Sei sulla pagina 1di 18

VELOCITY-DISTRIBUTION IN PRESSURIZED PIPE FLOW USING CFD:

ACCURACY AND MESH ANALYSIS

NUNO M. C. MARTINS (1), NELSON J. G. CARRIÇO (2),


HELENA M. RAMOS (3) & DÍDIA I. C. COVAS (4)

(1) Instituto Superior Técnico, Universidade de Lisboa, Portugal, nunomiguelmartins@tecnico.ulisboa.pt (tel.+351 218 418 149; fax.+351 218 497 650)
(2) Instituto Superior Técnico, Universidade de Lisboa, Portugal, nelson.carrico@tecnico.ulisboa.pt
(3) Instituto Superior Técnico, Universidade de Lisboa, Portugal, helena.ramos@tecnico.ulisboa.pt
(4) Instituto Superior Técnico, Universidade de Lisboa, Portugal, didia.covas@tecnico.ulisboa.pt

Abstract

The aim of the current paper is the development and application of a systematic and comprehensive approach for
obtaining the most efficient meshes, described in terms of dimensionless parameters, for modelling pressurized water
flows in pipes using Computational Fluid Dynamics (CFD). This analysis is carried out for the combination of three
dimensionless mesh parameters: a , the dimensionless mesh size in the axial direction expressed in terms of number of
diameters; c , the dimensionless mesh size in the circumferential direction expressed in terms of the pipe perimeter; and
r , the ratio between the first layer thickness (FLT) and the thickness of the viscous sublayer for a specific turbulent flow.
A four-step systematic approach has been used to obtained the most efficient meshes: i) the mesh is generated in a
meshing software using the 3-D fluid domain which is defined in a computer-aided design software; ii) steady state fluid
flow is simulated by using the CFD until a convergence criteria is met; iii) the accuracy of the numerical results is
evaluated by comparing the computed velocity profiles with exact or semi-empirical solutions; iv) the previous three
steps are repeated for different combinations of mesh parameters and for two flows (laminar and turbulent) and the most
efficient meshes are obtained by the compromise between the maximum accuracy and the minimum computational effort.
The use of the most efficient meshes in other pipe flows is discussed.

Keywords: Computational fluid dynamics, efficient mesh, mesh generation, pipe flow and velocity profile.

1. Introduction

Pressurized pipe flow occurs in urban water, industrial and irrigations systems, in which fluid is typically water.
Although it could also be wastewater, oil or, even, gas. This paper focuses only on water flows in circular ducts (i.e., pipes);
other types of cross sections are not considered herein. Steady state fluid flow in pressurized pipes is well-known, not
being necessary to use advanced computational techniques for describing it. The on-going research aims, at a later stage,
to use Computational Fluid Dynamics (CFD) modelling for fluid flow simulation under unsteady state conditions. The
reason is that the velocity profile inverts during transient flows and the momentum dissipation cannot be described by
steady state formulations that assume a uniform velocity distribution. CFD modelling will contribute for a better
understanding of this phenomenon as well as of the mechanisms of momentum dissipation in transient flows both due
to the wall shear stress and to the fluid turbulence.

Governing equations of pressurized flows are derived from the mass, the momentum and the energy conservation
principles. Since temperature variations are insignificant in most engineering situations of liquid flows, such as water, the
energy equation can be neglected. Navier-Stokes (N-S) equations can be derived by adding the constitutive equations of
isotropic and Newtonian fluids to the momentum equation. The N-S equations together with the mass conservation
equation (continuity equation) and appropriate boundary conditions provide a complete mathematical description of
fluid flow motion (Foias et al., 2001; Young et al., 2011).

Exact mathematical solutions for the N-S equations cannot be obtained, except for very specific cases for which the
achieved results have proven to be in close agreement with collected data. Thus, the N-S equations have to be numerically
resolved. Since computer capabilities have increased in the last decades, a new branch of fluid mechanics, called
Computational Fluid Dynamics (CFD), has emerged (Olsen, 2000). In CFD, the common approach consists of dividing
the fluid domain into a number of topologically simple regions. Three dimensional (3-D) fluid domain is divided into
polyhedrons (called cells), each polyhedron is limited by faces with vertices (called nodes), forming a mesh. The N-S
equations are solved for each cell. This requires a substantial amount of computing resources and it would be impractical
for hand-solution purposes (Olsen, 2000). CFD is a mature technology, even though there is still a substantial amount of
theoretical development necessary before becoming a robust and reliable engineering tool. Currently, CFD is used to

1
analyse different physical problems. The huge computational costs for 3-D flow simulations are still a major limiting
factor for using CFD tools in industry, in particular for unsteady calculations (Knopp, 2006).

In recent years, CFD has been applied in several hydraulic engineering topics, namely: i) lake circulation (Lavelli et al.,
2002; Olsen et al., 2000; Simons, 1974); ii) flow in rivers (Baranya and Józsa, 2006; Fischer-Antze et al., 2008; Li and Xie,
2012; Olsen and Skoglund, 1994; Olsen and Stokseth, 1995); iii) flow around obstacles (Harlow and Welch, 1965; Mudgal
and Pani, 1998; Seed and Limited, 1997); iv) sediment deposition and transport (Huggins et al., 2004; Lenzi et al., 2006;
Olsen and Skoglund, 1994; Shams et al., 2002); v) local scour (Olsen and Kjellesvig, 1998; Roulund et al., 2005); vi) channel
morphology (Wu et al., 2000); vii) spillways (Andersson et al., 2010; Chanel and Doering, 2008; Huggins et al., 2004; Li et
al., 2010; Olsen and Kjellesvig, 1998); viii) reservoirs (Gidaspow et al., 2013; Olsen, 1999); ix) algae movements (Olsen et
al., 2000); x) turbines runners (Carija et al., 2008; Date and Akbarzadeh, 2010; Nilsson and Davidson, 2003; Ramos et al.,
2009; Wang et al., 2012); xi) laminar flow (Sahu et al., 2009; Weston et al., 1998); xii) transient analysis (Li et al., 2010; Martins
et al., 2012; Zhou et al., 2011); xiii) multiphase flow (Cheng et al., 2007; Hirt and Nichols, 1981; Li and Vasquez, 2012; Li et
al., 2010; Martins et al., 2012; Zhou et al., 2011).

One of the most important requirements, in CFD simulation, is the mesh generation. To produce accurate results, it is
essential to understand the influence of the mesh size. Analysing different parameters of the mesh and evaluating the
numerical results by comparing with classical solutions allows to obtain a suitable mesh for a specific problem with
acceptable computing effort and numerical accuracy (Ahmad et al., 2010; Versteeg and Malalasekera, 2007).

The current paper aims at the development and application of a systematic and comprehensive approach for obtaining
the most efficient meshes, described in terms of dimensionless parameters, for modelling pressurized water flows in pipes
using ANSYS Fluent. The most efficient meshes are the ones that meet two contradictory criteria: the maximum accuracy
and the minimum computational effort (i.e., computational time). A four-step approach is followed. In the first step, a
3 - D fluid geometry is drawn using a Computer-Aided Design (CAD) software and a meshing software is used to generate
a mesh. Three mesh parameters are used and combined for the analysis namely: a , the dimensionless mesh size in the
axial direction expressed in terms of number of diameters; c , the dimensionless mesh size in the circumferential direction
expressed in terms of the pipe perimeter; r , the ratio between the first layer thickness (FLT) and the thickness of the
viscous sublayer for a specific turbulent flow. The second step consists of the simulation of fluid flow for each generated
mesh using ANSYS Fluent; the realizable k-turbulence model is used to model 3-D turbulence. In the third step, the
velocity profiles in a fully developed flow are compared with theoretical solutions – namely the axi-symmetric velocity-
distribution by Hagen- Poiseuille formulation for laminar flows (White, 2003) and the linear law for the viscous sublayer
and the power-law in turbulent layer for turbulent flows (Çengel et al., 2012; Schlichting, 1955) – and errors are computed
(e.g., absolute errors, quadratic errors or weighed errors). Finally, the most efficient meshes for laminar and for turbulent
flows are obtained. The application of obtained meshes to other pipes and flows is discussed, as well as the use of the
proposed comprehensive approach for obtaining efficient meshes in other engineering problems (e.g., flows in flumes,
tanks or around obstacles).

2. Governing equations

2.1 Navier-Stokes equations

The governing equations that describe the motion of fluids are derived from the conservation principles of mass and
momentum. The mass conservation is expressed by the continuity equation:

D uk
0 [1]
Dt xk

where is fluid density, t is time, and uk is fluid velocity in the xk spatial coordinate (with k 1,2,3 ). This equation is
the general form of the mass conservation equation valid for both incompressible and compressible fluids.

The momentum conservation can be expressed in terms of the Cauchy stress tensor, , being known as the Cauchy
equations:

Dui ij
fi [2]
Dt xj

where f represents the body unit forces acting on the fluid (i.e. gravity acceleration) and x j is the spatial coordinates in
the three dimensions.

Further transformations of the conservation of momentum equation need additional parameters and assumptions.
Rheology theory relates the stress tensor to the velocity field for different materials through the stress-strain law and other

2
constitutive equations. For Newtonian fluids, the stress-strain law is linear. Additionally, for isotropic, homogenous and
Newtonian fluids, the stress tensor is expressed in terms of the velocity field by:

uk ui uj
ij p ij ij
xk xj xi
[3]
where p is the thermodynamic pressure (equal to static pressure, for motionless fluids), ij is the Kronecker symbol, is
the fluid viscosity or the shear viscosity coefficient, and is the second viscosity (for liquids, 2 / 3 ). Inserting the
stress-strain law given in Eq.[3] in the momentum equation (Eq.[2]), yields the general form of the N-S equations for
isotropic and Newtonian fluids:

Dui p uk ui uj
fi [4]
Dt xi xi xk xj xj xi

Eq.[1] is a scalar equation while Eq.[4] is vector equation. For three-dimensional flows described in Cartesian coordinates,
there are four coupled differential equations and four unknowns (u, v, w, p) . For problem completeness, initial and
boundary conditions, in space and time, need to be specified.

For incompressible fluids, for which density is constant in space and time, (x , t ) 0, the continuity equation is reduced
to the divergence-free condition:

uk
0 [5]
xk

Taking into consideration Eq.[5], the N-S equations for an incompressible fluid flow yield:
2
Dui p ui
fi [6]
Dt xi x j2

The N-S equations contain terms that can be neglected in large parts of the flow domain. This allows the equations to be
simplified, and herewith to reduce the effort in solving them. The terms that describe the viscous shear stresses offer such
a possibility for simplification (Veldman, 2011). When considering inviscid fluids, the free-divergence is retained but the
momentum equation changes, resulting in the Euler equations for ideal fluids (i.e., inviscid and incompressible fluids):

Dui p
fi [7]
Dt xi

The N-S equations are strictly a statement of the conservation of momentum. As referred, initial and boundary conditions
are needed to fully describe fluid motion. The final form of the N-S equations, depending on the assumptions and
simplifications made, may be linear, quasi-linear or non-linear, parabolic or hyperbolic in nature. Consequently, the
simplifications made have a major bearing on numerical techniques that can be employed in the model (Denton and
Dawes, 1998). The numerical resolution of the N-S equations is the basis for the CFD simulations.

2.2 Turbulence models

Osborne Reynolds (1883) carried out an experiment to that laminar flow is characterized by smooth streamlines and
highly-ordered motion, while, in turbulent flow, secondary random motions are superimposed on the main flow and
there is an exchange of fluid from one adjacent region to another. This extra transport mechanism of momentum is due
to the random motion of the fluid. Between laminar and turbulent, there is a transitional flow which is not yet well
understood.

Turbulent flows are complex: finer features, called turbulent eddies, arise in all directions in the turbulent flow field; these
eddies are unsteady, three-dimensional, random and swirling. Direct Numerical Simulation (DNS) is a numerical
technique that attempts to resolve the unsteady motion of all the scales of the turbulent flows. The size difference and the
time scale between the largest and the smallest eddies can be of several orders of magnitude, making DNS calculations
complex and time consuming, requiring fine and fully three-dimensional meshes as well as high performance computers
with a considerable amount of Central Processing Unit (CPU). For today’s computers, DNS simulations are not feasible
for practical purposes in engineering. Therefore, simplifications are needed. Most engineering simulations of turbulent
flows are carried out by using Reynolds-Averaged Navier-Stokes (RANS) methods. In RANS methods, the turbulent
fluctuations are time-averaged; however, nothing that occurs on a scale below the size of the mesh is resolved, such as
turbulent eddies. Additional models must be added to the RANS equations in order to take into account the enhanced
mixing and diffusion caused by turbulent eddies (Anderson, 1995).

The RANS equations are time-averaged equations of motion for fluid flow. In order to be able to take time-average, the
instantaneous values (e.g. velocity) are decomposed in two parts, the mean the value and the fluctuating value. With the

3
approach of the eddy viscosity principle after Boussinesq, the general time-averaged N-S equations (or RANS equations)
are given by:

Dui p ui
fi ui u j [8]
Dt xi xj xj

The Reynolds stress tensor or turbulent stresses, , is:

ij ui u j [9]

where the primes denote fluctuating velocity components, u j , and the over bar indicates the time average of the product
of two fluctuating velocity components. Since the Reynolds stress tensor is symmetric, six additional unknowns are
introduced into the problem. These unknowns are modelled in various ways by turbulence models. In laminar flows, the
fluctuations do not exist and ij 0.

RANS equations cannot be directly solved for the mean flow, since the Reynolds stresses are unknown and non-linear,
leading to a closure problem (Fursikov and Imanuvilov, 1999). The equations for turbulence stresses can be derived but
the derivation involves further unknown moments. This continues with equations for moments of a given order
depending on new moments up to a higher order, leading to an infinite system of equations known as the ‘Friedman-
Keller system’. For practical applications, closing the system at some finite order is needed, which is called ‘the closure
problem’. Many semi-empirical formulations to describe Reynolds stresses, in terms of average velocity gradients, are
used to provide mathematical closure to the equations of motion. Such models are called the turbulence models.

Turbulence models can be classified in different types, namely: RANS, Large Eddy Simulation (LES), Detached Eddy
Simulation (DES) and DNS. RANS can be algebraic models (e.g., Cebeci-Smith model, Baldwin-Lomax model, Johnson-
King model, and roughness-dependent model), one-equation models (e.g., Prandtl’s, Baldwin-Barth, Spalart-Almaras,
Rahman-Siikonen-Agarvan), two-equation models such as k - models (e.g., standard, realisable, RNG) and k - models
(e.g., Wilcox’s and Wilcox’s modified, SST) and Reynolds stress models. The most popular are the two-equation models,
k - and k - . These models add two more transport equations, which must be solved simultaneously with the mass
conservation equation and linear momentum equations. Two additional boundary conditions and initial conditions need
to be specified for turbulence properties (Bradshaw and Huang, 1995; Pope, 2001).

The turbulence model used in this research is the realizable k - model. Among the turbulence models available in the
ANSYS Fluent software (e.g., k - , Reynolds stress, SAS, DES, LES), the two equations models are considered as the most
successful for practical engineering purposes, and the one that requires less computational resources for simulating
parietal flows is the realizable k - model (Anderson et al., 2001). The term realizable means that the model satisfies certain
mathematical constraints on the Reynolds stresses, consistent with the physics of turbulent flows. In comparison to the
standard k - model, the realizable k - contains an alternative formulation for the turbulent viscosity, k , and a modified
transport equation for the dissipation rate, , that has been derived from an exact equation for the transport of the mean-
square vorticity fluctuation .

All turbulence k - models use the Boussinesq hypothesis to relate the Reynolds stresses with the mean velocity gradients:

ij 2
uiu j 2 T Sij kS [10]
3 ij

where the turbulent kinetic energy, k , is defined by:

1
k uu [11]
2 i i
and T is the kinetic eddy viscosity assumed as an isotropic scalar quantity, and Sij is the mean strain-rate tensor given by:

1 ui uj
Sij [12]
2 xj xi

The main advantage of using of the Boussinesq hypothesis is the low computational cost associated with the computation
of the turbulent viscosity.

The additional transport equations – one associated to the turbulence kinetic energy, k , and the other to the turbulence
dissipation rate, – specific for the realizable k - model are the following:

t k
( k) ( ku j ) Gk Gb YM Sk [13]
t xj xj k xj

and

4
2
t
( ) ( uj ) C1S C2 C1 C3 Gb S [14]
t xj xj xj k k

when t is the turbulent viscosity computed by C k 2 / and Gk represents the generation of turbulence kinetic
t
energy due to the mean velocity gradients, Gb is the generation of turbulence kinetic energy due to buoyancy, YM
represents the contribution of the fluctuating dilatation in compressible turbulence to the overall dissipation rate,
respectively, Sk and S are user-defined source terms, k and are the turbulent Prandtl number for k and , C 2 , C1
and C 3 are constants, and the remaining parameters are given by C1 max 0.43, /( 5) , S( k / ), S 2SijSij .

The disadvantage of the Boussinesq hypothesis is that it assumes that t is an isotropic scalar quantity, which is not
strictly true. However, this assumption typically works well for shear flows dominated by only one of the turbulent shear
stresses. This covers most flows, such as wall boundary layers, mixing layers and jets.

In the k - models, the k equation is solved in the whole domain including the wall-adjacent cells. The boundary condition
for k imposed at the wall is k / y 0, where y is the distance from wall measured in the radial direction. The
production of kinetic energy , Gk , and its dissipation rate, , at the wall-adjacent cells, which are the source terms in the
k equation, are computed on the basis of the local equilibrium hypothesis. Under this assumption, the production of k
and its dissipation rate are assumed to be equal in the wall-adjacent control volume.

3. Velocity-distribution laws in pipe flows

3.1 Introduction

The theory and applications of turbulent fluid motions were significantly advanced by Ludwig Prandtl in his ‘boundary
layer theory’. Prandtl (1949) showed that the flow around a smooth body could be divided into two regions: i) a region
close to the boundaries, where fluid viscosity governs flow and for which the N-S equations simplify to what it is currently
referred to as ‘the boundary layer’ equations; and ii) a region away from the boundaries, where the fluid behaviour is
practically inviscid, and the fluid velocity can be approximated by a solution of the Euler equations (Monty, 2005). Later,
Prandtl’s students continued his work in fluid mechanics, namely: Paul Blasius, Johann Nikuradse, Hermann Schlichting
and Theodore von Kármán. Blasius (1950) formulated the velocity distribution for the laminar boundary layer.
Nikuradse (1950) conducted extensive pipe flow experiments including investigations in smooth and rough wall pipe
flow and conceived the so-called sand grain roughness (Anderson, 2005). Schlichting (1955) studied turbulent rectangular
duct flow using highly accurate experimental techniques and von Karman (1963) analysed the vortex formation in the
wake of a cylinder, initiating the analysis of fluid turbulence. The mentioned works led to the modern methods for
estimating velocity-distribution in pipes.

Viscous flows in pipes can be either laminar or turbulent. Fluid motion in a laminar flow is dominated by the viscous
forces, while in turbulent flow, the fluid motion is influenced by the viscous (friction) forces near the wall and inertial
forces away from the boundaries. In pipe flows, these conditions are associated with a dimensionless parameter, the
Reynolds number, Re=UD/ , in which U is the mean velocity, D is the inner pipe diameter and is the kinematic
viscosity. Laminar flow occurs for small Reynolds numbers (below 2000) and turbulent flows for larger (typically, larger
than 4000), in-between, a transitional zone exists in which the flow does not have purely laminar or turbulent
characteristics. Depending on the flow characteristics (laminar or turbulent), different velocity-distribution laws through
pipes have been developed, namely, exact solutions for laminar flows and empirical or semi-empirical for turbulent flows.
All solutions consider that the boundary condition at the wall is described by the ‘no-slip condition’, that is the relative
velocity between the wall and the fluid is set to zero. Solutions for pipe flows are presented in the next sections.

3.2 Exact solution for laminar flow in pipes

Hagen-Poiseuille flow is a laminar viscous stationary flow through a constant cross-section pipe (with Re 2000) . The
equations governing the Hagen–Poiseuille flow can be directly derived from the N-S equations, in cylindrical coordinates,
by considering the following assumptions: i) the flow is steady; ii) the radial and circumferential components of the fluid
velocity are null; and iii) velocity distribution is axisymmetric and fully developed (Acheson, 1990; Çengel et al., 2012;
White, 2003). For these conditions, the axial component of velocity distribution along the pipe radius, u , is described by:

r2
u(r ) 2U 1 [15]
R2

and the corresponding viscous shear stress, , is given by:

du r dp
(r ) [16]
dr 2 dx

5
dp
where U is the mean velocity, r is the radial distance from the pipe axis, R is the pipe radius, the viscosity, is the
dx
pressure gradient along the pipe axis in x direction. Note that the velocity profile is parabolic.

3.3 Semi-empirical solutions for turbulent flow

Most flows in engineering practice are turbulent, and thus it is important to understand how turbulence affects wall shear
stress. Unlike laminar flow for which the theory is nowadays well-developed, the expressions for the velocity profile in a
turbulent flow are based on both theoretical analysis and measurements, and thus these have a semi-empirical nature
with constants determined from experimental data.

For a fully developed turbulent pipe flow, the velocity profile is axisymmetric and with an almost uniform-shape with a
sharp drop near the pipe wall. Turbulent flow can be divided in four regions, characterized by the distance from the wall.
A thin layer next to the wall where viscous effects are dominant is the viscous sublayer; the velocity profile in this layer is
nearly linear, and the flow is streamlined, similar to laminar flow. Next to the viscous sublayer is the buffer layer, for which
turbulent effects become significant, but the flow is still dominated by viscous effects. Above the buffer layer is the overlap
layer, also called the inertial sublayer, in which the turbulent effects are much more significant, but still not dominant. The
turbulent layer occupies the remaining part of the flow, in which turbulent effects dominate over viscous effects
(Çengel et al., 2012).

Flow characteristics are quite different in each of the regions, being difficult to find an analytic expression for the velocity
profile for the entire pipe cross-section, as for laminar flow. The alternative is to identify the key parameters and functional
forms using dimensional analysis, and to use experimental data to determine the values of those parameters.

The thickness of the viscous sublayer is very small (


v 5 / v* ), though large velocity gradients are observed. The wall
dampens any eddy motion and the flow behaves as in laminar conditions where shear stress is proportional to the fluid
viscosity (as in Eq.[16]). Considering that velocity changes from zero to nearly the core region value, the velocity profile
is almost linear, as confirmed by the experiments, and the velocity gradient is almost constant, equal to du / dy u /y .
Accordingly, wall shear stress can be expressed by:

du u u w u
w [17]
dy y y y
in which u is axial velocity and y is the distance from the wall ( y R r , for a circular pipe). By defining the friction
velocity by u* w / and by replacing it in the previous equation, the velocity profile in the viscous sublayer is obtained
and described in terms of dimensionless parameters u and y as follows:

u u yu*
u*2 u y [18]
y u*

in which u u / u* and y yu* / . This equation is known as the ‘law of the wall’ and, for smooth-wall pipes, applies
for y 5 (as shown in Figure 1). In this region, the inertia terms and the turbulent fluctuations are negligible.

In the turbulent layer, the experimental data have shown that the velocity is proportional to the logarithm of distance and
the velocity profile is described by:

1
u ln y B
[19]
where and B are constants whose values are experimentally determined, being typically 0.40 and 5.0, respectively. This
is known as the ‘logarithmic-law’, whose width is Reynolds number dependent and typically lies in the range of
30 y 350 . In this region, the turbulent shear stress dominates over the laminar shear stress.

Numerous other empirical velocity profiles exist for describing velocity profiles in turbulent pipe flow. The simplest and
the most well-known is the ‘power-law’ velocity profile (as shown in Figure 1):
1/n 1/n
u y u r
1
umax R umax R
[20]
where n is a Reynolds number-dependent constant given by the following equation:

U 2n2
[21]
umax (n 1)(2n 1)

According to Eq.[20], velocity profile becomes flatter with the increasing Re (Schlichting, 1955).

6
In the interim region between the viscous sublayer and the turbulent layer, where the effects of the viscosity and turbulent
are equally important – the buffer layer for 5 y 30 – velocity distribution, cannot be accurately described by none of
the previous equations – Eq.[18] and Eqs.[19]-[20] (cf. Figure 1):

Viscous sublayer Buffer layer Turbulent layer

20

16
Eq.[20]

12
Eq.[18]
+

8
u

4 Velocity distribution law


CFD

10-1 100 101 102 103


+
y

Figure 1. Law of the wall

4. CFD simulations

4.1 Main stages

CFD simulations can be divided in three main stages: pre-processing; simulation; and post-processing. The pre-processing
stage includes: geometry definition, mesh generation and definition of the boundary conditions and solver parameters.
The simulation stage consists of using the constructed model for running the simulations. In the post-processing stage,
obtained results are analysed and visualized in different forms (e.g., x -y graphics, contours, velocity vectors and stream
lines). The main procedures of the pre-processing stage are described in the following paragraphs.

4.2 Geometry definition and mesh generation

The geometry definition consists of specifying the shape of the physical boundaries of the fluid. A CAD software is used
to draw a cylindrical pipe in 3-D. This cylindrical shape has 4 m length and 0.02 m of pipe inner diameter. According to
Çengel et al. (2012), the 4 m pipe length, that corresponds to 200 times of the pipe inner diameter, ensures the establishment
of the fully developed flow in both laminar (> 70 times the diameter) and turbulent flows (>10 times the diameter).

After geometry definition, CFD simulation requires the division of the fluid domain into a number of smaller, non-
overlapping sub-domains resulting in a mesh of cells. The equations that describe flow properties and motion are
numerically solved in each of the defined mesh cells. The final solution accuracy is strongly dependent on the number of
cells in the mesh within computational domain (Tu et al., 2008). A larger number of cells generally lead to a more accurate
solution, but it can be more restrictive to model larger systems, since it needs more computation resources that are not
always available. The capacity of a computer, or even a cluster, can be easily overcome with a large number of cells in a
mesh. In this paper, being the mesh generation one of the most important steps of the pre-processing stage, a mesh
independence analysis is carried out to obtain the most efficient meshes that correspond to the compromise between the
maximum accuracy and the minimum computational effort.

A mesh generator software, included in the ANSYS package, is used. The pipe mesh is described by three parameters:
‘Sweep’, the mesh size in the axial direction of the pipe; ‘Mesh Maximum Size’ (MMS), the mesh size in circumferential
direction in the pipe wall; and ‘First Layer Thickness’ (FLT), the mesh size in the layer near wall in the radial direction
(the mesh increases from the wall to the pipe axis with a geometric growth rate equal to 1.24), as shown in Figure 2. For
each combination, a mesh is generated and solved by ANSYS Fluent for the same initial conditions and for steady state
conditions.

7
Figure 2. Parameters definition.

4.3 Boundary conditions and viscous models

Appropriate boundary conditions that describe the physics of the fluid flow need to be defined in the CFD solver. It is
considered that the pipe has an inlet and an outlet as shown in Figure 3. The inlet velocity is 0.070 m/s for laminar flow
(Re=1393) and 0.354 m/s for turbulent flow (Re=7040). The outlet pressure is of 190 kPa for both flows (i.e., laminar and
turbulent). Pipe walls are considered as hydraulically smooth so the boundary condition at the wall is described by the
‘no-slip condition’, that is the relative velocity between the wall and the fluid is set to zero.

Velocity=0
Wall Pressure=190 000 Pa

Flow direction
Inlet Outlet

Wall
Velocity(laminar)=0.070 m/s
Velocity=0
Velocity(turbulent)=0.354 m/s

Figure 3. Boundary conditions.


For laminar flows, Eq.[16] is used to relate velocity with the shear stress and, for turbulent flows, the realizable k - model
is used to describe turbulence.

4.4 Solver parameters definition

CFD simulations considered are run for laminar and turbulent flows in steady state, considering that fluid is viscous,
incompressible and monophasic and that the fluid motion is an isothermal process (without heat transfer). The numerical
technique used by ANSYS Fluent is the finite volume method. The flow solver uses the Semi-Implicit Method for Pressure-
Linked Equations (SIMPLE) algorithm, which is a pressure-based solver considering the relationship between velocity
and pressure corrections to carry out mass conservation and to obtain the pressure field.

Convergence can be assess by progressively track the imbalances that are accentuated by the advancement of numerical
calculations of algebraic equations, through each iteration step. These imbalances measure the overall conservation of
flow properties: commonly known as residuals. A converge solution is achieved when the residuals fall below the
convergence criterion or tolerance that is pre-set inside the solver iterative parameters. The convergence criterion (relative
error of the calculated parameters), due to the flow characteristics, is equal to 10-6.

5. The proposed approach

A systematic and comprehensive approach is proposed for obtaining the most efficient meshes when using CFD for
describing flow conditions in any engineering situation. The approach is applied to pressurised water pipe flow in steady
state conditions. The most efficient meshes are the ones that meet two contradictory criteria: the maximum accuracy (i.e.,
the minimum error) and the minimum computational effort (i.e., the minimum computational time). The proposed
approach consists of a four-step procedure as depicted in Figure 4.

8
Figure 4. Proposed comprehensive approach for obtaining the most efficient mesh.

Step 1 consists of the definition of the 3-D fluid domain geometry. This can be carried out using a CAD software. Different
meshes are generated using a software for different combinations of the main parameters that describe the fluid geometry.
The range of values for each parameter should be defined. Each parameter should be combined with values from other
parameters and, for each combination, a mesh is generated. If necessary, other combinations of parameters could be added
to further complementing the analysis.

For the particular case of pressurized pipe flow, three dimensionless mesh parameters have been defined:

 a , mesh size in the axial direction of the pipe described in terms of number of diameters;

 c , mesh size in the circumferential direction described in terms of the perimeter; and

 r , ratio between the FLT and the thickness of the viscous sublayer1.

The considered range for each dimensionless mesh parameter used in the simulation is presented in Table 1.

Each parameter was combined with the lowest value of the other two parameters: thirty-eight meshes were generated for
laminar flow and ninety-two for turbulent flow.
Table 1. Range of values of dimensionless mesh parameters

Range of values
Dimensionless mesh parameters
Laminar flow Turbulent flow

a = mesh size in the axial direction described in terms of number of


[0.10; 5.50]
diameters, a Sweep / D

c = mesh size in the circumferential direction described in terms of the


[0.016; 0.041]
perimeter, c MMS / D

r = ratio between the first layer thickness (FLT) and the thickness of the
[0.0054; 0.1084] [0.0005; 0.1626]
viscous sublayer for turbulent flow in the radial direction, r FLT / yy
5

Note: a=axial; c=circumferential; r=radial.

Step 2 consists of running the fluid flow simulation by using the CFD software until specified convergence criteria is met
(see section 4.4), for each generated mesh and for the solver options considered (see section 4.3). In the current case, the
procedure is repeated for two inlet boundary conditions associated to laminar and turbulent (as explained in section 4.3):
the inlet velocity of 0.07 m/s for laminar flow and 0.354 m/s for turbulent flow. Within the range of considered values, a

1 Although the viscous layer in not defined for laminar flow, the thickness of the viscous sublayer calculated for the turbulent flow analysed
was used in the definition of this dimensionless parameter (i.e., for mean velocity = 0.354 m/s).

9
step variation is considered in order to cover the whole range. A total of 130 simulations are run (38 for laminar and 92
for turbulent flows), considered sufficient to satisfy the previous assumption in the current case.

Step 3 consists of the evaluation of the performance of used meshes. This can be carried out by comparing obtained
numerical results with reliable collected data or with exact or empirical solutions obtained by other authors. Errors should
be computed: these can be absolute errors, quadratic errors or weighed errors, depending on the nature of the problem.
In the current case, obtained velocity profiles are compared to the exact theoretical solution for laminar flows (the axi-
symmetric velocity-distribution by Hagen- Poiseuille formulation) and to semi-empirical solution for turbulent flows (i.e.,
the linear law for the viscous sublayer and the power-law in turbulent layer).

Additionally, two accuracy indices are computed and applied to a set of cylinders centred in the pipe axis and equally
spaced (10-3) along the pipe radius (a total 1000 cylinders). These cylinders are created due to the irregularity of the pipe
mesh near the pipe axis to allow the comparison of results for meshes with different dimensions.

The first index is the Relative Error ( RE) used to evaluate the accuracy of each value of the velocity in the cross-section:

Xn Yn
REn
Xn
[22]
in which REn is the Relative Error at cylinder n, Xn is the exact or semi-empirical value of the velocity profile at cylinder
n (m/s), Yn is the calculated value of the velocity profile by CFD numerical modelling at cylinder n (m/s).

The second index is the Nash-Sutcliffe coefficient ( NSE) which is a dimensionless parameter that determines the relative
magnitude of the residual variance compared to the measured (or reference) data variance (Nash and Sutcliffe, 1970).
NSE is used to evaluate different models, in terms of accuracy of simulated flow compared to measured flow (Falcão et
al., 2013; Moriasi et al., 2007). This index is calculated weighting the square differences of the parameter (i.e., velocity)
with the corresponding cross-sectional area of the cylinder:
N
(Xn Yn )2 * An
n 1
NSE 1 N
(Xn X )2 * An
n 1 [23]
in which, NSE is the dimensionless average square error (-), X is the mean value of the exact or semi-empirical velocity
profile (m/s), N is total number of cylinders (N=1000), An is the cross-sectional area of the cylinder n (m²). It should be
highlighted that the cylinders may not correspond to the actual CFD cells defined in the radial direction. In this case,
interpolations are necessary to calculate the velocity values in the intermediate cylinders.

NSE ranges between and 1 , being the desirable value NSE=1 (i.e., the maximum value). In the current case, 1-NSE
index is used with the aim of determining the Pareto frontier for the two criteria for obtaining the most efficient mesh, in
which the aim is the minimization of both criteria (i.e., number of mesh nodes and 1-NSE). 1-NSE ranges between
0 and + , being 1-NSE=0 the best value.

For laminar flows, 1-NSE index is calculated for all cylinders, describing the global simulation accuracy. For turbulent
flows, results of CFD in buffer region will be neglected, as neither Eq.s[18] nor [20] are accurate enough to describe
velocity profile in this region. Thus, error analysis for determining the most efficient meshes will focus only on the viscous
sublayer in which the law of the wall is applied ( y 5) described by the Eq.[18], and the turbulent layer described by the
power-law (y 30) by Eq.[20]. For this reason, a global index value should be computed for assessing the accuracy of
each simulation. This index is computed by the weighted sum of the 1-NSE index in both regions
(0 y 5 and y 30) by the following equation:

1-NSE global w (1-NSE)y 5 (1 w) (1-NSE)y 30


[24]
in which w stands for the weight attributed to the viscous sublayer ( y 5) . A sensitivity analysis of the weight will be
carried out to assess to which extent the most efficient meshes are affected by the accurate calculations in the viscous
sublayer or in the turbulent layer.

The previous three steps are repeated for different mesh configurations (parameter configuration) and flow regimes
(laminar and turbulent).

Finally, Step 4 consists of obtaining the most efficient meshes by the compromise between the maximum accuracy
(evaluated by 1-NSE) and the minimum computational effort (evaluated by the number of CFD nodes). In this study, the
number of nodes is used as a mean of comparing the required computer time (computational effort), as these two items
are directly related. The higher the number of nodes is, the higher the corresponding simulation time becomes.

10
It should be highlighted that the proposed methodology does not use any optimization algorithm, because: i) each
simulation is very time consuming and using an optimization algorithm would require a huge amount of computational
time, or the solution would be stuck on a local minimum, if a gradient method were used; and ii) the aim of the current
research is to identify the most efficient mesh within a certain range of parameter values and not the global optimal mesh.

As the most efficient meshes are different for laminar and for turbulent flows, results will be presented and discussed
separately.

6. Results and discussion

6.1 Laminar flow results

The velocity-distribution computed by Hagen-Poiseuille law – Eq.[15] – is compared with the calculated values by CFD
modelling. Thirty-eight simulations have been carried out. RE performance index is calculated using CFD results

Velocity profiles computed by CFD and obtained by the exact solution are presented in Figure 5a) for the most accurate
simulation (the lowest 1-NSE). Hardly any differences between the two profiles can be observed. The relative errors for
each cross-section radius (assuming axi-symmetry of the velocity profile) are presented in the scatter plot in Figure 5b)
for the analysed parameter combinations. Relative errors are below 1% for all simulations. Higher dispersion of RE values
is observed, when varying the c parameter (circle markers in Figure 5b). On the overall, the following conclusions can
be drawn:

 CFD velocity profiles obtained for all simulations are in a good agreement with the exact solution for laminar
flow, for any combination of dimensionless parameters.

 c parameter is of the utmost importance for the overall accuracy of the model, as the lowest value of c lead
to the lowest relative error.

 Good agreements near the wall and in the core region of the flow are observed for all flows, occurring the
largest relative errors for r / R 0.4 , with no apparent physical justification.

a) b)

Figure 5. Laminar flow results: (a) Computed velocity profiles and the exact solution by Eq.[15] for the most accurate
simulation. (b) RE indices for a , c and r parameters variation in the cross-section.

Results of the combination of three mesh parameters for laminar flow for the 38 simulations have been ordered by the
decreasing number of nodes and presented in Figure 6, in terms of value the number of nodes (secondary vertical axis)
an the corresponding 1-NSE value (main vertical axis) as a function of the dimensionless parameters. A sensitivity analysis
is carried out by varying each dimensionless parameter and maintaining the other two constants and equal to the
minimum values.

The mesh parameter a (mesh size in the axial direction expressed in terms of number of diameters) is varied between
0.1 and 5.5 and has been combined with the lowest values of c (0.016) and r (0.0054), as depicted in Figure 6a). Results
have shown that: as the number of nodes of the mesh decreases from 5.84×106 to 0.11×106, the index 1-NSE has almost no
variation; and the highest 1-NSE value (4.72×10-5) was obtained for the smallest number of nodes (108 040).

11
Results of the variation of the r dimensionless parameter (size of the mesh in radial direction) between 0.0054 and 0.1084,
combined with the lowest values of a and c , are presented in Figure 6c). The increase of the dimensionless parameter
r results on the increase the 1-NSE index (from 4.57×10-5 to 8.36×10-5) and, thus, the lowest accuracy of the mesh,
corresponds to the lowest number of nodes.

Figure 6b) shows the variation of the mesh parameter c combined with the lowest values of a 0.10 and r 0.0054 . In
this case, the increase of the mesh parameter c , from 0.016 to 0.041, results on a decrease of the number of nodes and on
an increase in the 1-NSE index (i.e., lower accuracy). The highest 1-NSE value (1.47×10-4) was obtained for the smallest
number of nodes (2 553 276).

(a) (b)

c)

Figure 6. Laminar flow results in terms of 1-NSE index and number of nodes for the dimensionless parameters
variation: a) a [0.10 : 5.50] , c 0.016 , r 0.0054 ; b) a 0.10 , c [0.016 : 0.041] , r 0.0054 ; c) a 0.10,
c 0.016 , r [0.0054 : 0.1084].

Results of combinations of parameters are plotted in Figure 7 in terms of the 1-NSE and the number of nodes. Figure 7 a)
presents the results for the 38 simulations ordered by the decreasing number of nodes. Additionally, as the number of
nodes decreases from 5.84×106 to 0.11×106 (i.e., 50 times), 1-NSE index only increases 3.22 times (1.47×10-4 to 4.57×10-5).
This figure shows that high accuracies (low 1-NSE index values) occur for both high and low number of nodes, depending
on the combination of parameters. However, results accuracy are highly dependent on the size of the first layer thickness
(FLT), being the best results obtained for the lowest values of this parameter. This is because the radial mesh is defined
as a function of this parameter and it is of the utmost importance to have very thin layers close to the pipe wall where the
velocity gradients are higher. The thinner the mesh close to the wall is, the most accurate the numerical solutions obtained
are.

Figure 7b) plots the values of the 1-NSE index versus the number of nodes. The imaginary lower line that limits the
obtained pool of solutions (called non-dominated solutions) is known as the Pareto-frontier and corresponds to the set of
the most efficient meshes analysed (Ngatchou et al., 2005). These meshes correspond to the ones that meet two
contradictory objectives: the maximum accuracy and the minimum computational effort (i.e., computational time). A set
of ‘efficient meshes’ are circled in Figures 7a) and 7b) that correspond to the solutions with the smallest number of nodes
(less than 0.5 million) and the highest accuracy (1-NSE index below 5×10-5).

12
Question remains of which meshes should be chosen. A selection criterion has been established: the best solutions are the
ones that correspond to the shortest distance of the Pareto frontier to the minimum of 1-NSE index and minimum number
of nodes of laminar flow simulations. The most efficient meshes are presented in Table 2 and highlighted in Figures 7c)
and 7d). These meshes have the same parameter values c 0.016 and r 0.0054 , being the a parameter varying
between the maximum values analysed, 5.5 and 4.
a) b)

Detail c) Detail d)

Figure 7. The most efficient mesh for laminar flow: a) Number of nodes and 1-NSE index versus simulation; b) number
of nodes versus 1-NSE index; c) and d) numbered by ranking and most efficient meshes identify.

Table 2. Most efficient meshes (Laminar flow)

Ranking a c r Nodes 1-NSE

1 5.2 113 880 4.679×10-5


2 4.0 146 000 4.686×10-5
0.016 0.0054
3 4.6 128 480 4.717×10-5
4 4.9 119 720 4.720×10-5

The four most efficient meshes for laminar flow are quite similar, being described by c 0.016 , r 0.0054 and a
between 4.0 and 5.2, and correspond to a number of nodes between 110-150 thousand and 1-NSE index below 1%. These
meshes have the lowest values of the dimensionless parameters in the radial ( r ) and circumferential ( c ) directions,
showing that the accuracy is very much dependent of the mesh definition in these directions. The most accurate solutions
have the thinnest mesh near the pipe wall where the velocity gradients are higher and where derivatives should be more
accurately calculated. This result is independent of the mesh definition in the axial direction (as a has quite high values
in the range defined).

6.2 Turbulent flow results

The velocity-distribution in the viscous sublayer ( 0


y 5 ) computed by the law of the wall – Eq.[18] – is compared with
the predicted values of the CFD modelling, as presented in Figure 8a) for one of the meshes generated. CFD velocity

13
profiles ensure a reasonably good agreement with Eq.[18]. Ninety-two simulations have been carried out. Results of the
RE performance index are plotted in Figure 8b) for all parameter combinations ( a , c , r ). RE values vary between 0
and 0.22 depending on the radius and on the combination of parameters. The range of RE values obtained for the r
variation are significantly higher than the ones obtained for the other parameters variation for the whole viscous sublayer,
indicating that this parameter significantly affects the accuracy of the results. The largest RE values occur for r 0.982
and 0.988 (i.e. the cylinder close to the wall). For each simulation, the average RE value varies between 0.07 and 0.15.
a) b)

Figure 8. Results for the turbulent flow in the viscous sublayer ( 0 y 5 ): a) velocity profiles for r 0.0054 ,
c 0.016 and a 0.10 and b) RE values for different dimensionless mesh parameters combinations ( a , c, r )

The velocity-distribution in the turbulent layer y 30 computed by the power law – Eq.[20] – is compared with the
predicted values of the CFD modelling, as presented in Figure 9a). RE performance index is used to evaluate in terms of
accuracy of the simulated data by CFD to Eq.[20]. Results are plotted in the cross section and presented in Figure 9b) for
all parameter combinations ( a , c , r ). The r dimensionless mesh parameter variation is responsible for the largest
range of the group of RE values along the cross section. There is a good agreement in the core of the fluid flow for all
parameters combination, as the RE is below 0.06 for all the mesh parameters combination.
a) b)

Figure 9. Results for the turbulent flow in the turbulent layer ( y 30 ): a) velocity profile for r 0.0054 , c 0.016
and a 0.1 and b) RE index for different parameters combinations ( a , c , r ).

Results of CFD modelling in buffer layer have been neglected and 1-NSE index has been computed separately for the viscous
sublayer and the turbulent layer. A global index described by Eq. [24] has been calculated, to obtain a single error for meshed
comparison. A sensitivity analysis of the weight attributed has been carried out considering different scenarios: Scenario
A ( wA 5.44%) , Scenario B ( wB 25%) , Scenario C ( wC 50%) and Scenario D ( wD 75%) . The lowest weight

14
( wA 5.44%) corresponds to the ratio of the cross-sectional area of the viscous sublayer and the total cross-sectional area
of both layers.

The values of the 1-NSEGlobal index versus the number of nodes for the different weight combinations are plotted in Figure
10a). The most efficient meshes hardly change with the chosen scenario. A detail of the results for the four most efficient
meshes are presented in detail in Figure 10 b) and corresponding parameters combination in Table 3. The most efficient
mesh presents the lowest value of r (0.0005) considered in this analysis and medium values of c and a dimensionless
parameters. The next three most efficient meshes have low and constant values of c and r , 0.016 and 0.0016 respectively,
and a variation between 1.6 and 2 of a parameter. This situation was also observed for the laminar flow.
a) Detail b)

Figure 10. Global results for the turbulent flow for Scenarios A, B, C and D: a) number of nodes versus 1-NSEGlobal index;
b) most efficient meshes identification.

Table 3. Efficient meshes (Turbulent flow)

(1-NSE)Global (1-NSE)Global (1-NSE)Global (1-NSE)Global


Ranking a c r Nodes
Scenario A Scenario B Scenario C Scenario D

1 1.0 0.016 0.0005 266 800 7.608×10-2 6.050×10-2 4.059×10-2 2.068×10-2


2 2.0 141 800 9.782×10-2 7.775×10-2 5.211×10-2 2.647×10-2
3 1.6 0.032 0.0016 177 250 9.824×10-2 7.809×10-2 5.234×10-2 2.658×10-2
4 1.7 167 324 9.839×10-2 7.821×10-2 5.242×10-2 2.662×10-2

The most efficient mesh for turbulent flow is described by c 0.016 , r 0.0005 and a 3.5 and 5.2, which
correspond to 266 800 nodes and 1-NSEGlobal index from 7.608×10-2 to 2.068×10-2. Once again, this mesh has the lowest
values defined in the radial and circumferential directions, due to the same reason as stated before for laminar flows: the
highest velocity gradients are observed close to the pipe wall, being more accurately computed the thinner the layers near
the wall are.

7. Conclusions

A larger number of cells generally leads to more accurate solutions, though it might be more restrictive to model larger
systems, since more computation resources are needed. The capacity of a computer, or even a computer cluster, can be
easily overcome with a large number of cells or nodes in a mesh. To overcome this difficulty, a systematic and
comprehensive approach for obtaining the most efficient meshes, described in terms of dimensionless parameters, for
modelling water flows has been proposed and applied to flows in pressurized pipes. The efficient mesh is the one that
allows a balance between the maximum accuracy and the minimum computational effort (less number of nodes). Three
dimensionless mesh generation parameters ( a , c , r ) have been defined to describe pipe geometry. A range of values
was defined for the three parameters and each parameter value was combined with the lowest value of the other two
parameters. Several meshes were defined and used for simulating laminar and turbulent flows. Thirty eight meshes were
defined for laminar flow and 92 for turbulent flow. Obtained velocity profiles were compared with the exact Hagen-
Poiseuille solution for laminar flow and semi-empirical solution for turbulent flow. Relative and global errors were

15
computed for each mesh. The most efficient meshes were obtained for the minimum error and the minimum number of
nodes, described by a ‘Pareto frontier’.

Results obtained include two sets of the most efficient meshes: four meshes for laminar flow and another four for turbulent
flow. The best ranked meshes (Rank 1) for laminar and turbulent flows have the same value of parameter in the
circumferential, c 0.016 (the minimum) and the minimum values considered for each case for the radial direction
( r 0.0054 and
r 0.0005 , for laminar and turbulent flows, respectively); these meshes require less than 10% of the
maximum number of nodes analysed (5.84×106) and have a global error 0.1% higher than the lowest obtained. These
meshes corresponds to the lowest analysed values of c and r considered, emphasising the importance of defining
thinner meshes near boundaries where velocity gradients are higher (i.e., near the pipe wall). Should other smaller meshes
had been defined in these directions, most probably these would have been the best ranked ones, through with
significantly higher computational time.

This study is the first step of the on-going research at Instituto Superior Técnico about the “Frictional and mechanical
dissipation of energy in transient pipe flows”, and it aims at using CFD for simulating hydraulic transients in pipes, for
which the accuracy of the model should be maximum and the number of nodes minimum. In transient flows, velocity
profile inverts and the momentum dissipation cannot be described by steady state formulations that assume a uniform
velocity distribution. CFD modelling will contribute for a better understanding of this phenomenon, as well as of the
mechanisms of momentum dissipation in transient flows both due to wall shear stress and to fluid turbulence.

The next step in this research is to model a longer straight pipe system using the efficient mesh parameters for steady
state conditions and, later, for transient state simulations. At the later stage, the analysis will be extended to a long coiled
pipe.

Acknowledgements

The authors would like to thankfully acknowledge Fundação para a Ciência e Tecnologia (FCT) for the project
PTDC/ECM/112868/2009 “Friction and mechanical energy dissipation in pressurized transient flows: conceptual and
experimental analysis” for funding the current research in terms of experimental work and grants hold.

References

Ansys Fluent, ANSYS Fluent Theory Guide, Release-14.5 (2012). in.


Acheson DJ (1990) Elementary fluid dynamics, Oxford University Press
Ahmad NE, Abo-Serie EF and Gaylard A (2010) Mesh Optimization for Ground Vehicle Aerodynamics. CFD
Letters 2 54-65
Anderson JD (1995) Computational Fluid Dynamics - The Basics with Applications, McGraw-Hill
Anderson JD (2005) Ludwig Prandtl's Boundary Layer. Phisics Today 42-48
Anderson WK, Newman JC, Whitfield DL and Nielsen EJ (2001) Sensitivity Analysis for Navier-Stokes
Equations on Unstructured Meshes Using Complex Variables. {AIAA} Journal 39
Andersson AG, Lundström K, Andreasson P and Lundström TS (2010) Simulation of free surface flow in a
spillway with the rigif lid and volume of fluid methods and validation in a scale model. in Pereira JCF and
A Sequeiras (eds), V European Conference on Computational Fluid Dynamics ECCOMAS CFD 2010, Lisbon,
Portugal.
Baranya S and Józsa J (2006) Flow analysis in river Danube by field measurement and 3D CFD turbulence
modelling. Periodica Polytechnica Ser. Civ. Eng. 50 (1) 55-68
Blasius H (1950) The Boundary Layers in Fluid with Little Friction.
Bradshaw P and Huang GP (1995) The Law of the Wall in Turbulent Flow. Proceedings: Mathematical and
Physical Sciences 451 (1941) pp. 165-188 DOI: 10.1098/rspa.1995.0122.
Carija Z, Mrsa Z and Fucak S (2008) Validation of Francis Water Turbine CFD Simulations. Strojarstvo 50 (1)
5-14
Çengel YA, Cimbala JM and Turner RH (2012) Fundamentals of Thermo-Fluid Sciences, McGraw-Hill Higher
Education
Chanel PG and Doering JC (2008) Assessment of spillway modeling using computational fluid dynamics.
Can J Civil Eng 35 (12) 1481-1485 DOI: 10.1139/l08-094.
Cheng Y, Li J and Yang J (2007) Free surface-pressurized flow in ceiling-sloping tailrace tunnel of
hydropower plant: simulation by VOF model. Journal of Hydraulic Research 45 (1) 88-99
Date A and Akbarzadeh A (2010) Design and analysis of a split reaction water turbine. Renewable Energy 35
(9) 1947-1955
Denton JD and Dawes WN (1998) Computational fluid dynamics for turbomachinery design. Proceedings of
the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science 213 (2) 107-124

16
Falcão AP, Mazzolari A, Gonçalves AB, Araújo MAVC and Trigo-Teixeira A (2013) Influence of elevation
modelling on hydrodynamic simulations of a tidally-dominated estuary. Journal of Hydrology (0) 152-164
Fischer-Antze T, Olsen NRB and Gutknecht D (2008) Three-dimensional CFD modeling of morphological
bed changes in the Danube River. Water Resour Res 44 (9) DOI: 10.1029/2007wr006402.
Foias C, Manley O, Rosa R and Temam R (2001) Navier-Stokes Equations and Turbulence, Cambridge
University Press
Fursikov AV and Imanuvilov OY (1999) Exact controllability of the Navier-Stokes and Boussinesq equations.
Russian Mathematical Surveys 54 (3) 565
Gidaspow D, Li F and Huang J (2013) A CFD simulator for multiphase flow in reservoirs and pipes. Powder
Technol 242 2-12 DOI: 10.1016/j.powtec.2013.01.047.
Harlow FH and Welch JE (1965) Numerical Calculation of Time-Dependent Viscous Incompressible Flow of
Fluid with Free Surface. The Physics of Fluids 8 (12) 2182-2189
Hirt CW and Nichols BD (1981) Volume of Fluid (VOF) Method for the Dynamics of Free Boundaries. Journal
of Computational Physics 39 201-225
Huggins DL, Piedrahita RH and Rumsey T (2004) Analysis of sediment transport modeling using
computational fluid dynamics (CFD) for aquaculture raceways. Aquacult Eng 31 (3-4) 277-293 DOI:
10.1016/j.aquaeng.2004.05.007.
Karman TV (1963) Aerodynamics, McGraw-Hill Inc.,US, ISBN 007067602X.
Knopp T (Year) Model-consistent universal wall-functions for RANS turbulence modelling. (2006).
Lavelli A, Boillat J-L and Cesare GD (2002) Numerical 3D modelling of the vertical mass exchange induced
by turbidity currents in Lake Lugano (Switzerland). in The Fifth International Conference on Hydroscience and
Engineering, Warsaw, Poland.
Lenzi MA, Mao L and Comiti F (2006) Effective discharge for sediment transport in a mountain river:
Computational approaches and geomorphic effectiveness. Journal of Hydrology 326 (1-4) 257-276 DOI:
10.1016/j.jhydrol.2005.10.031.
Li H and Vasquez SA (2012) Numerical simulation of steady and unsteady compressible multiphase flows.
in The ASME International Mechanical Engineering Congress and Exposition IMECE2012, Houston, Texas, USA.
Li J, Liu F and Yang J (2010) Optimization and analysis of flow characteristics in a pratical spillway design
based on the VOF model. in Pereira JF and A Sequeiras (eds), V European Conference on Computational Fluid
Dynamics ECCOMAS CFD 2010, Lisbon, Portugal.
Li J, Wu P and Yang J (2010) CFD numerical simulation of water hammer in pipeline based on the Navier-
Stokes Equation. in V European Conference on Computational Fluid Dynamics ECCOMAS CFD 2010, Lisbon,
Portugal.
Li YH and Xie LQ (2012) 2-D CFD Model for Free Surface River Flow with Tilt Rigid Leave of Submerged
Vegetation. Appl Mech Mater 212-213 332-335 DOI: DOI 10.4028/www.scientific.net/AMM.212-213.332.
Martins SC, Martins NM, Ramos HM and Almeida AB (2012) Liquid flow and entrapped air behaviours in
an experimental set-up using CFD analysis. in Pressure Surge 11, Lisbon, Portugal, pp. 505-516.
Monty JP (2005) Developments in smooth wall turbulent duct flows.
Moriasi D, Arnold J and Van M (2007) Model evaluation guidelines for systematic quantification of accuracy
in watershed simulations.
Mudgal BV and Pani BS (1998) Flow around obstacles in plane turbulent wall jets. J Wind Eng Ind Aerod 73
(3) 193-213 DOI: Doi 10.1016/S0167-6105(97)00288-2.
Nash JE and Sutcliffe JV (1970) River flow forecasting through conceptual models part I — A discussion of
principles. Journal of Hydrology 10 (3) 282 - 290
Ngatchou P, Zarei A and El-Sharkawi MA (Year) Pareto Multi Objective Optimization. (2005), 84-91.
Nikuradse J (1950) Laws of Flow in Rough Pipes. 1292.
Nilsson H and Davidson L (2003) Validations of CFD against detailed velocity and pressure measurements
in water turbine runner flow. International Journal for Numerical Methods in Fluids 41 (8) 863-879 DOI:
10.1002/Fld.472.
Olsen NRB (1999) Two-dimensional numerical modelling of flushing processes in water reservoirs. Journal
of Hydraulic Research 37 (1) 3-16
Olsen NRB (2000) CFD Algorithms for Hydraulic Engineering, Department of Hydraulic and Environmental
Engineering The Norwegian University of Science and Technology
Olsen NRB, Hedger RD and George DG (2000) 3D Numerical Modeling of MICROCYSTIS Distribution in a
Water Reservoir. Journal of Environmental Engineering 126 (10) 949-953
Olsen NRB and Kjellesvig HM (1998) Three-dimensional numerical flow modelling for estimation of
spillway capacity. Journal of Hydraulic Research 36 (5) 775-784
Olsen NRB and Skoglund M (1994) Three-dimensional numerical modeling of water and sediment flow in
a sand trap. Journal of Hydraulic Research 32 (6) 833-844
Olsen NRB and Stokseth S (1995) Three-dimensional numerical modelling of water flow in a river with large
bed roughness. Journal of Hydraulic Research 33 (4) 571-581
Pope SB (2001) Turbulent Flows. Meas. Sci. Technol. 12 (11) 2020

17
Prandtl L (1949) Report on Investigation of Developed Turbulence.
Ramos HM, Borga A and Simão M (2009) New design solutions for low-power energy production in water
pipe systems. Water Science and Engineering 2 (4) 69-84
Reynolds O (1883) An Experimental Investigation of the Circumstances Which Determine Whether the
Motion of Water Shall Be Direct or Sinuous, and of the Law of Resistance in Parallel Channels. Philosophical
Transactions of the Royal Society of London 174 pp. 935-982
Roulund A, Mutlu BS, Fredsoe J and Michelsen J (2005) Numerical and experimental investigation of flow
and scour around a circular pile. Journal of Fluid Mechanics 534 351-401
Sahu M, Khatua KK, Patra KC and Naik T (2009) Developed Laminar Flow In Pipe Using Computational
Fluid Dynamics. in.
Schlichting H (1955) Boundary-Layer Theory, McGraw-Hill
Seed DJ and Limited HR (1997) River Training and Channel Protection: Validation of a 3D Numerical Model,
Hydraulics Research Limited
Shams M, Ahmadi G and Smith DH (2002) Computational modeling of flow and sediment transport and
deposition in meandering rivers. Advances in Water Resources 25 (6) 689-699 DOI: 10.1016/S0309-1708(02).
Simons TJ (1974) Verification of Numerical Models of Lake Ontario: Part I. Circulation in Spring and Early
Summer. Journal of Physical Oceanography 4 507-523
Tu J, Yeoh GH and Liu C (2008) Computational Fluid Dynamics - A Practical Approach, Elsevier
Veldman A (2011) Boundary Layers in Fluid Dynamics. in, pp. 93.
Versteeg HK and Malalasekera W (2007) An Introduction to Computational Fluid Dynamics - The Finite Volume
Method, Pearson Education Limited
Wang J-F, Piechna J and Muller N (2012) A novel design of composite water turbine using CFD. Journal of
Hydrodynamics, Ser. B 24 (1) 11-16
Weston SJ, Wood NB, Tabor G, Gosman AD and Firmin DN (1998) Combined MRI and CFD analysis of fully
developed steady and pulsatile laminar flow through a bend. Jmri-J Magn Reson Im 8 (5) 1158-1171 DOI: DOI
10.1002/jmri.1880080523.
White FM (2003) Fluid Mechanics, McGraw-Hill, ISBN 9780072402179.
Wu W, Rodi W and Wenka T (2000) 3D Numerical Modeling of Flow and Sediment Transport in Open
Channels. Journal of Hydraulic Engineering 126 (1) 4-15
Young DF, Munson BR, Okiishi TH and Huebsch WW (2011) A brief Introduction to Fluid Mechanics, Wiley
Zhou L, Liu D-y and Ou C-q (2011) Simulation of Flow Transients in a Water Filling Pipe Containing
Entrapped Air Pocket with VOF Model. Engineering Applications of Computational Fluid Mechanics 5 (1) 127-
140

18

Potrebbero piacerti anche