Professional Documents
Culture Documents
Development of Computational Fluid Dynamics Based Artificial Neural Network Metamodels For Wastewater Disinfection
Development of Computational Fluid Dynamics Based Artificial Neural Network Metamodels For Wastewater Disinfection
Development of Computational Fluid Dynamics Based Artificial Neural Network Metamodels For Wastewater Disinfection
A Thesis
Submitted to the Faculty
of
Drexel University
by
Wangshu Wei
in partial fulfillment of the
requirements for the degree
of
Doctor of Philosophy
September 2018
iii
To My Family
iv
ACKNOWLEDGEMENTS
First of all, I want to thank my academic advisors, Professors Bakhtier Farouk, and
Charles Haas. I would not be successful without their kind support, guidance and
mentoring. With their consideration, I conquered all the challenges on the road. With
their help, I went through the most difficult time of my life during this period. With their
advice, I corrected the direction of the path of my life. They are not only my teachers in
Secondly, I am very very grateful to my parents. They sacrificed their life to realize my
dream. Their support was the strongest motivation for me to achieve what I believed.
During the past five years, we were ten thousand kilometers apart, we did not have much
time to be together. Their perseverance was the example that I am following all the time.
At last, I would say thankyou to my beloved wife, Yue Wang. It was her love, companion
and caring that made this thesis possible. Her love was my courage when facing
predicament, her companion was my consolation in this foreign place, her caring was my
Thank PeroxyChem (Mr. Philip Block ), the department of Mechanical Engineering and
Mechanics and the L.D. Betz Chair Funds (via Professor Charles Haas) for financial
support.
v
Table of Contents
ACKNOWLEDGEMENTS .......................................................................................................... iv
LIST OF FIGURES...................................................................................................................... vii
LIST OF TABLES......................................................................................................................... xi
ABSTRACT .................................................................................................................................. xii
I. INTRODUCTION .................................................................................................................... 1
II. BACKGROUND AND LITERATURE REVIEW .............................................................. 5
2.1 CFD techniques in wastewater process ............................................................................. 5
2.2 Turbulent modeling in CFD ............................................................................................... 8
2.3 Applications of PAA in wastewater treatment................................................................ 13
2.4 ANN Based Metamodel ..................................................................................................... 16
III. COMPUTATIONAL FLUID DYNAMICS MODELING .............................................. 19
3.1 Introduction ....................................................................................................................... 19
3.2 Mathematical model .......................................................................................................... 19
3.3 Turbulent Modeling .......................................................................................................... 21
3.4 Grid Generation and Independence Study ..................................................................... 24
3.5 Validation of CFD models................................................................................................. 26
IV. MIXING OF TWO GASEOUS FLOWS AND MISCIBLE LIQUIDS IN A T-
JUNCTION ......................................................................................................................... 32
4.1 Introduction ....................................................................................................................... 32
4.2 Problem Statement ............................................................................................................ 34
4.3 Mathematical Formulation ............................................................................................... 38
4.4 Numerical Scheme, Boundary and Initial Conditions ................................................... 41
4.5 Results and Discussions..................................................................................................... 44
V. THERMAL MIXING OF TWO LIQUIDS IN A T-JUNCTION ..................................... 67
5.1 Introduction ....................................................................................................................... 67
5.2 Problem Description.......................................................................................................... 68
5.3 Methodology....................................................................................................................... 70
5.4 Results and Discussions..................................................................................................... 77
5.5 Conclusions ........................................................................................................................ 89
VI. CFD SIMULATIONS ON PILOT SCALE WASTEWATER TREATMENT
REACTOR WITH THE STUDY OF MIXING ENHANCEMENT....................................... 90
vi
LIST OF FIGURES
Figure 1 Techniques for metamodeling (Wang and Shan, 2007) .................................................. 18
Figure 2 The grid generation of a T-junction pipe (with 900,000 cells) ....................................... 25
Figure 4 Comparison of methane concentration from the simulations and the experiments at R =
11.0, d/D = 0.014, z/D = 3.0, Rem = 21,000, Rej = 4000, Sct = 0.9 ............................................. 27
Figure 5 Comparison of grayscale images (a) experiment (b) CFD simulation at 15:00 Sec ....... 29
Figure 6 T-junction simulation with inclined jet inlet (a) 3-degree forward; (b) 3-degree backward
....................................................................................................................................................... 31
Figure 8 Schematic of 3-D T-junction geometry considered (not for scale) ................................. 35
solution cases along the axial length for d/D = 0.1 and 0.2 ........................................................... 44
Figure 11 Contour plot of normalized (a) methane, Sc = 0.75; (b) PAA, Sc = 700 concentration at
Figure 12 Plot of degree of laminar mixing along the downstream length for the air-methane and
the water-PAA solution cases, d/D = 0.1 and 0.2, Rem = 200, Rej = 40 ...................................... 48
Figure 13 Comparison of methane concentration from the simulations and the experiments
(Forney and Lee, 1982, Forney, Noureddine and Hanh, 1996) at R = 11.0, d/D = 0.014, z/D = 3.0,
Figure 15 Contour plots of methane concentration at center vertical plane at different times d/D =
Figure 16 Contour plot of methane concentration at the exit plane, time average from 3.600 sec to
3.620 sec; R = 2.45 and d/D = 0.2, Rem = 21,000, Rej = 6000, Sct = 0.9 .................................... 53
Figure 17 Comparison of liquid-liquid mixing and gas-gas mixing at R = 2.45, d/D = 0.2, Rem =
21000, Rej = 6000 Sct = 0.9 and R = 3.73, d/D = 0.1, Rem = 21000, Rej = 4000 Sct = 0.9 ........ 55
Figure 18 Contour plots of PAA solution concentration at center vertical plane at different times
(a) t =20.0 s (b) t =20.5 s (c) t = 21.0 s (d) t = 21.5 s and (e) t = 22.0 s; ....................................... 57
Figure 19 Contour plot of PAA concentration at the exit plane, with R = 3.73 and d/D = 0.1, time
average period: 49.80 sec to 50.00 sec, Rem = 21,000, Rej = 4000, Sct = 0.9 ............................. 58
Figure 20 Comparison of grayscale images of the side-view instantaneous concentration field for
R = 3.06, d/D = 0.1667, Rem = 20,850, Rej = 10,630. Screenshot of simulation is taken at 15.00
sec .................................................................................................................................................. 60
Figure 21 Instantaneous velocity field on the center plane of the tee mixer, R = 3.06, d/D =
0.1667, Rem = 20,850, Rej = 10,630. Screenshot of simulation is taken at 15.00 sec .................. 62
Rem=20,850, Rej=10,630. Time averaging period from the present CFD simulation: 22.0 sec to
Figure 23 Contour plots of tracer (sodium fluorescein) concentration at center vertical plane at
different times (a) t =15.0 s (b) t =15.5 s (c) t = 16.0 s (d) t = 16.5 s and (e) t = 17.0 s; d/D =
Figure 24 Schematic of 3-D T-junction geometry considered (not for scale) ............................... 69
Figure 25 Grid Generation of the T-junction (a)prospective view; (b) Cross-sectional view of the
T-junction ...................................................................................................................................... 74
Figure 26 y+ at the top wall of the main pipe after the jet nozzle at y = 0.0381m. ....................... 76
ix
Figure 27 Visualized image of the interface between two flows in Z/A = 0 from (a) Experiment;
Figure 28 Mean velocity vector diagram in the symmetric plane of Z/A = 0. (The solid lines show
the time-averaged interface between the main and branch flows obtained by the flow
Figure 29 Profiles of the mean normalized temperature at (a) x = 0; (b) x = 1.0; (c) x = 3.0 ....... 81
Figure 30 Temperature contour at center plane at (a) t = 15.00 sec; (b) t = 15.30 sec; (c) t = 15.60
sec; (d) t = 15.90 sec; (e) t = 16.20 sec; at TR = 1.28, VR = 3.0. .................................................. 83
Figure 31 Time-averaged temperature contour plot at exit plane (15 pipe diameter downstream).
Figure 32 Temperature contour at center plane at (a) t = 15.00 sec; (b) t = 15.30 sec; (c) t = 15.60
sec; (d) t = 15.90 sec; (e) t = 16.20 sec; at TR = 1.09, VR = 3.0. .................................................. 86
Figure 33 Time-averaged temperature contour plot at exit plane (15 pipe diameter downstream).
Figure 34 Comparisons of the degree of mixing between TR = 1.28 and TR = 1.09, where (a) VR
Figure 39 (a) mechanical diamgram of baffle plate. (b) CFD geometric model of baffle plate .... 93
Figure 40 Contour plots of PAA concentration at 5.0 sec simulating time at (a) center plane; (b)
exit ................................................................................................................................................. 94
Figure 41 Instanteous contour plot of PAA in the first channel of the serpentine contactor at 9.0
sec. ................................................................................................................................................. 95
x
Figure 42 Three jet inlet configurations. (a) Inserted straightly (b) inserted with 90-degree turn
Figure 43 Contour plots of the tracer concentration at center vertical plane (a) Inserted straightly
(b) inserted with 90-degree turn downstream; (c) inserted with 90-degree turn upstream............ 97
Figure 44 Contour plots of tracer concentration at cross-sectional planes (a) Inserted straightly (b)
inserted with 90-degree turn downstream; (c) inserted with 90-degree turn upstream ................. 98
Figure 45 Contour plots of tracer concentration at the outlet (a) inserted vertically; (b) inserted
with a 90-degree turn downstream; (c)inserted with a 90-degree turn upstream ........................ 100
Figure 47 The plot of pressure head (a) inserted vertically; (b) inserted with a 90-degree turn
Figure 48 (a): top view; (b) prospective view of a full-scale reactor. ......................................... 106
Figure 49 (a) Contour plot of PAA Residual concentration; (b) Contour plot of N survival ratio at
Figure 57 Validation plot with experiments in the pilot scale reactor. A 1 log experimental error
LIST OF TABLES
Table 1 Listing of the main variables ............................................................. 37
Table 2 Listing of the main variables for the mixing problems ...................... 70
Table 10 Parameter table for the experiment after scale-up ......................... 125
Table 13 Average bias and Mean squared error in the sensitivity study ...... 127
xii
ABSTRACT
The mechanisms of the mixing of two streams are investigated numerically. The transport
processes of both mass and heat transfer in the mixing of either two gases or two miscible
liquids are studied with 3-D time-dependent computational fluid dynamics (CFD)
simulations. The flow fields are calculated for the T-junction (of two circular cross-
section pipes that meet orthogonally at a junction) which is one of the most common
mixing devices in the study of mass and heat transfer. For turbulent flow regimes, the
applied in the simulation to minimize numerical diffusion. The results obtained by the
numerical simulations are verified with available experimental data in the literature for
specific water characteristics. On-site physical modelling on pilot scale has been an
important and usual method to validate the kinetics and the performance for a particular
system. However, physical modelling is expensive and time consuming. CFD technique
more and more achievable. With proven accuracy on predicting the performance of full
scale facilities, CFD is now playing a significant role in the wastewater treatment
can be achieved within hours on PCs, which significantly lower the cost of undertaking
with multiple non-linear features. With a parametric study of CFD simulations providing
variety contactors with different geometries, inlet conditions and chemical kinetics. In
this thesis, the results of microorganism inactivation and PAA decay in full scale
contactors using three-dimensional CFD models are presented. Then the results of the
metamodels were validated with both CFD simulations and pilot scale experiments.
1
I. INTRODUCTION
Mixing of fluid streams plays a key role in many areas in the industry and in nature. An
optimum condition of mixing is desired for enhancement of species mixing, heat and mass
transfer, combustion and other chemical reactions. For example, good mixing can promote
One of the most encountered mixing devices is a T-junction pipe. The T-junction containing
a transverse pipe and a 90º branch is widely used in mixing of two fluids (in gaseous or
liquid form). The geometry can be utilized in a chemical reactor, combustion chamber,
piping system, etc. Gas mixture flows appear in diverse applications. Examples include
pollutant dispersion, chemical processing, and combustor mixing and reaction. Previous
discussion was carried out on the mixing of two gases (air and methane) with no heat
transfer considered in a T-junction (Forney and Lee, 1982). However, it is also common
that the two streams are having different temperature. For example, cool air flow is mixing
with a hot air flow in an HVAC unit (Hirota, Mohri, Asano and Goto, 2010). Liquid-liquid
mixing problems with temperature variations are also common in industry and in nature. A
previous study of liquid-liquid mixing problem was the mixing of water with water-sodium
solution in a T-junction (Pan and Meng, 2001). T-junctions are also used for the mixing of
wastewater and disinfectant in wastewater treatment process, the mixing of hot water with
Vaidyanathan, 2010), etc. Therefore, the hot water mixing with cooling fluid is considered
The mixing of two flows in a T-junction mixing chamber contains mass and heat transport
processes, thus the process is complex. Previous studies on T-Junction mixing showed the
mixing behaviors and flow structures between the main flow and the jet in detail (Wei,
Farouk and Haas, 2017). With increasing power of computing, CFD simulations are applied
in the present study to predict the flow properties and mixing characteristics of two fluids
computational cost and flow resolution, LES is becoming one of the most popular turbulent
modeling techniques in CFD. LES has been proven to be accurate on both mass transfer
In this study, CFD simulations of the mixing of two gaseous flows is carried out first for
validation purpose. In the gaseous mixing case, a cold air flow is introduced from the main
inlet, a stream of hot air is injected from the jet flow inlet. The main flow Reynolds number
is denoted by Rem, inlet jet Reynolds number is denoted by Rej. The mixing of two streams
is studied under turbulent flow conditions, where Rem ~ 15,000. After the fidelity of the
CFD model is validated, another series of simulations are focused on the mixing of two
water streams, i.e. cool water flow mixing with hot disinfectant solution (0.3% water-
peracetic acid mixture) to represent the scenario in a pilot scale wastewater treatment
reactor. In the liquid-liquid study, cold water flow is introduced from the main pipe inlet
and hot water is injected from the jet flow inlet. The mixing of the two liquid flows are
Peracetic acid (PAA) is an efficient disinfectant for its wide spectrum of antimicrobial
industry has been emerging in recent years. The characteristics of PAA for microbial
wastewater disinfection such as ease of implementing treatment (without the need for
products, no quenching requirement (i.e., no de-chlorination), short contact time, are all
desirable to the wastewater treatment industry, which makes PAA gain its popularity in
prevent their use in optimal design. An artificial neural network (ANN) based metamodel is
wastewater treatment process would reduce the computation cost dramatically (Hoque, Farouk
and Haas 2010; Wei, Farouk and Haas 2016). Therefore, in this study, ANN based metamodel to
approximate CFD results was assessed. The ANN metamodel was trained from a series of CFD
design space was sampled using a quasi-random sampling technique. Quasi-random sampling is
one of the ‘space-filling’ designs (Wang and Shan, 2007). To generate space filling designs
for computer experiments or numerical simulations, sampling methods are applied to select
the sequence (Hoque, 2011). 40 different CFD cases on an 11-dimensional space are obtained
and used as input to the training process of metamodel development. Metamodels were
developed to predict disinfectant residual concentration and microbial inactivation rate on full
scale reactors. The performance of the ANN based metamodel is evaluated by comparison to
4
approximation method to a high dimensional nonlinear system, the ANN based metamodel shows
its ability to provide an efficient yet accurate solution to the wastewater disinfection process with
PAA.
5
After being applied for over twenty years, CFD has made significant contributions to the
design and optimization in the wastewater disinfection industry. CFD has been used, not
only in flow simulations, but also in reacting processes such as disinfection, and in the
efficiency, and flexibility that can be used to resolve most kinds of problems, it is still
research that applied CFD, the common stages of a complete disinfection process
simulation are: flow simulation, tracer transport simulation, reaction process simulation,
In order to study the hydraulic efficiency of the water treatment system, tracer transport
simulation is applied. In the simulation, tracer is always treated as an inert particle. Either
equation for concentration has been applied. However, latter method is more widespread
since most CFD codes are based on Eulerian system. For reaction process simulation, the
main purposes are to predict inactivation, and also potentially formation of disinfectant
byproducts. With this purpose, a transport equation of all species has to be solved. In the
The first two-phase flow CFD code for ozone disinfection came into being at the end of
1990s (Cockx et al., 1999; Ta and Hague, 2004; Bolaños et al., 2008; Bartrand et al.,
2009; Talvy et al., 2011) Within last two decades, the development of CFD has been
conducted to solve more complicated problem, increase the accuracy, and save the
computation costs. Not only simulating flows in lab-scale contactors, but also applied to
full-scale reactors, CFD became the most reliable and most widely used method in this
field comparing to other modeling methods such as plug flow reactor (PFR) or
completely mixed flow reactor (CMFR) (Zhang et al., 2014). For parametric studies,
there are many parameters that would affect the disinfection process. The parameters
include: pH, temperature, flow rate, bubble size variation (for ozone disinfection), inlet
zones are the key to achieve the optimal efficiency. Increasing the number of baffles,
rearranging the reactors, adjusting the locations of inlet, outlet, and diffusers would make
contributions.
Though CFD techniques have been being developed rapidly in last two decades, there are
still several challenges that we are facing. These challenges include: 1) unsteady flow
uncertainty of inactivation kinetics, and 5) closure problem for chemical source terms.
These challenges may cause a lot of inaccuracies that could limit their applicability.
In order to overcome unsteady flow structure effect, the method has to be chosen
Wols et al. (2010b) found that RANS wrongly predict local flow features around a UV
7
lamp which is a typical unsteady flow structure in a flow around a blunt body. LES was
employed and matched the experimentally measured velocity profile better than RANS.
Zhang et al. (2014b) investigated a baffled contactor and a column contactor which are
typically used for ozone and chlorine disinfection and showed that LES is a more reliable
strategy than RANS in simulating tracer transport in column contactors due to its ability
to better predict the spatial transition to turbulence characterizing the flow. Therefore, in
further study, higher resolution approaches such as LES, detached LES, or even DNS
should be involved if computational cost could be affordable. The multi-phase flow effect
acid disinfection, as opposed to the ozone disinfection process. Most previous ozone
studies Cockx et al., 1999; Ta and Hague, 2004; Bolaños et al., 2008; Talvy et al., 2011).
neglected this effect because of two reasons: 1) unknown parameters, such as bubble size
distribution, mass transfer coefficients, models for closure of the two-phases, and so on;
cost. However, this effect could not be neglect especially when studying in a column
contactor. The gas flows from bottom of the column upwards to the top because of
pressure gradient, while water goes against the direction that the ozone flows in. Thus, it
can impact the overall flow significantly. In future study, more simulations and
experiments of this effect need to be done so that we can understand the role of a gas
flow in disinfection process much better. The complexity of reaction system is caused by
the complex reaction mechanisms in the disinfection process. For example, once ozone
pumped into wastewater flow, many reactions are triggered. There is not only the ozone
8
decay reaction, a reaction between dissolved ozone and natural organic matters (NOM),
For flow simulation, the biggest problem is how to describe a turbulent flow. Typical
Reynolds number in turbulent flow region for an internal flow is above 2,000, which is
most common scenario in the wastewater treatment process. Turbulent flow is very
complicated: it has a wide range of spatial and temporal scales from large spatial scales
that can be comparable to the size of flow, to the small scales even less than millimeters.
A similar situation arises for the temporal scales. Because of the random fluctuation in
modeling will help to reduce the difficulty of predicting the turbulent region. There are
DNS has the best accuracy of the three, because it solves Navier-Stokes equation directly
on the all spatial and temporal scales. However, DNS has a higher demand on fine grids
model, i.e. the grid for DNS should contain about Re9/4 points (Pope, 2000). For example,
if the Reynolds number for the mean flow is around 20,000, then it requires a meshing
with 4×109 cells, which make a great rise on computational time would take thousands of
hours computing on a PC. So far, we still cannot afford the computational cost today.
LES is an alternative to DNS, which has a good accuracy and an ability to predict the
spatial transition to turbulence characterizing the flow. RANS have a good computational
In RANS methods, there are two major families, they are: k-ε and k-ω Family. k-ε
family has different models such as: Standard k-ε Model, RNG K-ε Model, Realizable k-
2017 (ESI-Group, 2017), the k-ω Family contains such as k-ω SST, Wilcox etc. The
main idea is to describe the flow field with two introduced variables: one is for the
turbulent kinetic energy, k, and the other is either for the rate of dissipation, ε., or for the
There are two transported variables in k-ϵ model. The first variable determines the energy
in the turbulence and is called turbulent kinetic energy (k). The second variable is the
turbulent dissipation (ε) which describes how much the turbulent kinetic energy will
dissipate to the surrounding fluid. The idea of k-ϵ model is Reynolds decomposition,
which mathematically separates the time-averaged and fluctuating parts of the flow. The
ui = ui + ui ' (0.1)
1 2 2 2
k= (u + v + w ) (0.2)
2
∂( pk ) ∂( pu j k ) ∂ µt ∂k
+ = ρ P − ρε + (µ + ) (0.3)
∂t ∂x j ∂x j σ x ∂x j
and
10
∂( pε ) ∂( pu jε ) ρ Pε ρε 2 ∂ µt ∂ε
+ = Cε1 − Cε 2 + ( µ + )ε (0.4)
∂t ∂x j k k ∂x j σ ∂x j
∂u ∂u 2 ∂um 2 ∂um
P =ν t i + j − δ − k
∂x ∂x 3 ∂x ij 3 ∂x (0.5)
j i m m
C µ = 0.09,
C ε1 = 1.44,
Cε 2 = 1.92,
σ x = 1.0,
σ s = 1.3
small scales from the solution to the Navier-Stokes equations. Because the principal
difficulty in simulating turbulent flows comes from the wide range of length and time
scales, this operation makes turbulent flow simulation cheaper by reducing the range of
scales that must be resolved. The LES filter operation is low-pass, meaning it filters out
the scales associated with high frequencies. A LES filter can be applied to a spatial and
temporal field and the filter size is decided by maximum grid size and time step. Any
spatial fluctuation smaller than grid size will be filtered out, and similarly, any temporal
11
frequency higher than time step frequency will be filtered out as well. Therefore, filtered
∂ ui
=0 (0.6)
∂ xi
∂ ui ∂ 1 ∂p
+ (ui u j ) = − + ν∇ 2 ui + f i (0.7)
∂ t ∂x j ρ ∂ xi
Furthermore, the filtered advection term can be split up, following Leonard (Leonard,
1974),
u i u j = τ ijr + u i u j (0.8)
where τ ijr is the residual stress tensor. The filtered momentum equations become
∂ ui ∂ 1 ∂p ∂ ∂ τ ijr
+ (u i u j ) = − + 2ν S ij − (0.9)
∂t ∂x j ρ ∂ xi ∂x j ∂x j
Applying the Smagorinsky model (Smagorinsky, 1963) for sub-grid scale turbulence,
1
τ ijr − τ iiδ ij = 2ν t S ij (0.10)
3
ν t ~ ∆2 2S ij S ij (0.11)
ν t = cs2∆2 2S ij S ij (0.12)
where cs is called Smagorinsky constant, cs = 0.1 has been chosen in the present
simulations.
As shown clearly above, RANS is solving the averaged Navier-Stokes equation which
gives a much better efficiency but at the same time, it fails to predict the fine structure of
the flow. LES on the contrary, is able to capture the large scale structure in the flow
which results in a much better flow realization. Wols et al. (2010) showed that the
accuracy of RANS is weakened when dealing with unsteady flow structure. In this
research, RANS wrongly predicted local flow features around a UV lamp which was a
typical unsteady flow structure in a flow around a blunt body. In this case, LES was
employed and matched the experimentally measured velocity profile better than RANS.
In order to overcome unsteady flow structure effect, the method has to be chosen
appropriately for each particular problem. Zhang et al. (2014b) investigated a baffled
contactor and a column contactor which are typically used for ozone and chlorine
disinfection. Results showed that LES is a more reliable strategy than RANS in
simulating tracer transport in column contactors due to its ability to better predict the
spatial transition to turbulence characterizing the flow. Therefore, in further study, higher
resolution approaches such as LES, detached LES, or even DNS should be involved if
Peracetic acid (PAA) is the peroxide of acetic acid (AA). The product is
hydrogen peroxide (HP, H2O2), PAA (CH3CO3H) and water, as shown in the following
where CH3CO2H = acetic acid, CH3CO3H = peracetic acid; H2O2 = hydrogen peroxide.
Industrially, PAA is mass-produced from the reaction of acetic acid with hydrogen
1991). The PAA solution is always a mixture of PAA, AA and HP. Although hydrogen
French, 1989b; Fraser et al., 1984). It was found that HP required much larger doses than
It has been shown that PAA produces no known toxic disinfection by-products (DBPs) in
water after disinfection (Baldry and Fraser, 1988; Monarca et al., 2001; Monarca et al.,
2002). This is one of the most important advantages of PAA over other typically used
sodium hypochlorite, chlorine dioxide, or ozone (Monarca et al., 2002). Once the
series of different reactions. There are three major reactions in which PAA is consumed
reaction with microorganisms (Antonelli et al., 2013; Gehr et al., 2002; Yuan et al.,
1997).
Liberti and Notarnicola (1999) identified the PAA disinfectant activity as consequence of
the release of active oxygen, which likely oxidizes sensitive sulfhydryl and sulfur bonds
Pelczar and Chan (1986) the whole inactivation process is composed of four different
mechanisms:
Gehr et al. (2002) concluded that for municipal primary effluents, PAA doses of 2 to 6
mg/L were needed to achieve 1000 colony forming units (CFU) per 100 ml of fecal
coliforms with a contact time of 60 min. For the secondary effluents, lower PAA doses of
0.6 to 4 mg/L were required to achieve 1000 CFU per 100 ml of fecal coliform. Lefevre
et al. (1992) recommended a dose in the range between 5 mg/L to 7 mg/L and 60 minutes
contact time in order to achieve effluent total coliform concentrations of less than 1000
CFU/100 mL and effluent fecal coliform concentrations of less than 100 CFU/100 mL.
15
PAA was also suggested to be applied with ultraviolet (UV) irradiation for wastewater
wastewater was in contact with PAA for a relatively longer time first, and then being
achieved.
Several factors that affecting PAA disinfection processes were identified. PAA can be
temperature during disinfection with PAA (Stampi et al., 2001). Another feature to be
considered is pH. With a lower pH value, PAA showed with greater activity (Colgan and
Gehr, 2001). An ideal environment for PAA against fecal coliforms was under neutral or
Major disadvantages associated with PAA disinfection are the increases of organic
content in the effluent due to acetic acid and thus in the potential for microbial regrowth
(Lazarova et al. 1998). Another drawback to the use of PAA is its high cost, which is
partly due to limited production capacity worldwide. However, if the demand for PAA
increases, especially from the wastewater industry, mass production capacity might also
be increased in the future, thus lowering the cost. In such a case, in addition to having
environmental advantages, PAA may also become cost-competitive with chlorine. Thus,
simulation that serves as a surrogate for the more computationally intensive simulations.
Metamodeling allows for wider exploration of the input variables, improves the
understanding of the model to be generated and enables further studies for solution
optimization (Tabach, et al., 2007, Broad, et al., 2005). Areas that metamodels play a role
2. Design space exploration: The design space is explored to enhance the engineers’
metamodel;
Meta modeling evolved from classical design of experiments theory (Wang and Shan,
composite design, Box-Behnken design and Plackett-Burman design (Box, et al., 1978,
Montgomery, 2005). These methods tend to spread the sample points around boundaries
of the design space and leave a few at the center of the design space. For computer
experiments space filling design are recommended (Wang and Shan, 2007, Sacks, et al.,
17
1989). These are the orthogonal arrays, various Latin hypercube designs, Hammersley
sequences and uniform designs (Wang and Shan, 2007). Different types of metamodels
can be created such as regression models, artificial neural networks, radial basis
functions, and Kriging stochastic model (Wang and Shan, 2007). Hammersley sequences
and uniform designs belong to a more general group called low discrepancy sequences.
Hammersley sampling has been found to provide better uniformity than Latin
uniform in one dimensional projection while the other methods tend to be more uniform
in the entire space. The number of sampling points or the appropriate sampling size is
Metamodels have been used in a wide variety of applications, including calibrating water
Levandler and Payraudeau (2007) developed a metamodel for optimizing the design of
storm water detention basins. Galleli and Soncini – Sessa (2010) applied the
metamodeling design approach for the design of release policies in reservoir systems
serving irrigation districts by modeling the dynamics of the irrigation demand through
metamodels. The metamodeling approach was also applied to a water quality and
hydraulics – based optimization problem for water distribution systems with the objective
al., 2005). The developed metamodel was approximately 700 times faster than the
simulation model and similar optimal solutions were obtained (Broad, et al., 2005).
18
ANNs have increasingly been applied instead of the conventional regression models
complex, nonlinear functions than other forms (such as polynomials). Tabach et al. (2007)
Figure below summarizes the techniques for metamodeling (Wang and Shan, 2007).
identify an efficient set of computer runs or numerical experiments and then apply
(Wang and Shan, 2007, Simpson, et al., 2001). Furthermore, a better solution of producing
a metamodel of CFD results with ANNs has emerged with the Hoque’s study (Hoque,
2011). In the study, it was shown that the ANN based metamodeling performed better than
MODELING
3.1 Introduction
The numerical simulations were conducted using the commercial code CFD-ACE+ (ESI
Group, 2017). It was necessary for the current situation to be able to track particles in a
transient flow. CFD-ACE+ had the required features. Computational fluid dynamics
(CFD) is a mathematical modeling procedure whereby the fluid parameters are calculated
domains in space (grid size) and/or time (size of time step). These algebraic equations
The CFD model applied to simulate the wastewater disinfection process with PAA
comprised of two main sections: transport equations and chemical reactions. The sections
below:
Continuity Equation
∂ui
=0 (1.1)
∂xi
20
Momentum Equation
∂ui ∂ 1 ∂p ∂u ∂u
+ (ui u j ) = − +ν i + j + Fi (1.2)
∂t ∂x j ρ ∂xi ∂x ∂x
j i
Species Equation
∂ci ∂c ∂ ∂ci
+ uj i − Di = Si (1.3)
∂t ∂x j ∂x j ∂x j
To simplify the chemical kinetics, two reactions are mainly considered in the disinfection
process. They are: the reaction of PAA decay and the inactivation of microorganisms (N),
A recent batch study(Tizzoni, 2016) shows that PAA decays via immediate demand plus
first order decay in wastewater, and the inactivation of microorganisms with PAA follows
Hom’s power law (Hom, 1972). Therefore, the chemical kinetics may be written as,
d [PAA]
= −kd [PAA] (1.6)
dt
d[ N ]
= −mk *[PAA]n t m−1[ N ] (1.7)
dt
21
where kd is the decay rate constant of PAA, k* is the Hom’s law rate constant, m and
Since an implicit form of equation (3.6) is needed to apply in CFD simulation, the Hom’s
power law can be transformed to an equivalent form only in state variables(Haas and
Joffe, 1994):
d[ N ]
= − mk
1
* m
n
[PAA] m [ − ln(
(
[ N ] 1− 1 m
)]
)
(1.8)
dt [ N ]0
Two turbulent modeling techniques are considered in the research. For full scale
wastewater treatment simulations, a k-ϵ model is applied. On the other hand, in order to
realize the flow field with higher fidelity, LES is chosen in studies of T-junction mixing.
k-ϵ model
The flow velocity ui is to be decomposed as:
ui = ui + ui ' (1.9)
'
where ui is the averaged laminar velocity, ui is turbulent fluctuation.
1 2 2 2
k= (u + v + w ) (1.10)
2
∂( pk ) ∂( pu j k ) ∂ µt ∂k
+ = ρ P − ρε + (µ + ) (1.11)
∂t ∂x j ∂x j σ x ∂x j
22
and
∂( pε ) ∂( pu jε ) ρ Pε ρε 2 ∂ µt ∂ε
+ = Cε1 − Cε 2 + ( µ + )ε (1.12)
∂t ∂x j k k ∂x j σ ∂x j
∂u ∂u 2 ∂um 2 ∂um
P =ν t i + j − δ − k
∂x ∂x 3 ∂x ij 3 ∂x (1.13)
j i m m
The five constants used in this model are as follows (ESI-Group, 2017):
C µ = 0.09,
C ε1 = 1.44,
Cε 2 = 1.92,
σ x = 1.0,
σ s = 1.3
∂ ui
=0 (1.14)
∂ xi
∂ ui ∂ 1 ∂p
+ (ui u j ) = − + ν∇ 2 ui + f i (1.15)
∂ t ∂x j ρ ∂ xi
23
Furthermore, the filtered advection term can be split up, following Leonard (Leonard,
1974),
u i u j = τ ijr + u i u j (1.16)
where τ ijr is the residual stress tensor. The filtered momentum equations become
∂ ui ∂ 1 ∂p ∂ ∂ τ ijr
+ (u i u j ) = − + 2ν S ij − (1.17)
∂t ∂x j ρ ∂ xi ∂x j ∂x j
Applying the Smagorinsky model (Smagorinsky, 1963) for sub-grid scale turbulence,
1
τ ijr − τ iiδ ij = 2ν t S ij (1.18)
3
ν t ~ ∆2 2S ij S ij (1.19)
ν t = cs2∆2 2S ij S ij (1.20)
where cs is called Smagorinsky constant, cs = 0.1 has been chosen in the present
simulations.
24
Mesh generation was carried out using CFD-Geom (ESI-Group, 2017), which is capable
to generate both structured and unstructured mesh. For a structured messing, codes use
3D domain as in Figure 2. Because the simplicity of the model geometries, most of the
grid generations are structured messing. The models in the study of mixing enhancement
with different inlet configurations (in Chapter 6) are combined with both techniques.
means that the numerical results of the simulation would not change with
increasing/decreasing the grid size which minimizing the error due to the numerical error.
For example, a model with coarsened mesh could have a larger truncation error.
Sometimes the error is too large to neglect. Therefore, to perform a grid independence
study is necessary.
In the T-junction mixing study, four CFD model were generated with 300,000; 600,000;
900,000 and 1,200,000 structured cells respectively. Figure 2 below showed the grid
(a)
(b)
Following the experimental work of Forney et al. (Forney and Lee, 1982, Forney,
Noureddine and Hanh, 1996), simulations of mixing of two gases (air and methane) in a
T-junction were carried out for grid independence study by carefully matching the
sectional maximum concentrations were collected for comparison. The results are shown
in Figure 3.
It is clear that the model with 300,000 performed most poorly among four simulations. It
over-predicted the mixing thoroughness. However, when increasing the grid size from
900,000 to 1,200,000, there was very little difference between the two simulations.
Therefore, the model with 900,000 structured cells was used in the T-junction
simulations.
necessary step. One way to validate the CFD simulations is to compare the CFD
predictions with experimental results. In this study, the T-junction CFD model is test with
experimental measurements (Forney and Lee, 1982; Pan and Meng, 2001).
27
Forney et al. discussed the optimum mixing conditions of two gaseous flow in a T-
junction. In the experiments (Forney and Lee, 1982, Forney, Noureddine and Hanh,
1996), the two flow streams are air and methane-air mixture. The main flow Reynolds
number Rem = 21,000, and the jet flow Reynolds number Rej varies from 4,000 to 8,000.
Figure 4 below shows the normalized concentration profile of methane (c/cm) along the
center vertical line at the axial location where z/D = 3 and R (jet-to-pipe inlet velocity
ratio) = 11.0.
Figure 4 Comparison of methane concentration from the simulations and the experiments
at R = 11.0, d/D = 0.014, z/D = 3.0, Rem = 21,000, Rej = 4000, Sct = 0.9
28
Pan and Meng (Pan and Meng, 2001) conducted two sets of experimental measurements
on T-junction mixing in liquid-liquid flows. The main flow is water and the jet flow is a
sodium fluorescein solution in water. The Reynolds number of the main flow is fixed at
20,850 in both cases; Reynolds numbers of the jet flow are 10,630 and 17,500 (Pan and
Meng, 2001). Simulations were carried out for the case with main flow Reynold number
The Figures below are the comparison of grayscale images of the side-view instantaneous
concentration field for R = 3.06, d/D = 0.1667, Rem = 20,850, Rej = 10,630. Screenshot of
Sec
As shown above, there is noticeable difference between the two plots. CFD is not able
to predict the fine flow structure. Because of the nature of LES, small structures are filtered
out. Furthermore, a roll-up vortex on the left of the jet in the experiment plot (marked by
red frame) did not show up in CFD simulations. These differences might be caused by
imperfect geometry lineup, or the residual swirl in either the jet or the main flow, etc. The
numerical errors contribute to the difference as well, for example, LES filtered out small
structures, which is one of the reason that simulation predicts less diffusion. Besides, the
difficulty on matching experimental time and simulation time could be another factor as
well. Though the simulation conditions were carefully matched with the experimental
conditions (Pan and Meng, 2001), the predicted concentration field lacks some of the
details found in the experimental work. The LES filters out the sub-grid level fluctuations.
The alternate here is to carry out the direct numerical simulations (DNS) which can become
prohibitively expensive in time and resources. In addition, the simulations are for ideal
30
in most experiments. These could contribute to the mismatch of the finer details of the
To test out the assumptions above, a series of 2-D cases with imperfect geometries were
carried out. In the plots below, two cases were considered first with inclined jet inlet. This
could be a common scenario in laboratory. Bartend et al., found a perfect CFD model of a
vertical reactor was not providing a good prediction. In order to get a better correlations
with his experiment, 10 inclination on the vertical tank was made (Bartend, 2009).
In this case, we first looked at the two cases with 30 inclination forward and backward
(a)
31
(b)
Figure 6 T-junction simulation with inclined jet inlet (a) 3-degree forward; (b) 3-degree
backward
With only 30 difference, there is some significant difference between the two results.
Especially in the case with 3-degree forward inclination, the roll-up vortex showed up.
Furthermore, when installed the T-junction in the laboratory, the intersection of the two
pipes might not be smooth ideally. Therefore, to add a small roughness by blocking the
The simulation showed that the development of the roll-up vortex and providing more
flow structure. The observation from Figure 6 to 7 supported the previous assumption.
32
4.1 Introduction
Mixing of fluid streams plays a key role in many areas in the industry and in nature. An
optimum condition of mixing is desired for enhancement of species mixing, heat and mass
transfer, combustion and other chemical reactions. For example, good mixing can promote
junction geometry is widely used in industry and research for mixing of two dissimilar
Gas mixing problems are common in engineering. Gas mixture flows appear in diverse
mixing and reaction. The first systematic study of gas mixing with a T-junction was
conducted by Chilton and Genereaux (Chilton and Genereaux, 1930) who used smoke
right angle configurations were as effective as any other geometry for good mixing. Chilton
and Genereaux also found that when the ratio of the velocity of the secondary flow to the
velocity of the main flow was 2-3 for their geometries, satisfactory mixing was obtained in
2-3 pipe diameters. Narayan (Narayan, 1971) used quantitative methods to measure the
degree of mixing of air-carbon dioxide feed streams in three pipeline mixers. Narayan
(Narayan, 1971), like Chilton and Genereaux (Chilton and Genereaux, 1930), found it was
33
possible to achieve quality mixing in a few diameters with perpendicular jet injection
devices but that parallel flow devices required up to 250 pipe diameters.
Following the work of Forney et al. (Forney and Lee, 1982, Forney, Noureddine and
Hanh, 1996) we consider the mixing of two gases (air and methane) in a T-junction.
one of the most basic devices in microfluidics. Due to the small size of the device,
pressure driven flows in simple channels (i.e. with smooth walls) are laminar and mostly
uniaxial, so that the confluent liquids tend to flow side by side. As the Schmidt number
(the ratio between kinematic viscosity ν and molecular diffusivity between two species
D1-2, ν/D1-2) is typically large for two miscible liquids – the mixing between two parallel
laminar liquid streams is weak. The subscripts 1 and 2 in D1-2 stand for the main flow and
the jet flow components. T-shaped micromixers have been investigated extensively in
recent years, as they are quite suitable to carry out fundamental studies to understand
Liquid-liquid mixing problems are also common in industry and in nature. Recently liquid-
liquid mixing problems with T-junctions have been intensely studied for problems in
microchannels (Raj, Mathur and Buwa, 2010). CFD techniques have been widely applied
in simulating wastewater disinfection process in the past decade since they can provide the
(Zhang, Tejada-Martínez and Zhang, 2014). T-junctions are also used for the mixing of
of the disinfectant and the wastewater is vital in process performance. Treatment techniques
vary in industrial, municipal and agricultural wastewater treatment process, due to the
34
technologies being applied. Each technique has its own particular usage. In all cases of
volume of fluid. This current work is motivated by the use of peracetic acid (PAA) as a
disinfectant.
Computational fluid dynamic (CFD) simulations are carried out in the present study to
predict the flow properties and mixing characteristics of two fluids (gas-gas or liquid-
liquid) at a 90o T- junction of two pipes of dissimilar diameters. Following the work of
Forney et al. (Forney and Lee, 1982, Forney, Noureddine and Hanh, 1996), we first
consider the mixing of two gases (air and methane) in a T-junction. For the turbulent flow
field predictions the LES technique is used to predict the spatial-temporal evolution of the
mixing process. A schematic of the mixing problem considered in the T-junction is shown
below (Figure 8). The two flows are labeled as the main flow (in the larger pipe with
diameter D) and the jet flow (in the smaller pipe with diameter d).
35
The main pipe with D = 11.43 cm is placed horizontally along the z-axis. The jet pipe
with smaller diameter is on the top and placed vertically along the x-axis. The diameter of
the inlet jet can be varied in order to control the jet-to-pipe diameter ratio i.e. d/D. The
distance from injecting point to the outlet is 15 times of main pipe diameter to ensure the
In general, the larger pipe contains the main solvent and a solute is injected into the main
flow via the T-junction. Forney et al. (Forney and Lee, 1982) investigated gas-phase mixing
in a T-junction. In a succeeding paper (Forney, Noureddine and Hanh, 1996), they showed
that an optimized mixing condition can be determined by varying the geometry of the T-
junction, i.e. the jet-to-pipe diameter ratio d/D and inlet velocity ratio R = U j / U m ). Based
on the jet and the main flows, we define two Reynolds numbers viz. Rej ( U j d / ν and Rem
( U m D / ν , where U j
and U m are the mean inlet velocity components for the jet and main
flows respectively and ν is the kinematic viscosity of the mixing components (assumed to
be the same). The laminar Schmidt number ν /D1-2 is around 0.7 for gas-gas mixing (say
36
air/methane mixing) problems; however, the corresponding laminar Schmidt number is two
to three orders higher in liquid-liquid mixing problems indicating a low degree of mixing
between two liquids in laminar flows (Raj, Mathur and Buwa, 2010). In general, the
laminar mixing characteristics of two gases beyond the T-junction is stronger than that of
two liquids. For turbulent flows however, the Schmidt number is defined as ν t/Dt,1-2 where
ν t is the so-called eddy diffusivity and Dt,1-2 is the (sub-grid scale) turbulent mass
In this study, for gas-gas mixing in a T-junction, we consider the mixing of air and methane
(0.3% methane) following the work of Forney and co-workers (Forney and Lee, 1982). A
series of three-dimensional CFD/LES simulations were carried out to simulate the air-
methane experiments. To simulate the liquid-liquid mixing, we consider two cases: (a)
mixing of water with a PAA-water solution and (b) mixing of water with a sodium
fluorescein solution in water (Pan and Meng, 2001). Wastewater contains microorganisms
(e.g., coliforms and E.coli) and the inactivation of microorganisms can be achieved by PAA
(Wei, Farouk and Haas, 2016). The case of mixing water with a sodium fluorescein solution
in water (Pan and Meng, 2001) is of particular interest as the experimental study provides
CFD simulations of mixing of two gases is considered first. In the gas-gas mixing case,
air flows in the main pipe, and 0.3% methane-air mixture is injected from the jet flow
inlet (see Table 1). The jet-to-pipe diameter ratio, d/D is chosen from following values:
0.014, 0.025, 0.1, and 0.2. The mixing of air and methane is studied under both laminar
and turbulent flow conditions. In laminar flows, we consider Rem ~ 200 and Rej ~ 40. In
mixing
With the same geometry, two additional series of simulations were carried out on the
mixing of two miscible fluids, i.e. (a) water and peracetic acid solution (PAA) and (b)
water and sodium fluorescein solution in water. In the liquid-liquid study, water flows in
the main pipe and PAA solution (or sodium fluorescein solution in water) is injected from
the jet flow inlet. The mixing of the two fluids (gas-gas or liquid-liquid) are studied in
laminar and turbulent flow conditions. The Reynolds number ranges (for the laminar and
turbulent flows) in liquid-liquid case are kept in the same range as in the gas-gas mixing
simulations. For the water- sodium fluorescein solution in water (Pan and Meng, 2001),
the case with Rem = 20,850, and Rej = 10,630 was considered for comparison with the
present predictions. All parameters for the mixing problems considered are specified in
38
the Table 1. In the case of PAA, reactions were neglected due to the relatively short
Laminar flows
The governing equations of mass, momentum and concentration for laminar mixing
problem in a T-junction are described below. It is assumed that the kinematic viscosities of
Conservation of mass:
∂ui
=0 (2.1)
∂xi
Conservation of momentum
∂ui ∂ 1 ∂p
+ (ui u j ) = − +ν∇2ui (2.2)
∂t ∂x j ρ ∂xi
Concentration conservation
∂ci ∂c ∂ 2 ci
+ u j i = D1− 2 (2.3)
∂t ∂x j ∂x j ∂x j
where the index ‘i’ can be either 1 or 2 for the normalized mass concentrations (c1 or
c1 + c2 = 1.0 (2.4)
39
Turbulent flows
For simulating turbulent flows in the T-junction, we need to invoke a turbulence model.
There are three most popular turbulence models: the Reynolds-averaged Navier–Stokes
(RANS) equations, the Large-eddy Simulation (LES) and the Direct Numerical
requirements, we have used the LES model to simulate the turbulent mixing problems in
the T-junction.
Compared to DNS, LES solves a “filtered” set of Navier-Stokes equation instead (Yang,
scales from the solution to the Navier-Stokes equations. Because the principal difficulty
in simulating turbulent flows comes from the wide range of length and time scales, this
operation makes turbulent flow simulation cheaper by reducing the range of scales that
must be resolved. The LES filter operation is low-pass, meaning it filters out the scales
associated with high frequencies. A LES filter can be applied to a spatial and temporal
field and the filter size is decided by maximum grid size and time step. Any spatial
fluctuation smaller than grid size will be filtered out, and similarly, any temporal
frequency higher than time step frequency will be filtered out as well. Therefore, filtered
∂ ui
=0 (2.5)
∂ xi
40
∂ ui ∂ 1 ∂p
+ (ui u j ) = − + ν∇ 2 ui + f i (2.6)
∂ t ∂x j ρ ∂ xi
Furthermore, the filtered advection term can be split up, following Leonard(Leonard,
1974),
where τ ijr is the residual stress tensor. The filtered momentum equations become
∂ ui ∂ 1 ∂p ∂ ∂ τ ijr
+ (u i u j ) = − + 2ν S ij − (2.8)
∂t ∂x j ρ ∂ xi ∂x j ∂x j
Applying the Smagorinsky model (Smagorinsky, 1963) for sub-grid scale turbulence,
1
τ ijr − τ iiδ ij = 2ν t S ij (2.9)
3
ν t ~ ∆2 2S ij S ij (2.10)
41
ν t = cs2∆2 2S ij S ij (2.11)
where cs is called Smagorinsky constant, cs = 0.1 has been chosen in the present
simulations.
∂ ci ∂ ∂q j
+ (u j ci ) = D1− 2 ∇ 2 c i − (2.12)
∂t ∂x j ∂x j
where D1-2 is diffusion coefficient, and qj is the sub-grid scale term, representing a
mass flux term which is critical to study mixing characteristics (Calmet and Magnaudet,
1997) . Applying the Smagorinsky model as the sub-grid scale model to the filtered
conservation equation,
νt ∂c
qj = (2.13)
ScSGS ∂x j
Schmidt number Sct). A value of Sct = 0.9 is used in the present simulations.
The time-dependent simulations were carried out using the CFD-ACE+ software (CFD-
ACE, 2016) with a time step equal to 10-3 sec was used for the gas mixing cases and a
time step equal to 10-2 sec was set for the liquid mixing cases. A second order blended
upwind scheme was applied for the convective-diffusive terms and Crank-Nicolson
scheme with a blending factor of 0.7 was used for the temporal terms in the governing
42
equations. The internal relaxations for velocities and passive scalar were 0.2. The
convergence criterion for each variable is the same (set to 0.001 for the present
The grid generation (see Figure 9 below for a sample grid structure used in the study) was
2016). To balance the computation cost and the resolution of an eddy structure, 900,000
(a)
43
(b)
Figure 9 Grid Generation of the T-junction with d/D = 0.1
(a) cross-sectional view passing through the jet flow pipe; (b) perspective view.
At inlets of the domain, normal velocities in the profile of 1/7th power law were specified
for turbulent simulations. For laminar mixing on the other hand, a flathead velocity
profile was assigned. No-slip conditions were assigned for the walls. Initially, the main
pipe and was filled up with air (gas-gas flows) or water (liquid-liquid flows), while the jet
solution in water (Pan and Meng, 2001). At t ≥ 0.0 s, jet tracers, air-methane mixture or
PAA solution or sodium fluorescein solution in water was injected from jet inlet and air
or water started flowing from main inlet as well. The simulations were carried out for a
period of three times the theoretical residence time to diminish the effect of initial
For laminar cases, the simulations were obtained for the geometry shown in Figure 8.
Mixing of air-methane (0.3% methane by volume in the jet flow) and water-PAA solution
(0.3% by volume of the jet flow) were studied with low Reynolds numbers, with Rem =
200, and Rej = 40 with jet-to-pipe inlet velocity ratio R = 1.0, and d/D = 0.1 and 0.2. Plots
of normalized maximum concentration of the solute cm - methane (for the gas-gas flows)
and PAA solution (for the water-PAA solution) cases are shown in Figure 10 below.
solution cases along the axial length for d/D = 0.1 and 0.2
45
From Figure 10, it is evident that the mixing is faster for the gas-gas case (Sc = 0.7)
compared to the liquid-liquid case (Sc = 700.0) in laminar flows. The present simulations
also provide the nature of the 3-D mixing process in the flow geometry.
For the case described in Figure 8, the spatial distribution of the mixing components for
the air-methane the water-PAA solutions are shown in Figures 11 (a) and (b) respectively.
While the mixing process is quite vigorous for the air-methane case as evidenced in Figure
11 (a), the mixing of PAA solution in water is significantly less intense as shown in Figure
11 (b).
(a)
46
(b)
Figure 11 Contour plot of normalized (a) methane, Sc = 0.75; (b) PAA, Sc = 700 concentration at
For the laminar liquid-liquid mixing, the PAA solution contours at downstream planes
tend to form a horse-shoe type vortex structure that retains its identity further
(Karthick Selvam, 2017) had observed similar vortex structure at the near field of the T-
junction.
To evaluate the difference in the behavior of the two mixing cases, i.e. air-methane and
std (ci )
φ= (14)
mean (ci )
47
where std (ci) is the standard deviation of the tracer concentration on a given cross-section
plane; mean (ci) is the averaged tracer concentration on the same plane. A higher value of
φ would thus indicate a lower level of mixing. Figure 12 below shows the variation of φ
as a function of the normalized axial length (z/D ranging from 0.0 to 15.0) for the gas-gas
and liquid mixing cases for d/D values of 0.1 and 0.2 respectively. The degree of mixing is
good for the gas-gas (Sc = 0.75) mixing cases for d/D = 0.1 and d/D = 0.2 cases. On the
other hand, the water-PAA mixing is poor under laminar flow conditions for d/D = 0.1 and
d/D = 0.2 cases. From comparisons above, it is clear that Sc value is affecting the mixing
of two streams. With about three orders of magnitude difference in Sc value, the flow with
d/D=0.2
0.2
d/D=0.1
0.1
Figure 12 Plot of degree of laminar mixing along the downstream length for the air-methane and
the water-PAA solution cases, d/D = 0.1 and 0.2, Rem = 200, Rej = 40
Forney et al. discussed the optimum mixing conditions of two gaseous flow in a T-
junction. In the experiments (Forney and Lee, 1982, Forney, Noureddine and Hanh,
1996), the two flow streams are air and methane-air mixture. The main flow Rem =
21,000, and the jet flow Rej varies from 4,000 to 8,000. Figure 13 below shows the
normalized concentration profile of methane (c/cm) along the center vertical line at the
49
axial location where z/D = 3 and R (jet-to-pipe inlet velocity ratio) = 11.0. The results
Figure 13 Comparison of methane concentration from the simulations and the experiments
(Forney and Lee, 1982, Forney, Noureddine and Hanh, 1996) at R = 11.0, d/D = 0.014, z/D = 3.0,
and R = 20, d/D = 0.014, z/D = 3.0, Rem = 21,000, Rej = 7700, Sct = 0.9
In Figure 13, the predictions agree well with the measured values reported in (Forney and
Lee, 1982, Forney, Noureddine and Hanh, 1996). However, slight deviations are
noticeable. This may be caused by a possible systematic error on locating sampling points
50
during the experiment; as well as numerical errors viz. mesh not fine enough, under-
Figure 14 below shows the normalized concentration profile of methane along the center
horizontal line where R and d/D are 3.73 and 0.1 respectively. The results were time-
averaged with last 0.02s in the simulation, which were the average of 20 time steps.
with R = 3.73, d/D = 0.1, Rem = 21,000, Rej = 6000, Sct = 0.9
Both the plots in Figure 14 show good agreement of CFD predictions with the
experimental data (Forney and Lee, 1982, Forney, Noureddine and Hanh, 1996), which
indicate the fidelity of the current CFD grid generation together with the applied
51
A series of contour plots (at different times with a 0.01 s interval) displaying the mixing
behavior of air-methane mixing (with R = 3.73 and d/D = 0.1) as predicted by the LES
are shown in Figures 15 (a) – (e). The boxed zones in the figures show the progression
(a)
(b)
(c)
52
(d)
(e)
Figure 15 Contour plots of methane concentration at center vertical plane at different times d/D =
(a) t =3.005s (b) t = 3.015s (c) t = 3.025s (d) t = 3.035s and (e) t = 3.045s;
From the simulations, it appears that the jet is unable to fully penetrate the main pipe
diameter, however, the large eddy structures of the methane jet can be identified
downstream. The boxed zones in Figures 15 (a) – (e) shows the progression and decay of
a mixing structure as it flows downstream. It is also interesting to note that the upper part
of the main pipe has little or no methane concentration in the downstream region.
53
Figure 16 shows the time-averaged contour plot of methane concentration from 3.600 sec
to 3.620 sec at the exit plane of the geometry given by Forney et al. (Forney and Lee,
1982, Forney, Noureddine and Hanh, 1996) . It is interesting to note that the degree of
mixing of methane at the exit plane is rather non-uniform and shows wide variations
(perhaps caused by the presence of eddy structures). Though the methane is injected from
the top region of the main pipe, the resulting flow field produces a methane rich region at
the bottom region of the main pipe at the exit which is at 15-pipe-diameter downstream
from the jet pipe. Recall Fig 1, the exit plane is marked in red on the schematic.
Figure 16 Contour plot of methane concentration at the exit plane, time average from 3.600 sec to
3.620 sec; R = 2.45 and d/D = 0.2, Rem = 21,000, Rej = 6000, Sct = 0.9
For turbulent mixing of two liquid streams, the simulations were carried out for the
geometry shown in Figure 8. Turbulent mixing of water-PAA solution were studied with
54
Rem = 21,000, and Rej = 6,000 with jet-to-pipe inlet velocity ratio R = 2.0, and d/D = 0.2.
For the turbulent mixing of water-sodium fluorescein solution in water, we carried out the
simulations following the parameter values as specified in the paper by Pan and Wang
(Pan and Meng, 2001), in which cases Rem = 20,850, and Rej = 10,630 respectively.
Figure 17 below shows the cm (maximum concentration of the solute) for liquid-liquid
laminar cases described earlier. For the results shown in Figure 12, R = 2.45, and d/D = 0.2
with Rem = 21000, Rej = 6000 Sct = 0.9, and R = 3.73, d/D = 0.1 with Rem = 21000, Rej =
4000 Sct = 0.9. The corresponding results for turbulent mixing of the air-methane case is
also shown in the Figure 17 by dash lines. Since the effect of the laminar diffusivity plays
a minor role in the turbulent mixing problems, the mixing characteristics for turbulent
Figure 17 Comparison of liquid-liquid mixing and gas-gas mixing at R = 2.45, d/D = 0.2, Rem =
21000, Rej = 6000 Sct = 0.9 and R = 3.73, d/D = 0.1, Rem = 21000, Rej = 4000 Sct = 0.9
Recalling the laminar flow mixing results (for air-methane and water-PAA solution) shown
in Figures 11, the turbulent flow mixing results (gas-gas mixing vs. liquid-liquid mixing)
Figures 18 below show a series of contour plots (at different times with a 0.5 s interval)
displaying the water-PAA mixing (with R = 3.73 and d/D = 0.1). The series of plots are
given at 0.5 s interval. For the mixing of water and PAA solution, the main flow Rem =
21,000, and the jet flow Rej ~ 4,000, d/D = 0.1, and R =3.73. In Figure 15 (shown
earlier), for gas-gas turbulent mixing, the series of plots are given at 0.01 s interval. Even
56
though the Rem and Rej for the gas-gas and the liquid-liquid cases are the same, the U m
for the liquid-liquid case is about an order of magnitude smaller than in the gas-gas case
as the kinematic viscosity of air is about an order of magnitude higher than that of water.
The growth and evolution of the large eddies are less pronounced in the liquid-liquid
(a)
(b)
(c)
57
(d)
(e)
Figure 18 Contour plots of PAA solution concentration at center vertical plane at different times
(a) t =20.0 s (b) t =20.5 s (c) t = 21.0 s (d) t = 21.5 s and (e) t = 22.0 s;
Because the residence time of water-PAA solution mixing ( U m = 0.11 m/s) is much
larger than residence time of air-methane mixing ( U m =2.0 m/s), the liquid-liquid case
was computed for a larger time to ensure the flow-field reached quasi-steady state. Each
simulation is conducted from t = 0.00 sec to three times of the theoretical residence time
to reach the quasi-steady state. For example, the theoretical residence time of a water-
PAA solution mixing case is about 16.6 sec, then numerical calculation is carried up to
50.0 sec. Even though the Rem and Rej for the gas-gas and the liquid-liquid cases are the
58
same, the U m for the liquid-liquid case is about an order of magnitude smaller than the
gas-gas case as the kinematic viscosity of air is about an order of magnitude higher than
that of water.
Figure 19 shows the time-averaged contour plots of PAA concentration from 49.80 sec to
50.00 sec at the exit plane of the geometry for the turbulent mixing of water-PAA
solution. It is interesting to find that the degree of mixing of PAA solution at the exit
plane is again good only for the bottom half of the pipe cross-section. Comparing with
gas mixing results shown in Figure 16, the liquid mixing shows more uniform and less
Figure 19 Contour plot of PAA concentration at the exit plane, with R = 3.73 and d/D = 0.1, time
average period: 49.80 sec to 50.00 sec, Rem = 21,000, Rej = 4000, Sct = 0.9
59
Pan and Meng (Pan and Meng, 2001) conducted two sets of experimental measurements
on T-junction mixing in liquid-liquid flows. The main flow is water and the jet flow is a
sodium fluorescein solution in water. The Reynolds number of the main flow is fixed at
20,850 in both cases; Reynolds numbers of the jet flow are 10,630 and 17,500 (Pan and
Meng, 2001). Simulations were carried out for the case with main flow Reynold number
Figures 20 (a) and (b) show the grayscale image of the instantaneous concentration
field for R = 3.06. The image covers a cross section at the symmetry plane in the cross-jet
region. The jet can be found entering from the top and turns to the right with the pipe flow.
The simulation results in Fig 20 (b) showed lack of structure compared to the experimental
measurements. The inflow side (left of the jet) did not show any sign of roll-up vortex
either. These differences might be caused by imperfect geometry lineup, or the residual
swirl in either the jet or the main flow, etc. The numerical errors contribute to the difference
as well, for example, LES filtered out small structures, which is one of the reason that
simulation predicts less diffusion. Besides, the difficulty on matching experimental time
and simulation time could be another factor as well. Though the simulation conditions were
carefully matched with the experimental conditions (Pan and Meng, 2001), the predicted
concentration field lacks some of the details found in the experimental work. The LES
filters out the sub-grid level fluctuations. The alternate here is to carry out the direct
numerical simulations (DNS) which can become prohibitively expensive in time and
resources. In addition, the simulations are for ideal conditions, assuming no upstream or
laboratory disturbances. Such effects are unavoidable in most experiments. These could
60
contribute to the mismatch of the finer details of the experimental and computational results
presented here.
R = 3.06, d/D = 0.1667, Rem = 20,850, Rej = 10,630. Screenshot of simulation is taken at 15.00
sec
61
Figure 21 (a) shows the instantaneous two-dimensional velocity field measurements (Pan
and Meng, 2001) on the center plane of the T-junction (side view, covering the pipe
diameter). The prediction from CFD simulation in Fig 14 (b) shows some fluctuation,
however (in part due to averaging) it does not provide as many fine structures as found in
the experimental measurement. The darker part on Fig 14 (b) shows high velocity region,
Figure 21 Instantaneous velocity field on the center plane of the tee mixer, R = 3.06, d/D =
0.1667, Rem = 20,850, Rej = 10,630. Screenshot of simulation is taken at 15.00 sec
Figure 22 (a) below shows the measured (Pan and Meng, 2001) time-averaged tracer
concentration at the vertical center plane, where R = 3.06, Rem=20,850, Rej=10,630. The
contours from CFD simulation shown in Figure 22 (b) predict a similar distribution of
tracer concentration as found in the measurements. The simulation, however, predicts less
diffusion of the tracer compared to the experimental measurement. Some probable causes
might again be slight misalignment in the experimental setup, or the residual swirl in either
the jet or the main flow, numerical errors introduced by the differencing (spatial and
Rem=20,850, Rej=10,630. Time averaging period from the present CFD simulation: 22.0 sec to
22.2 sec
64
Finally, Figures 23 below show a series of contour plots (at different times with a 0.05 s
interval) displaying the water-sodium fluorescein solution mixing (with R = 3.06 and d/D
= 0.1667). For the mixing of water-sodium fluorescein solution, the main flow Rem =
20,850, and the jet flow Rej = 10,630, d/D = 0.1667, and R = 3.06, d/D = 0.1667. Even
though the Rem for the water-PAA solution and the water-sodium fluorescein solution
cases are nearly the same (Figure 18 vs. Figure 23), the Rej for two cases are significantly
different.
(a)
(b)
(c)
65
(d)
(e)
Figure 23 Contour plots of tracer (sodium fluorescein) concentration at center vertical plane at
different times (a) t =15.0 s (b) t =15.5 s (c) t = 16.0 s (d) t = 16.5 s and (e) t = 17.0 s; d/D =
Comparing the results shown in Figure 18 (a) – (e), (Rej = 4,000), with the above results
in Figures 23 (a) – (e), (Rej = 10,750), it is evident that the downstream eddy structures
4.6 Conclusions
The transport processes that are involved in the mixing of either two gases or two miscible
liquids in a T-junction mixer are investigated. The turbulent flow fields are calculated for
the T-junction with the LES model employing CFD-ACE+. In the mathematical model the
66
both space and time. The results obtained by numerical simulations are compared with two
well-defined experiments (Forney and Lee, 1982, Pan and Meng, 2001) for gas-gas and
inlet flow is designed to penetrate to the opposite pipe wall in the mixer.
The comparison between laminar air-methane and water-PAA solution cases showed that
with similar Rem and Rej, the air-methane case (with a smaller Sc) has improved mixing
beyond the T-junction, compared to the water-PAA solution case (with a higher Sc). LES
simulations for turbulent mixing showed differences between the mixing behavior of gas-
gas and liquid-liquid streams for similar geometrical and flow conditions. However, the
mixing of two turbulent flow streams (indicated by the maximum concentration value of
the solute cm) downstream of a T-junction are not affected by the Sc values even when Sc
for the gas and liquid-liquid streams vary by about three orders of magnitude. Cross-
sectional contour plots show that even with such high Reynolds number (as found in this
study), there is still a relative large concentration gradient at the pipe exit (~15 pipe
diameters). Additional mixing devices like baffle plates are used in the industry to improve
mixing. For the liquid-liquid cases, the comparisons of flow structures between CFD
predictions and experimental measurements (Pan and Meng, 2001) showed qualitative
agreement only.
67
JUNCTION
5.1 Introduction
One of the most encountered mixing devices is a T-junction pipe. The T-junction containing
a transverse pipe and a 90º branch is widely used in the mixing of two fluids (in gaseous
or liquid form). The geometry can be utilized as a chemical reactor, combustion chamber,
piping system, etc. Gas mixture flows appear in diverse applications. Examples include
pollutant dispersion, chemical processing, and combustor mixing and reaction. Previous
discussion was carried out on the mixing of two gases (air and methane) with no heat
transfer considered in a T-junction (Forney and Lee, 1982). However, it is also common
that the two streams are having different temperature. For example, cool air flow is mixing
with a hot air flow in an HVAC unit (Hirota, Mohri, Asano and Goto, 2010). Liquid-liquid
mixing problems with temperature variations are also common in industry and in nature. A
previous study of liquid-liquid mixing problem was the mixing of water with water-sodium
solution in a T-junction (Pan and Meng, 2001). T-junctions are also used for the mixing of
wastewater and disinfectant in wastewater treatment process, the mixing of hot water with
Vaidyanathan, 2010), etc. Therefore, the hot water mixing with cooling fluid is considered
The mixing of two flows in a T-junction mixing chamber contains mass and heat transport
processes, thus the process is very complicated. It is necessitated to entirely reveal the
detailed characteristic and the mechanism of the mixing of hot and cold flows in the T-
68
junction. Previous studies on T-Junction mixing showed the mixing behaviors and flow
structures between the main flow and the jet in detail (Wei, Farouk and Haas, 2017). With
are applied in the present study to predict the flow properties and mixing characteristics of
two fluids (gas-gas or liquid-liquid) at a 90o T- junction of two pipes as well. As a balance
on computational cost and flow resolution, large eddy simulation (LES) is becoming one
of the most popular turbulent modeling technique in CFD. LES has been proven to be
In this study, CFD simulations of the mixing of two gaseous flows is carried out first for
validation purpose. In the gaseous mixing case, a cold air flow is introduced from the
main inlet, a stream of hot air is injected from the jet flow inlet. Main flow Reynolds
number is denoted by Rem, inlet jet Reynolds number is denoted by Rej. The mixing of
two streams is studied under turbulent flow conditions, where Rem ~ 15,000. After the
fidelity of the CFD model is validated, another series of simulations are focused on the
mixing of two water streams, i.e. cool water flow mixing with hot disinfectant solution
(0.3% water-peracetic acid mixture) to represent the scenario in a pilot scale wastewater
treatment reactor. In the liquid-liquid study, cold water flow is introduced from the main
pipe inlet and hot water is injected from the jet flow inlet. The mixing of the two liquid
CFD simulations are carried out in the present study to predict the flow properties and
mixing characteristics of two fluids at a 90o T- junction of two pipes of dissimilar diameters.
Previous works in the T-junction mixing problem were either focused on velocity ratio
69
(Sroka and Forney, 1989) or temperature difference (Hirota, Mohri, Asano and Goto,
2010). In this study, both species and temperature are considered in a T-junction. For the
turbulent flow field predictions, the LES technique, which has been proven to be a better
process. A schematic of the mixing problem considered in the T-junction is shown below.
The two flows are labeled as the main flow (in the larger pipe with diameter D) and the jet
flow (in the smaller pipe with diameter d). The main pipe with D = 7.62 cm is placed
horizontally along the x-axis. The jet pipe with smaller diameter is on the top and placed
The diameter of the inlet jet is 1.27 cm making the diameter ratio d/D = 1/6. The distance
from injecting point to the outlet is about 15 times of main pipe diameter to ensure the flow
is fully developed when it exits. The main inlet and jet velocity is marked by Um and Uj
respectively. The ratio between the two velocities is denoted by VR, where VR = UM / UJ.
70
Similarly, the temperatures for the two inlets are Tm and Tj, the ratio of the two is TR, where
TR = Tm/ Tj. Table 1 below lists out the major variables in this study.
5.3 Methodology
Mathematical Model
For simulating turbulent flows in the T-junction, we need to invoke a turbulence model.
We have used the Large-eddy simulation (LES) model to simulate the turbulent mixing
problems in the T-junction. LES solves a “filtered” set of Navier-Stokes equation (Yang,
scales from the solution to the Navier-Stokes equations. Because the principal difficulty in
simulating turbulent flows comes from the wide range of length and time scales, this
operation makes turbulent flow simulation cheaper by reducing the range of scales that
must be resolved. The LES filter operation is low-pass, meaning it filters out the scales
associated with high frequencies. A LES filter can be applied to a spatial and temporal field
and the filter size is decided by maximum grid size and time step. Any spatial fluctuation
smaller than grid size will be filtered out, and similarly, any temporal frequency higher
71
than time step frequency will be filtered out as well. The operation of “filtering” applied to
the flow is a convolution with a spatial filter over entire domain, which is,
f ( x, t ) = ∫ f (ξ , t ) G ( x − ξ , ∆ ) d ξ (5.1)
D
where f is filtered value of a field variable f, ∆ is the filter width and G( x − ξ , ∆)dξ is
In the simulation,
1 if x − ξ ≤ ∆
G( x − ξ , ∆) = ∆ 2 (5.2)
0 otherwise
With this operation, variables such as velocity u were divided into large scale which is
denoted by an overbar and sub grid scale (SGS) denoted by a prime, which is,
u = u +u' (5.3)
Therefore, the filtered governing equations for a Newtonian incompressible flow become
as following:
∂ρ ∂
+ ( ρ ui ) = 0 (5.4)
∂t ∂xi
∂ui ∂ 1 ∂p ∂u ∂u j
+ (u i u j ) = − +ν ( i + ) + Si (5.5)
∂t ∂x j ρ ∂ xi ∂ x j ∂ xi
72
Furthermore, the filtered advection term can be split up, following Leonard (Leonard,
1974),
where τ ijr is residual stress tensor, then the filtered momentum equations become
∂ ui ∂ 1 ∂p ∂ ∂τ ijr
+ (u i u j ) = − + 2ν S ij − (5.7)
∂t ∂x j ρ ∂xi ∂x j ∂x j
1 ∂ ui ∂ u j
S ij = ( + ) (5.8)
2 ∂ x j ∂ xi
1
τ ijr − τ iiδ ij = 2ν t S ij (5.9)
3
ν t ~ ∆2 2S ij S ij (5.10)
ν t = cs2∆2 2S ij S ij (5.11)
where cs is called Smagorinsky constant, cs = 0.1 has been chosen in this simulation.
73
∂ ( ρ h) ∂ ∂ ∂T
+ ( ρ hu j ) = ( k eff ) (5.12)
∂t ∂x j ∂x j ∂x j
where keff is effective coefficient that includes turbulent mixing contribution in addition to
µ t Cp
k eff = k + (5.13)
Prt
The grid meshing is generated in the CFD-GEOM. To balance the computational cost and
the resolution of flow characteristics, a grid with 900,000 structured cells is addressed to
this model. In order to capture the small structure at near field, more condensed meshing
is assigned to the region (see Fig. 2(b)). Then the length scale of the grid is gradually
(a)
(b)
Figure 25 Grid Generation of the T-junction (a) prospective view; (b) Cross-sectional view of the
T-junction
75
When generating the meshing grid, as suggested by Zang (Zang, 1991), near-wall
configuration should have the first node being located between 0 < y+< 1, and more than
three nodes should be placed with 0 < y+ < 10 at the same time, where y+ stands for the
τw
y
ρ
y+ = (5.14)
υ
The plot in Figure 26 shows the y+ at the top wall of the main pipe along after the jet
nozzle, where y = 0.0381 m. It is noticeable that in this model, each first node near the
top wall is below 1.0, which follows the suggestion from Zang(Zang, 1991).
76
Figure 26 y+ at the top wall of the main pipe after the jet nozzle at y = 0.0381m.
The simulations are carried out in transient mode with a time step equal to 10-3 sec. A
second order blended upwind scheme is applied for the convective-diffusive terms and
Crank-Nicolson scheme with a blending factor of 0.7 is used for the temporal terms in the
governing equations. The internal relaxations for velocities and enthalpy are 0.2. The
At inlets of the domain, normal uniform velocities are specified. No-slip conditions are
assigned for the walls. Initially, the main pipe is filled up with cold water, while the jet
inlet is specified with hot flow (mixture of water with tracer). At t = 0.0 s, hot flow is
introduced from jet inlet and cold flow is injected from main inlet as well. The simulation
is carried out for three times of the theoretical residence time to diminish the effect of
77
initial conditions and ensure the flows becomes quasi-steady(Wei, Farouk and Haas,
Validation
Previous experimental work on a T-junction geometry was carried out by Hirota et al.,
(Hirota, Mohri, Asano and Goto, 2010). In their experiments, the temperature of cold and
hot air flow is 12ºC and 60ºC respectively. By matching the experimental parameters, a
(a)
78
(b)
Figure 27 Visualized image of the interface between two flows in Z/A = 0 from (a) Experiment;
As shown in Figure 27, the interface of hot and cold fluid predicted by CFD simulation
matches the experimental measurement. Because of the nature of LES, small structures
were filtered out. This is the reason that CFD didn’t provide enough information on those
small fluctuations. Furthermore, mean velocity vector plot shows agreement between the
CFD simulation and experimental result as well. The interface line in CFD is about the
same location as in experiments. Besides, CFD simulation also predicts the recirculation
(a)
(b)
Figure 28 Mean velocity vector diagram in the symmetric plane of Z/A = 0. (The solid lines show
the time-averaged interface between the main and branch flows obtained by the flow
Figures 29 below are comparing the time averaged temperature profile at various
locations on the center plane between CFD simulation predictions and experimental
measurements. The time averaged temperature profile plots show that the prediction from
CFD simulation is agreeing with experiments fairly well. These comparisons from
(a)
(b)
81
(c)
Figure 29 Profiles of the mean normalized temperature at (a) x = 0; (b) x = 1.0; (c) x = 3.0
In Figures 30 (a) to (e) below, a series of contour plots displaying the characteristics of
the mixing of cold water with hot water-PAA solution (with TR = 1.28 and VR = 3.0) as
(a)
(b)
(c)
(d)
83
(e)
Figure 30 Temperature contour at center plane at (a) t = 15.00 sec; (b) t = 15.30 sec; (c) t = 15.60
From the simulations, it appears that the jet is unable to fully penetrate the main pipe
diameter at the near field. The large eddy structures of the hot liquid jet can be identified
downstream. The boxed zones in Figures 30 (a) – (e) shows the progression and decay of
a mixing structure as it flows downstream. It is noticeable that the upper part of the main
pipe has little heat transfer until 7-8 main pipe diameters downstream the nozzle. On the
other hand, the jet flow takes about 4 pipe diameters to reach the bottom.
When looking at the exit plane, which locates at about 15 main pipe-diameter
downstream after the jet nozzle, the time averaged temperature contour in Figure 31
shows rather non-uniform and shows wide variations of 6 degree Kelvin from lowest
Figure 31 Time-averaged temperature contour plot at exit plane (15 pipe diameter downstream).
TR = 1.28, VR = 3.0.
Another series of CFD simulation with TR = 1.09 and VR = 3.0 shows the comparable
results. By observation, the mixing is not through at the upper half of the pipe neither.
However, the jet flow reaches the bottom of the pipe at about 2 pipe diameter which is
(a)
(b)
(c)
(d)
86
(e)
Figure 32 Temperature contour at center plane at (a) t = 15.00 sec; (b) t = 15.30 sec; (c) t = 15.60
Figure 33 Time-averaged temperature contour plot at exit plane (15 pipe diameter downstream).
TR = 1.09, VR = 3.0.
87
The plot in Figure 32 shows the time-averaged temperature contour plot with TR = 1.09,
VR = 3.0, which has a similar pattern as in Figure 30. It is observed that the hot flow is
accumulating at the second half of the pipe at the exit, despite slightly smaller density.
Comparing with TR = 1.28, the mixing of TR = 1.09 shows more uniform and less
std (ci )
φ= (5.15)
Mean(ci )
where std(ci) represents the standard deviation of the concentration at certain cross-
As shown in Figure 34 (a) and (b), with same velocity ratio, the simulation with higher
temperature difference has a slightly smaller degree of mixing, which indicates a better
(a)
(b)
Figure 34 Comparisons of the degree of mixing between TR = 1.28 and TR = 1.09, where (a) VR
5.5 Conclusions
temperature main flow and a hot vertical jet in a T-junction were carried out. LES with
sub-grid scale modeling has shown to reveal the characteristics of the mixing process
successfully. CFD simulations show the effects due to density difference cannot be
overlooked in the study where temperature difference exists. The CFD predictions for
measurement is good regarding both the temperature field and the velocity field.
Further simulations show that the shear layer formed in the near field after jet injection is
essential for the thermal mixing. With the increase of the temperature ratio between the
two flows, the gradient of the shear increases correspondingly, which promotes the
kinetics and hydraulics of the wastewater treatment processes. With geometric scaling-
up, the performance of such disinfectant in a full-scale contactor can be predicted. The
pilot-scale serpentine reactor shown in Fig. 35 was built by PeroxyChem, LLC. This
facilities for studying the kinetics of disinfecting local wastewater with PAA.
On this pilot-scale reactor, the inlet pipe is a 90o T-junction, shown in Fig 36. The
0.635-centimeter-diameter injector which injecting PAA into the center of the main pipe
at a very slow speed (~0.001m/s) located in the middle. Total length of the T-junction
pipe is 5 meters.
Based on the geometry provided, a 3-D CFD geometric model with 600,000 Cartesian
cells first was created as shown in Fig 37 (a) and (b). Square shaped pipe addressed in
this model to approximate circular pipe could reduce computational cost considerably.
The time-dependent simulations were carried out with a time step equal to 0.1 sec. A
second order blended upwind scheme was applied for the convective-diffusive terms and
92
Crank-Nicolson scheme with a blending factor of 0.7 was used for the temporal terms in
the governing equations. The internal relaxations for velocities and passive scalar are 0.2.
The convergence criterion was set to 0.001. Main pipe inlet had a mixture of water and
microbial with a velocity of 0.77 m/s. The injection pipe has 10 mg/L PAA with a
(a)
(b)
(b) Zoomed-in view of near field of the injector. The serpentine contactor was a 12-inch
diameter PVC pipe with 3 180° turns. Four channels formed a 90 ft long contact pass to
retain the contact time. Each turn has two 90° bends and one 19-inch straight section, in
this way, it could be fitted on a trailer truck to relocate conveniently. Over all dimensions
of the device was 22×4×4 ft. Meshing for the serpentine is shown in Fig 38.
93
Because of the difference in diameter, the flow pass from inlet T-junction would be
expanded. A ¼-inch thick baffle plate placed 4 inches from the expansion. The baffle
plate contained 72 0.5-inch-diameter holes, shown in Fig 39(a). To simplify the model,
the circular holes were accommodated to 15 1-inch-long square shaped structure. The
(a) (b)
Figure 39 (a) mechanical diamgram of baffle plate. (b) CFD geometric model of baffle plate
94
After simulating three residence times, the system became quasi-steady. Contour plots of
PAA concentration on center plane and exit cross-section showed the structure of the
mixing in pipeline. From the plots, it was obvious that within a 5-meter pipeline, PAA
was not mixing well with water. Further mixing enhancement method such as static
(a)
(b)
Figure 40 Contour plots of PAA concentration at 5.0 sec simulating time at (a) center plane; (b)
exit
95
Same numerical scheme was carried out in this simulation. Fig 40 showed the PAA
Figure 41 Instanteous contour plot of PAA in the first channel of the serpentine contactor at 9.0
sec.
For mixing enhancement study with different inlet configurations, three types of jet inlet
are proposed, they are: (a) inserted straightly to the middle of the main pipe (b) inserted
with 90-degree turn downstream; (c) inserted with 90-degree turn upstream as in Figure
42 below:
(a)
96
(b)
(c)
Figure 42 Three jet inlet configurations. (a) Inserted straightly (b) inserted with 90-degree
Figures 43 and 44 below show contour plots of the tracer concentration at center vertical
(a)
(b)
(c)
Figure 43 Contour plots of the tracer concentration at center vertical plane (a) Inserted
straightly (b) inserted with 90-degree turn downstream; (c) inserted with 90-degree turn
upstream
98
(a)
(b)
(c)
straightly (b) inserted with 90-degree turn downstream; (c) inserted with 90-degree turn
upstream
99
From both comparisons, the model with a 90-degree turn upstream jet inlet has the best
mixing performance among the three. The jet flow in the T-junction with a vertical jet
inlet is injected to the bottom. Carrying a large momentum with high velocity, it hits on
the bottom and reflected upward. In the case with 90-degree turn to downstream, on the
other hand, the system is lack of turbulence, which results from the interaction of main
flow and jet flow, so that the mixing performed most poorly. When examining the outlet
(a)
(b)
100
(c)
Figure 45 Contour plots of tracer concentration at the outlet (a) inserted vertically; (b)
inserted with a 90-degree turn downstream; (c) inserted with a 90-degree turn upstream
As shown in this plot, after 10-pipe diameter downstream, the degree of mixing in the
case with 90-degree turn upstream is much smaller than the other two simulations.
large pressure head requires a larger capacity pump, and a larger energy consumption.
(a)
(b)
(c)
Figure 47 The plot of pressure head (a) inserted vertically; (b) inserted with a 90-degree
The pressure head of each case is 914.4 Pa, 539.5 Pa and 1,829 Pa respectively. The
pressure head of the case with the best mixing results appears to be the largest. This is
because the energy is dissipated due to the intense turbulence and so that it requires the
pressure head.
103
6.5 Conclusions
In this study, we showed the simulation of a pilot scale wastewater treatment reactor.
However, the mixing of wastewater with disinfectant (PAA solution) did not well-mixed
condition even after 20-pipe-diameter in the main pipe. Therefore, a study of mixing
enhancement with different inlet configuration was carried out. It was shown that the one
with a 90-degree turn upstream provided the best performance of mixing. However, it
METAMODEL ON WASTEWATER
7.1 Introduction
Duggirala, Thompson, Zhao and Halaby, 1997) or (in drinking water) minimize the
formation of disinfection byproducts (Craun, Bull and Symons, 1994). This inactivation
must be achieved at a reasonable cost, and also must be capable of meeting the objective
(minimize toxicity) given the variability of water or wastewater quality and disinfection
kinetics.
While there is a growing literature to show that computational fluid dynamics (CFD) can
systems(Zhang, Tejada-Martínez and Zhang, 2014, Greene, Farouk and Haas, 2004),
these models frequently require hours of computer time to run for a particular
configuration. These speeds are inadequate to serve as an engine for the optimal
In prior works, it has been shown in other applications that a metamodel could be used to
approximate CFD results over a range of physical and design variables(Hoque, Farouk
and Haas, 2011, Hoque, Farouk and Haas, 2011, Hoque, Farouk and Haas, 2010).
105
and have gained interest over the past two decades(Barton, 1994). The objective of this
work was to develop a proof of concept for the use of artificial neural network based
wastewater disinfection contactor where peracetic acid was used as the disinfectant.
7.2 Methodology
The development of the present metamodel requires following five steps: (1) identify 11
input variables, the 11-D performance space and 2 outputs, i.e. PAA residual
concentration and the number of survival microorganisms; (2) determine the range of
each variable in interest; (3) generate the dataset by using a quasi-random sampling
technique to cover most scenarios in the design and operational space; (4) conduct
numerical experiments, i.e. 3-D CFD simulations configured with each set of parameters;
and finally (5) train the metamodel with numerical results and determine a configuration
As shown in Figures 48 (a) and (b), a full-scale wastewater treatment reactor is modeled
as an open rectangular tank with dimensions of length L, width W and depth D. An inlet
pipe of wastewater is placed horizontally near the bottom and an inlet tube for PAA is
located on top, dispensing disinfectant to the top surface. A weir at certain height locates
at the outlet wall to maintain a certain depth of the flow. Different number of baffles
forming a serpentine control the residence time (θ). Here b is to stand for the number of
baffles. The baffle slit width (bx) is described by the ratio with total length (L).
106
Moreover, there is a sharp edge weir designed near the entrance to form a mixing
(a)
(b)
To simplify the chemical kinetics, two reactions are mainly considered in the disinfection
process. They are: the reaction of PAA decay and the inactivation of microorganisms (N),
A recent batch study(Tizzoni, 2016) shows that PAA decays via immediate demand plus
first order decay in wastewater, and the inactivation of microorganisms with PAA
follows Hom’s power law(Hom, 1972). Therefore, the chemical kinetics may be written
as:
d [PAA]
= − k d [PAA] (4.3)
dt
d[ N ]
= − mk * [PAA]n t m −1[ N ] (4.4)
dt
where kd is the decay rate constant of PAA, k* is the Hom’s law rate constant, m and n
Since an implicit form of equation (4) is needed to apply in CFD simulation, the Hom’s
power law can be transformed to an equivalent form only in state variables(Haas and
Joffe, 1994):
contains both physical and chemical parameters, which divided into three types:
hydraulic characteristics (flow rate, residence time), reactor dimensions and PAA decay
section, both PAA residual concentration and the number of microorganisms can be
written as:
[N ]
f (log( )) = f (Q,θ , L,W , D, b, bx, kd , k * , m, n,[PAA]0 ) (4.6)
[ N ]0
D by expression:
D=
( Qgθ )
( LgW ) (4.8)
[N]
f (log( )) = f (Q,θ , L ,W , b, bx, kd , k * , m, n,[PAA]0 ) (4.9)
[ N ]0 W (b + 1) D
per day (MGD) range. The range of the flow rate Q in this study is set from 10 to 100
MGD. The hydraulic residence time of a contactor varies between 15 and 75 mins. On
the other hand, the geometry of a contactor can vary from a square to a long narrow
rectangle (from top view). Therefore, the ratio of l/w is limited within the range of 1.0 to
12.0; while w/(b+1)/d is limited within between 0.8 to 2.0. The number of baffle varies
from 1 to 10. The baffle slit width varies from 5% to 15% of the length of the reactor.
Typically, the applied PAA dose has a range of 0.5-10 mg/L. The range of the
For numerical experiments, space-filling designs play a significant role. The quality of
the space-filling affects the performance of the developed model. Four types of space-
filling designs have been used more often in the literature: orthogonal arrays, Latin
hypercube designs, Hammersley sequences, and uniform designs(Wang and Shan, 2007).
Hammersley sequence sampling (HSS) has been found to provide better uniformity than
example. 40 data points are generated for each variable by the HSS. Detail of the
flow in the reactor always maintains a minimum scour velocity to ensure microorganisms
Q
V= (4.11)
W
Dg
b +1
111
Any dataset with scour velocity smaller than 0.3 m/s is discarded and replaced by a
resampled dataset.
Since water flow can be treated as incompressible flow, the continuity equation can be
written as:
112
∂u ∂v ∂w
+ + =0 (4.12)
∂x ∂y ∂z
∂u ∂u ∂u ∂u 1 ∂p
+u +v +w =− + ν∇ 2u + f x
∂t ∂x ∂y ∂z ρ ∂x
∂v ∂v ∂v ∂v 1 ∂p
+u +v +w = − +ν∇ 2 v + f y (4.13)
∂t ∂x ∂y ∂z ρ ∂y
∂w ∂w ∂w ∂w 1 ∂p
+u +v +w =− + ν∇ 2 w + f z
∂t ∂x ∂y ∂z ρ ∂z
Turbulence Modeling
which is in turbulent flow region for an internal flow. Because of the random fluctuation
modeling will help to reduce the difficulty of predicting the turbulent region. There are
equations, Large-eddy Simulation (LES) and Direct Numerical Simulation (DNS). k-ϵ
There are two transported variables in k-ϵ model. The first variable determines the energy
in the turbulence and is called turbulent kinetic energy (k). The second variable is the
turbulent dissipation (ε) which describes how much the turbulent kinetic energy will
dissipate to the surrounding fluid. The idea of k-ϵ model is Reynolds decomposition,
which mathematically separates the time-averaged and fluctuating parts of the flow. The
ui = ui + ui ' (4.14)
113
'
where ui is the averaged laminar velocity, ui is turbulent fluctuation.
1 2 2 2
k= (u + v + w ) (4.15)
2
microorganisms in wastewater. Hence, there are two major reactions in this study, which
Both of the two reactants in equation (1) and (2) can be predicted by solving chemical
(4.16)
∂[ N ] ∂[ N ] ∂[ N ] ∂[ N ] ∂ 2 [N] ∂ 2 [ N ] ∂ 2 [ N ]
+u +v +w = DN −Water { 2 + + } + SN (4.17)
∂t ∂x ∂y ∂z ∂x ∂y 2 ∂z 2
becomes:
One of the advantages of CFD in post-processing is that it is very easy to visualize the
data. In the first run as the based model, the values of each variable were chosen as
shown in Table 7.
Variable Value
Q 30 MGD
θ 30 min
L/W 2.75
W/(b+1)/D 0.667
b 9
bx 10%
[PAA] 0 10 mg/L
kd 0.028/min
k* 2.289 min-m (mg/L)-n
m 0.468
n 0.574
Figure 49 (a) shows a contour plot of PAA concentration on the middle horizontal plane.
The contour plot of MO survival ratio on the same plane is shown in Figure 49 (b).
115
(a)
(b)
Figure 49 (a) Contour plot of PAA Residual concentration; (b) Contour plot of N survival ratio at
As shown on Figure 49(a), about 60% of PAA has been consumed during the process
To ensure the flow in the time-dependent simulation has reached the quasi-steady state,
the survival ratio at the exit is tracked during the entire simulation. Plot of survival ratio
As shown on Figure 50, the plot of N survival ratio has a dramatic decrease in the second
residence time. And the curve flats out around 90 minutes, which is about three
in this research are carried out from zero to three theoretical residence time.
117
Inspired by neural system in animals, ANN is a structure that data flows between neurons
with different weights and biases (Hopfield, 1988). There are three layers in general,
which are input layer, hidden layer and output layer respectively. Each neuron in the
hidden layer represents a nonlinear transfer function (Basheer and Hajmeer, 2000). In this
study, hyperbolic tangent function is addressed. The number of the nodes in the hidden
and bias values to best fit the data. Figure 51 shows the structure of the ANN model.
To evaluate the performance of an ANN based metamodel, the following diagnostic plots
were employed: metamodel prediction values (ANN fitted values) versus observed values
(numerical simulation results). A plot appears close to a 45º line indicating an accurate
model. The sum of squares of errors (SSE) is an important aspect to assess the
SSE = ∑ (Y − Yo )
2
(4.19)
Besides SSE, there are two more criteria for assessing the adequacy of fit of a model,
which are, Cross Validation SSE (CV) and Akaike Information Criterion (AIC) (Akaike,
1978). Cross-validation is conducted by parsing the data into two groups. Firstly, the
metamodel is trained or fitted to one of the groups. The fitted model then is used to
predict the target values of the other group of data. This test is repeated for all groups of
evaluations were conducted to analyze the artificial neural networks. For AIC, this
criteria weights the overall error in the outputs to the number of parameters in the model
where n is the number of data points, sse is the sum squared error, and K is the
Both criteria are used to test if the data was over-fitted. SSE decreases monotonically
with the increase of the number of nodes in the hidden layer. The cross-validation SSE
initially decreases, but then is expected to increase when over-fitting occurs. AIC, which
correct for additional parameters and is expected to behave in a similar fashion to CV.
Training ANN model with 11 features of interests as input, starting from 1 node in the
hidden layer, the survival ratio of survival ratio and PAA residual as two outputs
119
separately, table 8 and 9 record the change of each criterion for survival ratio and PAA
Table 8 shows that the model with three nodes in the hidden layer provides the best
performance with respect to both AIC and CV. In Table 9 on the other hand, there is a
discrepancy between AIC and Cross Validation wherein there is no optimal AIC
achieved. This study accords more significance to the results of cross validation in the
interpretation. Therefore, the model with three nodes in the hidden layer is chosen.
Numerical Validation
In order to test and validate the metamodels for MO and PAA, an additional 15 CFD runs
have been conducted. These 15 CFD cases are derived by the same quasi-random
sampling technique to generate points distinct from the 40 in the training set. Table 10
Figure 54 and 55 below display ANN fitted values of MO survival ratio and PAA
residual concentration versus the values from CFD simulations respectively. SSE for each
target is 0.8605, 2.1143 respectively. Then the mean squared error (MSE) of two models
are 0.2395, and 0.375 respectively, which can be considered with a reasonable accuracy
pilot scale reactor(Santoro, 2005). The configuration of the pilot scale reactor is shown
on Figure 56.
125
Figure 56 The configuration of the pilot scale reactor used in the experiment
This pilot scale reactor has geometric similarity to the configurations simulated in the
CFD study. However, it is substantially smaller. The pilot scale reactor treats 0.0025
MGD, far below the range of the CFD study design space. Before utilizing the
used with a scaled-up flow of 25.2 MGD. The other parameters after scale-up are shown
in the table below. The process kinetics were also those used from Santoro’s study16.
L/w w/(b+1)/d Q θ b bx
∗ m n [PAA] 0
Then applying the parameter set to the metamodel, the validation plot is obtained by
Figure 57 Validation plot with experiments in the pilot scale reactor. A 1 log experimental error is
The error bars on Figure 57 showed the typical error in measuring the number of
experimental measurements. But with considering the error bars, there are still 10 cases
out of 15 cases overlapping the ideal fitting line. When computing Reynolds number
(Re), the pilot-scale flow is about 5,300; on the other hand, the scaled-up model is around
127
250,000. The Froude number (Fr) for each set up is 0.008 and 0.036 respectively. Such
difference in both Re and Fr may have contributed to the discrepancy that was observed
To determine the influence of scale factor in this comparison, a sensitivity study on scale-
up factor (Sf) is carried out. Besides the original case with Sf = 10,000, two more cases
are considered: one with Sf = 3,000 (which represents about the lower limit of size in the
ANN metamodel development); the other one with Sf = 30,000. The parameters after
scale-up are shown in table 12 (Sf = 3,000) and table 13 (Sf = 30,000).
L/w w/(b+1)/d Q θ b bx
2.38 2.1 7.56 MGD 30 min 3 10%
∗ m n [PAA] 0
-1
0.015 min 1.75 0.57 0.54 Varying
L/w w/(b+1)/d Q θ b bx
2.38 2.1 75.6 MGD 30 min 3 10%
∗ m n [PAA] 0
-1
0.015 min 1.75 0.57 0.54 Varying
After applied these parameter sets into metamodel prediction, the results of sensitivity
Table 14 Average bias and Mean squared error in the sensitivity study
As shown on Table 14, the prediction with Sf = 3,000 has the smallest average bias and
the least mean squared error. With increasing Sf value, difference in both Re and Fr
(between experimental pilot and the ANN model) is getting larger, resulting in increasing
7.7 Conclusions
order to provide a relatively reasonable size data set to apply in the ANN based
metamodels train process, which are for predicting N survival ratio and PAA residual
Evaluated by multiple criterions such as SSE, cross validation and AIC, the metamodels
with three nodes in the hidden layer were shown to be the best models for predicting N
survival ratio and PAA residual concentration. 15 more CFD simulations conducted for
validation showed that the metamodels can provide a reasonable accuracy of prediction
with a lot less time consuming compared with full-scale numerical simulations.
wastewater treatment reactor due to its effectiveness. Comparing the mathematical model
with a full scale transient CFD simulation, the computation time for a reasonable
the industrial use. A wide adaptable metamodel is suitable for any similar wastewater
treatment reactor configuration. For example, at given geometry and chemistry situations,
by varying the configuration of the reactor such as baffle number, baffle slit width,
129
residence time and flow rate, ANN based metamodels will provide a reasonable
Nomenclature
b = number of baffle;
L = length of reactor;
N = number of microorganisms;
Q = Flow rate;
Re = Reynold’s number;
130
Γ = molecular diffusivity.
131
FUTURE RESEARCH
8.1 Summary
The transport processes that are involved in the mixing of either two gases or two miscible
liquids in a T-junction mixer are investigated. The turbulent flow fields are calculated for
the T-junction with the LES model employing CFD-ACE+. In the mathematical model the
both space and time. The results obtained by numerical simulations are compared with two
well-defined experiments (Forney and Lee, 1982, Pan and Meng, 2001) for gas-gas and
inlet flow is designed to penetrate to the opposite pipe wall in the mixer.
The comparison between laminar air-methane and water-PAA solution cases showed that
with similar Rem and Rej, the air-methane case (with a smaller Sc) has improved mixing
beyond the T-junction, compared to the water-PAA solution case (with a higher Sc). LES
simulations for turbulent mixing showed differences between the mixing behavior of gas-
gas and liquid-liquid streams for similar geometrical and flow conditions. However, the
mixing of two turbulent flow streams (indicated by the maximum concentration value of
the solute cm) downstream of a T-junction are not affected by the Sc values even when Sc
for the gas and liquid-liquid streams vary by about three orders of magnitude. Cross-
sectional contour plots show that even with such high Reynolds number (as found in this
study), there is still a relatively large concentration gradient at the pipe exit (~15 pipe
diameters). Additional mixing devices like baffle plates are used in the industry to
132
improve mixing. For the liquid-liquid cases, the comparisons of flow structures between
CFD predictions and experimental measurements (Pan and Meng, 2001) showed
qualitative agreement only. CFD simulations show the effects due to density difference
cannot be overlooked in the study where temperature difference exists. The CFD
predictions for validation showed that the agreement with previously published
experimental measurement is good regarding both the temperature field and the velocity
field. Further simulations show that the shear layer formed in the near field after jet
injection is essential for the thermal mixing. With the increase of the temperature ratio
between the two flows, the gradient of the shear increases correspondingly, which
a relatively reasonable size data set to apply in the ANN based metamodels training
process, which are for predicting N survival ratio and PAA residual concentration. The
Evaluated by multiple criterions such as SSE, cross validation and AIC, the metamodels
with three nodes in the hidden layer were shown to be the best models for predicting N
survival ratio and PAA residual concentration. 15 more CFD simulations conducted for
validation showed that the metamodels can provide a reasonable accuracy of prediction
with a lot less time consuming compared with full-scale numerical simulations.
wastewater treatment reactor due to its effectiveness. Comparing the mathematical model
133
with a full scale transient CFD simulation, the computation time for a reasonable
the industrial use. A wide adaptable metamodel is suitable for any similar wastewater
treatment reactor configuration. For example, at given geometry and chemistry situations,
by varying the configuration of the reactor such as baffle number, baffle slit width,
residence time and flow rate, ANN based metamodels will provide a reasonable
• The simulations could be run in cluster machines with parallel processing which
• In the current study of T-junction mixing LES was applied with the Smagorinsky
model coupled with 2nd order differentiating method, future simulations could use the
dynamic model instead and a higher order scheme such as 3rd order or Quadratic
• When comparing the CFD simulation results with experimental works, it might
not be a good idea to compare the flow contour or velocity vector. Even a very small
time-averaged quantities are available, validations should be carried out with mean
quantities.
134
shaped pipes, varying the angle of jet pipe with main pipe is to be discovered. For
• In current study, the simulations were carried out with k-ϵ model. In the future, a
• In the future study, other parameters such as temperature, pH, or gravity should be
• We show that the ANN based metamodeling can provide reasonably accurate
predictions with a small database. In future study, when facing the similar scenario, i.e.
limited data points of a high-dimensional system, to develop the ANN based metamodel
• To determine the number of the hidden nodes, the criteria do not always coincide.
The weight of each criterion should be evaluated individually. For example, in this study,
the cross-validation values did not follow the same trend as AIC values. To fulfill the
LIST OF REFERENCES
[1] Aieta E M, Berg J D, Roberts P V, and Cooper R C. Comparison of chlorine dioxide
and chlorine in wastewater disinfection. Journal of Water Pollution Control Federation,
pages 810-822, 1980.
[2] Akaike, H., 1978, "On the likelihood of a time series model", Statistician, 27, pp. 217-
235
[3] Alasri A, Roques C, Michel G, Cabassud C, Aptel P. Bactericidal properties of
peracetic acid and hydrogen peroxide, alone and in combination, and chlorine and
formaldehyde against bacterial water strains. Can J Microbiol 1992;38:635 – 42.
[4] Anders, U. and Korn, O., 1999, "Model Selection in Neural Networks", Neural
Networks, 12, pp. 309-323
[5]Antonelli, M., Turolla, A., Mezzanotte, V., & Nurizzo, C. (2013). Peracetic acid for
secondary effluent disinfection: a comprehensive performance assessment.
[6] Baldry MGC. The bactericidal, fungicidal, and sporicidal properties of hydrogen
peroxide and peracetic acid. J Appl Bacteriol 1983;54: 417 – 23.
[7] Baldry MGC, French MS. Activity of peracetic acid against sewage indicator
organisms. Water Sci Technol 1989a;21:1747 – 9.
[8] Baldry MGC, French MS. Disinfection of sewage effluent with peracetic acid. Water
Sci Technol 1989b;21:203 – 6.
[9] Barton, R. R. (1994). "Metamodeling: A state of the art review." Simulation
Conference Proceedings, 1994. WinterLake Buena Vista, FL, USA, 237-244.
[10] Bartrand, T.A., 2006. High Resolution Experimental Studies and Numerical
Analysis of Fine Bubble Ozone Disinfection Contactors (Doctoral dissertation).
Available from ProQuest Dissertations & Theses Database. (UMI No. 3239812).
[11] Bartrand, T.A., Farouk, B., Haas, N.C., 2009. Countercurrent gas/liquid flow and
mixing: implications for water disinfection. Int. J. Multiph . Flow. 35 (2), 171-184.
[12] Basheer, I. A., and Hajmeer, M. (2000). "Artificial neural networks: fundamentals,
computing, design, and application." Journal of Microbiological Methods, 43, 3-31.
[13] Blatchley, E., Hunt, B., Duggirala, R., Thompson, J., Zhao, J., and Halaby, T.
(1997). "Effects of Disinfectants on Wastewater Effluent Toxicity." Water Res., 31, 8.
[14] Block SS. Disinfection, sterilization, and preservation. 4th ed. Philadelphia:
Lea&Febiger Pubs; 1991.
[15] Calmet, I., and Magnaudet, J., 1997, "Large-eddy simulation of high-Schmidt
number mass transfer in a turbulent channel flow," Physics of Fluids, 9, p. 438.
136
[16] Chilton, T. H., and Genereaux, R. P., 1930, "The Mixing of Gases for Reaction,"
AIChE Transuctions, 25, p. 103.
[17] Colgan S, Gehr R. Disinfection. Water Environ Technol 2001;13:29 – 33.
[18] Craun, G. F., Bull, R. J., and Symons, J. M. (1994). "Balancing chemical and
microbial risks of drinking water disinfection, Part 1. Benefits and potential risks."
Journal of Water Supply: Research and Technology - AQUA, 43, 8.
[19] E. ToolBox, Water - Thermal Conductivity, in,
https://www.engineeringtoolbox.com/water-liquid-gas-thermal-conductivity-temperature-
pressure-d_2012.html, 2018.
[20] ESI-Group, 2016, "CFD-ACE, 2016, "ESI-US RD (2016) CFD-ACE+ modules
manual V2016. Huntsville, AL," ESI Group.
[21] ESI-Group, 2017, "CFD-ACE, 2017, "ESI-US RD (2017) CFD-ACE+ modules
manual V2017. Huntsville, AL," ESI Group.
[22] Fraser JAL, Godfree AF, Jones F. Use of peracetic acid in operational sewage sludge
disposal to pasture. Water Sci Technol 1984;17:451 – 66.
[23] Forney LJ, Lee HC. Optimum Dimensions for Pipeline Mixing at a T-Junction.
AiChE Journal. 1982;28(6):8.
[24] Forney, L. J., Noureddine, N., and Hanh, X. V., 1996, "Optimum jet mixing in a
tubular reactor," AIChE Journal, 42(11), pp. 3113 - 3122.
[25] Gehr R, Cochrane D, French M. Peracetic acid as a disinfectant for municipal
wastewaters: encouraging performance results from physicochemical as well as
biological effluents. Proc of the US water environment federation disinfection
conference; 2002.
[26] Greene, D. J., Farouk, B., and Haas, C. N. (2004). "CFD Design Approach for
Chlorine Disinfection Processes." J Am Water Works Assoc. (96), 13.
[27] Haas, C. N. and Joffe, J., 1994, "Disinfection under Dynamic Conditions:
Modification of Hom's Model for Decay", Environ. Sci. Technol., 28, pp. 1367-1369
[28] Hirota, M., Mohri, E., Asano, H., and Goto, H., 2010, "Experimental study on
turbulent mixing process in cross-flow type T-junction," International Journal of Heat
and Fluid Flow, 31, pp. 776-784.
[29] Hopfield, J. J. (1988). "Artificial neural networks." IEEE Circuits and Devices
Magazine, 4(5), 3 - 10.
[30] Hoque, S., Farouk, B., and Haas, C. N. (2011). "Development of metamodels for
predicting aerosol dispersion in ventilated spaces." Atmos Environ., 45, 12.
[31] Hoque, S., Farouk, B., and Haas, C. N. (2011). "Development of Artificial Neural
Network Based Metamodels for Inactivation of Anthrax Spores in Ventilated Spaces
Using Computational Fluid Dynamics." J Air Waste Manag Assoc., 61, 15.
137
[32] Hoque, S., Farouk, B., and Haas, C. N. (2010). "Multiple Linear Regression Model
Approach for Aerosol Dispersion in Ventilated Spaces Using Computational Fluid
Dynamics and Dimensional Analysis." Journal of Environmental Engineering, 136(6),
638-649.
[33] Lazarova V, Janex ML, Fiksdal L, Oberg C, Barcina I, Pommepuy M. Advanced
wastewater disinfection technologies: short and long term efficiency. Water Sci Technol
1998;38:109 – 17.
[34] Lefevre F, Audic JM, Ferrand F. Peracetic acid disinfection of secondary effluents
discharged off coastal seawater. Water Sci Technol 1992;25:155 – 64.
[35] Leonard A. On the Energy Cascade in Large-Eddy Simulations of Turbulent Fluid
Flows. Adv. Geophys. 1974;18:12
[36] Liberti L, Notarnicola M. Advanced treatment and disinfection for municipal
wastewater reuse in agriculture. Water Sci Technol 1999;40:235 – 45.
[37] Liberti L, Lopez A, Notarnicola M, Barnea N, Pedahzur R, Fattal B. Comparison of
advanced disinfecting methods for municipal wastewater reuse in agriculture. Water Sci
Technol 2000;42:215 – 20.
[38] Monarca S, Feretti D, Zerbini I, Zani C, Alberti A, Richardson SD, et al. Studies on
mutagenicity and disinfection by-products in river drinking water disinfected with
peracetic acid or sodium hypochlorite. Proc of IWA world conference, Berlin, Germany;
2001.
[39] Monarca S, Richardson SD, Feretti D, Grottolo M, Thruston AD, Zani C, et al.
Mutagenicity and disinfection by-products in surface drinking water disinfected with
peracetic acid. Environ Toxicol Chem 2002;21:309 – 18.
[40] Narayan, B. C., 1971, "Experimental Study of the Rates of Turbulent Mixing in Pipe
Flow," M.S. Thesis, University of Tulsa, Tulsa, OK
[41] Naik-Nimbalkar VS, Patwardhan AW, Banerjee I, Padmakumar G, Vaidyanathan G,
Thermal mixing in T-junctions, Chemical Engineering Science, 65(22) (2010) 5901-
5911.
[42] Nguyen, N. T., and Wu, Z. G., 2005, "Micromixers - a review," Journal of
Micromechanics and Microengineering, 15, pp. 1-16.
[43] Pan G, Meng H. An Experimental Study of Turbulent Mixing in a Tee Mixer Using
PIV and PLIF. AIChE Journal. 2001;47(12)
[44] Raj R, Mathur N, Buwa VV. Numerical Simulations of Liquid−Liquid Flows in
Microchannels. Ind. Eng. Chem. Res. 2010;49(21):10606-10614
[45] Rajala-Mustonen RL, Toivola PS, Heinonen-Tanski H. Effect of peracetic acid and
UV irridation on the inactivation of coliphages in wastewater. Water Sci Technol
1997;35:237 – 41.
138
[60] Wols, B.A., Uijttewaal, W.S., Rietvelda, L.C., Stellinga, G.S., van Dijka, J.C.,
Hofman, J.A.M.H., 2008b. Residence time distributions in ozone contactors. Ozone Sci.
Eng. J. Int. Ozone Assoc. 30 (1), 49-57.
[61] Wols, B.A., Hofman, J.A.M.H., Uijttewaal, W.S.J., Rietveld, L.C., van Dijk, J.C.,
2010a. Evaluation of different disinfection calculation methods using CFD. Environ.
Model. Softw. 25, 573e582.
[62] Wols, B.A., Shao, L., Uijttewaal, W.S.J., Hofmana, J.A.M.H., Rietveld, L.C., van
Dijk, J.C., 2010b. Evaluation of experimental techniques to validate numerical
computations of the hydraulics inside a UV bench-scale reactor. Chem. Eng. Sci. 65 (15),
4491-4502.
[63] Wols, B.A., Uijttewaal, W.S.J., Hofmana, J.A.M.H., Rietveld, L.C., van Dijk, J.C.,
2010c. The weaknesses of a k-epsilon model compared to a large-eddy simulation for the
prediction of UV dose distributions and disinfection. Chem. Eng. J. 162 (2), 528-536.
[64] Wols, B.A., Hofman, J.A.M.H., Beerendonk, E.F., Uijttewaal, W.S.J., van Dijk, J.C.,
2011. A systematic approach for the design of UV reactors using computational fluid
dynamics. AIChE J. 57 (1), 193-207.
[65] Wols, B.A., Hofman-Caris, C.H.M., Harmsen, D.J.H., Beerendonk, E.F., van Dijk,
Chan, J.C., Blatchley, E.R., 2012. Comparison of CFD, biodosimetry and Lagrangian
actinometry to assess UV reactor performance. Ozone Sci. Eng. 34(2), 81-91.
[66] Xu P, Janex M L, Savoye P, Cockx A, and Lazarova V. Wastewater disinfection by
ozone: main parameters for process design. Water Research, 36(4):1043-1055, 2002.
[67] Yang Z. Large-eddy simulation: Past, present and the future. Chinese Journal of
Aeronautics. 2014;28(1):14.
[68] Yuan Z, Ni Y, van Heiningen ARP. Kinetics of peracetic acid decomposition: Part I.
Spontaneous decomposition at typical pulp bleaching conditions. Can J Chem Eng
1997;75:37 – 41.
[69] Zang, T. A., 1991, "Numerical simulation of the dynamics of turbulent boundary
layers: perspectives of a transition simulator," Philos. Trans. Roy. Soc. Lond. A,
336(1647), pp. 95-102.
[70] Zhang J, Tejada-Martínez AE, Zhang Q. Developments in computational fluid
dynamics-based modeling for disinfection technologies over the last two decades: A
review. Environmental Modelling & Software 2014;58:71-85.
[71] Zhiyin, Y., 2015, "Large-eddy simulation: Past, present and the future," Chinese
Journal of Aeronautics, 28(1), pp. 11-24.
140
VITA
Wangshu Wei was born in the city of Shenyang, China. He started his academic career in
received his bachelor’s degree in mechanical engineering in 2012. Then he continued his
study towards a doctoral degree in Drexel University since the Fall of 2013. The research
area he focused on was the 3-dimensional numerical flow realization with mass and heat
that affecting the mixing behavior of fluids, his research was to provide a solution to
LIST OF PUBLICATIONS
b). Wei W, Farouk B, Haas C N. Mixing of Gases and Miscible Liquids in a T-Junction.
ASME 2017 International Mechanical Engineering Congress and Exposition. Tampa, FL,
US, 2017
c). Wei W, Farouk B, Haas C N. Gas mixing vs. liquid mixing in a T-junction: a
d). Wei W, Farouk B, Haas C N. Development of a CFD Based Artificial Neural Network