 50% overlap, computed the power spectra at each and
of 𝑎𝑖 (𝑡) as, averaged; this process smooths out the spectra. As can be seen, the re-
𝑖=𝑛
∑ ∑ 𝑗=𝑛
𝑖=𝑛 ∑ sults for three window lengths are similar. We use a window size of
𝑃 ′ (𝐱, 𝑡) = 𝑃 (𝐱, 𝑡) − 𝑃 (𝐱, 𝑡) ≈ 2 𝑃𝑖0 (𝐱)𝑎𝑖 (𝑡) + 𝑃𝑖𝑗 (𝐱)𝑎𝑖 (𝑡)𝑎𝑗 (𝑡), (3) approximately 82 impeller rotations for all subsequent power number
𝑖=1 𝑖=1 𝑗=1 spectra.
(𝑗≠𝑖)
Interestingly, the frequencies of 1.5𝐹𝑁 and 3.0𝐹𝑁 corresponding
where 𝑃𝑖𝑗 are known as pressure components. The governing equations to the two leading pairs of modes are not present. In addition, the
and boundary conditions for 𝑃𝑖𝑗 are also derived in Appendix A. This spectrum contains peaks at many low frequencies and in particular at
expansion for 𝑃 ′ (𝐱, 𝑡) contains terms that are both linear and quadratic ∼ 0.04𝐹𝑁 and ∼ 0.43𝐹𝑁 . Similar low frequency oscillations in the power
in 𝑎(𝑡). In previous studies (Noack et al., 2005) the linear terms have number were observed in Rave et al. (2021) for the case of a Rushton
been found to be dominant throughout the domain. The analysis of the turbine of diameter T/3 and clearance 0.15T in a baffled tank. In the
data for our case reveals a similar behaviour, in particular in the vicinity dataset of 210 rotations, the period corresponding to the slowest fre-
of the blades, see section 6 below. quency of ∼ 0.04𝐹𝑁 is repeated approximately 8 times. We examine the
In the above context, we seek an estimate 𝑎𝑒 (𝑡) of 𝑎(𝑡) from a set absence of the 1.5𝐹𝑁 and 3.0𝐹𝑁 frequencies and the origin of the small
of pressure fluctuation measurements 𝒔(𝑡) = (𝑃1′ (𝑡), … , 𝑃𝑝′ (𝑡))𝑇 taken at 𝑝 frequencies below.
sensor points located at 𝒙1 , … , 𝒙𝑝 within the stirred tank. From (3) we The power number signal for individual blades and blade contribu-
can write tions for 5 rotations is shown in Fig. 3. The time signal for the combined
power number from the three blades has fewer periodic features com-
𝒔(𝑡) = 𝑺 𝑝 𝒂(𝑡) + 𝒈(𝑡), (4)
pared with the time signals of the individual blades. This is confirmed
where the matrix 𝑺 𝑝 can be obtained from 𝑃𝑖0 (𝒙𝑗 ) and 𝒈(𝑡) arises from by inspecting the spectra of the power numbers of individual blades
the quadratic terms, the fact that we are considering a finite number and blade combinations shown in Fig. 4. All spectra (with the exception

3
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 2. Power number for Re=600.

sidering three adjacent blades, both the 3.0𝐹𝑁 and 1.5𝐹𝑁 signal vanish,
while the 2.2𝐹𝑁 remains, but is weak. Finally, when considering all 6
blades, all frequencies greater than 1.0𝐹𝑁 vanish.
A similar frequency cancellation was observed in Basbug (2017), for
a 4-bladed impeller. More specifically, the 1.70𝐹𝑁 frequency observed
for individual blades cancelled out in the power number because the
pressure distribution of adjacent blades was exactly 180◦ out of phase.
Notably, when breaking the symmetry of the blades by considering frac-
tal blades, the authors found that this frequency was restored.
Unfortunately, the cancellation of the dominant frequencies of
1.5𝐹𝑁 and 3.0𝐹𝑁 in the 𝑁𝑝 (𝑡) indicates that it is not possible to re-
cover the evolution of the large scales by measuring the power number
fluctuation. However, in the case of a different geometry and in partic-
ular, one in which the symmetry is (even slightly) broken, the approach
that we demonstrate below in section 6 using point pressure data could
Fig. 3. Power number signal for individual blades and blade combinations. potentially also be applied using power number fluctuation data.

of the all-blades spectrum) are averaged using the rotational symmetry 5.2. Power number decomposition
of the system. For a single blade, there are peaks mainly at frequen-
cies 1.5𝐹𝑁 , 3.0𝐹𝑁 , 2.2𝐹𝑁 . The set of 1.5𝐹𝑁 , 3.0𝐹𝑁 and 2.2𝐹𝑁 matches The instantaneous pressure can be written in terms of 𝑎𝑖 (𝑡) as (see
the frequencies of the first, second and fourth POD mode pairs respec- proof in Appendix A),
tively. It can also be seen that the frequencies of 1.5𝐹𝑁 and 3.0𝐹𝑁 are 𝑖=𝑛
∑ ∑ 𝑗=𝑛
𝑖=𝑛 ∑
present in all spectra from pairs of adjacent blades. On the other hand, 𝑝(𝐱, 𝑡) ≈ 𝑃00 (𝐱) + 2 𝑃𝑖0 (𝐱)𝑎𝑖 (𝑡) + 𝑃𝑖𝑗 (𝐱)𝑎𝑖 (𝑡)𝑎𝑗 (𝑡). (6)
the 2.2𝐹𝑁 frequency vanishes when considering two adjacent blades or 𝑖=1 𝑖=1 𝑖=1

two blades separated by 180◦ , but increases to become stronger than the Due to the presence of the quadratic terms 𝑃𝑖𝑗 𝑎𝑖 (𝑡)𝑎𝑗 (𝑡), the mean pres-
1.5𝐹𝑁 and 3.0𝐹𝑁 signals when the separation angle is 120◦ . When con- sure has contributions from not just the mean flow mode 𝑃00 , but also

Fig. 4. Spectra of power numbers for individual blades and blade combinations.

4
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 5. Power number due to diagonal pressure components.

from all of the diagonal pressure components 𝑃𝑖𝑖 , see equation (A.25). Contours of linear pressure components (that have one index equal
Hence, the mean power is given by, to 0, see (A.20)) at a vertical plane bisecting two blades and a horizontal
plane 0.353T from the tank bottom √ are shown
√ in Figs. 8 and 9 respec-
𝑖=𝑛
∑ 𝑖=𝑛
∑ ¨
𝑃 𝑜𝑖 𝑎𝑖 (𝑡)𝑎𝑖 (𝑡) tively. In the vertical plane, 𝜆1 𝑃01 and 𝜆2 𝑃02 are anti-symmetric
𝑁𝑝 ≈ =− ⃗
𝑃𝑖𝑖 𝑢⃗ ⋅ 𝑑𝑠
𝑖=0 𝜌𝑁 3 𝐷5 𝑖=0 𝜌𝑁 3 𝐷5 about the impeller disk but the second pair of modes does not have this
𝑖𝑚𝑝𝑒𝑙𝑙𝑒𝑟
(7) symmetry. In the horizontal plane, the structure sizes for the first two
𝑖=𝑛
∑ ¨
𝜆𝑖 modes again appear to be twice the size of those corresponding to the
=− ⃗
𝑃𝑖𝑖 𝑟𝜔𝑛̂ ⋅ 𝑑𝑠,
𝜌𝑁 3 𝐷5 second pair of modes.
𝑖=0 𝑖𝑚𝑝𝑒𝑙𝑙𝑒𝑟

where 𝜆𝑖 is the eigenvalue of the 𝑖-th POD mode (including the 0-th) and 5.3. Analysis of power number spectrum using pressure components
𝑛̂ is the unit vector perpendicular to the blade surface. The distribution
of the power number components and their cumulative percentage with In Fig. 4 it was shown that peaks in the spectra of the power number
respect to the total power number are shown in Fig. 5. The pressure at 1.5𝐹𝑁 and 3.0𝐹𝑁 were present for individual blades, but these peaks
component 𝑃00 is found to account for 97% of the total power number, vanished when all blades were considered. This suggests that there is a
the first pair of POD modes 0.6%, the second pair of POD modes 0.06% cancellation between blades. To investigate this further, we examine the
and other modes < 0.03%. The first 300 modes contain 99.7% of the total spectra of the power number computed from the linear pressure com-
power number. ponents of the first pair of modes. All the spectra were found to have
The contribution of 𝑃00 to the power number (97%) is larger com- a single peak at frequency 1.5𝐹𝑁 . The values of these peaks for differ-
pared with that of the 0-th velocity mode (time-average velocity) to the ent blade combinations are shown in the bar chart of Fig. 10. When
kinetic energy (86%). Recall that the power number is obtained by in- considering the power number from all blades, the peak is reduced by
tegrating the pressure difference between the two sides of each blades. more than 2 orders of magnitude with respect to an individual blade
Since higher order modes are approaching homogeneous turbulence, (compare the leftmost and rightmost bars).
we would expect their contributions to the average power number to The time signal of the power number component is shown in Fig. 11
be negligible (because the time average pressure distributions on either for 2 rotations. The amplitude of the oscillation is significantly reduced
side of a blade would be approximately equal). when the contribution from three adjacent blades is considered, in
Contour plots of the mean pressure and diagonal partial pressures agreement with Fig. 10. From the contours of 𝑃01 and 𝑃02 on the im-
on a vertical plane bisecting the blades are shown in Fig. 6. The mean peller surface (Figs. 12a and 12b) the phase difference between adjacent
pressure is primarily contained in the 𝑃00 . In regions away from the blades can be seen, leading to cancellation when considering the aggre-
impeller 𝑃00 almost exactly matches the mean pressure. On the other gate power number for 3 blades. A similar feature to the first pair of
hand, in areas close to the impeller height where the other pressure modes is observed by inspecting the spectra of the partial power num-
components have large magnitude, there are noticeable differences. The ber computed from 𝑃03 and 𝑃04 . The large magnitude peak at 3.0𝐹𝑁 for
contours of 𝜆1 𝑃11 and 𝜆1 𝑃22 are symmetric about the blade disk height, individual blades is significantly suppressed (by a factor of over 400)
but this symmetry is broken for the third and fourth components, 𝜆3 𝑃33 when the total power number is considered (results are not shown for
and 𝜆4 𝑃44 . brevity).
Contour plots of the mean pressure and diagonal pressure compo- The linear pressure components corresponding to velocity modes 1-
nents on a horizontal plane 0.353T above the tank bottom (plane of 4 cannot explain the origin of the low frequency peaks in the spectra
maximum RMS velocity) are shown in Fig. 7. As for the vertical plane, of the power number, Fig. 2b, therefore higher order modes have to be
the largest contribution to the mean pressure comes from 𝑃00 , but with considered. The 5th and 6th, and 7th and 8th velocity modes are also
noticeable contributions from other modes as the trailing vortices get paired and the power number computed from the corresponding linear
further from the blades. The distributions of 𝜆1 𝑃11 , 𝜆2 𝑃22 , 𝜆3 𝑃33 and pressure components is again found to have negligible contribution.
𝜆4 𝑃44 have small differences across the pressure and suction sides of The next mode, 9, is found to be unpaired, with a low frequency
the blades, compared with the 𝑃00 term, which is in agreement with the peak at ∼ 0.04𝐹𝑁 . Modes with low frequencies have been observed in
earlier finding that the largest contribution to the power number comes previous studies (Nikiforaki et al., 2003; Ducci et al., 2008; Hasal et al.,
from 𝑃00 . The length scales of the structures corresponding to the first 2004; Roy and Acharya, 2011, 2012; Lavezzo et al., 2009) and have
pair of modes, 𝜆1 𝑃11 and 𝜆2 𝑃22 , are observed to be twice compared with been associated with macro-instabilities. Contours of the velocity mag-
those corresponding to the second pair, 𝜆3 𝑃33 , 𝜆4 𝑃44 . nitude on a horizontal plane at 0.353T above the bottom of the tank

5
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 6. Contour plots of mean and diagonal pressure components on a vertical section bisecting the blades.

Fig. 7. Contour plots of the mean and diagonal pressure components on a horizontal section 0.353T from the tank bottom.

6
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 8. Contour plots of the linear pressure components on a vertical section bisecting the blades.

and on a vertical plane bisecting the blades are shown in Figs. 13a and It should be noted that the spectrum of a single POD mode can have
13b respectively. The mode has a 60◦ rotational symmetry. multiple frequency peaks and that a physical phenomenon with a par-
The spectrum of the power number due to the linear pressure com- ticular frequency can be spread across multiple modes. In addition, the
ponent 𝑎9 (𝑡)(𝑃09 + 𝑃90 ) is shown in Fig. 14a. There is a peak at 0.04𝐹𝑁 , ordering of the POD modes is based on their contribution to the ki-
the same as the one observed in the global power number, Fig. 2b). netic energy in the whole tank, and this order may not be the same
The contribution of this mode to the power number can be explained in the vicinity of the impeller. An approach to extract velocity modes
by examining the spatial distribution of 𝑃09 mode. Contour plots at a that beat at particular frequencies is the Dynamic Mode Decomposition
horizontal plane 0.353T above the bottom of the tank, on a vertical (DMD) (Schmid, 2010) or Spectral POD (Taira et al., 2017; Towne et
plane bisecting the blades, and on the surface of the impeller are shown al., 2018); one can then obtain the corresponding DMD or Spectral POD
in Figs. 13c, 13d and 13e respectively. The 60◦ symmetry means that pressure modes. This is a very promising research direction, but it is
the power number contributions from the individual blades will be in beyond the scope of the present work.
phase, i.e. synchronised, and so will sum up to contribute to the total
power number without cancellation. Furthermore, considering a single 6. Reconstruction of large scale vortical structures from surface
blade, the pressure on either side has different sign, giving a significant pressure measurements
contribution to the torque and therefore the power number.
The spectrum of the power number computed from the linear pres- In this section, pressure recordings taken at a few sensor points on
sure components corresponding to the first 300 modes is shown in the impeller surface and vessel wall are used to reconstruct the evolu-
Fig. 14b. It matches well with the spectrum of the global power number tion of the temporal coefficients for the first two POD modes for the
(blue line), suggesting that the majority of the contribution is captured design operating condition 𝑅𝑒 = 600 and at two off-design conditions,
in the first 300 linear terms. The discrepancies arise from contributions 𝑅𝑒 = 500 and 𝑅𝑒 = 700.
of higher linear pressure components and also from the quadratic com- As we are constructing a linear estimator for the evolution of the
ponents. temporal POD coefficients from pressure readings, it is important to
The spectrum of 𝑁𝑝 (𝑡) computed from the 30 modes with the great- consider the relative sizes of the linear and quadratic pressure compo-
est contribution to RMS of 𝑁𝑝 (𝑡) is shown in Fig. 14c. While some peaks nents. We find that the magnitudes of the linear components 𝑃01 and
in the full spectrum are reproduced, a significantly simpler spectrum 𝑃02 (shown in Figs. 8, 9 and 12) is approximately double that of the
is obtained. In particular, the peak at ∼ 0.43𝐹𝑁 is not as distinct. This largest cross term 𝑃12 (shown in Fig. 15) in the fluid domain. However,
suggests that while individual modes may have small contributions to these two modes have a substantially larger footprint in the vicinity of
the kinetic energy, their contribution should not be neglected as their the blades than 𝑃12 , with a maximum magnitude over 10 times greater
combination can give noticeable peaks in the spectrum of the power than that of 𝑃12 . This suggests that, for points chosen on the surfaces of
number. the blades, the assumption of a linear relationship between the pressure

7
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 9. Contour plots of the linear pressure components on a horizontal section 0.353T from the tank bottom.

Fig. 10. Spectrum amplitudes (at the frequency of 1.5𝐹𝑁 ) of the power number
Fig. 11. Signal of power number component obtained from 𝑎1 (𝑡)(𝑃01 + 𝑃10 ) +
component obtained from 𝑎1 (𝑡)(𝑃01 + 𝑃10 ) + 𝑎2 (𝑡)(𝑃02 + 𝑃20 ) for different blade
𝑎2 (𝑡)(𝑃02 + 𝑃20 ) for different blade combinations.
combinations.

evolution and the evolution of the temporal coefficient of the first two 12 input locations. The pressure time series are pre-processed by sub-
tracting the mean and dividing by the standard deviation at each sensor
POD modes, i.e. equation (4), is valid.
location.
First, a single sensor point located at the maximum of the pressure
The quality of the reconstruction is quantified by the percentage fit
RMS on the impeller surface is considered. This was found to be on
defined as,
the suction edge of the blade, indicated with a green dot in Fig. 16(b).
( )
Next, exploiting the rotational symmetry of the geometry, 5 additional ||𝑎𝑖 (𝑡𝑘 ) − 𝑎𝑖,𝑒 (𝑡𝑘 )||
points are placed at 60◦ intervals from the first point. Finally, we add a FIT𝑖 [%] = 100 × 1 − , (8)
further 6 points based upon the location of the maximum of the pres- ||𝑎𝑖 (𝑡𝑘 ) − 𝑎𝑖 (𝑡𝑘 )||
sure RMS on the outer wall of the tank, indicated with a magenta dot where || ⋅ || is the 2-norm of a vector, the subscript 𝑒 denotes the esti-
in Fig. 16(b) and again exploit the rotational symmetry to get 5 more mated coefficient, and the overbar denotes time-average (here equal to
points. Combining these 6 points with the 6 points on the blade sur- 0). A large value of FIT𝑖 (max 100) indicates good matching between
faces, we consider the final case of input data from pressure sensors at 𝑎𝑖 (𝑡) and 𝑎𝑖,𝑒 (𝑡). For both the training and validation dataset, the model

8
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 12. Impeller surface contour plots of 2 linear pressure components corresponding to the first pair of POD modes.

Fig. 13. Contour plots of velocity magnitude and linear pressure component corresponding to Mode 9.

9
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 14. Power number spectra from contributions of different linear pressure components.


Fig. 15. Contours of 𝑎21 𝑎22 𝑃12 .

is initialised with zero conditions, and the FIT is evaluated after discard- time series for the pressure shows that, while having a strong periodic
ing the first 5 rotations. It should be noted that in some of the results nature, the signal is chaotic.
below, the model spin-up takes longer than 5 rotations and as such the The reconstruction results for 𝑅𝑒 = 600 are shown in Fig. 18. The
fit percentage can be considered as a measure of both model accuracy training and validation fit values for modes (1,2) were found to be
and speed of convergence to the correct state.
(49%,48%) and (53%,52%) respectively. These values reflect that while
the period of both modes is predicted correctly, the amplitude is under-
6.1. Reconstruction using pressure data from a single point on the blade
estimated locally. Increasing the model spin-up time by discarding the
surface located at maximum pressure RMS
first 20 rotations increased both the training and validation fit values by
The pressure time signal and spectrum at the point of maximum approximately 7-9%. This suggests that with substantial spin-up time, a
RMS are shown in Fig. 17. The spectrum shows a strong peak at 1.5𝐹𝑁 reasonable fit quality of the reconstructed signal to the DNS signal can
corresponding to the frequency of the first two pairs of modes. The be achieved.

10
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 16. Contours of the (a) Mean and (b) RMS Pressure on the impeller surface. The locations of max RMS on the impeller and the outer tank wall are shown in
green and magenta dots respectively. The coordinates (𝑟, 𝜃, 𝑧) are (0.125T, 57◦ , 0.304T) and (0.500T, 74◦ , 0.372T) respectively.

Fig. 17. Pressure at max RMS on the impeller surface.

6.2. Reconstruction using pressure data from six points on the blade surface 84%) fits for the training dataset and (78%,77%) fits for the validation
dataset, see Fig. 20. Both the period and the magnitude of the signals
The results from 6 sensors on the blades surface are shown in Fig. 19. are predicted well and this is reflected in the fit values.
A very good fit is obtained, 76% for both modes for the training dataset For the 𝑅𝑒 = 700 case the model output matches the DNS values
and (76%,75%) for the validation dataset (only the latter dataset is very well, with fits of (79%,77%). In the 𝑅𝑒 = 500 case the reconstruc-
shown). Both the period and the magnitude of the signals are relatively tion is also good with fits of (66%,68%). Notably, the small phase shift
well predicted and this is reflected in the higher fit values. The esti- between the DNS and reconstructed signals is no longer present. This
mated signal converges to the true one considerably more quickly than suggests that the addition of probe points on the outer walls corrects
the single point case. for the phase difference. A potential explanation for this behaviour is
For the 𝑅𝑒 = 700 case, Fig. 19(d), the model output matches the that the vortex in the 𝑅𝑒 = 500 case is shed slightly earlier than for the
DNS values very well, with fits of (80%,79%). In the 𝑅𝑒 = 500 case, 𝑅𝑒 = 600 and 𝑅𝑒 = 700 cases, and reaches the outer wall slightly later.
Fig. 19(c), the fit is relatively low, with values of (32%,31%). This is Hence, this suggests that the vortex propagation speed for the 𝑅𝑒 = 500
due to the small phase difference between the true and reconstructed case is slightly slower compared to the other two cases.
signal. Introducing a delay of 5 snapshots to the reconstructed signal
increases the fit values to (70%,69%), suggesting that both the period
6.4. Reconstruction of instantaneous pressure
and the amplitude of the signal are in fact reconstructed well. The better
performance of the reconstruction with an increased Reynolds number
suggests that the coupling of the pressure signal at the blade to the Using the mean pressure and the pressure components that only de-
evolution of the structures represented by the modes increases as the pend on 𝑎1 (𝑡) and 𝑎2 (𝑡), the instantaneous pressure can be reconstructed
impeller speed increases. by applying equation (6)). Contours of the true and estimated pressure
on a vertical slice midway between the impeller blades at the same time
6.3. Reconstruction using pressure data from 12 points instant are shown in Fig. 21. The pressure field is reproduced well in
the vicinity of the impeller, with the large structures appearing in the
We now consider the case of 12 points, with 6 locations at max reconstructed pressure field. However, away from the impeller, the re-
pressure RMS on the blades and 6 more at the max pressure RMS on constructed pressure looks smooth compared with the true field which
the outer tank wall. In this case a good fit is obtained for both the has random, small scale, features. This is expected as these features are
training and validation dataset for the base 𝑅𝑒 = 600 case with (84%, represented by higher order modes that are truncated.

11
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 18. Estimated (orange) and true (blue) coefficients for modes 1 and 2 using a single probe point located at the max pressure RMS (5th order model).

7. Conclusions to be on the suction side of the blade. Exploiting the rotational symme-
try, 5 additional points were placed 60◦ apart. Similarly, the first point
It is customary in the literature to decompose the velocity field into on the tank wall was placed at the maximum fluctuating pressure RMS,
a set of POD modes. In this paper, we extend the decomposition to the with 5 further points placed symmetrically around the periphery.
pressure field. More specifically, we express this field as the sum of a In the base 𝑅𝑒 = 600 case, the temporal coefficients were recon-
component due to the mean flow plus components that are linear and structed using a single data point, with the period correctly identified.
quadratic with respect to the temporal coefficients of the velocity POD However, the validation fit values were (53%, 52%) for modes (1,2) due
modes. The pressure modes corresponding to the combinations of the to the slow convergence of the model to the DNS results. It should be
first four velocity POD modes in a stirred tank were visualised for the noted that unlike the velocity-based reconstruction method employed
first time. The contributions of the diagonal pressure components to the in Mikhaylov et al. (2021), the input data at each probe point consists
power number were examined, and it was found that 97% comes the of just one (instead of 3) input signals. Increasing the number of points
𝑃00 mode, that originates from the time-average flow field. to 6 or 12 led to an significant increase in the fit values to (76%, 75%)
The spectra of the global and blade power numbers were examined. or (78%, 77%) respectively.
It was found that although the frequencies 1.5𝐹𝑁 and 3.0𝐹𝑁 (corre- For the off-design 𝑅𝑒 = 700 case, similar results were obtained to
sponding to the first four modes) were present in the power numbers the base case for both 6 and 12 points, with fit values (80%,79%) and
of individual blades, they were absent in the global one. Further ex- (79%,77%) respectively. For the 𝑅𝑒 = 500 case, and measurements at 6
amination of the power number computed from the linear pressure sensor points, the fit was (32%,31%), with the magnitude and period re-
components revealed that the contributions from the first two pairs of constructed correctly, but with a small phase shift. However, for the 12
modes cancelled out when considering all blades together. The first sig- point case the reconstruction was good, with fit values of (66%, 68%),
nificant contribution to the power number fluctuation came from the with no phase difference between the true and reconstructed signals.
unpaired mode 9, that accounts for the low frequency peaks in the For the present geometry, the frequencies for the first two pairs of
spectrum. The latter was reproduced well by considering the first 300 modes were not present in the power number. It would be interesting to
linear pressure modes, suggesting that the dominant frequencies are dis- examine whether this would also hold for other geometries, for example
tributed over many POD modes. a differently sized or shaped impeller or a baffled tank. If the dominant
The temporal coefficients of the leading two modes were then re- frequencies do not cancel out, it may be possible to reconstruct the flow
constructed using the N4SID algorithm using pressure measurements at within the stirred tank using only external measurements of the global
1 and 6 points on the impeller blade surfaces and 12 points (that com- power number.
prise the 6 points on the impeller surface plus 6 additional points on The extension of the reconstruction method to use pressure data
the outer tank wall). The first point on the impeller surface was placed increases its applicability due to the non-intrusive nature of such mea-
at the location of maximum RMS of fluctuating pressure that was found surements. However, this comes with a small performance degradation.

12
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 19. Estimated (orange) and true (blue) coefficients for modes 1 and 2 using 6 probe points based on max PRMS (6th order model).

The examination of the Power number revealed some low frequency A.1. Equation for the pressure field
components, which require further exploration. It would be also very in-
teresting to exploit the spatial and temporal dynamics of the identified The governing equations in a rotating reference frame take the form
large scale structures and perform mixing simulations with parameters
(timing, location etc) guided by this information. 𝜕𝐮 1
+ (𝐮 ⋅ ∇)𝐮 = − ∇𝑝+𝜈∇2 𝐮 − [𝝎 × (𝝎 × 𝐱) + 2𝝎 × 𝐮], (A.1a)
𝜕𝑡 𝜌
Declaration of competing interest ∇ ⋅ 𝐮 = 0. (A.1b)

The authors declare that they have no known competing financial The centrifugal and pressure gradient terms in (A.1a) can be com-
interests or personal relationships that could have appeared to influence bined to define the modified pressure (Davidson, 2015),
the work reported in this paper.
1
𝑃̂ = 𝑝 − 𝜌(𝝎 × 𝐱)𝟐 , (A.2)
2
Data availability
thus

Data will be made available on request. 𝜕𝐮 1


+ (𝐮 ⋅ ∇)𝐮 = − ∇𝑃̂ + 𝜈∇2 𝐮 − 2𝝎 × 𝐮. (A.3)
𝜕𝑡 𝜌
Acknowledgements Taking the divergence of (A.3) yields

The authors are grateful for computational resources made avail- 𝜕(∇ ⋅ 𝐮) ∇2 𝑃̂
+ ∇ ⋅ (𝐮 ⋅ ∇𝐮) = − + 𝜈∇ ⋅ (∇2 𝐮) − ∇ ⋅ (2𝝎 × 𝐮). (A.4)
able through the UK Turbulence Consortium (EP/R029326/1) and the 𝜕𝑡 𝜌
UK Materials and Molecular Modelling Hub (EP/P020194/1). The first Setting the transient and the viscous terms to 0 (due to the incompress-
author, KM, would like to acknowledge funding from the EPSRC CDT ibility condition) results in the simplified pressure Poisson equation
for Fluid Dynamics Across Scales (EP/L016230/1). GP is supported by (SPPE) (Sani et al., 2006). However, for numerical reasons, we retain
EPSRC grants EP/X017273/1, EP/W001748/1 and EP/W00481X/1. the viscous term (Sani et al., 2006), (Guermond et al., 2006). This yields
the consistent pressure Poisson equation (CPPE),
Appendix A. Decomposition of pressure fields
∇2 𝑃̂
∇ ⋅ (𝐮 ⋅ ∇𝐮) = − − ∇ ⋅ (2𝝎 × 𝐮) + 𝜈∇ ⋅ ∇2 𝐮, (A.5)
In this Appendix we decompose the pressure fields (instantaneous, 𝜌
mean and fluctuating) in components that are related to the POD coef- which upon rearrangement becomes,
ficients of the velocity field. We also derive the governing equation and
boundary conditions for each pressure component. ∇2 𝑃̂ = −𝜌∇ ⋅ (𝐮 ⋅ ∇𝐮) − 𝜌∇ ⋅ (2𝝎 × 𝐮) + 𝜇∇ ⋅ ∇2 𝐮. (A.6)

13
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. 20. 12 probe points based on max PRMS 6th order model.

Fig. 21. Pressure contours after 40 rotations in the validation dataset using 6 points on the blade surface located at max pressure RMS.

𝜕 𝑃̂ 𝜕2 𝑢
A.2. Boundary conditions for the pressure equation = 𝜇 2𝑛 . (A.9)
𝜕𝑛 𝜕𝑛
For the outer (peripheral) wall, 𝒏 is in the radial direction, 𝝎 in the
Projecting equation (A.3) onto the unit vector normal to the wall, 𝐧,
axial direction (i.e. 𝝎 = 𝜔̂𝐳 , where 𝐳̂ is the unit vector along 𝑧) and 𝐮 in
gives the tangential direction, hence the Coriolis term simplifies to −2𝑅𝜔2 ,
where 𝑅 is the vessel radius. In this case, the boundary condition be-
𝜕𝑢𝑛 1 𝜕 𝑃̂ 𝜕 2 𝑢𝑛
+ 𝐧 ⋅ (𝐮 ⋅ ∇𝐮) = − +𝜈 − 2𝐧 ⋅ (𝝎 × 𝐮), (A.7) comes,
𝜕𝑡 𝜌 𝜕𝑛 𝜕𝑛2
where 𝑢𝑛 = 𝐮 ⋅ 𝐧. At the wall, the normal velocity 𝑢𝑛 = 0 and the first 𝜕 𝑃̂ 𝜕2 𝑢
= −𝜌𝐧 ⋅ (𝐮 ⋅ ∇𝐮) + 𝜇 2𝑛 − 2𝜌𝑅𝜔2 . (A.10)
term vanishes, giving the boundary condition, 𝜕𝑛 𝜕𝑛
The advection term in cylindrical coordinates takes the form,
𝜕 𝑃̂ 𝜕2 𝑢
= −𝜌𝐧 ⋅ (𝐮 ⋅ ∇𝐮) + 𝜇 2𝑛 − 2𝜌𝐧 ⋅ (𝝎 × 𝐮). (A.8) ( )
𝜕𝑛 𝜕𝑛 𝜕𝑢 𝑢 𝜕𝑢 𝜕𝑢 𝑢2
−𝜌𝐧 ⋅ (𝐮 ⋅ ∇𝐮) = −𝜌 𝑢𝑟 𝑟 + 𝜃 𝑟 + 𝑢𝑧 𝑟 − 𝜃 . (A.11)
For stationary walls, this simplifies to, 𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑧 𝑟

14
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

Fig. A.22. Contours of instantaneous pressure obtained (a) directly from DNS and (b) from solving the Poisson equation. Results are shown on a vertical plane
midway between two blades.

The first three terms on the right hand side are zero, as 𝑢𝑟 = 0 and 𝑢𝑧 = 0 ∇2 𝑃̂ = −𝜌∇ ⋅ (𝒖̄ ⋅ ∇𝒖) ̄ + 𝜇∇ ⋅ ∇2 𝒖̄
̄ − 𝜌∇ ⋅ (2𝝎 × 𝒖)
at the wall. Bearing in mind that 𝑢𝜃 = −𝑅𝜔, equation (A.11) reduces to, 𝑖=𝑛
∑ [ ]
− 𝑎𝑖 (𝑡) 𝜌∇ ⋅ (𝒖̄ ⋅ ∇(𝝓𝐢 )) + 𝜌∇ ⋅ (𝝓𝐢 ⋅ ∇(𝒖))
̄ + 𝜌∇ ⋅ (2𝝎 × 𝝓𝐢 ) − 𝜇∇ ⋅ ∇2 𝝓𝐢
−𝜌𝐧 ⋅ (𝐮 ⋅ ∇𝐮) = 𝜌𝑅𝜔2 , (A.12) 𝑖=1
𝑖=𝑛 ∑
∑ 𝑗=𝑛
( )
and hence the boundary condition at the outer wall is given by, − 𝜌𝑎𝑖 (𝑡)𝑎𝑗 (𝑡)∇ ⋅ 𝝓𝐢 ⋅ ∇(𝝓𝐣 ) .
𝑖=1 𝑗=1
𝜕 𝑃̂ 𝜕2 𝑢
= 𝜇 2𝑛 − 𝜌𝑅𝜔2 . (A.13) (A.18)
𝜕𝑛 𝜕𝑛
For the top and bottom walls, 𝒏 = 𝐳̂ and 𝒏 = −̂𝐳 respectively, and since 𝝎 Due to the linearity of the Poisson equation (A.18), the solution
is also along the 𝑧 direction, the Coriolis term is zero, and the boundary 𝑃̂ (𝐱, 𝑡) can be decomposed as,
condition becomes, 𝑗=𝑛
𝑖=𝑛 ∑

𝑃̂ (𝐱, 𝑡) = 𝑃̂𝑖𝑗 (𝐱)𝑎𝑖 (𝑡)𝑎𝑗 (𝑡)
𝜕 𝑃̂ 𝜕2 𝑢
= −𝜌𝐧 ⋅ (𝐮 ⋅ ∇𝐮) + 𝜇 2𝑛 . (A.14) 𝑖=0 𝑖=0
𝜕𝑛 𝜕𝑛 (A.19)
𝑖=𝑛
∑ 𝑖=𝑛 ∑
∑ 𝑗=𝑛
The advection term in cylindrical coordinates is given by, = 𝑃̂00 (𝐱) + 2 𝑃̂𝑖0 (𝐱)𝑎𝑖 (𝑡) + 𝑃̂𝑖𝑗 (𝐱)𝑎𝑖 (𝑡)𝑎𝑗 (𝑡),
( ) 𝑖=1 𝑖=1 𝑖=1
𝜕𝑢 𝑢 𝜕𝑢 𝜕𝑢
𝜌𝐧 ⋅ (𝐮 ⋅ ∇𝐮) = 𝜌 𝑢𝑟 𝑧 + 𝜃 𝑧 + 𝑢𝑧 𝑧 = 0, (A.15)
𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑧 where 𝑃̂𝑖𝑗 are known as modified pressure components (Noack et al.,
because 𝑢𝑧 = 0 and 𝑢𝑟 = 0 at the top and bottom walls. Hence, the bound- 2005)). Similarly, the pressure components 𝑃𝑖𝑗 (𝐱) are defined so that,
ary condition reduces to that of stationary walls, 𝑗=𝑛
𝑖=𝑛 ∑

𝑝(𝐱, 𝑡) = 𝑃𝑖𝑗 (𝐱)𝑎𝑖 (𝑡)𝑎𝑗 (𝑡)
𝜕 𝑃̂ 𝜕2 𝑢
= 𝜇 2𝑛 . (A.16) 𝑖=0 𝑖=0
(A.20)
𝜕𝑛 𝜕𝑛 𝑖=𝑛 𝑗=𝑛
𝑖=𝑛 ∑
∑ ∑
The Poisson equation (A.6) and the boundary conditions (A.13) and = 𝑃00 (𝐱) + 2 𝑃𝑖0 (𝐱)𝑎𝑖 (𝑡) + 𝑃𝑖𝑗 (𝐱)𝑎𝑖 (𝑡)𝑎𝑗 (𝑡).
(A.16) can then be used to obtain 𝑃̂ relative to a reference pressure. In 𝑖=1 𝑖=1 𝑖=1

this study, we set the reference pressure at a point located at the tank From the definition (A.2), we have 𝑃𝑖𝑗 = 𝑃̂𝑖𝑗 for 𝑖, 𝑗 ≠ 0. For the mean
center line, at a distance 0.005T from the bottom wall. term with 𝑖 = 𝑗 = 0,
The Poisson equation is solved numerically using an adapted version
1
of the pressure correction solver of the Pantarhei code that is used for 𝑃̂00 = 𝑃00 − 𝜌(𝝎 × 𝒙)2 . (A.21)
the main simulations. The pressure is normalised by 0.5𝜌𝑣2𝑡𝑖𝑝 throughout, 2
following Lane et al. (2001). To check the numerical implementation, The pressure component due to time-average flow, 𝑃̂00 (𝐱), can be ob-
we compare the instantaneous pressure field obtained with the method tained from,
presented in this section to the one provided by the DNS; the results are
identical as shown in Fig. A.22. ∇2 𝑃̂00 = −𝜌∇ ⋅ (𝝓𝟎 ⋅ ∇(𝝓𝟎 )) − 𝜌∇ ⋅ (2𝝎 × 𝝓𝟎 ) + 𝜇∇ ⋅ ∇2 𝝓𝟎 , (A.22)
the linear pressure component 𝑃𝑖0 (𝐱) = 𝑃̂𝑖0 (𝐱) from,
A.3. Decomposition of pressure into pressure components
∇2 𝑃𝑖0 = ∇2 𝑃0𝑖 =
Substituting the truncated velocity POD expansion (2) into (A.6) we 1[ ] 1
get, − 𝜌∇ ⋅ (𝝓𝟎 ⋅ ∇(𝝓𝐢 )) + 𝜌∇ ⋅ (𝝓𝐢 ⋅ ∇(𝝓𝟎 )) − 𝜌∇ ⋅ (𝝎 × 𝝓𝐢 ) + 𝜇∇ ⋅ ∇2 𝝓𝐢 ,
2 2
𝑗=𝑛
𝑖=𝑛 ∑ 𝑖=𝑛 𝑖=𝑛 (A.23)
∑ ( ) ∑ ∑
∇2 𝑃̂ = − 𝜌𝑎𝑖 𝑎𝑗 ∇ ⋅ 𝝓𝐢 ⋅ ∇(𝝓𝐣 ) − 𝜌𝑎𝑖 ∇ ⋅ (2𝝎 × 𝝓𝐢 ) + 𝜇𝑎𝑖 ∇ ⋅ ∇2 𝝓𝐢 , and finally the quadratic component 𝑃𝑖𝑗 (𝐱) = 𝑃̂𝑖𝑗 (𝐱), where 𝑖, 𝑗 ≠ 0, from
𝑖=0 𝑗=0 𝑖=0 𝑖=0
(A.17) ∇2 𝑃𝑖𝑗 = ∇2 𝑃𝑗𝑖 = −𝜌∇ ⋅ (𝝓𝐢 ⋅ ∇(𝝓𝐣 )). (A.24)
where both indices 𝑖 and 𝑗 start from 0, and hence include the mean Finally, taking the time-average of (A.19) we get,
flow. The right hand side can be rearranged to reveal explicitly the
𝑖=𝑛

contributions from the mean velocity field and the terms that are linear
𝑃̂ (𝐱, 𝑡) = 𝑃̂00 (𝐱) + 𝑃𝑖𝑖 (𝐱)𝑎2𝑖 (𝑡), (A.25)
and quadratic with respect to 𝑎𝑖 (𝑡),
𝑖=1

15
K. Mikhaylov, S. Rigopoulos and G. Papadakis Chemical Engineering Science 279 (2023) 118881

because 𝑎𝑖 (𝑡)𝑎𝑗 (𝑡) = 0 (𝑖 ≠ 𝑗 and 𝑖, 𝑗 ≥ 1). Thus the pressure fluctuations Ducci, A., Yianneskis, M., 2007. Vortex tracking and mixing enhancement in stirred pro-
can be written in terms of the velocity POD coefficients 𝑎𝑖 (𝑡) as cesses. AIChE J. 53 (2), 305–315.
Ducci, A., Doulgerakis, Z., Yianneskis, M., 2008. Decomposition of flow structures in
𝑖=𝑛
∑ ∑ 𝑗=𝑛
𝑖=𝑛 ∑ stirred reactors and implications for mixing enhancement. Ind. Eng. Chem. Res. 47
𝑃 ′ (𝐱, 𝑡) = 𝑃 (𝐱, 𝑡) − 𝑃 (𝐱, 𝑡) = 2 𝑃𝑖0 (𝐱)𝑎𝑖 (𝑡) + 𝑃𝑖𝑗 (𝐱)𝑎𝑖 (𝑡)𝑎𝑗 (𝑡). (A.26) (10), 3664–3676.
𝑖=1 𝑖=1 𝑗=1 Guermond, J.L., Minev, P., Shen, J., 2006. An overview of projection methods for incom-
(𝑗≠𝑖) pressible flows. Comput. Methods Appl. Mech. Eng. 195 (44), 6011–6045.
Guzmán-Inigo, J., Sodar, M.A., Papadakis, G., 2019. Data-based, reduced-order, dynamic
A.4. Boundary conditions for the pressure components estimator for reconstruction of nonlinear flows exhibiting limit-cycle oscillations.
Phys. Rev. Fluids 4, 114703.
Hasal, P., Fort, I., Kratena, J., 2004. Force effects of the macro-instability of flow pattern
For the top and bottom walls the boundary condition for the modi-
on radial baffles in a stirred vessel with pitched-blade and Rushton turbine impellers.
fied pressure is Chem. Eng. Res. Des. 82 (9), 1268–1281.
Holmes, P., Lumley, J.L., Berkooz, G., Rowley, C.W., 2012. Turbulence, Coherent Struc-
𝜕 𝑃̂ 𝜕2 𝑢
= 𝜇 2𝑛 , (A.27) tures, Dynamical Systems and Symmetry, 2nd ed. Cambridge University Press, Cam-
𝜕𝑛 𝜕𝑛 bridge.
Hosseini, Z., Martinuzzi, R.J., Noack, B.R., 2015. Sensor-based estimation of the velocity
and substituting the pressure and velocity expansions we obtain,
in the wake of a low-aspect-ratio pyramid. Exp. Fluids 56 (1), 1–16.
∑ 𝑗=𝑛
𝑖=𝑛 ∑
𝜕 𝑃̂𝑖𝑗 𝑖=𝑛
∑ 𝜕 2 𝜙𝑖,𝑛 Janiga, G., 2019. Large-eddy simulation and 3D proper orthogonal decomposition of the
𝑎𝑖 (𝑡)𝑎𝑗 (𝑡) = 𝑎𝑖 (𝑡)𝜇 , (A.28) hydrodynamics in a stirred tank. Chem. Eng. Sci. 201, 132–144.
𝑖=0 𝑗=0
𝜕𝑛 𝑖=0 𝜕𝑛2 Lane, G.L., Rigby, G.D., Evans, G.M., 2001. Pressure distribution on the surface of Rush-
ton turbine blades-experimental measurement and prediction by CFD. J. Chem. Eng.
where 𝜙𝑖,𝑛 denotes the projection of the velocity POD mode 𝜙𝑖 on the Jpn. 34 (5), 613–620.
wall normal. The boundary condition for 𝑃̂00 is, Lavezzo, V., Verzicco, R., Soldati, A., 2009. Ekman pumping and intermittent particle
resuspension in a stirred tank reactor. Chem. Eng. Res. Des. 87 (4), 557–564.
𝜕 𝑃̂00 𝜕 2 𝜙0,𝑛 Mayorga, C., Morchain, J., Liné, A., 2022. Reconstruction of the 3D hydrodynamics in
=𝜇 . (A.29) a baffled stirred tank using Proper Orthogonal Decomposition. Chem. Eng. Sci. 248,
𝜕𝑛 𝜕𝑛2
117220.
As the right hand side of (A.28) contains only linear terms in 𝑎𝑖 (𝑡), Mikhaylov, K., Rigopoulos, S., Papadakis, G., 2021. Reconstruction of large-scale flow
for the relation to hold at all times, the boundary condition for the structures in a stirred tank from limited sensor data. AIChE J. 67 (10), e17348.
Nikiforaki, L., Montante, G., Lee, K., Yianneskis, M., 2003. On the origin, frequency and
quadratic component must be
magnitude of macro-instabilities of the flows in stirred vessels. Chem. Eng. Sci. 58
𝜕𝑃𝑖𝑗 (13), 2937–2949.
= 0, where 𝑖, 𝑗 ≠ 0. (A.30)
𝜕𝑛 Noack, B.R., Afanasiev, K., Morzyński, M., Tadmor, G., Thiele, F., 2003. A hierarchy of
low-dimensional models for the transient and post-transient cylinder wake. J. Fluid
As 𝑃0𝑖 = 𝑃𝑖0 , the boundary conditions for the linear components are Mech. 497, 335–363.
given by Noack, B.R., Papas, P., Monkewitz, P.A., 2005. The need for a pressure-term representa-
tion in empirical Galerkin models of incompressible shear flows. J. Fluid Mech. 523,
𝜕𝑃0𝑖 𝜕𝑃𝑖0 1 𝜕 2 𝜙𝑖,𝑛 339–365.
= = 𝜇 , where 𝑖 > 0. (A.31)
𝜕𝑛 𝜕𝑛 2 𝜕𝑛2 Picard, C., Delville, J., 2000. Pressure velocity coupling in a subsonic round jet. Int. J.
Heat Fluid Flow 21 (3), 359–364.
For the outer moving walls, the boundary condition for 𝑃 is Rave, K., Lehmenkühler, M., Wirz, D., Bart, H.-J., Skoda, R., 2021. 3d flow simulation of a
baffled stirred tank for an assessment of geometry simplifications and a scale-adaptive
𝜕𝑃 𝜕 2 𝑢𝑛
=𝜇 − 𝜌𝑅𝜔2 . (A.32) turbulence model. Chem. Eng. Sci. 231, 116262.
𝜕𝑛 𝜕𝑛2 Roy, S., Acharya, S., 2011. Perturbed turbulent stirred tank flows with amplitude and
mode-shape variations. Chem. Eng. Sci. 66 (22), 5703–5722.
For the linear and quadratic components the boundary conditions re-
Roy, S., Acharya, S., 2012. Effect of impeller speed perturbation in a Rushton impeller
main the same, i.e. (A.31) and (A.30), respectively. However, for the stirred tank. J. Fluids Eng. 134 (06), 061104.
mean term it must be modified to incorporate the effect of the moving Sani, R.L., Shen, J., Pironneau, O., Gresho, P.M., 2006. Pressure boundary condition for
wall, i.e. the time-dependent incompressible Navier–Stokes equations. Int. J. Numer. Methods
Fluids 50 (6), 673–682.
𝜕 𝑃̂00 𝜕 2 𝜙0,𝑛 Scargiali, F., Tamburini, A., Caputo, G., Micale, G., 2017. On the assessment of power
=𝜇 − 𝜌𝑅𝜔2 . (A.33) consumption and critical impeller speed in vortexing unbaffled stirred tanks. Chem.
𝜕𝑛 𝜕𝑛2
Eng. Res. Des. 123, 99–110.
These boundary conditions match those found in the literature for 𝜔 = 0, Schmid, P.J., 2010. Dynamic mode decomposition of numerical and experimental data.
see Noack et al. (2005). J. Fluid Mech. 656, 5–28.
Steiros, K., Bruce, P.J.K., Buxton, O.R.H., Vassilicos, J.C., 2017. Power consumption and
form drag of regular and fractal-shaped turbines in a stirred tank. AIChE J. 63 (2),
References
843–854.
Taira, K., Brunton, S.L., Dawson, S.T., Rowley, C.W., Colonius, T., McKeon, B.J., Schmidt,
Arndt, R.E.A., Long, D.F., Glauser, M.N., 1997. The proper orthogonal decomposition of O.T., Gordeyev, S., Theofilis, V., Ukeiley, L.S., 2017. Modal analysis of fluid flows: an
pressure fluctuations surrounding a turbulent jet. J. Fluid Mech. 340, 1–33. overview. AIAA J., 4013–4041.
Assirelli, M., Bujalski, W., Eaglesham, A., Nienow, A., 2002. Study of micromixing in a Tamburini, A., Gagliano, G., Micale, G., Brucato, A., Scargiali, F., Ciofalo, M., 2018. Direct
stirred tank using a Rushton turbine: comparison of feed positions and other mixing numerical simulations of creeping to early turbulent flow in unbaffled and baffled
devices. Chem. Eng. Res. Des. 80 (8), 855–863. stirred tanks. Chem. Eng. Sci. 192, 161–175.
Assirelli, M., Bujalski, W., Eaglesham, A., Nienow, A., 2005. Intensifying micromixing in Tamura, Y., Suganuma, S., Kikuchi, H., Hibi, K., 1999. Proper orthogonal decomposition
a semi-batch reactor using a Rushton turbine. Chem. Eng. Sci. 60 (8–9), 2333–2339. of random wind pressure field. J. Fluids Struct. 13 (7), 1069–1095.
Assirelli, M., Bujalski, W., Eaglesham, A., Nienow, A.W., 2008. Macro- and micromixing Taylor, J.A., Glauser, M.N., 2004. Towards practical flow sensing and control via POD
studies in an unbaffled vessel agitated by a Rushton turbine. Chem. Eng. Sci. 63 (1), and LSE based low-dimensional tools. J. Fluids Eng. 126, 337–345.
35–46. Towne, A., Schmidt, O.T., Colonius, T., 2018. Spectral proper orthogonal decomposition
Basbug, S., 2017. Flow field and mixing in stirred vessels with regular and fractal im- and its relationship to dynamic mode decomposition and resolvent analysis. J. Fluid
pellers. PhD thesis. Imperial College London. Mech. 847, 821–867.
Borée, J., 2003. Extended proper orthogonal decomposition: a tool to analyse correlated Van Overschee, P., De Moor, B., 1994. N4SID: subspace algorithms for the identification
events in turbulent flows. Exp. Fluids 35 (2), 188–192. of combined deterministic-stochastic systems. Automatica 30 (1), 75–93.
Davidson, P.A., 2015. Turbulence: An Introduction for Scientists and Engineers. Oxford
University Press.

16

You might also like