Optimisation of Mixing Performance of Helical Ribbon Mixers For High Throughput Applications Using Computational Uid Dynamics

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953

Contents lists available at ScienceDirect

Chemical Engineering Research and Design

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

Optimisation of mixing performance of helical


ribbon mixers for high throughput applications
using computational fluid dynamics

O. Mihailova a,∗ , T. Mothersdale a , T. Rodgers b , Z. Ren b , S. Watson a ,


V. Lister c , A. Kowalski a
a Unilever Research and Development, Quarry Road East, Port Sunlight CH62 4ZD, United Kingdom
b School of Chemical Engineering and Analytical Science, The University of Manchester, Oxford road, Manchester
M13 9PL, United Kingdom
c Siemens PLM Software, 200 Shepherds Bush Road, W6 7NL, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: The work presented focuses on the optimisation in a 1 L vessel using an anchor with a helical
Received 31 October 2017 ribbon design using CFD and a learning algorithm, optimised by minimisation of the torque
Received in revised form 25 January output of the mixer and the homogeneity of the mixture in the vessel after a defined mixing
2018 time. The results were successfully validated experimentally using Electrical Resistance
Accepted 30 January 2018 Tomography (ERT) and direct torque measurements.
Available online 14 February 2018 The study determined that the height of the mixer is a key factor in the performance
of the mixer, with other significant factors present, but with a lower impact. For the case
Keywords: of torque, all design features of the mixer which increase the size, i.e., surface area acting
CFD against motion, were found to be significant in increasing the modelled torque response. The
Optimisation Auger screw was found to have no significant impact on either mixing and torque response.
Mixing The results illustrate the capability of optimisation algorithms to achieve results com-
Helical ribbon mixers parable to those achieved experimentally, while assessing a significantly larger number of
Rapid prototyping design options and optimising for several performance indicators simultaneously.
© 2018 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

1. Introduction An example of such high throughput device is the Geoff


formulation robot (LabMan, UK), located at the Materials Inno-
Due to several factors mixing becomes more difficult as the vation Factory at the University of Liverpool. This automated
equipment is scaled down, therefore traditional mixer geome- formulator is designed to produce up to 20 distinct 1 L formu-
tries might prove inefficient in a smaller vessel. However, lations in a 24-h window. However, as the formulator targets
achieving efficient mixing in smaller vessels is highly desir- customers from across the FMCG industry, it is often tasked
able, especially for high-throughput automated systems used with processing complex non-Newtonian fluids (Balzer et al.,
for research and development. Such systems are expected to 1995). The standard agitation geometry on the robotic formu-
produce high quality outputs at scales as small as 100 mL, lator includes a single stage pitched blade turbine, a 2-arm
while formulating products with complex rheological prop- anchor agitator and a high shear device. This geometry can
erties, unlocking research and development capability and find it challenging to effectively blend more viscous and com-
integrating such capabilities into automated high-throughput plex fluids, reducing the throughput capability of the platform
formulation platforms. (Ameur, 2016), and was therefore targeted for improvement
using a range of techniques.


Corresponding author.
E-mail address: olga.mihailova@unilever.com (O. Mihailova).
https://doi.org/10.1016/j.cherd.2018.01.053
0263-8762/© 2018 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953 943

The application of helical ribbon mixers to enhance the be used directly to manufacture prototypes for experimental
bulk mixing of non-Newtonian fluids in cylindrical vessels is validation.
well known and widely researched, with the various designs
being the recommended solution for achieving high homo- 2. Materials and methods
geneity in such systems (Ayazi Shamlou and Edwards, 1985;
Brito-De La Fuente et al., 1997; Doraiswamy et al., 1994; Gijón- 2.1. Starting geometry and testing platform
Arreortúa and Tecante, 2015; Zhang et al., 2008). Therefore,
helical mixers were chosen as the lead option for improv- The robotic platform mixer assembly is held on the lid of the
ing mixing on the small scale high throughput formulator individual processing stations and consists of 3 components
platform, with the aim to improve the mixing performance which promote mixing: propeller turbine, high shear mixer
of the anchor assembly, while keeping the geometry rela- and an anchor-scraper, a typical example of which is shown
tively simple, to ensure effective cleaning using the automated in Fig. 1a. Further to the mixer elements, there are dip in tem-
cleaning protocols on the robot. However, due to a number perature and pH sensors are present on the module lid. This
of design constraints resulting from the integration of the mixer assembly was designed with a wide range of product
mixer into an existing platform a standard mixer design could rheologies and processing requirements in mind, while still
not be implemented and therefore an optimisation study was allowing the geometry to be easily cleaned in between inves-
performed, where the maximising the homogeneity of the tigated formulations.
final mixture after a given mixing time and minimising the For the purposes of optimisation of the robotic formula-
torque response of the mixer were the optimisation parame- tor through formulation cycle time reduction as well as to
ters. improve the mixing performance in more challenging formu-
Computational fluid dynamics (CFD) is a powerful tool in lations (e.g., high viscosity and non-Newtonian liquids) several
mixing research, as it allows to assess and compare the per- modifications have been investigated, predominately aimed at
formance of various mixer geometries without the need for optimisation of turbine and anchor geometries. For the pur-
experimentation (Han et al., 2012; Vicum et al., 2004). CFD pro- poses of this study the focus is solely on the modifications
vides a further benefit of tracking the evolution of the process to the anchor-scraper, where the inspiration for the modi-
across the volume of the vessel, which is often not practical fications of the simple two arm anchor has come from the
when studying real industrial systems experimentally, due to popular helical ribbon mixers which have been shown to pro-
the impermeable nature of the construction materials of the vide enhanced mixing in fluids exhibiting a wide range of
vessels and often the fluids within. In addition, it is possible to rheological properties.
couple a CFD modelling tool with an optimisation tool, where All design modifications are assessed using the purpose
a target can be set, for example to minimise energy consump- built stand-alone mixing unit, which replicates all the design
tion of the mixer, and some features of the design can be varied features of the robotic mixer platform, but the speed and
by the optimiser to achieve the set goal (Hanada et al., 2016; runtime of the mixer is controlled manually, as opposed to
Palacz et al., 2016). For the purposes of this study Siemens via a pre-set programme. In addition, the anchor shaft has
STAR-CCM+ CFD package and HEEDS optimisation package been modified with a quick-release clasp, allowing to swap
were used. However, to ensure the quality of the CFD predic- out mixer geometries without interfering with the rest of the
tions, it is advisable to carry out experimental validation on machine. Furthermore, the lid lowering mechanism on the
select cases, to ensure model compliance. Mixing is often diffi- stand-alone rig ensured that the geometry is brought into the
cult to validate, due to the challenges posed by the equipment, same position with respect to the sample pot, with a consis-
as discusses previously. Nevertheless, a number of techniques tent immersion level and orientation.
exist which allow process tracking by measuring some form However, due to the presence of other elements in the
of tracker or difference between the materials being blended, mixer volume (i.e., temperature and pH probes), it is not
such as planar laser image fluorescence (PLIF) (Ramsay et al., possible to accommodate the traditional central shaft ribbon
2016), Particle Tracking Velocimetry (PTV) (Alberini et al., 2017), support, typical for helical ribbon mixers, as shown in Fig. 1b.
Positron Emission Particle Tracking (PEPT) (Mihailova et al., Therefore, the helical ribbons presented in this study were
2015), Magnetic Resonance Imaging (MRI) (McCarthy et al., supported solely by the anchor arms, where additional anchor
2002) and Electrical Resistance Tomography (ERT) (Rodgers arms are introduced to provide support in the designs with
and Kowalski, 2010), which used for validation purposes in increasing number of turns of the ribbons.
this work.
3D printing has become a mature technology which is used 2.2. Computational fluid dynamics (CFD) modelling
across industries for the rapid manufacture of both proto-
types for illustrative purposes and functional parts in a range Siemens STAR-CCM+ software was used to carry out the mod-
of materials (Chua et al., n.d.). The validation work for this elling of 98 mixer designs differentiated by seven key design
project was made possible through rapid prototyping using features: (1) height of the anchor/helix assembly, (2) width of
3D printing, in particular selective laser sintering (SLS), which the helix, (3) thickness of the helix, (4) number of helix turns,
allowed the production of the complex mixer geometries in- (5) number of anchor arms, (6) height of the central auger
house, providing the opportunity to test and evaluate the new screw, and (7) width of the central auger screw. The number
designs quickly. of anchor arms is varied to accommodate the increase in the
The combination of the abovementioned techniques number of helical ribbon turns, as described previously. How-
allowed to design, analyse and validate a wide range of mixer ever, the two are independent variables, i.e., some of designs
design alternatives to improve the mixing performance of the tested can have the same number of helical ribbon turns, but
high-throughput formulator platform. Moreover, the CFD and a different number of anchor arms, or vice versa.
the rapid prototyping approaches were found to be exception- The simulation was set to assess the mixer performance
ally complimentary, as the models generated using CFD could in a fluid with the rheological profile based on that of a typ-
944 Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953

Fig. 1 – (a) Typical anchor scraper geometry. (b) Typical helical ribbon geometry.

ing the mixing duty, with the high shear homogeniser and the
turbine impeller present in the system, but static.
To assess the mixing performance of each anchor design
tracers were introduced into the simulation, where the tracer
is modelled as a liquid with the same properties as the bulk,
but is tagged with a different concentration parameter, similar
to injecting a liquid containing a dye into the bulk. Homogene-
ity describes the distribution of this tracer in a vessel volume,
where if the tracer is distributed equally, the resulting num-
ber is 1, while homogeneity of less than 1 implies some local
concentration gradients, with lower numbers indicating an
increasingly poorly mixed system, as calculated using Eq. (2).
Fig. 2 – Flow curve of the fluid used in the CFD optimisation

study. |c − |V
¯ c
Homogineity of  = 1 − c
 (2)
2||
¯
c
Vc
ical commercially available shampoo formulation, typically
fitting a Carreau model, shown in Eq. (1), where eff is the
where ¯ is the volume average of , c is the value for the
apparent viscosity (Pa s), 0 is the viscosity at low viscosity,
selected scalar in a cell and Vc is the cell volume.
i.e., the Newtonian plateau (Pa s), ∞ is the viscosity at infi-
As part of the study the torque response of the mixer was
nite shear rate (Pa s), ˙ is the shear rate (s−1 ), n is the Power
also modelled and was used as a metric of the size of the
law index (–) and  is the relaxation time (s). The fluid has a
geometry, it can be defined as the mass resistance to mixing,
distinct Newtonian plateau at lower shear rates, with the aver-
and for the case of the same fluid used larger designs can be
age viscosity of 10 Pa s, and a shear thinning region at higher
expected to exhibit higher resistance and result in a higher
shear rates, as shown in Fig. 2, where the coefficients are as
torque response, due to the increase in the surface area nor-
follows, 0 = 9.3 Pa s, ∞ = 0 Pa s, n = 0.18 and  = 0.037 s. Such
mal to the direction of motion of the anchor (Ameur, 2016;
rheological behaviour is typical of personal and household
Kuhs et al., 2017). As the final optimised design was due to be
care products, as well as some foods and other fast-moving
installed on an automated formulation platform which is also
consumer goods (FMCGs).
self-cleaning using a spray down approach, it is crucial to keep
the complexity of the anchor-helix assembly low, as a bulkier
2 (n−1)/2
eff = ∞ + (0 − ∞ )(1 + ()
˙ ) (1) design could be more challenging to clean and additionally
impede the cleaning of the instruments contained within the
The fluid behaviour was based on the above model, by solv- assembly (e.g., pH metre or high shear mixer). Therefore, for
ing the Navier–Stokes equations numerically using the widely this application it is desirable to keep the torque low, while
used SIMPLE algorithm (Ferziger and Perić, 2002), assuming maximising the homogeneity of the system at the end of the
laminar flow. mixing operation.
The simulation was set to run for a finite number of time The torque response was calculated based on the forces
steps, corresponding to 10 min of the mixer running in real exerted on the surfaces due to shear and pressure in the sys-
time, at the end of which the homogeneity of the system was tem, where the total force on a surface is computed as follows:
assessed. For these simulations, all the elements of the mixer
assembly were present, to allow for the assessment of any  pressure

dead-zones caused by the presence of obstructions, such as f = Ff + Fshear
f
· nf (3)
the pH metre. However, only the anchor assembly was provid- f
Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953 945

After the flow field was converged based on asymptotic


stopping criteria for the torque (±0.01 Nm) and minimum cri-
teria for the continuity residual the tracer was introduced by
solving the scalar transport equation incorporating diffusion,
convection and accumulation. Time was then advanced using
a 1st order implicit methodology using a time step size of
0.005 s.

2.3. Mixer geometry optimisation

The Siemens HEEDS optimisation package was used to analyse


the outputs of the STAR-CCM+ simulations and inform design
changes based on the defined constraints. As mentioned pre-
viously, the goal was to arrive at the anchor/ribbon design
which would maximise the mixing in the system, while keep-
ing the size of the agitator simple, to ease cleaning. Therefore,
the optimisation package was set to maximise homogeneity
and while reducing the torque response, with the schematic
of the optimisation loop illustrated in Fig. 4. The optimisation
process is learning, basing the next design iteration on the out-
Fig. 3 – An example of the mesh used to model the torque
comes of the previous designs, targeting the desired goal, but
and mixing response, where it can be seen the finer mesh
limited by the number of iterations predefined by the model
around the moving mixer elements, i.e., anchor and screw,
set-up, N, where for the purposes of this study N = 98 (Mariani
and coarser mesh in the bulk of the vessel.
et al., 2017).
It must be noted, that due to the number of design features
pressure
where Ff (N) and Fshear
f
(N) are pressure and shear which are being varied during this study it is possible that
forces on the surface face, respectively, and f is the surface more than one design can achieve the desired combination of
face and nf are a user specified direction in which to compute low torque and high mixing performance.
the force.
The force that the fluid exerts on the surface, the pressure 2.4. Statistical analysis
pressure
force vector (Ff ), is computed as follows:
The CFD output data was fitted to a standard least squares
pressure model to investigate the correlation between the mixer perfor-
Ff = (pf − pref )af (4)
mance and the individual features of the mixer design, such
as anchor height, number of anchor arms and the number of
where pf is the face static pressure (Pa) and af is the face
helix turns. The least squares approach allows to identify out-
area vector (m2 ).
liers and define the strength of the correlation between input
While the shear force exerted on the surface by the fluid,
and response variables (Robinson and Cleary, 2012).
the shear force vector (Fshear
f
) (Pa), is computed as show in Eq.
(5).
2.5. Experimental validation

Fshear
f
= −Tf · af (5)
2.5.1. Mixer prototype construction for validation
A selection of designs modelled using the STAR-CCM+ and
where Tf is the stress tensor (Pa) at face f. HEEDS packages were chosen for experimental validation.
Torque can be further calculated when the forces on the These anchor/ribbon designs were 3D printed using Acryloni-
individual mixer faces are known. trile Butadiene Styrene (ABS) plastic through Selective Laser
Sintering (SLS), as 3D printing allowed for rapid prototype
 =r·f (6) manufacture in comparison to steel. Furthermore, the use of
plastic components allowed the application of Electrical Resis-
where  is the torque (N m) and r is the distance of the tance Tomography (ERT) as one of the validation techniques,
element for which the force was calculated from the shaft (m), whereas the use of a metal mixer elements would interfere
which is based on the distance of each mesh element from the with the tracking of the conductivity throughout the system,
shaft. which the technique relies upon.
An example of the polyhedral mesh used in the STAR-CCM+ During the selective laser sintering (SLS) process, small
is shown in Fig. 3, where it can be seen that the mesh is finer grains of the material are fused together using a laser before
around the areas where the mixer is present, i.e., anchor and a new layer of granular material is deposited on top of the
central screw, to closely track the torque and mixing response. existing structure and process is repeated. SLS is limited to
The mesh is coarser in the bulk of the mixer, which allows to which plastics can be used with the technique, with ABS being
reduce the computational power required to complete each one of the most common ones. It is recognised that ABS does
simulation step. The average mesh count is 10 million cells, not provide the same physical properties as stainless steel
with a standard deviation of 2 million, which varies depending which was used for the CFD, however, due to the rigidity of the
on the individual geometry of the of the anchor assembly. The plastic this was not thought to have had an effect on the mix-
model is steady state with a frozen rotor, also known as the ing performance. The effects of using a different construction
moving reference frame (MRF). material could, nevertheless, influence the power draw of the
946 Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953

Fig. 4 – Schematic of the HEEDS optimisation loop.

mixer, due to the differences in density and surface roughness filled with 780 g of the shampoo, which corresponds to the fill
between the two materials. Despite these differences, it was level of 800 mL, which is recommended for the mixing mod-
concluded that when conducting the torque validation using ule used and is the same as the fill level used in the CFD
ABS parts, the trends between the experimental and the CFD simulation. To reproduce the conditions of the model, all the
values are expected to be consistent, while the actual values elements of the mixer assembly were present in the system
may differ due to the construction material. during validation (e.g., temperature probe).
For the torque acquisition step, the mixer assembly was
lowered into the pot containing the shampoo and the motor
2.5.2. Torque validation
was set to run at 20 rmp, the same rotational speed as that
Using the stand alone mixing module it was possible to
used in the CFD. The torque response was recorded every 0.1 s
carry out the torque validation using a direct measurement
and each trial ran for 5 min.
obtained from the torque metre (E200 ORT, Sensor Technolo-
The acquired torque data was analysed, where the first
gies, UK) attached to the anchor motor, capable of measuring
minute of the measurement was discarded, as during this time
the torque response in the range of 0–200 mN m with 0.5%
the readings could be affected by any air trapped in the sys-
accuracy, as defined by the manufacturer. The sensor can oper-
tem when the mixer was lowered in. The remaining data was
ate at a higher torque range, of up to 600 mN m, where within
averaged to provide the value of the torque response of the
this range the accuracy can be reduced, however, this allows
mixer, the standard deviation of the results was also found,
to accommodate the wide range of geometry designs, includ-
addressing the level of noise in the data.
ing more complex geometries which result in higher torque
response.
For torque validation studies, basic commercially available 2.5.3. Mixing validation
shampoo formulations were used, with a known viscosity of The validation of the mixing performance of the different
10 Pa s at 25 ◦ C, the same as the viscosity of the fluids in the anchor/ribbon designs was carried out using Electrical Resis-
CFD optimisation study. The temperature of the shampoo was tance Tomography (ERT), where a conductive tracer, in this
maintained at the desired level by immersing the sample pot case a 25% (w/w) NaCl solution, was added to the system
into a jacketed vessel connected to a temperature controlled and the local changes in conductivity indicate the degree of
bath circulator (Thermo Scientific, UK). Each sample pot was spreading of the tracer and how well it has mixed with the
Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953 947

conductivity in the entire area within the tank vessel, giving


localised information during mixing (Yang and Peng, 2003).
The adjacent measurement approach was used, where a
fixed current is injected between two adjacent electrodes and
voltage is measured across all remaining adjacent electrode
pairs, resulting in a total of 520 independent voltage measure-
ments per current injection. The resulting data was processed
using the Electrical Impedance Tomography and Diffusion
Optical Tomography Reconstruction Software (EIDORS) in
MATLAB (Adler and Lionheart, 2006). The sensor model in built
in MATLAB was meshed into 14,147 finite elements, where the
conductivity was calculated for each element based on the 520
voltage readings using the L-curve Tikhonov method (Hansen
and O’Leary, 1993).
The uniformity of the voltage measured between different
electrode pairs is then used to estimate how well the high
conductivity NaCl tracer is dispersed across the vessel volume.

3. Results and discussions

Fig. 5 – The model of the ERT mixing vessel, where the 3.1. CFD and mixer geometry optimisation
green rectangles correspond to the positions of the
electrodes and the mesh of 14,147 finite elements used for The results of the optimisation study are summarised in Fig. 6,
conductivity analysis is shown. where the torque response of each geometry modelled is
presented on the x-axis (N m), with the homogeneity of the
mixture delivered by each design at the end of the simulation
bulk of the system fluid. Tracking these changes is made pos- is shown on the y-axis. It can be noted that the results are lim-
sible by using the voltage recorded across electrode pairs in ited by what can be described as a Pareto front, shown in red
the ERT cage, where the voltages can in turn be sued to pro- in the figure, where during the optimisation process no design
duce 2- or 3D reconstructions of the conductivity distribution could be found, within the limits of the defined design fea-
(Polydorides and Lionheart, 2002; Vauhkonen et al., 2001). tures, which would lie outside of this front, hence the Pareto
For the purposes for the validation a bespoke ERT Per- front defines the optimum solution, within the set constraints
spex vessel was manufactured at the University of Manchester (Khorram et al., 2014). As described above, the highly desir-
which replicates the dimensions of the standard 1 L sample able features of the mixer for the application on the high
pots used for the formulation work on the high throughput throughput formulation platform are high homogeneity at the
platform. To enable ERT the technique 6 planes of electrodes, end of the process, with low torque response. Therefore, any
containing 16 electrodes each, were installed in on the walls of of the designs located towards the top right corner of the
the vessel, of which only 5 bottom planes were used, due to the Pareto front, before the homogeneity appears to rapidly drop
fill level of the vessel. Fig. 5 shows the model of the vessel used, for lower torque designs, would be highly desirable options
where the green rectangles correspond to the positions of the for the high throughput platform. The designs circled in blue
electrodes. Using the ITS P2000 ERT system (Industrial Tomog- in Fig. 6 are shown in detail in Fig. 7, with design parameters
raphy Systems, UK) the boundary voltages between electrode listed in Table 1, these were chosen for experimental valida-
pairs mounted on the wall of the tank vessel were measured. tion purposes. Here some of the designs can be seen to be very
The change in boundary voltage is related to the change in similar, for example design 22 and 67 are identical, bar the size

Fig. 6 – Representation of the torque responses and homogeneity of the mixture for all the geometries modelled. The line is
added to illustrate the Pareto front, while the designs chosen for experimental validation are circled.
948 Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953

Fig. 7 – 7 designs chosen for experimental validation.

Table 1 – Details of geometry designs used for the validation study.


Design no. Number of Helix height Number of Helix pitch (◦ ) Number of Screw width Area of Area of helix
anchor arms (cm) helix turns screw turns (cm) anchor (m2 ) (m2 )

19 3 6.54 1.5 15.6 2 0.9 6.4E−04 2.1E−01


22 4 9.00 0.5 49.1 2 0.9 1.4E−03 4.5E−02
43 4 6.54 1.5 15.6 2 0.9 1.0E−03 2.1E−01
67 2 9.00 0.5 49.1 1 0.3 7.2E−04 4.5E−02
75 1 8.85 1 29.6 2 0.4 4.0E−04 7.9E−02
81 2 6.00 0.5 37.6 2 0.4 5.4E−04 3.1E−02
93 1 7.08 0.5 42.2 2 0.9 2.3E−04 2.7E−02

of the central Auger screw. Referring to Fig. 6 it can be seen be clearly seen that the mixing performance of the shorter
that while the larger Auger screw does not appear to provide mixer is inferior to that of the full height one. This suggests
a significant improvement in the homogeneity of the mixture, that when selecting the anchor/helical ribbon design for such
it does negatively affect the torque response, adding approxi- applications, the mixer should span as much of the height of
mately 25% torque for design 22, compare to 67. This suggests the vessel as possible, without introducing other phenomena,
that the presence of the Auger screw does not enhance the such as aeration.
mixing significantly, and when taking into the account the It can also be noted that designs 19 and 43 are almost identi-
geometry cleaning considerations, should be omitted from the cal, with design 43 having an additional anchor arm, compared
final design. The height of the anchor/helix assembly can be to design 19, with no other differences in anchor height, helix
seen to play a significant role in improving the homogeneity or auger screw properties. However, both the homogeneity
of the mixture, as can be seen when comparing the response and the torque response of the two designs are quite differ-
for designs 67 and 81, where full height and 2/3 height mixers ent, which cannot be explained by the geometrical differences
are compared, which otherwise have similar features. It can between the two designs alone, and can expected to not be
Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953 949

Fig. 8 – Residuals from the initial least squares correlation for (a) torque and (b) homogeneity.

consistent with the experimental validation set. Both these


Table 2 – Effect and significance of different design
designs represent the extremes of the experimental space, as parameters on homogeneity. The effects of parameters
these have most the parameters set to maximum, with the marked with (*) are statistically significant at 95%
largest Auger screw, the maximum number of helical ribbon confidence limit.
elements and the maximum number of turns of the helical Term Estimate Standard t ratio Probability > |t|
ribbon, as well as the maximum number of anchor arms (4) in coefficient error
design 19, and 3 arms in design 43. It can be argued that the
Height of Mixer 18.7037 0.5811 32.19 <0.0001*
model was incapable of accurately calculating the outputs of
Width of Helix 0.0767 0.0070 11.03 <0.0001*
interest for these designs due to the complexity. No. of helix turns 0.0550 0.0133 4.12 <0.0001*
Screw width 0.0098 0.0030 3.26 0.0016*
3.2. Statistical analysis No. of anchor arms −0.0212 0.0067 −3.17 0.0022*
Helix thickness −0.0258 0.0092 −2.81 0.0064*
While comparing specific features of select designs can pro- Anchor width −0.0256 0.0123 −2.07 0.0416
Screw height −1.2819 0.6609 −1.94 0.0561
vide some information about the effect of the individual
No. of screw turns −0.0237 0.0161 −1.47 0.1456
design parameters on mixing and torque, systematic statisti-
cal analysis can provide quantitative insight into these effects.
Therefore, least squares analysis was applied to model the at the end of the run. Therefore, further study of the individual
effects of the mixer design parameters on the torque and design parameters were considered next.
homogeneity results of the 100 designs modelled, using JMP Based on the conclusions from the surface area mod-
analytical package (SAS Institute Inc.). Firstly, to test for out- els shown previously, homogeneity and torque models were
liers the residuals of the model were considered, as shown in created using the discrete mixer design features, such as
Fig. 8(a), where for homogeneity, any design with a residual of helix height and number of turns, for example, and only
±0.3 was excluded from future analysis (Couturier et al., 2016). direct correlations of parameters considered, i.e., no combi-
For the torque residuals, a similar procedure was performed, nation effects were included, and neither were non-linear
leading to excluding two designs, both with the residuals of correlations. Fig. 10 illustrates least squares models for the
over −0.05 N m, which can clearly be seen in Fig. 8(b). These homogeneity (a) and torque response (b), where the x axis cor-
outliers included mixer design 43, which was suspected to responds to the output of the least squares model, while the
be an outlier based on the inconsistency in results between y axis corresponds to the results obtained using CFD.
designs 43 and 19. A total of 4 outliers were removed using Both least squares models provide correlations with high
this approach, which were further discovered to have poor R2 values, 0.97 for both homogeneity and torque, and a low
correlation with experimental results, during validation. residual mean square error, 0.0545 and 0.0069 respectively,
It can be expected that the torque response of the mixer suggesting that not only a strong correlation between the CFD
will be directly correlated with the surface area of the mixer data and the statistical model, but also low deviation from the
normal to the liquid, i.e., displacing the liquid during motion. identity line for individual points on the plot.
This is confirmed by least square analysis where the variables When looking at the individual effects of the different
defining torque are the anchor and the helical mixer areas. As terms on the expression characterising homogeneity several
can be seen from Fig. 9(a) the model based on these variables terms can be identified as statistically significant, as can be
provides an acceptable fit between actual (CFD) and predicted seen in Table 2. Based on the p-test values, shown in the 5th
(least squares) values, with the corresponding p-test values for column of the table, it can be concluded that the height of
both parameters lying well within the significant range. How- the anchor/mixer assembly, width of the helix and the num-
ever, the quality of fit suggests that additional parameters, or ber of helix turns are all significant factors, while the width
level of granularity might be necessary to predict the torque of the screw, the number of anchor arms and the thickness of
response better. It can be further seen from Fig. 9(b) the com- the helix are borderline. The coefficient is the highest for the
bination of the two area parameters fails to predict the quality anchor height, suggesting variations that that design param-
of the mixer performance, i.e., the homogeneity of the mixture eter would have the greatest effect on the mixing efficiency.
950 Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953

Fig. 10 – Mixing performance (a) and torque (b) least


squares models based on individual mixer design features,
where “actual” corresponds to the CFD calculated results,
while “predicted” corresponds to the results calculated by
Fig. 9 – Torque (a) and mixing performance (b) least squares the least squares model.
models based on the anchor and helical mixer surface
areas, where “actual” corresponds to the CFD calculated
Table 3 – Effect and significance of different design
results, while “predicted” corresponds to the results parameters on torque. The effects of parameters marked
calculated by the least squares model. with (*) are statistically significant at 95% confidence
limit.
The remaining significant terms have relatively small coeffi- Term Estimate Standard t Ratio Probability > |t|
cients, suggesting any changes to those design features would error
not affect the performance of the mixer as much. Therefore, in
No. of anchor arms −0.0127 0.0009 −13.93 <0.0001*
order to improve mixing the height of the mixer is to be max- Width of Helix −0.0131 0.0009 −13.86 <0.0001*
imised, while the central screw does not provide any mixing No. of helix turns −0.0215 0.0018 −11.87 <0.0001*
benefit. It can also be noted that the width of the anchor does Height of Mixer −0.8847 0.0788 −11.22 <0.0001*
not appear to have a significant impact, and in fact, reduc- Screw width −0.0010 0.0004 −2.45 0.0165*
ing the number of anchor arms has a positive effect on the Anchor width −0.0021 0.0017 −1.34 0.1828
Helix thickness −0.0012 0.0012 −0.93 0.3544
mixture homogeneity. This suggests that here the anchor only
No. of screw turns 0.0015 0.0022 0.67 0.5039
acts as a supporting structure for the helical ribbon, but does
Screw height −0.0470 0.0896 −0.52 0.6015
not significantly contribute to mixing. Based on the above, it
can be concluded that the thickness of the anchor as well as
the number of arms should be minimised, whenever possi- Similarly, the coefficients defining the significance of the
ble. However, caution is advised, as based on the CFD data parameters for the torque response are shown in Table 3. Here
the designs providing the best torque/homogeneity outputs the significant parameters are the number of anchor arms,
could have anchors too thin to ensure structural integrity, or the width of the helix, the number of helix turns and the
insufficient anchor arms to support the helix effectively, e.g., height of the anchor/helix mixer assembly, with the central
1 anchor arm in some designs. screw width being borderline and the remaining terms are
Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953 951

Fig. 11 – The comparison between the experimentally measured torque and torque calculated using CFD. Here the
individual points represent the different geometry designs, with a linear trendline showing the correlation between the
experimentally measured torque (x axis) and that calculated using CFD (y axis), with the CFD results shown to exceed the
experimental results by approximately 20%. Geometry 19 is shown to be an outlier and is marked in red. A dashed x = y line
is added for reference, representing a theoretical perfect agreement between experimental and CFD results.

Fig. 12 – The comparison between the experimentally measured mixing time and the homogineity of the fluid in the vessel
at the end of the CFD simulation run.

not significant. This is to be expected, as all of the signifi- consistent offset of approximately 20% between the two val-
cant terms can be linked to substantially changing the size of ues, with the CFD values higher than the experimental results.
the mixer, and therefore affecting the torque measurement. This can be attributed to a number of factors, like the construc-
Ideally, to achieve the best design within the desired param- tion material of the anchor, where ABS is lighter than steel and
eters, the height of the mixer assembly is to be kept at a also exhibits different surface properties, e.g., roughness.
maximum height which does not introduce aeration, while Alternatively, it is possible that the fluid viscosity was not
all the remaining parameters which increase torque should perfectly matched between the model and the experiment,
be minimised wherever possible. either due to error in off-line viscosity measurements or tem-
perature control during the experiment.
Nevertheless, this error appears to be systematic, and does
3.3. Experimental validation
not prevent from concluding that trend in the torque response
based on CFD is consistent with experimental measurements.
3.3.1. Torque
Geometry 19 must also be highlighted as an outlier, this
For torque validation, the mixer geometries which were 3D
was expected based on the observations of the Pareto front
printed were attached to the standalone mixer station and set
results and the least squares model used to determine the
to run at constant rpm, consistent with the rmp of the CFD
significant design parameters. Here, geometries 19 and 43 are
model, immersed in the silicone oil or shampoo with a defined
known to be very geometrically similar, and are expected to
viscosity, also matching that of the model.
provide similar torque response, which can be seen from the
The comparison between the torque response delivered
experimental results, with the recorded torque of 0.1297 and
by the model and that obtained experimentally is shown in
0.1254 N m respectively for the two geometries. However, the
Fig. 11, where 6 out of the 7 designs tested show agreement
CFD results show a dramatic difference, with geometry 19
between the model and experimental data. However, there is
952 Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953

yielding a torque response of 0.320 N m, almost double of the In turn, torque was determined to be highly dependent on
0.164 N m for geometry 43. It is reasonable to conclude that the the height of the mixer assembly, the number of mixer arms,
value obtained for geometry 43 is the correct one, as it falls the number of helix turns and the width of the helix, where
in line with the results for other geometries used in the val- all these parameters also significantly influence the size of the
idation study. The reason for the torque being overpredicted mixer, which in turns is directly linked to increased torque in
for geometry 19 could be a result of the mistake within the real systems.
software when setting up the CFD parameters. It is can therefore be advised to maximise the height of the
mixer whenever practical, while reducing the size of the other
design features.
3.3.2. Mixing
The use of CFD for the estimation of the effect of differ-
Mixing was validated using ERT, where a conductive tracer was
ent mixer design features on the two key parameters, torque
added to the bulk of less conductive liquid and the conductiv-
and mixing performance, has been shown to be an effective
ity across the volume of the vessel was recorded over time,
approach for rapid optimisation of a mixer design for a given
until homogeneity was reached. As the homogeneity of the
set of constraints.
mixture in the CFD predicted results was measured at the end
of the simulation, as opposed to continuously over time, it
cannot be directly compared with the mixing times required Acknowledgements
to achieve a fully mixed system, which is calculated using
ERT. However, it is still possible to draw parallels between the The authors would like to thank Innovate UK for the financial
homogeneity and mixing time, as shown in Fig. 12. Here, for support provided for this work as a part of the Embed-
mixer designs which have achieved high homogeneity at the ding Manufacturing Development into Formulation Research
end of the CFD simulation maintain a plateau at low mix- (EMFormR) project (EP/L505778/1). The authors would also like
ing times, as no differentiation can be made between designs to thank the Unilever Port Sunlight R&D rapid prototyping
which completed the mixing process at the end of the sim- team for help with 3D printing of mixer parts.
ulation run. However, as the CFD predicted homogeneity at
the end of simulation allotted time declines the experimental
time required to achieve a fully homogeneous mixture can be References
observed to increase, illustrating that it takes longer to achieve
a homogeneous mixture for the mixer designs which perform Adler, A., Lionheart, W.R.B., 2006. Uses and abuses of EIDORS: an
poorly. extensible software base for EIT. Physiol. Meas. 27, S25–S42,
http://dx.doi.org/10.1088/0967-3334/27/5/S03.
Here design 43 appears to be an outlier, as the actual mixing
Alberini, F., Liu, L., Stitt, E.H., Simmons, M.J.H., 2017. Comparison
time required to achieve a fully mixed system using this design
between 3-D-PTV and 2-D-PIV for determination of
is lower than that predicted using the model, however, this hydrodynamics of complex fluids in a stirred vessel. Chem.
design was excluded from the least squares model based on Eng. Sci. 171, 189–203,
the model residuals, and is therefore expected to perform out http://dx.doi.org/10.1016/j.ces.2017.05.034.
of line with the other designs. Ameur, H., 2016. Effect of some parameters on the performance
of anchor impellers for stirring shear-thinning fluids in a
cylindrical vessel. J. Hydrodyn. 28, 669–675,
4. Conclusions http://dx.doi.org/10.1016/S1001-6058(16)60671-6.
Ayazi Shamlou, P., Edwards, M.F., 1985. Power consumption of
helical ribbon mixers in viscous newtonian and
STAR-CCM+ CFD package and HEEDS optimisation package to
non-newtonian fluids. Chem. Eng. Sci. 40, 1773–1781,
define the optimal design for a small scale 1 L mixer anchor
http://dx.doi.org/10.1016/0009-2509(85)80040-3.
mixer. Over 100 designs were processed by the CFD package Balzer, D., Varwig, S., Weihrauch, M., Ag, H., Marl, D., 1995.
changing a number of geometrical parameters, such as the Viscoelasticity of personal care products. Colloids Surf. A 99,
height of the anchor/helix mixer assembly, the number of 233–246, http://dx.doi.org/10.1016/0927-7757(95)03144-3.
anchor arms and the number of turns of the helix, as well Brito-De La Fuente, E., Choplin, L., Tanguy, P., 1997. Mixing with
as the effect of the central Auger screw. The model provided helical ribbon impellers. Chem. Eng. Res. Des. 75, 45–52,
http://dx.doi.org/10.1205/026387697523381.
outputs on the torque response and the mixing performance
Chua, C.K., Wong, C.H., Yeong, W.Y., n.d. Standards, quality
of each design, which were validated experimentally, using a control, and measurement sciences in 3D printing and
stand-alone module of a high throughput platform formula- additive manufacturing.
tor, equipped with a torque sensor and an ERT conductivity Couturier, R., Ziane Khodja, L., Guyeux, C., 2016. TSIRM: a
measurement system. two-stage iteration with least-squares residual minimization
The outputs of the model were analysed by fitting the data algorithm to solve large sparse linear and nonlinear systems.
to a standard least squares model, identifying several out- J. Comput. Sci. 17, 535–546,
http://dx.doi.org/10.1016/j.jocs.2016.10.017.
liers and defining the key characteristics of the mixer designs
Doraiswamy, D., Grenville, R.K., Etchells, A.W., 1994. Two-score
which affect mixing and torque. It is desirable to reduce torque years of the Metzner-Otto correlation. Ind. Eng. Chem. Res. 33,
while increasing mixing efficiency, both from a power effi- 2253–2258, http://dx.doi.org/10.1021/ie00034a001.
ciency perspective and in order to ensure ease of cleaning. Ferziger, J.H., Perić, M., 2002. Computational Methods for Fluid
It was found that the height of the mixer assembly is the Dynamics, 3rd ed. Springer Berlin Heidelberg, Berlin,
key driver to the mixers performance, however, the anchor Heidelberg, http://dx.doi.org/10.1007/978-3-642-56026-2.
Gijón-Arreortúa, I., Tecante, A., 2015. Mixing time and power
itself does not contribute to improving mixing, with the helical
consumption during blending of cohesive food powders with
ribbon features, such as width and number of turns playing a
a horizontal helical double-ribbon impeller. J. Food Eng. 149,
significant role in achieving a highly homogeneous mixture. 144–152, http://dx.doi.org/10.1016/j.jfoodeng.2014.10.013.
In addition, it was determined that the central Auger screw Han, Y., Wang, J.J., Gu, X.P., Feng, L.F., 2012. Numerical simulation
does not have a statistically significant effect on mixing. on micromixing of viscous fluids in a stirred-tank reactor.
Chemical Engineering Research and Design 1 3 2 ( 2 0 1 8 ) 942–953 953

Chem. Eng. Sci. 74, 9–17, Polydorides, N., Lionheart, W.R.B., 2002. A Matlab toolkit for
http://dx.doi.org/10.1016/j.ces.2012.02.018. three-dimensional electrical impedance tomography: a
Hanada, T., Kuroda, K., Takahashi, K., 2016. CFD geometrical contribution to the Electrical Impedance and Diffuse Optical
optimization to improve mixing performance of axial mixer. Reconstruction Software project. Meas. Sci. Technol. 13,
Chem. Eng. Sci. 144, 144–152, 1871–1883, http://dx.doi.org/10.1088/0957-0233/13/12/310.
http://dx.doi.org/10.1016/j.ces.2016.01.029. Ramsay, J., Simmons, M.J.H., Ingram, A., Stitt, E.H., 2016. Mixing
Hansen, P.C., O’Leary, D.P., 1993. The use of the L-curve in the performance of viscoelastic fluids in a Kenics KM in-line static
regularization of discrete ill-posed problems. SIAM J. Sci. mixer. Chem. Eng. Res. Des. 115, 310–324,
Comput. 14, 1487–1503, http://dx.doi.org/10.1137/0914086. http://dx.doi.org/10.1016/j.cherd.2016.07.020.
Khorram, E., Khaledian, K., Khaledyan, M., 2014. A numerical Robinson, M., Cleary, P.W., 2012. Flow and mixing performance in
method for constructing the Pareto front of multi-objective helical ribbon mixers. Chem. Eng. Sci. 84, 382–398,
optimization problems. J. Comput. Appl. Math. 261, 158–171, http://dx.doi.org/10.1016/j.ces.2012.08.044.
http://dx.doi.org/10.1016/j.cam.2013.11.007. Rodgers, T.L., Kowalski, A., 2010. An electrical resistance
Kuhs, M., Moore, J., Kollamaram, G., Walker, G., Croker, D., 2017. tomography method for determining mixing in batch addition
Predicting optimal wet granulation parameters for with a level change. Chem. Eng. Res. Des. 88, 204–212,
extrusion-spheronisation of pharmaceutical pellets using a http://dx.doi.org/10.1016/j.cherd.2009.08.003.
mixer torque rheometer. Int. J. Pharm. 517, 19–24, Vauhkonen, M., Lionheart, W.R.B., Heikkinen, L.M., Vauhkonen,
http://dx.doi.org/10.1016/j.ijpharm.2016.11.057. P.J., Kaipio, J.P., 2001. A MATLAB package for the EIDORS project
Mariani, F., Risi, F., Grimaldi, C.N., 2017. Separation efficiency and to reconstruct two-dimensional EIT images. Physiol. Meas. 22,
heat exchange optimization in a cyclone. Sep. Purif. Technol. 107–111, http://dx.doi.org/10.1088/0967-3334/22/1/314.
179, 393–402, http://dx.doi.org/10.1016/j.seppur.2017.02.024. Vicum, L., Ottiger, S., Mazzotti, M., Makowski, Ł., Bałdyga, J., 2004.
McCarthy, K.L., Lee, Y., Green, J., McCarthy, M.J., 2002. Magnetic Multi-scale modeling of a reactive mixing process in a
resonance imaging as a sensor system for multiphase mixing. semibatch stirred tank. Chem. Eng. Sci. 59, 1767–1781,
Appl. Magn. Reson. 22, 213–222. http://dx.doi.org/10.1016/j.ces.2004.01.032.
Mihailova, O., Lim, V., McCarthy, M.J., McCarthy, K.L., Bakalis, S., Yang, W.Q., Peng, L., 2003. Image reconstruction algorithms for
2015. Laminar mixing in a SMX static mixer evaluated by electrical capacitance tomography. Meas. Sci. Technol. 14,
positron emission particle tracking (PEPT) and magnetic R1–R13, http://dx.doi.org/10.1088/0957-0233/14/1/201.
resonance imaging (MRI). Chem. Eng. Sci. 137, 1014–1023, Zhang, M., Zhang, L., Jiang, B., Yin, Y., Li, X., 2008. Calculation of
http://dx.doi.org/10.1016/j.ces.2015.07.015. metzner constant for double helical ribbon impeller by
Palacz, M., Smolka, J., Kus, W., Fic, A., Bulinski, Z., Nowak, A.J., computational fluid dynamic method. Chin. J. Chem. Eng. 16,
Banasiak, K., Hafner, A., 2016. CFD-based shape optimisation 686–692, http://dx.doi.org/10.1016/S1004-9541(08)60141-X.
of a CO2 two-phase ejector mixing section. Appl. Therm. Eng.
95, 62–69,
http://dx.doi.org/10.1016/j.applthermaleng.2015.11.012.

You might also like