Accepted Manuscript

Optimization of Pulsatile Flow and Geometry of a Convergent–Divergent Mi-

cromixer

Arshad Afzal, Kwang-Yong Kim

PII: S1385-8947(15)00886-4
DOI: http://dx.doi.org/10.1016/j.cej.2015.06.046
Reference: CEJ 13818

To appear in: Chemical Engineering Journal

Received Date: 19 March 2015

Revised Date: 5 June 2015
Accepted Date: 9 June 2015

Please cite this article as: A. Afzal, K-Y. Kim, Optimization of Pulsatile Flow and Geometry of a Convergent–
Divergent Micromixer, Chemical Engineering Journal (2015), doi: http://dx.doi.org/10.1016/j.cej.2015.06.046

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 This revised paper has been prepared for submission to
Chemical Engineering Journal

Optimization of Pulsatile Flow and Geometry of

a Convergent–Divergent Micromixer

Arshad Afzal and Kwang-Yong Kim*

Department of Mechanical Engineering, Inha University,

253 Yonghyun-Dong, Nam-Gu, Incheon 402-751, Republic of Korea

*Corresponding author: Kwang-Yong Kim

Fellow ASME, A. Fellow AIAA

Inha Fellow Professor

Tel.: +82-32-872-3096

Fax: +82-32-868-1716

E-mail: kykim@inha.ac.kr

June 5, 2015

0
 Abstract

The geometry and operating conditions of a micromixer with convergent–divergent sinusoidal

walls under pulsatile flow were optimized in this work to maximize the mixing performance.

Flow and mixing analyses were performed using the three-dimensional unsteady Navier–Stokes

equations with a convection–diffusion model for the species concentration. A time-averaged

mixing index was used to evaluate the mixing performance of the micromixer with a pulsatile

flow. The overall mixing index was selected as the objective function, and both geometric and

flow variables were used as the design variables. The ratio of the amplitude to the wavelength of

the sinusoidal walls and the throat width to depth of the convergent–divergent channel were

selected as the geometric variables, and the Strouhal number and the ratio of the pulsing

amplitude to the steady flow velocity of the pulsatile flow were selected as the flow variables.

Three different surrogate models were tested for the optimization to approximate the objective

function at a Reynolds number of 0.5. The results indicate that the Kriging model predicts the

best optimum design with a mixing index of 92.35% at the outlet of the micromixer.

Keywords: Micromixer; Convergent–divergent sinusoidal walls; Mixing index; Surrogate

modeling; Optimization.

1
1. Introduction

Microfluidic systems have gained widespread attention for their chemical and biological

applications over the past few decades. Micro-total analysis systems (µ-TAS) and lab-on-a-chip

platforms are widely used for sample preparation and analysis, drug delivery, and biological and

chemical synthesis. Micromixers are an important component in these applications [1–5]. Based

on their mixing mechanism, micromixers are classified into two types: active and passive.

Passive micromixers [4–7] are easy to fabricate and generally use geometry modification to

cause chaotic advection or lamination to promote the mixing of the fluid samples. The active

type uses moving parts or some external agitation/energy for the mixing. Magnetic energy,

electrical energy, pressure disturbance, and ultrasonic mixing can be used to stir the fluids [10–

14].

Many researchers have studied various geometries to design efficient passive

micromixers. For example, Aubin et al. [15] carried out a parametric study on a grooved

micromixer using the particle tracking approach. The results indicate that the groove dimensions,

i.e., width and depth, significantly affect the mixing performance. Using computational fluid

dynamics (CFD) simulations and the particle tracking technique, Wang et al. [16] generated

Poincare maps to study the chaotic flow in a staggered herringbone groove micromixer [6].

Ansari and Kim [17] performed a parametric investigation of the mixing of two fluids in a three-

dimensional serpentine microchannel. The mixing performance and pressure drop characteristics

were investigated in relation to two geometrical parameters, viz. the ratios of the channel height

to width and the length of the straight portion of an L-shaped unit to its channel width, at four

different Reynolds numbers. The results revealed that the mixing performance in the serpentine

channel was sensitive to the geometric parameters at various Reynolds numbers. The mixing

2
performance was found to increase with decreasing channel length of the L-shaped unit. For a

given Reynolds number, the pressure drop increased with both geometrical parameters. Ansari et

al. [18, 19] optimized the shape of a staggered herringbone groove micromixer using surrogate

models, such as a radial basis neural network (RBNN) and a response surface approximation

(RSA). These optimizations were performed with the mixing index as the objective function.

In recent studies, time-dependent pulsatile flows have been used by many researchers as

active mechanisms for fast and efficient mixing. In this approach, the flow is pulsed from one or

both of the inlets in a regular fashion. Glasgow and Aubry [20–22] conducted a series of

numerical simulations and experiments on pulsed flows for various channel geometries, viz.

straight, ribbed, and 3-D twisted channels, under pulsed flow conditions. The pulsing was

performed using a low-frequency sinusoidal flow superimposed on a steady flow. The results

indicated a significantly better mixing performance with the pulsed sinusoidal flow than without

pulsing in all tested micromixers. In addition, it was shown that geometrical modifications can be

implemented to further improve the mixing performance under pulsatile flow. In a later study,

Bottausci et al. [23] showed that high mixing performance can be achieved by a perturbation of

the main flow via secondary channels using oscillating sinusoidal jet flows. However, for

complete mixing, the secondary channels must be actuated by a strong sinusoidal flow (high

amplitude and frequency). In line with this observation, Afzal and Kim [24] tested various

micromixer geometries and showed that a convergent–divergent microchannel with sinusoidal

walls (space-periodic in the streamwise direction) shows the best coupling with pulsed sinusoidal

flows (time-periodic) for enhanced mixing performance. Additionally, the mixing performance

was found to be sensitive to the Strouhal number and velocity ratio, i.e., the ratio of the pulsing

amplitude to the steady flow velocity of the pulsatile flow.

3
 In the present work, a numerical optimization was performed to optimize the pulsing and

geometry of a convergent–divergent micromixer with pulsatile flow for enhanced mixing

performance at a fixed Reynolds number of 0.5. Surrogate modeling was used to approximate the

objective function. A total of four design variables, consisting of two geometric and two flow

variables, were employed for the optimization. The ratios of the amplitude to the wavelength of

the sinusoidal microchannel walls and the throat width to depth of the convergent–divergent

section were selected as the geometric design variables. Additionally, the ratio of the pulsing

amplitude to the steady flow velocity of the pulsatile flow and the Strouhal number defined using

the pulse frequency were taken as the pulsatile flow design variables. A preliminary parametric

study was performed prior to the optimization to investigate the mixing characteristics of the

convergent–divergent micromixer and obtain information on the sensitivity of the mixing

performance to the design variables.

2. Problem Formulation

The convergent–divergent micromixer used by Afzal and Kim [24] is shown in Fig. 1. The walls

of the convergent–divergent microchannel were generated using a sinusoidal function of the

form y = A sin(2πx/λ), where A and λ are the amplitude and wavelength, respectively. The

centerline of the channel is the axis of symmetry. The microchannel consists of two mixing units

linked in series. The values of the wavelength λ and the depth of the channel d measured in the z-

direction were fixed at 1 and 0.125 mm, respectively. The fluid samples entered at the inlets,

Inlet 1 and Inlet 2, which are joined to the main channel with a T-joint, as shown in Fig. 1. The

cross sections of the inlet channels are rectangles with dimensions of 0.100 mm × 0.125 mm.

4
The axial lengths of the connecting channel Lo, main channel Lc, and exit channel Le are 0.2, 2.0,

and 1.4 mm, respectively.

To analyze the flow and mixing characteristics inside the microchannel, numerical

simulations were performed using ANSYS CFX 12.1 [25], a commercial CFD package based on

the finite volume method. The mixing of two fluids with similar physical properties, e.g., density

and viscosity, is governed by the continuity, Navier–Stokes, and mass conservation equations

[26]:

∇.V = 0 , (1)

∂V 1
+ (V.∇) V = - ∇p +ν∇ 2 V , (2)
∂t ρ

∂Ci
+ (V.∇)Ci = α∇2 Ci (i = 1, 2), (3)
∂t

where V is the fluid velocity, ρ is the fluid density, ν is the kinematic viscosity, α is the

diffusivity coefficient, and Ci is the concentration (or mass fraction) of the ith fluid species. The

mass conservation equation for each fluid (Eq. (3)) employs the transient advection–diffusion

model for the species concentration field under the assumption that variations in the

concentration do not alter the density and viscosity of the fluid. In the present study, two similar

fluids with the same viscosity ν and fluid density ρ were used. The kinematic viscosity ν and

diffusivity coefficient α were set to 10-6 m2⋅s–1 (kinematic viscosity of water at room temperature)

and 10–10 m2⋅s–1 (typical diffusivity for microfluidic applications), respectively. To allow the

analysis of the mixing, the fluid that enters at Inlet 2 was assumed to be dyed.

Equations (1)–(3) were solved with the following boundary conditions. Time-dependent

pulsatile flows were introduced at the inlets with velocities

5
 Vi = Vs + Vo sin (2πft + φ) (i = 1, 2). (4)

Here, V1 and V2 are the velocities at Inlet 1 and Inlet 2, respectively, and have a phase difference

φ. Vs is the steady flow velocity. For the sinusoidal flow, Vo is the amplitude of the velocity, and

f is the pulse frequency. Zero static pressure was specified at the outlet. A no-slip condition was

applied at the walls. The concentration of the dyed fluid was set to 0 and 1 at Inlets 1 and 2,

respectively.

A second-order upwind differencing scheme was used to discretize the advection terms

in the governing equations. For transient calculations, the discretization in time was performed

using the second-order backward Euler scheme. The semi-implicit method for pressure linked

equations-consistent (SIMPLEC) algorithm [27] was used for pressure–velocity coupling. The

linearized algebraic system of equations resulting from the discretization was solved using a

multigrid accelerated incomplete lower–upper (ILU) factorization procedure. Multigrid

techniques significantly improved the convergence behavior.

Hexahedral cells were used to discretize the computational domain. In the authors’

previous work [24], grid dependency tests were performed for both spatial and temporal

resolutions. Among five different grid systems, the grid with 276,861 nodes was found to be

adequate with a relative mixing index error of less than 0.5% compared with a finer grid. Figure

2 shows an example of the grid system. For the unsteady analyses, 40 time steps were used to

capture the temporal behavior. The numerical solutions for the mass fraction were validated by

their agreement with the experimental results of Glasgow and Aubry [20] in the authors’

previous work [24]. The mass fraction distributions with and without pulsing exhibit

qualitatively similar mass fraction distributions as those obtained experimentally, as shown in

Fig. 3 [24].

6
 The criterion for convergence was a normalized root mean square (RMS) residual value

of 10–6. However, the pulsatile flows were computed over sufficiently long times to obtain

periodic solutions, i.e., solutions that did not change measurably from one cycle to the next. The

calculations were performed by an Intel Core i7 processor with eight CPUs and a clock speed of

2.94 GHz. The calculation time was typically 12–15 h for a single analysis.

A variance-based method was employed to evaluate the mixing performance of the

micromixer. The variance of the mass fraction of the mixture on a cross-sectional plane normal

to the flow direction can be expressed mathematically as [16]

σ = ∑Ni=1
 ,
(Ci )2 

(4)


where N is the number of sampling points on the plane, Ci is the mass fraction at sampling point

i, µ is the optimal mixing mass fraction (= 0.5, the mass fraction in the target case of equal

mixing of the two fluids), Vi is the velocity in the ith cell, and Vmean is the area-averaged fluid

velocity at the selected cross-sectional plane. Finally, the mixing index at any cross-sectional

plane perpendicular to the axial direction is defined as

(Ci ) 
∑N

2
  
MI = 1 −
i=1


. (5)

The value of MI varies from 0 (no mixing) to 1 (complete mixing). The overall mixing index MIo

was calculated 0.2 mm downstream from the second mixing unit (Fig. 1). In the transient

simulation, the MIo value was averaged over four values at quarter-cycle intervals.

3. Optimization

The single-objective optimization for the present problem can be formulated as

7
 Minimization of F(x) subject to LB ≤ x ≤ UB, x ∈ R,

where F(x) is a real-valued objective function and x is a vector of N design variables. LB and UB

are vectors for the lower and upper bounds of the design variables, respectively. In the present

study, surrogate modeling of the objective functions was used for this computationally expensive

problem. The various surrogate modeling techniques are described in this section.

3.1. Response surface approximation (RSA)

The RSA model [28] is a polynomial approximation of the objective function. The second-order

polynomial response surface model can be expressed as

N N N
y(x) = β0 + ∑ βi x i + ∑ βii x i 2 + ∑ βij x i x j (6)
i =1 i =1 i< j
.
The model includes, from left to right, an intercept, linear terms, quadratic interaction terms, and

squared terms. To determine the goodness of fit, Radj2 was used. For a good fit, Radj2 must be

close to 1.

3.2. Kriging (KRG) model

The Kriging (KRG) model [29] can be formulated as a combination of two components, the

global model and a systematic departure,

y( x ) = f ( x ) + Z ( x ) , (7)

where y(x ) is the unknown function to be estimated and f (x ) is a known function (usually a

regression function) representing the trend over the design space, also referred to as the global

model. The second part Z (x ) creates a localized deviation to interpolate the sampled data points

8
by quantifying the correlation of the points with a Gaussian correlation having a zero mean and

nonzero covariance.

3.3. Radial basis neural networks (RBNN)

The RBNN model [30] consists of a two-layered network with a hidden layer of radial basis

neurons and an output layer of linear neurons. The hidden layer performs a non-linear

transformation of the input space to an intermediate space using a set of radial basis units; the

output layer then implements a linear combiner to produce the desired targets. The linear model

f(x) for the function can be expressed as a linear combination of a set of N basis functions as

N
f (x) = ∑ w jφj (8)
j=1
,

where wj is the weight and φj is a basis function. The prediction ability of the network is stored

in the weights, which can be obtained from a set of training patterns. In the present study, the

customized RBNN function newrb, which is available in MATLAB [27], was used. The network

training is performed by changing the spread constant (SC) and error goal (EG) to adjust the

cross-validation error.

To find an optimum point on the constructed surrogate model, sequential quadratic

programming (SQP) was used as the search algorithm. In MATLAB, fmincon is the function for

SQP. The details of the optimization algorithm can be found in the authors’ previous works [18,

19].

9
3.4. Design variables and objective function

The dimensionless flow parameters that govern the physics in the micromixer with pulsatile flow

are the Reynolds number, the Strouhal number, and the ratio of the pulsing amplitude to the

steady flow velocity of the pulsatile flow, which are defined, respectively, as

Vs D
1. Reynolds number (Re) ≡ ,
ν

f D
2. Strouhal number (St) ≡ , and
Vs

Vo
3. Velocity ratio ≡ .
VS

Here, D is the hydraulic diameter of the inlet channel. Afzal and Kim [24] investigated the

mixing characteristics of the convergent–divergent micromixer under pulsatile flows defined by

the above dimensionless numbers. They showed that the mixing performance of the micromixer

is significantly affected by changes in the Strouhal number and velocity ratio of the pulsatile

flow. The Strouhal number was defined as the ratio of the flow time scale (D/Vs) to the pulsing

time period (1/f). It was found that there exits an optimum Strouhal number which corresponds to

maximum mixing performance at fixed Reynold number and velocity ratio. The development

and dissipation of puffs of fluids issuing from the inlet junction were strongly affected by the

Strouhal number, and mixing was significantly dependent on the structure of the puffs in the

microchannel [24]. However, the mixing index was almost invariant for flows with varying

Reynolds numbers in the range of Re = 0.25–4.0, at fixed Strouhal number of St = 0.278, and a

velocity ratio of Vo/Vs = 1.88 [24].

Based on these results, the Strouhal number and velocity ratio were selected as the design

variables of the flow condition for the optimization process. In addition, based on the results of

10
the preliminary parametric analysis, the ratios of the amplitude to the wavelength A/λ and the

throat width to depth w/d of the convergent–divergent section were selected as the geometry

design variables of the micromixer for the optimization process. Table I summarizes the ranges

for the design variables. The design ranges were determined considering the results of the

parametric analysis. In the authors’ previous study [20], it was found that the mixing index

increases with increasing velocity ratio. However, Glasgow and Aubry [16] suggested that there

is a maximum limit to the pulsing amplitude, beyond which the pulse would cause instability in

the flow structure in the mixing chamber and interfere with the proper operation of the

micromixer. In light of this observation, the upper bound of the velocity ratio was set to Vo/Vs =

4.50.

In this work, the objective function FM related to the mixing efficiency of the micromixer

was employed to optimize the micromixer. The objective function related to the mixing index,

which was to be maximized, was defined as

FM = MIo. (9)

4. Results and Discussion

Prior to the optimization process, a parametric study was performed with the design variables.

Figure 4 shows the variation of the mixing index with the Strouhal number for different velocity

ratios. The Reynolds number was fixed at 0.5. The values of the geometric parameters were

taken to be A/λ = 0.15 and w/d = 0.8. The mixing index was calculated by averaging four indices

at quarter-cycle intervals for a particular Strouhal number. It was observed that there exists an

optimum Strouhal number that yields a maximum mixing performance for each velocity ratio.

11
Lower velocity ratios yield lower mixing indices and smaller variation with the Strouhal number.

At lower velocity ratio (Vo/Vs = 0.93), the effect of pulsing leads to perturbation of the fluid

interface (Fig. 5) with lower overall mixing performance. However, at higher velocity ratios

(Vo/Vs ≥ 1.88), the breakdown of fluid interface and formation of puffs result in a rapid increase

in the mixing performance (Fig. 5). Beyond Vo/Vs = 1.88, the effect of the velocity ratio on the

mixing performance is restricted to high-Strouhal number flows (St ≥ 0.278), and the Strouhal

number corresponding to the maximum mixing index shifts from 0.139 to 0.278.

To illustrate the variation of the mixing behavior with the Strouhal number, mass fraction

contours on the central x–y plane (i.e., the plane at half of the channel depth, z = 0.0625 mm) are

plotted in Fig. 5 at time t = 0/T (data were collected after the pulsing was fully established) for

low (Vo/Vs = 0.93) and high (Vo/Vs = 3.76) velocity ratios. For Vo/Vs = 0.93, the mass fraction

distribution represents a small perturbation of the interface between the two fluids at St = 0.070.

As the Strouhal number increased (St = 0.278), rapid distortion of the interface took place, and

thus the enlargement of the interfacial area resulted in an increase in the mixing performance.

For Vo/Vs = 3.76, it was observed that the mixing proceeded via the formation of puffs along the

streamwise direction of the channel. As the Strouhal number was varied for 0.070 ≤ St ≤ 0.556,

large puffs were broken down to produce finer striations, and enhanced mixing was observed

[24]. With further increases in the Strouhal number (St = 0.556), zones of unmixed fluids in the

near-wall region led to a decrease in the mixing performance [24].

The effect of the amplitude of the sinusoidal walls on the mixing performance of the

convergent–divergent microchannel was investigated using the Strouhal number and velocity

ratio of the pulsatile flow, as shown in Fig. 6(a) and (b), respectively. The Reynolds number was

kept constant at Re = 0.5. For a fixed velocity ratio (Vo/Vs = 1.88), the variation of the mixing

12
index with the geometric parameter A/λ was observed for 0.070 ≤ St ≤ 0.556. In this range,

higher-amplitude microchannels resulted in improved mixing performance, especially for lower

St values, compared with their lower-amplitude counterparts. For St ≥ 0.278, only a small change

was observed in the mixing performance variation with the amplitude. On the other hand, at a

fixed Strouhal number (St = 0.278), the mixing performance increased with velocity ratios for all

values of the parameter A/λ. The mixing remained almost invariant with varying A/λ for velocity

ratios less than 1.8. Additionally, for Vo/Vs = 3.76, the variation in the mixing index was less

than 5% for A/λ = 0.1–0.2. In Figs. 7 and 8, the mass fraction contours on the central x–y plane

are plotted at time t = 0/T for three different Strouhal numbers and velocity ratios, respectively,

at low (A/λ = 0.1) and high (A/λ = 0.2) amplitudes. At a fixed velocity ratio (Vo/Vs = 1.88), Fig. 7

shows that the improved mixing performance in the Strouhal number range 0.070 ≤ St ≤ 0.278

results from an increase in the number of striations produced at higher amplitudes. A higher-

amplitude microchannel introduces a greater convergent–divergent slope, which results in an

increase in the width and a decrease in the thickness of the puffs and thus improved mixing

performance. At a fixed Strouhal number (St = 0278), Fig. 8 shows that the qualitative variations

of the mass fraction distribution with the velocity ratio are similar for two different values of A/λ.

The variation of the mixing behavior of the convergent–divergent microchannel with the

Strouhal number and velocity ratio of the pulsatile flow was investigated for three different

values of the ratio of the throat width to depth w/d, as shown in Fig. 9(a) and (b), respectively. In

Fig. 9(a), the effect of w/d on the mixing index is visible only for St ≥ 0.278, and contrary to the

variation with the parameter A/λ shown in Fig. 5, the mixing index decreases with increasing w/d

in this range of St at a fixed velocity ratio (Vo/Vs = 1.88). A similar trend was observed for the

variation with the velocity ratio at a fixed Strouhal number (St = 0.278) in Fig. 9(b). The value of

13
w/d does not affect the mixing index for low velocity ratios (0.46 ≤ Vo/Vs ≤ 0.94), which is

similar to the results shown in Fig. 9(a). Figs. 10 and 11 show the mass fraction contours on the

central x–y plane plotted at time t = 0/T for three different Strouhal numbers and velocity ratios,

respectively, at w/d = 0.8 and 1.2. The pattern of the mass fraction distribution is basically

unchanged with St and Vo/Vs for both w/d = 0.8 and 1.2. However, as Strouhal number increases,

the thickness of the unmixed zone near the wall increases, and the thickness also increases with

increasing w/d, as shown in Fig. 10. Thus, the case with St = 0.56 and w/d = 1.2 yields the

largest thickness of the unmixed zone near the wall. However, in Fig. 11, the unmixed zone

achieves a maximum thickness at the intermediate value Vo/Vs = 1.88 for w/d = 1.2. The

unmixed zone causes a reduction in the overall mixing index. In Fig. 9(a), the decrease in MIo

with Strouhal numbers beyond St = 0.278 is induced by the thickening of the unmixed zone

shown in Fig. 10.

The numerical optimization of the convergent–divergent micromixer with pulsatile flow

was performed to maximize the mixing index MIo using four design variables related to

geometric and flow conditions. Thirty-six design points were selected using the Latin hypercube

sampling (LHS) method [31], as listed in Table II. Figure 12 shows the dye mass fraction

distributions in the central x–y plane plotted at time t = 0/T for selected LHS designs. The

development of mixing inside the convergent–divergent microchannel is very sensitive to the

values of the design variables. The different surrogate models were formulated and trained using

the objective function values obtained from numerical simulations at the LHS-designed sites.

Using SQP, the optimum points were determined using the RSA, RBNN, and KRG models, as

shown in Table III. The CFD results at the predicted optimum points with different surrogate

models indicate that the best optimized point of the tested surrogate models was predicted with

14
the KRG model. The mixing indices for the reference and optimum designs predicted by the

KRG model are 0.9078 and 0.9235, respectively, with a relative increase of 1.7%. The maximum

relative errors of the predicted objective function values with reference to the CFD values were

5.16%, 2.24%, and 1.87% for the RSA, RBNN, and KRG models, respectively. Therefore, the

KRG model approximation of the objective function also shows the best accuracy in predicting

the optimum objective function value. The optimization results obtained with the KRG model

show that the optimum operating conditions of the pulsatile flow are St = 0.2492 and Vo/Vs =

3.2676, and the optimum geometric conditions are A/λ = 0.25 and w/d = 0.75 with a mixing

index MIo of 92.35%.

Figure 13 compares the reference design [24] and the optimum design predicted by the

KRG model in terms of the mixing performance and dye mass fraction contours. The optimum

design, which has a higher amplitude and thinner throat than the reference design, shows thinner

and longer puffs of fluids, and thus considerably higher mixing performance due to the larger

interfacial area between the two fluids. Figure 14 indicates that there is no qualitative change in

the dye mass fraction distribution with time in the central x–y plane (data are collected after the

pulsing is fully established) for the optimum design predicted by the KRG model.

5. Conclusion

A convergent–divergent micromixer coupled with a pulsed sinusoidal flow was optimized using

three-dimensional Navier–Stokes analysis and surrogate modeling. Each wall of the micromixer

was represented by two periods of a sinusoidal curve. Prior to the optimization process, the

effects of the design variables related to the geometry and flow conditions of the micromixer, viz.

the ratios of the amplitude to the wavelength of the sinusoidal walls A/λ and the throat width to

15
depth of the convergent–divergent section w/d, the Strouhal number St, and the velocity ratio

Vo/Vs, on the overall mixing index were analyzed at a fixed Reynolds number of Re = 0.5. At a

fixed velocity ratio (Vo/Vs = 1.88), the mixing index increased with increasing amplitude of the

sinusoidal walls and decreased with increasing throat width of the convergent–divergent channel.

Conversely, at a fixed Strouhal number (St = 0.278), the mixing performance remained almost

unchanged with varying A/λ over the range of considered velocity ratios. However, the mixing

performance decreased with increasing w/d for higher velocity ratios Vo/Vs ≥ 0.94. The

optimization was performed at Re = 0.5 with three different surrogate models, viz. the RSA,

RBNN, and KRG models, which were constructed based on the objective function values

calculated at 36 design points selected by LHS. Ultimately, the results show that the KRG model

predicted the best optimum point, i.e., it yields the highest overall mixing index (92.35%), and

also showed the highest accuracy in predicting the objective function value compared with the

CFD calculations (relative error of 1.87%). This optimum design demonstrates that the optimum

operating conditions of the pulsatile flow are St = 0.2492 and Vo/Vs = 3.2676 and the optimum

geometric conditions of the convergent–divergent microchannel are A/λ = 0.25 and w/d = 0.75.

Nomenclature

A Amplitude of sinusoidal walls (m)

f Frequency of pulsed sinusoidal flow (s-1)

Vs Steady flow velocity (m⋅s-1)

Vo Amplitude of pulsed sinusoidal flow (m⋅s-1)

d Depth of the channel (m)

16
Lo Length of the inlet channel (m)

Lc Length of the main channel (m)

Le Length of the exit channel (m)

Ci Species concentration

N Number of sampling points

w Throat width of the convergent–divergent section (m)

MIo Mixing index at exit

Re Reynolds number

St Strouhal number

x, y, z Spatial coordinates in the x-, y-, and z-directions

Greek letters

λ Pitch/wavelength (m)

ν Kinematic viscosity of fluid (kg/m⋅s)

ρ Fluid density (kg/m3)

α Diffusivity coefficient (m2⋅s–1)

φ Phase difference

Acknowledgment

This work was supported by the National Research Foundation of Korea (NRF) grant (No.

20090083510) funded by the Korean government (MSIP) through the Multi-Phenomena CFD

Engineering Research Center. The authors gratefully acknowledge this support.

17
References

[1] H. A. Stone and S. Kim, Microfluidics: Basic issues, applications and challenges, A.I.Ch.E.

J. 47 (2001) 1250-1254.

[2] T. Zhang, X. Zhang, X. Yan, L. Kong, G. Zhang, H. Liu, J. Qiu and K. L. Yeung, Synthesis

of Fe3O4@ZIF-8 magnetic core-shell microspheres and their potential application in a capillary

micro reactor. Chem. Eng. J. 228 (2013) 398-404.

[3] G. S Jeong, S. Chung, C. -B. Kim, S.-H. Lee, Applications of micromixing technology.

Analyst, 135 (2010) 460-473.

[4] D. R. Meldrum and M.R. Holl, Microfluidics; Microscale bioanalytical systems, Science

297 (2002) 1197-1198.

[5] T. N. Ng, X. Q. Chen and K. L. Yeung, Direct manipulation of particle size and morphology

of ordered mesoporous silica by flow synthesis. RSC Advances 5 (2015) 13331-13340.

[6] A. D. Stroock, S. K. Dertinger, A. Ajdari, I. Mezic, H. A. Stone, and G. M. Whitesides,

Chaotic mixer for microchannels, Science 295 (2002) 647-651.

[7] F. Jiang, K. S. Drese, S. Hardt, M. Kupper, F. Schonfeld, Helical flows and chaotic mixing

in curved micro channels, AIChE J. 50 (9) (2004) 2297-2305.

[8] Y. Lin, Numerical characterization of simple three-dimensional chaotic micromixer. Chem.

Eng. J. 277 (2015) 303-311.

[9] M. S. Cheri, H. Latifi, M. S. Moghaddam and H. Shahraki, Simulation and experimental

investigation of planar micromixers with short-mixing-length. Chem. Eng. J. 234 (2013) 247-

255.

[10] J. K. Chen and R. J. Yang, Electroosmotic flow mixing in zigzag microchannels,

Electrophoresis 28 (2007) 975-983.

18
[11] A. O. E. Moctar, N. Aubry, and J. Batton, Electro-hydrodynamic Micro-Fluidic Mixer,

Lab Chip 3 (2003) 273-280

[12] M. H. Oddy, J. G. Santiago, and J. C. Mikkelsen Electrokinetic Instability Micromixing,

Anal Chem 73 (2001), 5822-5832

[13] S. C. Jacobson, T. E. McKnight, and J. M. Ramsey, Microfluidic devices for

electrokinematically driven parallel and serial mixing, Anal. Chem. 71 (1999) 4455–4459.

[14] H. H. Bau, J. Zhong and M. Yi, A Minute Magneto Hydrodynamic (MHD) Mixer, Sens

Actuators B 79 (2001) 207-215.

[15] J. Aubin, D.F. Fletcher, C. Xuereb, Design of Micromixers using CFD modeling, Chem.

Eng. Sci. 60 (2005) 2503–2516.

[16] H. Wang, P. Iovenitti, E. Harvey, S. Masood, Numerical investigation of mixing in

microchannels with patterned grooves, J. Micromech. Microeng. 13 (2003) 801-808.

[17] M. A. Ansari and K.-Y. Kim, Parametric study on mixing of two fluids in a three-

dimensional serpentine microchannel, Chem. Eng. J. 146 (2009) 439-448.

[18] M. A. Ansari and K.-Y. Kim, Shape optimization of a micromixer with staggered

herringbone groove, Chem. Eng. Sci. 62 (2007) 6687-6695.

[19] M. A. Ansari and K.-Y. Kim, Application of the radial basis neural network to optimization

of a micromixer. Chem. Eng. Tech. 30 (7) (2007) 962-966.

[20] I. Glasgow and N. Aubry, Enhancement of Microfluidic Mixing using Time Pulsing, Lab

Chip, 3 (2003), 114-120.

[21] I. Glasgow, S. Lieber and N. Aubry, Parameters influencing pulsed flow mixing in

microchannels, Anal. Chem. 76 (2004), 4825-4832.

19
[22] I. Glasgow and N. Aubry, Pulsed flow mixing for BIOMEMS applications, 7th International

conference on Miniaturized Chemical and Biochemical analysis systems, California, USA,

October, 2003.

[23] F. Bottauski, C. Cardonne, C. Meinhart and I. Mezi´c, An ultrashort mixing length

micromixer: The shear superposition micromixer, Lab Chip 7 (2007) 396-398.

[24] A. Afzal and K.-Y. Kim, Convergent-divergent micromixer coupled with Pulsatile flow,

Sens Actuators B: Chemical, 211 (2015), 198-205.

[25] CFX-12.1, Solver Theory, ANSYS Inc., Canonsburg, Pennsylvania, 2006

[26] R. B. Bird, W. E. Stewart, E. N. Lightfoot, Transport Phenomenon (1960) Wiley, New

York, USA

[27] J. P. Van Doormaal, G. D. Raithby, Enhancement of the SIMPLE method for predicting

incompressible fluid flows, Numerical Heat Transf. 7 (1985) 147-163.

[28] R. H. Myers, and D. C. Montgomery, Response Surface Methodology: Process and Product

Optimization Using Designed Experiments (1995) Wiley, New York.

[29] J. D. Martin and T. W. Simpson, Use of Kriging models to approximate deterministic

computer models, AIAA Journal 43 (4) (2005) 853-863.

[30] C. Chen, C. F. N. Cowan, and P. M. Grant, Orthogonal least squares learning algorithm for

radial basis function networks, IEEE transaction of neural networks 2 (1991) 302-309.

[31] MATLAB R2014b, The Language of Technical Computing, Version 8.4, The MathWorks

Inc, 2014.

20
Figure captions:

Fig. 1. Geometry of convergent–divergent microchannel.

Fig. 2. Hexahedral grid system.

Fig. 3. Comparison of dye mass fraction distributions in the central x–y plane (i.e., the plane at
half of the channel depth, z = 0.0625 mm) determined by the present computation (left) and the
experiment by Glasgow and Aubry [15] (right) for two cases: (a) without pulsing at both inlets
(Re = 0.3) and (b) anti-phased pulsing (St = 0.13, Re = 0.3, and Vo/Vs = 7.5) [20].

Fig. 4. Variation of mixing index with Strouhal number for different velocity ratios at Re = 0.5.

Fig. 5. Dye mass fraction distributions in the central x–y plane at time t = 0/T at Re = 0.5; (a)
Vo/Vs = 0.93 and (b) Vo/Vs = 3.76.

Fig. 6. Variation of mixing index with (a) Strouhal number at Vo/Vs = 1.88 and (b) velocity ratio
at St = 0.278 for different A/λ at Re = 0.5.

Fig. 7. Dye mass fraction distributions in the central x–y plane at time t = 0/T for three different
Strouhal numbers at Re = 0.5 and Vo/Vs = 1.88; (a) A/λ = 0.1 and (b) A/λ = 0.2.

Fig. 8. Dye mass fraction distributions in the central x–y plane at time t = 0/T for three different
velocity ratios at Re = 0.5 and St = 0.278; (a) A/λ = 0.1 and (b) A/λ = 0.2.

Fig. 9. Variation of mixing index with (a) Strouhal number at Vo/Vs = 1.88 and (b) velocity ratio
at St = 0.278 for different w/d at Re = 0.5.

Fig. 10. Dye mass fraction distributions in the central x–y plane at time t = 0/T for three different
Strouhal numbers at Re = 0.5 and Vo/Vs = 1.88; (a) w/d = 0.8 and (b) w/d = 1.2.

Fig. 11. Dye mass fraction distributions in the central x–y plane at time t = 0/T for three different
velocity ratios at Re = 0.5 and St = 0.278; (a) w/d = 0.8 and (b) w/d = 1.2.

Fig. 12. Dye mass fraction distributions in the central x–y plane at t = 0/T in select LHS designs
(Table II).

Fig. 13. Dye mass fraction distributions in the central x–y plane at t = 0/T in reference [20] and
optimum designs.

Fig. 14. Change in dye mass fraction distribution in the central x–y plane with time at half of the
channel depth in the optimum design (St = 0.2492, Vo/Vs = 3.2676, A/λ = 0.25, and w/d = 0.75).

21
Table captions:

Table I: Design variables and ranges

Table II: Design variables and objective function values at design points
Table III: Results of optimization

22
Inlet 1 y

x
λ

w w
Outlet
Le
Li
Lc

Inlet 2 Fig. 1. Geometry of convergent–divergent microchannel.

23
Fig. 2. Hexahedral grid system.

24
Fig. 3. Comparison of dye mass fraction distributions in the central x–y plane (i.e., the plane at
half of the channel depth, z = 0.0625 mm) determined by the present computation (left) and the
experiment by Glasgow and Aubry [15] (right) for two cases: (a) without pulsing at both inlets
(Re = 0.3) and (b) anti-phased pulsing (St = 0.13, Re = 0.3, and Vo/ Vs = 7.5) [20].

25
Fig. 4. Variation of mixing index with Strouhal number for different velocity ratios at Re
= 0.5.

26
 St = 0.070

St = 0.278

St = 0.556
(a)

St = 0.070

St = 0.278

y
St = 0.556
(b)
x

Fig. 5. Dye mass fraction distributions in the central x–y plane at time t = 0/T at Re = 0.5; (a)
Vo/Vs = 0.93 and (b) Vo/Vs = 3.76.
27
 (a) (b)
Fig. 6. Variation of mixing index with (a) Strouhal number at Vo/Vs = 1.88 and (b) velocity
ratio at St = 0.278 for different A/λ at Re = 0.5.

28
 Vo/Vs = 1.88, St = 0.070

Vo/Vs = 1.88, St = 0.278

Vo/Vs = 1.88, St = 0.556

(a)

Vo/Vs = 1.88, St = 0.070

Vo/Vs = 1.88, St = 0.278

Vo/Vs = 1.88, St = 0.556

(b)
x

Fig. 7. Dye mass fraction distributions in the central x–y plane at time t = 0/T for three
different Strouhal numbers at Re = 0.5 and Vo/Vs = 1.88; (a) A/λ = 0.1 and (b) A/λ = 0.2.

29
 St = 0.278, Vo/Vs = 0.46

St = 0.278, Vo/Vs = 1.88

St = 0.278, Vo/Vs = 3.76 (a)

St = 0.278, Vo/Vs = 0.46

St = 0.278, Vo/Vs = 1.88

St = 0.278, Vo/Vs = 3.76 (b)

x

Fig. 8. Dye mass fraction distributions in the central x–y at time t = 0/T for three different
velocity ratios at Re = 0.5 and St = 0.278; (a) A/λ = 0.1 and (b) A/λ = 0.2.

30
 (a) (b)
Fig. 9. Variation of mixing index with (a) Strouhal number at Vo/Vs = 1.88 and (b) velocity
ratio at St = 0.278 for different w/d at Re = 0.5.

31
 Vo/Vs = 1.88, St = 0.070

Vo/Vs = 1.88, St = 0.278

Vo/Vs = 1.88, St = 0.556 (a)

Vo/Vs = 1.88, St = 0.070

Vo/Vs = 1.88, St = 0.278

Vo/Vs = 1.88, St = 0.556 (b)

x

Fig. 10. Dye mass fraction distributions in the central x–y plane at time t = 0/T for three
different Strouhal numbers at Re = 0.5 and Vo/Vs = 1.88; (a) w/d = 0.8 and (b) w/d = 1.2.

32
 Vo/Vs = 0.46, St = 0.278

Vo/Vs = 1.88, St = 0.278

Vo/Vs = 3.76, St = 0.278 (a)

Vo/Vs = 0.46, St = 0.278

Vo/Vs = 1.88, St = 0.278

Vo/Vs = 3.76, St = 0.278 (b)

x

Fig. 11. Dye mass fraction distributions in the central x–y plane at time t = 0/T for three
different velocity ratios at Re = 0.5 and St = 0.278; (a) w/d = 0.8 and (b) w/d = 1.2.

33
 DP 6: A/λ = 0.161, w/d = 1.01, Vo/Vs = 3.86, St = 0.273; MIo = 0.880

DP 7: A/λ = 0.124, w/d = 0.81, Vo/Vs = 2.44, St = 0.071; MIo = 0.574

DP 8: A/λ = 0.121, w/d = 1.17, Vo/Vs = 3.70, St = 0.139; MIo = 0.752

x
DP 29: A/λ = 0.175, w/d = 0.87, Vo/Vs = 0.81, St = 0.402; MIo = 0.341

Fig. 12. Dye mass fraction distributions in the central x–y plane at t = 0/T in select LHS
designs (Table II).

34
 Reference: A/λ = 0.15, w/d = 0.8, Vo/Vs = 3.70, St = 0.278; MIo = 0.924

x
x
Optimum: A/λ = 0.25, w/d = 0.75, Vo/Vs = 3.27, St = 0.249; MIo = 0.924

Fig. 13. Dye mass fraction distributions in the central x–y plane at t = 0/T in reference [20]
and optimum designs.

35
 t=T

t = 3T/4

t = T/2

t = T/4

y
t=0

Fig. 14. Change in dye mass fraction distribution with time in the central x–y plane in the
optimum design (St = 0.2492, Vo/Vs = 3.2676, A/λ = 0.25, and w/d = 0.75).
36
Table I: Design variables and ranges

Design variables

Bounds Flow Geometry

St Vo/Vs A/λ w/d

Lower (LB) 0.05 0.5 0.10 0.75

Upper (UB) 0.60 4.5 0.25 1.20

37
Table II: Design variables and objective function values at design points

Design Objective
Design variables
Point function

St Vo/Vs A/λ w/d MIo

1 0.1727 2.9156 0.1834 1.1231 0.8444
2 0.5951 0.9230 0.1043 1.0766 0.2967
3 0.3085 2.7258 0.1480 0.8072 0.8731
4 0.4659 1.1138 0.1067 1.0455 0.4065
5 0.0552 2.9912 0.1298 0.9871 0.5346
6 0.2731 3.8627 0.1612 1.0087 0.8799
7 0.0713 2.4379 0.1243 0.8113 0.5737
8 0.1394 3.7021 0.1216 1.1724 0.7525
9 0.4508 0.6012 0.1412 0.9269 0.2565
10 0.5467 3.4119 0.1024 1.1640 0.6526
11 0.5056 1.2283 0.1783 1.0237 0.4341
12 0.2191 3.0317 0.1575 0.9539 0.8738
13 0.3237 3.2891 0.1348 1.0185 0.8337
14 0.3256 2.8233 0.1544 0.9971 0.8235
15 0.3635 1.3737 0.1942 1.1926 0.5170
16 0.4383 2.6247 0.1755 1.0832 0.7089
17 0.2589 2.3193 0.1153 0.8304 0.8460
18 0.2438 3.1970 0.1257 1.1197 0.8426
19 0.3845 1.5848 0.1976 0.9422 0.6495
20 0.0966 2.2158 0.1808 0.8489 0.7016
21 0.2885 3.5000 0.1704 0.9211 0.8846
22 0.3462 1.7993 0.1307 1.0913 0.6884
23 0.2116 0.7343 0.1900 1.1433 0.3413
24 0.5626 0.9965 0.1109 0.8399 0.3718
25 0.5831 3.9364 0.1512 0.9067 0.7722
26 0.4002 3.5460 0.1864 0.8904 0.8429
27 0.0935 2.0587 0.1604 1.0613 0.6843
28 0.1562 1.4626 0.1673 0.9739 0.7316
29 0.4019 0.8062 0.1745 0.8662 0.3405
30 0.4275 2.0223 0.1115 1.1091 0.6475
31 0.1997 0.5410 0.1455 1.1445 0.2611
32 0.5103 2.5244 0.1421 0.8773 0.7498
33 0.1250 1.7028 0.1962 0.9591 0.7583
34 0.5299 1.5616 0.1368 1.1881 0.4557
35 0.1721 1.9113 0.1178 0.8843 0.7582
36 0.4928 3.7713 0.1653 1.0349 0.7658

38
Table III: Results of optimization

Design variables Objective function

Surrogate Flow Geometry Prediction CFD Error (%)

St Vo/Vs A/λ w/d

Ref. [20] 0.2780 3.703 0.1500 0.8000 - 0.9078

RSA 0.2741 3.2714 0.2500 0.7500 0.9640 0.9167 5.16

RBNN 0.2250 3.2597 0.2200 0.8528 0.9287 0.9084 2.24

KRG 0.2492 3.2676 0.2500 0.7500 0.9408 0.9235 1.87

39
Table I: Design variables and ranges

Design variables

Bounds Flow Geometry

St Vo/Vs A/λ w/d

Lower (LB) 0.05 0.5 0.10 0.75

Upper (UB) 0.60 4.5 0.25 1.20

Table II: Design variables and objective function values at design points

Design Objective
Design variables
Point function

St Vo/Vs A/λ w/d MIo

1 0.1727 2.9156 0.1834 1.1231 0.8444
2 0.5951 0.9230 0.1043 1.0766 0.2967
3 0.3085 2.7258 0.1480 0.8072 0.8731
4 0.4659 1.1138 0.1067 1.0455 0.4065
5 0.0552 2.9912 0.1298 0.9871 0.5346
6 0.2731 3.8627 0.1612 1.0087 0.8799
7 0.0713 2.4379 0.1243 0.8113 0.5737
8 0.1394 3.7021 0.1216 1.1724 0.7525
9 0.4508 0.6012 0.1412 0.9269 0.2565
10 0.5467 3.4119 0.1024 1.1640 0.6526
11 0.5056 1.2283 0.1783 1.0237 0.4341
12 0.2191 3.0317 0.1575 0.9539 0.8738
13 0.3237 3.2891 0.1348 1.0185 0.8337
14 0.3256 2.8233 0.1544 0.9971 0.8235
15 0.3635 1.3737 0.1942 1.1926 0.5170
16 0.4383 2.6247 0.1755 1.0832 0.7089
17 0.2589 2.3193 0.1153 0.8304 0.8460
18 0.2438 3.1970 0.1257 1.1197 0.8426
19 0.3845 1.5848 0.1976 0.9422 0.6495
20 0.0966 2.2158 0.1808 0.8489 0.7016
21 0.2885 3.5000 0.1704 0.9211 0.8846
22 0.3462 1.7993 0.1307 1.0913 0.6884
23 0.2116 0.7343 0.1900 1.1433 0.3413
24 0.5626 0.9965 0.1109 0.8399 0.3718
25 0.5831 3.9364 0.1512 0.9067 0.7722
26 0.4002 3.5460 0.1864 0.8904 0.8429
27 0.0935 2.0587 0.1604 1.0613 0.6843
28 0.1562 1.4626 0.1673 0.9739 0.7316
29 0.4019 0.8062 0.1745 0.8662 0.3405
30 0.4275 2.0223 0.1115 1.1091 0.6475
31 0.1997 0.5410 0.1455 1.1445 0.2611
32 0.5103 2.5244 0.1421 0.8773 0.7498
33 0.1250 1.7028 0.1962 0.9591 0.7583
34 0.5299 1.5616 0.1368 1.1881 0.4557
35 0.1721 1.9113 0.1178 0.8843 0.7582
36 0.4928 3.7713 0.1653 1.0349 0.7658
Table III: Results of optimization

Design variables Objective function

Surrogate Flow Geometry Prediction CFD Error (%)

St Vo/Vs A/λ w/d

Ref. [20] 0.2780 3.703 0.1500 0.8000 - 0.9078

RSA 0.2741 3.2714 0.2500 0.7500 0.9640 0.9167 5.16

RBNN 0.2250 3.2597 0.2200 0.8528 0.9287 0.9084 2.24

KRG 0.2492 3.2676 0.2500 0.7500 0.9408 0.9235 1.87

Highlights

- Operating conditions and geometry of a convergent–divergent micromixer with pulsed fl

ow was optimized at Re=0.5.
- The optimization was performed with three different surrogate models, viz. the RSA, RB
NN, and KRG models.
- The KRG model predicted the best optimum point with highest overall mixing index (92.
35%).
- The optimum design variables; St = 0.2492, Vo/Vs = 3.2676, A/λ = 0.25, and w/d = 0.75.

45