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

net/publication/327059961

Solid-Liquid Mixing in Stirred Vessels: Numerical Simulation, Experimental

Validation and Suspension Quality Prediction Using Multivariate Regression
and Machine Learning

Technical Report · August 2018

DOI: 10.13140/RG.2.2.11074.84164/1

CITATION READS

1 1,435

2 authors, including:

Corentin Macqueron
ORANO (ex-AREVA)
35 PUBLICATIONS 35 CITATIONS

SEE PROFILE

All content following this page was uploaded by Corentin Macqueron on 27 November 2020.

The user has requested enhancement of the downloaded file.

 Solid-Liquid Mixing in Stirred Vessels: Numerical Simulation,
Experimental Validation and Suspension Quality Prediction Using
Multivariate Regression and Machine Learning

Corentin MACQUERON
ORANO Projets
1 rue des hérons, 78180, Montigny-le-Bretonneux
[email protected]

Abstract
A two-phase solid-liquid stirred vessel modeling is performed with the FLUENT 15.0 software and
validated by comparison to experimental data (power number and solid volume fraction profiles). The
mixing quality, defined by the RSD (Relative Standard Deviation of the solid volume fraction), is studied
as a function of the main system parameters: stirring speed, vessel diameter, pitch blade angle, liquid
dynamic viscosity, and diameter, density and average volume fraction of the particles. On the basis of the
~300 numerical simulations performed with the experimentally validated model, a correlation to predict
the RSD as a function of the different parameters with an accuracy of ±10 points of RSD (expressed in
percentage) is proposed. Advanced machine learning techniques such as artificial neural networks can
achieve an accuracy an order of magnitude better. Finally, it is shown that the correlation, built on a
specific vessel and agitator design, can be applied as a pre-sizing to other designs with a reasonable
accuracy, provided those designs are of the same type (newtonian axial downward pumping with D/T = ½
and presence of baffles).

Keywords
Solid-Liquid – Mixing – Suspension – Two-phase – Correlation – Machine Learning – Deep Learning –
Neural Network – RSD – Euler Granular – Gidaspow – Brucato

1. Introduction
The solid-liquid mixing quality in a stirred tank as a function of the main system parameters (tank size
and design, stirrer design, suspension characteristics) is an important matter for the industry. Correlations
exist to predict the minimal stirring speed for bottom lifting or avoidance of settling (Zwietering, 1958;
Mersmann et al., 1998) but these correlations are only meant for ‘low’ speeds where the mixing quality is
generally poor. For ‘high’ speeds required to obtain a good mixing quality (defined by the RSD: Relative
Standard Deviation of the solid volume fraction), there are only scarce elements in the literature and the
construction of a general correlation is considered to be a difficult task (Kresta, 2016). To determine the
stirring configuration allowing for a high suspension quality, experimentations or simulations are hence
required. Real size experimentations are often not an option for different reasons, especially for very large
tanks. Small scale experimentations can help, but scale-up laws are not always clearly established
(Geisler et al., 1993; Montante et al., 2003; Dickey, 2013). Simulation is free from experimental
constraints, but is still an onerous, difficult and cautious tool. In this context, ORANO (formerly
AREVA) validated solid-liquid stirred tanks simulations performed with the FLUENT 15.0 software
(ANSYS, 2013b) by comparison to experimental results produced by dedicated mock-ups in the
framework of a preliminary study (Macqueron et al., 2015). In the present report, ORANO continues the
validation task by comparison to experimental results from the literature (Tsz-Chung Mak, 1992) and
present a numerical study to build a correlation aimed to predict the suspension quality as a function of
the main system parameters. This correlation is based on ~300 numerical simulations ran on the Tsz-
Chung Mak setup and is then tested against several other tank and impeller designs.

This technical report is an English follow-up of a previous study published in French in SFGP
(Macqueron and William, 2017).

1
2. Experimental setup
The experimental setup chosen to validate the simulations and to build the correlation is the one from
Tsz-Chung Mak because of its simplicity and because it is representative of many ‘typical’ industrial
stirring systems. The main tank characteristics are the following:

 diameter: T = 0.61 m
 height: H = T
 clearance: C = T/4
 torispherical base (R = T, r = T/10)
 4 baffles
 volume : 0.165 m3

Figure 1. Tank and impeller schematic

The main impeller characteristics (4 pitch blades turbine at 45°, single stage) are shown in table 1 and
figure 1. The power number Np is measured at 1.52 (with the mean suspension density).
Table 1. Impeller characteristics

D (m) W (m) χ (m) Hb (m) Hod (m)

3.10E-01 8.60E-02 3.40E-03 6.10E-02 5.70E-02

Tap water at ambient temperature is used as the liquid phase. The solid phase characteristics are the
following (glass (sand) beads, sphericity: 0.8):

 density: 2600-2650 kg/m3

 diameter: 150-210 µm (median value: 180 µm)
 concentration: 14% vol. (30% mass.)

The power is measured with a torque-meter (accuracy: ± 2%). The solid volume fraction φ is measured
with a conductivity probe shown on figure 2. This probe measures the solid concentration between 5%
and 50% vol. and is capable to detect a 1% variation.

PROBE

Figure 2. Conductivity probe at a T/6 distance to the wall

and at T/6.1, T/3.1, T/2, T/1.5 and T/1.2 from the surface

2
The solid volume fraction is measured at the 5 locations shown in red on figure 2. The experimental RSD
is calculated with these 5 measures according to the following formula (1):
1
1 1
𝑅𝑆𝐷 = ( ∑𝑛𝑖=1(𝜑𝑖 − 𝜑𝑚𝑒𝑎𝑛 )2 )2 (with n = 5) (1)
𝜑𝑚𝑒𝑎𝑛 𝑛−1

3. Modeling
3.1 Geometry and mesh

The geometry is built with the DesignModeler 15.0 software (ANSYS, 2013a) (figure 3). Because of the
symmetries, only one quarter of the tank is modeled. The mesh (~800 k cells) is built with the ANSYS
Meshing 15.0 software (ANSYS, 2013c) with boundary layers refinements on all the walls (5 layers, first
layer thickness: 1 mm) to ensure y+ (non-dimensional wall distance) values compatible with the selected
turbulence model and wall functions (mean y+ value: ~35, max y+ value: ~130) (figure 3).

Figure 3. Tank geometry and mesh

3.2 Hypotheses, physical models and numerical parameters

The main hypotheses, physical models and numerical parameters are the following:

 the system is in steady-state

 the impeller rotation is modeled with the Multiple Reference Frame (MRF) method (‘frozen
rotor’) (ANSYS, 2013b)
 turbulence is modeled with the k-ε realizable mixture approach with the enhanced wall
treatment option (ANSYS, 2013b)
 the solid phase is modeled with the ‘eulerian granular’ approach, with the Syamlal-O’Brien, Lun
et al. and Schaeffer models for granular viscosity (ANSYS, 2013b)
 the particles are perfectly spherical and their diameter is equal to the median value (180 µm)
 the particles density is equal to 2630 kg/m3 as assumed by Tsz-Chung Mak
 the tap water viscosity is 0.001 Pa.s (newtonian fluid) and its density is 1000 kg/m 3 (values at
~20°C (Sacadura, 2009))
 the drag coefficient is modeled with the Gidaspow approach (Gidaspow, 1994, ANSYS, 2013b).
No turbulent correction, such as the Brucato correction (Brucato et al., 1998), is taken into
account (see § 4.2.7 for details). The Gidaspow approach is compared to other approaches in
§ 4.2.8
 lift is neglected
 the turbulent dispersion force (TDF) is modeled with the Diffusion-in-VOF approach (ANSYS,
2013b)
 the probe is not modeled (it is assumed to be non-intrusive)

3
  the free surface is a plane wall with a zero shear stress
 the pressure-velocity coupling is the Coupled approach (ANSYS, 2013b)
 the discretization scheme is QUICK (ANSYS, 2013b)
 the system is solved using the pseudo-transient approach with automatic time-step (ANSYS,
2013b)

4. Simulations
4.1 Validation
4.1.1 Power number

The power numbers determined with the simulations using the calculated impeller torque are given in
table 2. The calculated values are very close to the measured values (maximal deviation is 2% taking into
account the measurement accuracy). The simulations appeared validated for dissipated power prediction.

Table 2. Power number

N (RPM) 60 90 120 150 180 210 240 270 300

Np (-) 1.51 1.51 1.52 1.53 1.56 1.58 1.58 1.58 1.58
Deviation from experiments -1% -1% 0% 1% 3% 4% 4% 4% 4%

4.1.2 Solid volume fraction

The solid volume fraction profiles from the simulations are compared to the measurements on figures 4
and 5. Some deviations are visible but the simulations appear to be capable to correctly reproduce the
suspension variations as a function of the stirring speed, both qualitatively and quantitatively. The
calculated RSD is very close to the measurements (maximal deviation is 5 points of RSD in %) as shown
on figure 6. The numerical results are less conclusive at ‘low’ speed (60 RPM) (figure 4). At such ‘low’
speed, deactivating the turbulent dispersion force (TDF), as proposed by Tamburini et al. (2014), allows
for a considerably better numerical prediction (figures 4 and 6). This 60 RPM speed being close and
smaller to the avoidance of settling speed proposed by Mersmann et al. (NAS (« Avoidance of Settling »),
72 RPM here), it seems that this NAS speed could be seen as a ‘speed criteria’ under which the turbulent
dispersion force should be deactivated. The ‘low speed’ regime thus appears different from the ‘high
speed’ regime (Tamburini et al., 2014). Further investigations should be performed but, from an
engineering point of view, the simulation can be considered validated as long as one stays in the ‘high
speed’ regime. The present report only aims at ‘good quality’ and ‘high speed’ predictions, so the
turbulent dispersion force will always be activated in the following. An example of a solid volume
fraction field obtained by the simulation is shown on figure 6.

Figure 4. Simulated and measured solid volume fraction profiles – 60 RPM

(with turbulent dispersion force on the left, without turbulent dispersion force on the right)

4
 Figure 5. Simulated and measured solid volume fraction profiles.
Left to right, top to bottom: 90, 120, 150, 180 RPM

Measurements

Simulations without TDF under NAS

Simulations with TDF

Figure 6. RSD versus stirring speed on the left (%).

Example of a solid volume fraction field (-) on the right (120 RPM)

4.2. Suspension quality

4.2.1. Tank diameter

Some results of the scale-up studies are shown on figure 7. The dependency in 0.814 exponent of the tank
diameter is close to the one found by Zwietering (0.85).

0.814

Figure 7. Tank diameter dependency

5
4.2.2 Particle diameter

Some results of the particle diameter studies are shown on figure 8. The dependency in -0.626 exponent
of the particle diameter is notably different from the one of Zwietering (-0.2). This highlights the
differences between the ‘low’ and ‘high’ speed physics.

0.626

Figure 8. Particle diameter dependency

4.2.3. Particle density

Some results of the particle density studies are shown on figure 9. The results are shown as a function of
the density difference over the liquid density, as in the Zwietering and Mersmann et al. formulae. The
dependency in -0.445 exponent of the particle diameter is very close to the one of Zwietering (-0.45).

0.445

Figure 9. Particle density dependency

4.2.4. Liquid viscosity

Some results of the liquid dynamic viscosity studies are shown on figure 10. The dependency in 0.171
exponent of the particle diameter is notably different from the one of Zwietering (0.1). This highlights the
differences between the ‘low’ and ‘high’ speed physics.

0.171

Figure 10. Liquid dynamic viscosity dependency

6
4.2.5. Solid volume fraction

Some results of the mean solid volume fraction studies are shown on figure 11. Figure 11 also shows that
the RSD variation reaches a maximum (‘bell curve’). The volume fraction for which this maximum is
reached is somewhat variable but is always around 7.5%. This dependency is here approximated by an
order 6 polynomial function (formula 11) and is very different from the one found by Zwietering which is
in -0.13 exponent. This, again, highlights the differences between the ‘low’ and ‘high’ speed physics.

Volume fraction (-)

Figure 11. Mean solid volume fraction dependency

A possible explanation of this somewhat surprising non-monotonous dependency could be an antagonistic

effect of the solid concentration. At ‘low’ concentrations (<~7.5%) the particles scarcely interact and the
more particles there are, the more difficult it is to mix them. At ‘high’ concentrations (>~7.5%), the
particles interact a lot because there are many of them, and these interactions increase the apparent fluid
viscosity, which is beneficial to the mixing homogeneity (figure 11). Additional tests not presented here
show that, above 7.5%, it is possible to correlate the RSD with the apparent viscosity calculated with the
Krieger and Dougherty formula (1959), which is an indication that the two parameters may indeed be
linked. It should nonetheless be noted that the terminal velocity of a particle in a still fluid is a strictly
monotonous function of the volume fraction (Richardson and Zaki, 1954), but still sedimentation and
stirred tanks are not the same. Further numerical and experimental investigations should be performed in
order to better understand this behaviour and to confirm or disconfirm the ‘bell curve’.

4.2.6. Pitch blade angle

The pitch blade angle results are shown on figure 12. The RSD dependency to the pitch blade angle being
quite weak when compared to the other parameters studies in this report, it is neglected in the correlation
detailed in § 5.1. The pitch blade angle has nonetheless a very important impact on the power
consumption, as shown in table 3. There must be an optimum that should be looked at.

0.2

Figure 12. Pitch blade angle dependency

Table 3. Power number

θ (deg) 30 35 40 45
Np (-) ~0.70 ~0.95 ~1.24 ~1.56

7
4.2.7. Turbulence, drag coefficient and the Brucato correction

When the turbulence is strong, the drag coefficient on ‘large’ particles is increased as observed in several
studies (Brucato et al., 1998; Pinelli et al., 2001; Montante et al., 2004). The so-called ‘Brucato’ and
‘Magelli’ correlations have been proposed to modify the drag coefficient to account for this effect. These
correlations appear to greatly improve the simulations accuracy in the aforementioned studies, whereas it
seems to have some flaws in other configurations and appear to be non-universal. For instance, Kasat et
al. (2008) used the Brucato correction but had to diminish its effect by a factor of ten to obtain correct
results. We were also unable to completely reproduce the results from Montante et al. (2004) and Liu and
Barigou (2013). In the latter case, we could accurately reproduce the experimental results for the 1 mm
particles configurations for which the Brucato correction has little to no effects, but we were unable to
reproduce the experimental results for the 3 mm particles configurations for which the Brucato correction
has major effects: the correction seems to increase the drag coefficient by an over-estimated factor,
resulting in over-suspended particles, especially in the lower part of the tank. Following the Kasat et al.
approach (2008), we tried to diminish the value of the ‘Brucato constant’. This approach had a great
impact on the results, but no value of the Brucato constant was able to produce reasonable results (figure
13). In our previous study (Macqueron and William, 2017), we had activated the Brucato correction for
all our calculations, but this had little to no effects in the tested configurations. In the present follow-up
report, with larger particles (up to 1750 µm), the Brucato correction has an important impact. We decided
to deactivate it, because, based on our observations, we were unsure of the correctness of its effects.
Furthermore, it tends to over-suspend the particles and hence to over-estimate the mixing quality of the
suspension, so we decided to stay on the safe side of the prediction. Further investigations should be
performed in order to better understand this behaviour.

Figure 13. Brucato constant (K) dependency

4.2.8. The Gidaspow drag correlation

It is sometimes said that the Euler-Granular approach is not significantly better than the simpler Euler-
Euler approach. Our sensitivity studies indeed showed that it is rather the drag coefficient correlation that
is of crucial importance to accurately mimic the physics of the solid suspension. The Gidaspow drag
correlation proved to be significantly better than other approaches (figure 14) and, as it is only available
in the Euler-Granular approach in FLUENT 15.0, we had to use the Euler-Granular approach even if it
was not for the Granular approach per se.

8
 Figure 14. Drag coefficient sensitivity studies.
(Left: volume fraction profiles for different drag approaches, center: volume fraction field for Euler-Granular
Schiller-Naumann, right: volume fraction field for Euler-Granular Gidaspow)

5. Prediction
The mixing quality (RSD) has been studied by performing ~300 numerical simulations with the validated
model. Some of these simulations are shown in § 4 and the total span of the study is the following:

 Stirring speed (N), from ~N AS to ~5 NAS

 Tank diameter (T), from 0.61 m to 6.50 m
 Pitch blade angle (θ), from 30° to 45°
 Particles diameter (Dp), from 40 µm to 1750 µm
 Solid-liquid relative density (Δρ/ρ), from 0.74 (ρS = 1740 kg/m3 (magnesium)) to 17.95
(ρS = 18 950 kg/m3 (uranium))
 Mean solid volume fraction (φ), from ~0.5% to 30%
 Liquid dynamic viscosity (μ), from 4.74.10-4 Pa.s to 1.7.10-3 Pa.s (newtonian fluid)

From now on, the RSD is calculated on the entire volume of the tank with the following formula (2):

1 1
𝑅𝑆𝐷 = √ ∭(𝜑(𝑉) − 𝜑𝑚𝑒𝑎𝑛 )2 𝑑𝑉 (2)
𝜑𝑚𝑒𝑎𝑛 𝑉

Among the ~300 calculations, roughly half has been performed with the configuration described in § 2 by
modifying only one parameter at a time to help visualize the variations, shown in § 4. The other half has
been performed by modifying several or all the parameters at a time in order to maximize the size of the
studied domain and to take into account the possible interactions between the parameters.

5.1 Multivariate regression

With the ~300 numerical simulation results, a correlation (formula 3) has been built by multivariate
regression (least square error) in order to predict the mixing quality. It appears that the multivariate
dependency of the RSD can be approximated with a reasonable accuracy by the product of monovariate
functions (formulae 6 to 12):
𝜔
𝑅𝑆𝐷 = 𝛼𝑒 𝛾𝛽 (in %) (3)

With:
𝛼 = 144,022 (4)
𝛾 = −3.36 × 10−3 (5)
𝜔 = 1.364 (6)
𝛥𝜌
𝛽 = 𝑁 × 𝑓1 (𝑇) × 𝑓2 (𝐷𝑝 ) × 𝑓3 ( ) × 𝑓4 (𝜇) × 𝑓5 (𝜑) (7)
𝜌
0.840
𝑓1 (𝑇) = 𝑇 (8)
𝑓2 (𝐷𝑝 ) = 𝐷𝑝 −0.702 (9)

9
 𝛥𝜌 𝛥𝜌 −0.449
𝑓3 ( ) = ( ) (10)
𝜌 𝜌
0.203
𝑓4 (𝜇) = 𝜇 (11)
𝑓5 (𝜑) = 69574.98𝜑 6 − 68045.96𝜑 5 + 26995.72𝜑 4 − 5516.34𝜑 3 + 610.77𝜑 2 − 33.42𝜑 + 1.36 (12)

Figure 15 shows the RSD values from the simulations as a function of the β variable.

Figure 15. Comparison between correlation predictions and 3D simulations

This correlation has been built on the Tsz-Chung Mak design only. Eventhough this design is a very
common one and makes the correlation widely applicable, its validity for other tank and impeller designs
is necessarily brought into question. In order to assess our correlation robustness, it was tested against
several other tank and impeller designs, shown on figure 16. All these designs have in common baffles, a
D/T = ½ ratio and a downward pumping axial flow, but they are different in terms of overall designs,
power numbers, pitch blade angle, blade numbers and H/T ratios. In these ~50 additional configurations,
the tank diameter is comprised between 0.69 m and 6.5 m, the particle diameter is comprised between
40 µm and 1000 µm, the mean solid volume is comprised between ~0.5% and 12%, the particle density is
comprised between 2500 kg/m3 and 4500 kg/m3, the liquid density is comprised between 992 kg/m3 and
1150 kg/m3 and the liquid dynamic viscosity is comprised between 6.5x10-4 Pa.s and 1.1x10-3 Pa.s.

Figure 16. Designs from left to right: HRB-1 – HRB-2 – RCB – HRB-3 – LIU
Power numbers from left to right: 0.62 – 1.27 – 0.28 – 0.18 – 1.6
H/T from left to right: 0.68 – 1.1 – 1.23 – 0.68 – 1

The numerical results obtained in these configurations are plotted against the β variable defined in § 5.1,
without any additional fitting (figure 17). The results are somewhat dispersed but are still in very good
agreement with our proposed correlation which can hence be considered validated for pre-sizing mixing
calculations for any newtonian downward pumping axial systems with D/T = ½ and baffles. Lower and
upper estimation curves are provided.

10
 Figure 17. Validation of the correlation for several tank and impeller designs
𝐿𝑜𝑤𝑒𝑟 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑖𝑜𝑛 = 500000 × 𝛽 −2,4

𝑈𝑝𝑝𝑒𝑟 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑖𝑜𝑛 = 700000 × 𝛽 −2,2 + 5

5.2 Machine Learning

After the multivariate regression presented in § 5.1, we decided to go further by using machine learning
techniques. (Technically, the aforementioned regression already fall in the ‘machine learning’ domain, in
its broader sense, but we prefer to use the term ‘machine learning’ for more advanced techniques.)

We hence built machine learning regressors such as k-Nearest-Neighbors, Support Vector Machine,
Decision Tree, Random Forest, Bagging, AdaBoost, Gradient Boosting, eXtreme Gradient Boosting and
Gaussian Process regressors using Scikit Learn (Pedregosa et al., 2011). We also built artificial neural
networks which appeared to be especially efficient using Keras (Chollet, 2015) and Deep Float
(Macqueron and Fragnaud, 2017).
Training, hyperparameters tuning and overfitting avoidance was performed by splitting the data into three
sets: a training set (75% of the data) to ‘teach’ the models, a validation set (25% of the data) for blind test
and hyperparameters tuning and a final set (5% of the data) for final model performance-overfitting
evaluation. For neural networks, early stopping combined with callback of the best fit is a major tool to
avoid overfitting.

The hyperparameters of the models were chosen by performing a parameter grid search and by choosing
the model with the lowest mean squared error on the validation set among the models having a
validation/training error ratio below a threshold equal to 5. The validation set was never given to the
models during the learning process, but as the hyperparameters were chosen to have the lowest mean
squared error on the validation set, some form of indirect and implicit ‘data leaking’ is introduced into the
process and the models could perform unreasonably well on the validation set and still generalize poorly
on completely unseen data. The final set, never seen by the models during both training and
hypeparameters tuning phases, ensures the generalization capabilities of the models.

In total, more than 200 000 machine learning models were tested. Figures 18 to 29 show the best model
for each type of regressor (regressor predictions vs FLUENT results, error histogram with gaussian fitting
showing the mean error (μ), the mean absolute error (MAE) and the standard deviation of the error (σ)).
Each model is compared to the linear model which serves as the reference model.

11
 Figure 18. Linear regressor

Figure 19. Polynomial regressor

Figure 20. Support Vector Machine regressor

12
Figure 21. k-Nearest-Neighbors regressor

Figure 22. Gaussian Process regressor

Figure 23. Decision Tree regressor

13
Figure 24. Random Forest regressor

Figure 25. Bagging regressor

Figure 26. AdaBoost regressor

14
 Figure 27. Gradient Boosting regressor

Figure 28. eXtreme Gradient Boosting regressor

Figure 29. Neural Network regressor

15
Figure 30 shows the performance of each model. It is clear that the neural network approach outperforms
every other machine learning regressor in this study (its accuracy is an order of magnitude better than the
correlation’s one). Having more neurons in the hidden layer (512) than the number of cases (~300) might
seem specious and prone to overfitting, but the very good performance on both validation and final sets
proves otherwise. Here, overfitting is avoided by early stopping during training and callback of the
optimal fit.

Figure 30. Regressor performance (comparison to the linear model which serves as the reference model)
(‘custom fit’: our correlation (§ 5.1))

The neural network generalization capabilities to other designs is however poorer than the correlation,
probably because, being able to learn much more accuracy and being only trained on one design, the
neural network becomes much more design-dependent than the correlation. When trained with the data
from all the designs, its mean accuracy deteriorates but its generalization capabilities increase and finally
outperform the correlation (figure 31).

Figure 31. Neural network performance on all the designs (scaled)

(final set not shown here but comparable to validation set)

Neural networks are prone to behave erratically outside of the domain they were trained on (figure 32
(Lohninger, 1999)). We hence tested our neural network inside and outside of its training domain to look
for possible unrealistic behaviour. Figure 33 shows the behaviour of the model around its convex hull
training domain center, variable by variable. Negative values are not possible and some behaviours are
questionable, but at least the neural network appears capable to keep a consistent behaviour for most

16
variables for both interpolation and extrapolation. Other regressors show clear signs of inconsistency
(even inside their training domain) (figure 34).

Whereas it would be wrong to draw any general conclusion, it is clear that, in our case, the neural
network approach completely outperforms the other tested machine learning techniques in terms of mean,
max and standard deviation error and interpolation-extrapolation behaviour.

Figure 32. An example of a neural network erratic behaviour in extrapolation (Lohninger, 1999)

Figure 33. Neural network behaviour inside and outside its training domain (inside: center square)

Figure 34. Various regressors behaviours inside and outside their training domain (inside: center square)

17
6. Conclusion
It is shown in this report that FLUENT 15.0 solid-liquid simulations in stirred tanks can accurately
reproduce dissipated power and solid volume fraction profiles when the stirring speed is above the
‘avoidance of settling’ speed as defined by Mersmann et al. (1998). Under this criterion, the simulations
deteriorates but can still produce good results by deactivating the turbulent dispersion force, as also
shown by Tamburini et al. (2014).

The mixing quality of the suspension was studied as a function of the main system parameters: stirring
speed, tank diameter, pitch blade angle, liquid viscosity, particles diameter, particles density and mean
solid volume fraction (~300 simulations with a validated model). Based on these results, a correlation
predicting the suspension quality defined by the RSD (Relative Standard Deviation, expressed in %) as a
function of the aforementioned parameters was built. It is also shown that this correlation, built for a
particular yet common tank and impeller design can produce relevant results for pre-sizing with other
designs, provided they are of the same type (downward pumping, newtonian liquid, with baffles and D/T
= ½ ratio). With advanced machine learning techniques such as neural networks the accuracy of the
prediction can be even better.

7. Perspectives
Further numerical and experimental investigations should be performed in order to better understand the
transition between ‘low’ and ‘high’ speed regimes, the effect of the turbulence on the drag coefficient
(‘Brucato correction’), the influence of the mean solid volume fraction (‘bell curve’) and to optimize the
tank and impeller design in order to obtain the best suspension quality with the minimal dissipated power.

Nomenclature

ADAM ADAptive Momentum

AI Artificial Intelligence
TDF Turbulent Dispersion Force
QUICK Quadratic Upstream Interpolation for Convective Kinematics
ReLU Rectified Linear Unit
RSD Relative Standard Deviation (-)
C clearance (m)
D impeller diameter (m)
Dp particle diameter (m)
H tank height (m)
Hb hub height (m)
Hod hub diameter (m)
MAE Mean Absolute Error (-)
N stirring speed (RPM, RPS)
NAS avoidance of settling stirring speed (RPM, RPS)
Np power number (-)
R, r curvature radius (m)
T tank diameter (m)
W blade width (m)
y+ non-dimensional wall distance (-)
α, β, γ correlation coefficients
φ solid volume fraction (-)
χ blade thickness (m)
μ dynamic viscosity or mean error (Pa.s) or (-)
ρS particle density (kg/m3)
ρL fluid density (kg/m3)
σ standard deviation of the error (-)
θ pitch blade angle (°)

18
 References
Brucato, A., Grisafi, F., Montante, G. (1998). Particle drag coefficients in turbulent fluids. Chemical Engineering
Science, vol. 53, n° 18, pp. 3295-3314.
Chollet, F. (2015). Keras: Deep Learning Library for Theano and TensorFlow. https://keras.io/
DesignModeler 15.0. (2013). Canonsburg: ANSYS.
Dickey, D. S. (2013). Ten Things You May Not Know About Liquid Mixing Scaleup. Chemical Engineering. New
York: McGraw Hill.
Fluent 15.0. (2013). Canonsburg: ANSYS.
Gidaspow, D. (1994). Multiphase Flow and Fluidization: Continuum and Kinetic Theory Description. Boston:
Academic Press.
Geisler, R. K., Buurman, C., Mersmann, A. (1993). Scale-up of the necessary power input in stirred vessels with
suspension, The Chemical Engineering Journal, vol. 51.
Hicks, T. G., Chopey, N. P. (2012). Handbook of Chemical Engineering Calculations. McGraw Hill.
Kasat, G. R., Khopkar, A. R., Ranade, V. V., Pandit, A. B. (2008). CFD simulation of liquid-phase mixing in solid-
liquid stirred reactor, Chemical Engineering Science, vol. 63, pp. 3877-3885.
Kresta, S. M., Etchells III, A. W., Dickey, D. S., Atiemo-Obeng, V. A. (2016). Advances in Industrial Mixing. A
Companion to the Handbook of Industrial Mixing. Hoboken: Wiley. p. 386.
Krieger, I. M., Dougherty, T. J. (1959). A mechanism for non-newtonian flow in suspensions of rigid spheres.
Transaction of the Society of Rheology, vol. 3, pp.137-152.
Liu, L., Barigou, M. (2013). Numerical modelling of velocity field and phase distribution in dense monodisperse
solid-liquid suspensions under different regimes of agitation: CFD and PEPT experiments. Chemical Engineering
Science.
Lohninger, H. (1999). Teach/Me Data Analysis. Berlin-New York-Tokyo: Springer-Verlag. ISBN 3-540-14743-8
Macqueron, C., William, J., Le Gall, R., Demarthon, R., Piot, G., Ragouilliaux, A. (2015). Waste homogenisation
tank: experimental validation of a numerical study. In: Proceedings of Global. Paris: SFEN, paper 5204.
Macqueron, C., William, J. (2017). Suspension diphasique liquide-solide en cuve agitée: une corrélation de prédiction
de la qualité du mélange sur la base de simulations numériques validées sur mesures expérimentales. Récents
Progrès en Génie des Procédés, n° 110, ISSN: 1775-335X, ISBN: 978-2-910239-85-5. Paris: SFGP.
Macqueron, C., Fragnaud, S. (2017). Deep Float: An Easy Deep Learning Tool in Your Internet Browser. Research
Gate. https://deep-float.herokuapp.com
Mersmann, A., Werner, F., Maurer, S., Bartosch, K. (1998). Theoretical prediction of the minimum stirrer speed in
mechanically agitated suspensions. Chemical Engineering and Processing, vol. 37 (issue 6), pp. 503-510.
Meshing 15.0. (2013). Canonsburg: ANSYS.
Montante, G., Pinelli, D., Magelli, F. (2003). Scale-Up Criteria for the Solids Distribution in Slurry Reactors Stirred
with Multiple Impellers. Chemical Engineering Science, vol. 58, pp. 5363-5372.
Montante, G., Bakker, A. (2004). Solid-Liquid Multiphase Flow Validation in Tall Stirred Vessels with Multiple
Impeller Systems. Fluent Technical Note 253.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss,
R., Dubourg, V. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, vol. 12,
pp. 2825-2830. https://scikit-learn.org/stable/
Pinelli, D., Nocentini, M., Magelli, F. (2001). Solids distribution in stirred slurry reactors: influence of some mixer
configurations and limits to the applicability of a simple model for predictions. Chem. Eng. Comm., 118, pp. 91-
107.
Richardson, J. F., Zaki, W. N. (1954). Sedimentation and fluidisation: part 1, Transactions of the Institution of
Chemical Engineers, vol. 32, pp. 35-41
Sacadura, J. F. (2009). Initiation aux transferts thermiques. 8ème édition. Paris: Tec&Doc, p. 430.
Tamburini, A., Brucato, A., Cipollina, A., Micale, G., Ciofalo, M. (2012). CFD Predictions of Sufficient Suspension
Conditions in Solid-Liquid Agitated Tanks. International Journal of Nonlinear Sciences and Numerical Simulation,
vol. 13 (issue 6), pp. 427-443.
Tamburini, A., Cipollina, A., Micale, G., Brucato, A., Ciofalo, M. (2014). Influence of drag and turbulence modelling
on CFD predictions of solid liquid suspensions in stirred vessels. Chemical Engineering Research and Design, vol.
92 (issue 6), pp. 1045-1063.
Tsz-Chung Mak, A. (1992). Solid-liquid mixing in mechanically agitated vessels. PhD. University of London.
Zwietering, T. N. (1958). Suspending of solid particles in liquid by agitators. Chemical Engineering Science, vol. 8
(Issues 3-4), pp. 244-253.

19

View publication stats