You are on page 1of 13

Powder Technology 261 (2014) 105117

Contents lists available at ScienceDirect

Powder Technology
journal homepage: www.elsevier.com/locate/powtec

Numerical prediction of the erosion due to particles in elbows


Gabriel Chucri Pereira, Francisco Jos de Souza , Diego Alves de Moro Martins
School of Mechanical Engineering, Federal University of Uberlandia, Av. Joo Naves de vila, 2121 Bloco 5P, 38400-902 Uberlandia, Minas Gerais, Brazil

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

Article history: Erosion by particles in process equipment is one of the major concerns in the oil industry. The transportation and
Received 2 August 2013 processing of oil and gas may involve eroding particles, such as sand and catalyst, which can cause damage to the
Received in revised form 18 March 2014 process equipment parts. Consequently, undesired maintenance operations are required, leading to unnecessary
Accepted 5 April 2014
costs. Also, there exists a risk of oil spill, which is extremely hazardous to the environment. This work is related to
Available online 15 April 2014
the investigation of numerical models for predicting erosion due to particles in an elbow pipe with a 90-degree
Keywords:
curvature angle. Wear can be easily identied in such geometry, which is commonly encountered in the oil indus-
Erosion prediction try. Four different correlations for the erosion rate were experimented with, namely Ahlert, Neilson and Gilchrist,
Erosion due to particles Oka and Zhang models. The input parameters for these empirical formulas were obtained from accurate CFD
CFD models for the gassolid ow within the bend. In order to assess the quality of the numerical predictions of
Oka correlation the erosion rate, experimental data was used. The effect of numerical parameters such as the number of compu-
tational particles, as well as the models for the coefcients of restitution and friction and surface roughness, was
evaluated. In general, it was found that the experimental correlations yield substantially different values for the
penetration ratio, although the erosion patterns generated by all of them are qualitatively similar. Despite the
complexity of the phenomenon, the Oka model produced results in close agreement with the experiments, sug-
gesting that it can be successfully employed in engineering applications along with accurate gassolid
simulations.
2014 Elsevier B.V. All rights reserved.

1. Introduction estimate the service life of a component, but also helps in nding geom-
etry spots in which severe erosion is more likely to occur.
In several engineering applications, a surface is attacked by solid par- In this context, the proper utilization of CFD models enables accurate
ticles carried by a uid, resulting in an undesired supercial wear to the simulation of the uid ow, such as oil or gas through pipes and bends.
component or piece of equipment. This kind of abrasive wear is named It is also possible to precisely predict the particle motion within the ow
erosion, and occurs frequently in industrial operations. by tracking the particles and their interactions with the walls and each
The erosion effects can be observed in the oil and gas industry, more other. Once the particle paths are known, empirical correlations relating
specically in the transportation and processing sectors, in which it the properties of both materials, i.e. the particle and the material being
affects directly the process efciency and safety. Sand particles for in- eroded, the particle velocity and angle of impact with the wear experi-
stance, generated when extracting oil, or catalyst particles, deliberately enced by the surface can be used to estimate the rate of erosion.
added for accelerating chemical reactions, act as threats to the integrity The main goal of this work is to investigate the application of CFD
of the transport and processing components of an oil facility such as and empirical erosion models to a geometry representative of those
pipes, cyclone separators and safety valves. The erosion in these systems commonly found in practical engineering applications; particularly in
and components might have catastrophic consequences, such as oil the oil and gas transport pipeline. The geometry investigated is a
spill, as well as a considerable liability to the company [31]. ninety-degree-elbow, where the erosion effects due to solid particle im-
Currently, it is of great interest for industries and technology centers pacts can be measured experimentally and calculated numerically. Ex-
to obtain reliable tools to predict and solve this issue, with the purpose perimental data collected by Chen et al. [5] was used to compare the
of saving time, resources and environmental complications due to po- simulation results. Four empirical correlations, proposed by Ahlert [1],
tential spillage. Obtaining an efcient method to determine erosion is Neilson and Gilchrist [23], Oka et al. [25] and Zhang et al. [32] were in-
essential for the prediction of component failure. It is also worth notic- vestigated. A sensitivity analysis regarding the number of computation-
ing that the accurate erosion prediction not only allows engineers to al particles, the restitution and the friction coefcient models was also
carried out. The nite-volume, unstructured code UNSCYFL3D, which
employs an EulerianLagrangian approach, was used to solve for the
Corresponding author. Tel.: +55 34 3239 4040 615. gasparticle ow. Turbulence effects are accounted for by the 2-layer
E-mail address: fjsouza@mecanica.ufu.br (F.J. de Souza). k-epsilon model. Generally speaking, it was found that the four

http://dx.doi.org/10.1016/j.powtec.2014.04.033
0032-5910/ 2014 Elsevier B.V. All rights reserved.
106 G.C. Pereira et al. / Powder Technology 261 (2014) 105117

correlations produce qualitatively similar results for the penetration 2.3. Inuence of the ow in the erosion
ratio, although the Oka et al. [24] model was seen to generate results
in better agreement with the experiments. The contours of the penetra- Fig. 1 shows four ow congurations commonly found in engineer-
tion ratio calculated numerically are presented, and a physical mecha- ing applications. The rst conguration illustrates an impinging jet,
nism for its shape is proposed. The effect of the surface roughness and which covers a wide range of applications, representing from research
coefcients of friction and restitution is also assessed. As expected, congurations to abrasion machining; Fig. 1b shows the ow congura-
these parameters are shown to substantially affect the particlegas tion found in ows over turbine blades and turbo machinery; Fig. 1c
ow, and therefore, the penetration ratio. shows the ow conguration that occurs in pneumatic transport of
solids and in piping; Fig. 1d represents the ow conguration found in
heat transfer devices [15].
2. Theoretical background The dynamic behavior of large and small particles is interpreted
briey in Fig. 1. The ability of a particle to respond to changes imposed
2.1. Erosion due to particles by the ow, and therefore, change its trajectory is characterized by the
number , which is dened by the ratio of two time scales that charac-
Erosion is dened as the wear resulted by the interaction between a terizes the dynamics of both solid and uid phases, respectively. In
solid surface and a uid ow containing abrasive particles with a certain Fig. 1, this number simply represents the particle dimension; for
speed, or the impact of free moving liquid (or solid) particles on a solid 1, particles have high momentum and respond slowly to ow chang-
surface [11]. We can divide the understanding of erosion in two major es; on the other hand, for 1, particles tend to follow the ow, being
parts, the rst being the determination of the uid ow conditions of an alternative to ow visualization. This is analogous to the Stokes num-
the number, direction, and velocity of the particles striking the surface. ber, classically used in particulate systems research.
The second part may be dened as the calculation of surface material re- The incident velocity magnitude of a particle depends on its interac-
moved, with the data acquired from the rst part. Clearly, the rst part tion with the uid, with other particles, and with the wall. The behavior
of the erosion process is characterized as a uid mechanics problem, of these interactions depends of the ow viscous regime (laminar or
with the uid ow transporting the particles into the surface, which de- turbulent), as well as the size, shape and density of particles. Interac-
nes the erosion wear. tions between particles are strongly related to the local particle concen-
Erosion wear is dependent of the number of particles striking a sur- tration, potentially causing low or high concentration regions.
face, as well as the physical quantities associated with it, such as particle
velocity and their direction relative to the surface to be struck. It is 3. Mathematical models
known that these quantities are noticeably determined by the ow con-
ditions. In other words, any minor change in the ow conditions such as As mentioned previously, the EulerLagrange approach is employed
viscous regime or temperature might bring large variations in the ero- in this investigation. Below the modeling of both phases is described,
sion rate. For example, in operations where the ow direction changes accordingly.
quickly such as turbine blade erosion is usually more severe than in a
straight run of piping. Other erosion-increasing factor is the local turbu- 3.1. Gas phase model
lence generated from roughened surface or misaligned parts [11].
A Reynolds-Averaged NavierStokes (RANS) approach is adopted in
this investigation. For a general, steady-state ow, the above-
2.2. Erosion mechanisms mentioned equations can be written in tensor notation as:

According to the literature, there are several ways to describe the


mechanism of erosion, as provided from different authors. Therefore, ui
0 1
it is difcult to establish only one mechanism as the most reliable and xi
real mechanism. The most used in the literature are the ones proposed
by Finnie [11] and Hutchings [16].
Finnie [11] proposed a mechanism of erosion in which the parti- " !#
cle acts as a miniature machine tool in which the surface material is   p ui u j
ui u j t : 2
cut, generating a chip. Also, for the erosion of ductile metals, at x j xi x j x j xi
oblique impact, this mechanism happens irrespective of its shape
and size.
Hutchings proposed a similar mechanism. However, he split the cut-
ting action into three different types, relying on the shape and the orien- The conservation equations for the continuity, velocity components
tation of the eroding particle. The rst type occurs when there is erosion and for the turbulence variables in steady state can be written generical-
by oblique impact of spherical particles, and the material is removed by ly as:
a plowing action, moving materials to the front and side of the particle.
The second and third types occur when there is the collision of angular !
 
shaped particles, and they differ from each other in the orientation of u j S : 3
the erodent particle as it strikes the target surface, as well as the direc- x j x j x j
tion of the particle during the contact with the surface; in other
words, if the particle rolls forward or backward during contact. Type I
cutting is dened when the particle rolls forward during the contact, By integrating the general conservation (Eq. (3)) over the control
and material is removed by repeated impacts on a prominent lip formed volume V, we obtain:
by the indenting angular particle. Type II cutting is dened when the
particle rolls backward, and the material is removed as if the erosion
was a machining operation, with the material being removed as a chip Z
! ! !
due to the fact that there's a sharp tip of the erodent particle, working A V  d A A grad  d A S dV: 4
as a machining tool [16]. V
G.C. Pereira et al. / Powder Technology 261 (2014) 105117 107

Fig. 1. Examples of ow congurations related to erosion due to impact by solid particles [15].

!
Note that, for the terms involving surface integrals in Eq. (4), the if Jf N 0. The vector dr L is directed from the geometric center of element L
Gauss Divergence Theorem was applied to convert the volume integrals to the face f center. (grad)rL is the reconstructed gradient at element L,
into surface integrals [10]: which is again computed by means of the Gauss Divergence Theorem:

Z
! 1 X ! 
!
dV A i i  d A : 5 gradr f A f 8
V f
V x i

where f is the average of the element centers sharing face f.


For the element L shown in Fig. 2, and located at the LHS of face f, the
It can be proven that the diffusive ux for face f is given by [17]
discretization of Eq. (4) yields:
! !
" ! !
#
X X   R L A f  A f ! !
Af  Af
Df f ! ! ! f grad  A f grad  e s ! : 9
J f f D f S V 6   A f  es
!
A f  es
L ds
f f

 
! In Eq. (9), !
e s is the unit vector connecting the centers of elements R
in which Jf is the mass ow rate, f !V f  A f , across face f, f the diffusion
!
!
coefcient at the that face and D f f grad f  A f is the diffusive ux and L, and e s ds
!. The rst term at the RHS of Eq. (14) is treated im-
!
across face f. The summations above apply to all the faces of element L. ds
!
A f is the normal area vector of face f, which is directed from the element
plicitly, whereas the remaining terms, which represent the secondary
L to the element R. diffusion, are calculated explicitly and therefore incorporated into the
Regarding the advective term in Eq. (6), when the rst-order up- source-term S in Eq. (6). The secondary diffusion is null for hexahedra
wind scheme is employed, f is assigned the value of the element center !
for instance, because vectors A f and ! e s are collinear. The gradient at
at element L if Jf is positive. Otherwise, the value of element R is set to
face f, grad, is calculated as the average of the gradients at the adjacent
the face. Because rst-order schemes are usually very diffusive for
elements. The treatment above is equivalent to the application of the
many applications of interest, a second-order upwind scheme was
second-order, centered differencing scheme in structured meshes and
used in this work:
is advantageous in the sense that it does not depend on the element
! shape.
f L gradrL  dr L 7
The numerical solution of the conservation equations for the mo-
mentum and turbulence, represented generically as Eq. (6), is accom-
plished by the computational code UNSCYFL3D. This in-house tool is
based on the nite volume method in unstructured three-dimensional
grids. The SIMPLE (Semi-Implicit Method for Pressure-Linked Equa-
tions) algorithm is used to couple the velocity and pressure elds. The
f
collocated arrangement is used for all variables, with the conventional
RhieChow interpolation scheme for the computation of the mass
R ow rate through each volume face. The discretization procedure de-
L scribed above generates a linear system of equations for at the ele-
ment centers. The biconjugate gradient and the algebraic multigrid
(AMG) methods are used to efciently solve the linear system resulting
from the discretization of each conserved variable. The main advantage
of this modeling is that it does depend on the element shape, as the data
Fig. 2. General control volume and nomenclature. structure is based on element faces. For each element face, arrays
108 G.C. Pereira et al. / Powder Technology 261 (2014) 105117

containing the number of L and R elements sharing it are created. As it Re y N Re y *, Re y * = 200 and viscosity-affected, Re y b 200. For the
be shown subsequently, all operations involved in transferring informa- one-equation model, the turbulent viscosity is computed from:
tion from the Eulerian to the Lagrangian frames and vice-versa utilize ei- p
ther face or element center data, and are therefore also irrespective of t;2layer C l k: 17
the number of faces each element contains. For storing the coefcients
of the linear systems for the velocity components, pressure correction The length scale in the equation above is computed as below:
and turbulence variables, the CSR (Compressed Sparse Row) format is  
used. For further information on the method, references Mathur and Re =A
l yC l 1e y : 18
Murthy [17] and Ferziger and Peric [10] are recommended.
In all the simulations carried out in this work only the steady-state
In UNSCYFL3D, both the standard k-epsilon and the one-equation
solution for uid was sought. The second-order upwind scheme was
model described above are solved over the whole domain, and the solu-
employed for the advective term, whereas the centered differencing
tions for the turbulent viscosity and the turbulence kinetic energy dissi-
scheme was used for the diffusive terms of the momentum equations
pation rate provided by both models are smoothly blended:
and turbulence model equations.
The standard k-epsilon model is the most widely known and exten- t t;standard 1 t;2layer : 19
sively used two-equation eddy viscosity model [2]. It was originally de-
veloped to improve the mixing-length model and to avoid the algebraic
A blending function, , is dened in such a way that it is equal to
prescription of the turbulence length scale in complex ows. Transport
unity far from walls and is zero very near walls. The blending function
equations are solved for two scalar properties of turbulence, the turbu-
used here is:
lence kinetic energy, k, and the dissipation rate of turbulent kinetic en-
ergy, epsilon: " 
!#
1 Rey Rey
  1 tanh : 20
"  # 2 A
u j k t k
P 10
x j x j k x j
The constant A determines the width of the blending function:

  0:20Rey
"  # A : 21
u j t 2 tanh0:98
C 1 PC 2 11
x j x j x j k k
The purpose of the blending function is to prevent solution diver-
gence when the solution from both the standard and the one-equation
where P is the production term, given by:
models do not match. The constants in the length scale formulas,
" ! # Eqs. (15) and (18), are taken from:
ui u j ui
P t : 12
u j ui u j 3=4
cl 0:4187c A 70 A 2cl : 22

The eddy-viscosity in the standard k-epsilon model is dened as a Since no wall-functions are used, it is very important to rene the
function of the turbulent kinetic energy and the turbulent dissipation grid so as to have y + b 1 in the rst element away from the wall and en-
rate as: sure accurate results for the uid ow. The models for both the uid and
2
particle have been validated in another publication [29].
k
t;standard C : 13
3.2. Particle motion model

Although widely used, the standard k-epsilon displays some weak-


As mentioned in the previous section, the dispersed phase is treated
nesses, such as the assumption that the ow is fully turbulent. To cir-
in a Lagrangian framework, in which each particle is tracked through
cumvent this issue, the 2-layer k-epsilon model was employed, as it
the domain and its equation of motion is based on Newton's second
can handle well both the core ow and the near wall region. Essentially,
law. The trajectory, linear momentum and angular momentum conser-
it consists in solving the standard model for the turbulent ow region
vation equations for a rigid, spherical particle can be written, respective-
and a one-equation model for the region affected by the viscosity. In
ly, as:
the one-equation k-epsilon model, the conservation equation for k is
retained, whereas epsilon is computed from: dxpi
upi 24
3=2
dt
k
: 14
l
!
dupi 3C D  
The length scale that appears in Eq. (14) is computed from: mp mp ui upi F si F ri 1 mp g i 25
dt 4p dp p
 
Re =A
l yC l 1e y : 15
dpi
Ip T i: 26
In Eq. (15), Rey is the turbulent Reynolds number, dened as: dt
p
y k In the above equations, ui = Ui + u'i are the components of the in-
Rey 16
stantaneous uid velocity. The average uid velocity Ui is interpolated
from the resolved ow eld, whereas the uctuating component u'i is
where y is the distance from the wall to the element centers. This calculated according to the Langevin dispersion model. dp is the particle
number is the demarcation of the two regions, fully turbulent if diameter and Ip = 0.1 mpd2p is the moment of inertia for a sphere. Unlike
G.C. Pereira et al. / Powder Technology 261 (2014) 105117 109

most commercial CFD codes, UNSCYFL3D solves for the particle rotation. tried out, and the best compromise between accuracy and cost was ob-
This is particularly important when dealing with large particles, which tained with the Sheppard's scheme. Basically, the velocity and vorticity
frequently collide with walls. components at the particle position are calculated by weighing the
The empirical correlation proposed by Schiller and Naumann [27] is neighboring element values with their inverse distances from their cen-
used to evaluate the drag coefcient past each particle: ters to the particle position. For integrating the ordinary differential
  Eqs. (24), (25) and (26), the analytical scheme was used for the linear
1 0:687
C D 24Rep 1 0:15Rep if Rep b1000 and angular velocities.
27
C D 0:44 if Rep N1000 : Upon a particle colliding with a wall, the new particle linear and an-
gular velocities after rebound are calculated according to the following
!  conservation equations [4]:nonsliding collision:
In Eq. (27), Rep is the particle Reynolds number Rep dp  u !u p =.
The calculation of the shear-induced lift force is based on the analyt-  2  
! ! ! ! ! !
ical result of Saffman [26] and extended for higher particle Reynolds u p u p 1 epar u 1 e u p  n n 34
7 pr
numbers according to Mei [22]:
h  i  
! 1=2 ! ! !
Fs 1:615dp Res C ls u up x 28 1 epar
! ! 10 ! !
  p p n x u pr : 35
! ! 7 dp
is the vorticity, Res d2p  = is the particle Reynolds number of the
shear ow and Cls = Fls/Fls,Saff represents the ratio of the extend lift force
to the Saffman force: Sliding collision:
2 3
    !
C ls 10:3314
0:5
e
0:1Rep 0:5
0:3314 if Rep b40 ! ! ! ! 6  u p  !7
u p u p 1 e u p  n 4 d   n 5 36
 U:b 29 !
up 
C ls 0:0524 Rep if Rep N40

is a parameter = 0.5Res/Rep(0.005 b b 0.4).  


! ! 5 ! ! ! !
The rotation-induced lift is computed based on the relation given by p p 1 e u p  n  d  n x u pr : 37
dp !
Rubinow and Keller [26], which was extended to account for the relative u  p

motion between particle and uid:


h  i In the above equations, the superscripts and + denote values
0 !
3 Rep x u !up prior to and after the collision, respectively, epar is the parallel restitu-
F r pdp C ! : 30 tion coefcient, e is the normal restitution coefcient and d is the dy-
8 Rer lr   !
 namic friction coefcient. n is the normal unit vector pointing
  outward of the element face being impacted. ! u pr is the relative velocity
! ! ! ! 2 !
In Eq. (30), 0:5  u p and Res dp  =. The lift coef- at the contact point:
cient Clr is obtained from the correlation proposed by [21]:
 
! ! ! dp !
Re u pr ! !
u p u p  n n x n: 38
C lr r if Rep b1 2 p
Rep
31
Re  0:522

Numerous experimental studies have shown evidence that wall
C lr r 0:178 0:822Rep if Rep N1:
Rep roughness is important in the particle behavior. Therefore, their inu-
ence must be included in the modeling. As demonstrated by Lan et al.
Also, the rotating particle experiences torque from the uid ow. [20] and Benson et al. [3], the wall roughness plays a vital role in the dis-
The correlation of Rubinow and Keller [25] was extended to account persion of particles in pneumatic transport systems. In order to account
for the relative motion between uid and particle at higher Reynolds for such effects, we implemented the model proposed by Sommerfeld
number: and Huber [28], to represent the effects of surface asperities on the par-
ticle ow. In summary, the wall roughness is simulated by assuming
dp  ! ! that the effective impact angle is composed of the geometric impact
5
!
T Cr   : 32 angle geometric added to a stochastic contribution due to wall roughness.
64

The coefcient of rotation, Cr, was obtained from the following corre- geometric  39
lation, derived from and the direct numerical simulations of Dennis et al.
[8]: This stochastic contribution is sampled from a Gaussian distribution
with a standard deviation , which depends on the structure of wall
64 roughness and particle size. Unfortunately, the value of must be cal-
Cr if Rer b32
Rer ibrated so as to provide the best agreement between the experimental
12:9 128:4 33
C r p if Rer N32: and simulated pressure losses.
Rer Rer When a structured grid is used, it is simple to determine the element
hosting the particle, as there exists a straightforward relationship be-
Forces such as Basset and virtual mass have been neglected. This is a tween the element index and its physical location. Because an unstruc-
reasonable assumption since the particle material density is over 1000 tured grid is used in this work, there is the need for a specic algorithm
times the gas density [6,7]. to locate the particle after its nal position is calculated by the integra-
The extension of the Euler-Lagrange approach to unstructured tion of Eq. (24). For that purpose, the particle-localization algorithm
meshes requires the use of accurate interpolation schemes, since in proposed by Haselbacher et al. [14] was implemented. This algorithm
the above equations the continuous phase properties must be deter- is based on tracking a particle along its trajectory by computing the in-
mined at the particle center. A few interpolation schemes have been tersections of the trajectory and the element faces. Since it does not
110 G.C. Pereira et al. / Powder Technology 261 (2014) 105117

depend on the element topology, it is suitable for use in unstructured in which erC and erD represent contributions from cutting and deforma-
grids. Furthermore, it was observed to be extremely faster than costly, tion respectively. The cutting erosion is modeled as a function of the in-
direct-search algorithms. cidence angle .
Due to the nature of the ow, as well as the particle size and concen- 8
tration, there was no need to solve the problem in a coupled approach. >
> 2 2
> u p cos sin
>
> 2
By using the segregated approach, the steady-state uid ow is solved < 0
b 0
separately from the particle ow, in a so-called one-way approach. 2C 45
>
> 2 2
This is justiable as the particle volume concentration in the incoming > u p cos
>
> N 0
:
ow is very low (less than 0.01%), as indicated in Elghobashi's diagram 2C
[9].
with 0 being the transition angle, normally set as 45, and C the cutting
3.3. Erosion models coefcient, specied as 3.332 107. Similarly, the deformation erosion
is given by
The erosion rate is dened as the mass of removed material per unit  2
of area per unit of time. It is calculated on the walls by accumulating the max up sinK; 0
damage each particle causes when colliding against the wall surface. It is erD 46
2D
given by:
where D is the deformation coefcient, set as 7.742 107, and K the
1 X
Ef m e 40 cut-off velocity, below which no deformation erosion occurs. In this
A f f r work, the cut-off velocity was set to zero.
The correlation proposed by Oka et al. [24] is:
in which Af is the face area, m is the particle mass ow rate represented !k !k
2 3
by each computational particle that collides with the face and er is the up Dp
er e90 g : 47
erosion ratio, consists in the ratio of mass of eroded material over uref Dref
mass of erodent material and must be computed by a correlation.
Four empirical models for the calculation of the erosion ratio were The reference velocity uref is set according to the user, depending on
tested in this work. It should be noticed that these models were imple- the situation, as well as the reference diameter Dref. The exponents k2
mented in UNSCYFL3D, working alongside the uid and particle models. and k3 are 2.325 and 3.26 104, respectively. The angle function g
Because the erosion is expressed in terms of the penetration ratio in () is dened as:
the reference experiments by Chen et al. [5], the erosion rates obtained
numerically are expressed in terms of this quantity in this work: n n2
g sin 1 1 H V 1 sin 48

Ef
Penetration ratio 41 with n1, n2 and HV material-dependant constants. For aluminum, the

m calculation method is described in Oka et al. [24]. The value of the latter
is identied by Oka et al. [24] as the Vickers hardness of the eroded ma-
wherem is the inlet sand mass ow rate (kg/s) and is the elbow material terial in units of GPa. By inspecting Eqs. (47) and (48), e90 is seen to be
density (kg/m3). The penetration ratio represents the thickness of mate- the reference erosion ratio at up = uref, DP = Dref and = 90 . The pur-
rial removed from the wall over the mass of sand injected into the ported strength of the Oka model is that the coefcients for a particular
elbow. combination of eroded material and eroding material can be derived
The correlation proposed by Ahlert [1] is: from more fundamental coefcients, which are specic to either the
!n eroded material or the eroding material. Hence, the fundamental coef-
up cients for sand can serve as a basis for both sand-steel erosion and sand-
er K F s f 42
uref aluminum erosion, for instance. The fundamental coefcients for the
eroding material, in turn, are shown to be derivable from measurable
in which K is a material-dependent constant, whose value was assigned properties of the eroding material such as its Vickers hardness.
2.388 107 in this work, Fs is a factor to account for the shape of the The correlation proposed by Zhang et al. [31] is:
particles, f() is a function expressing the dependency on the particle 0:59 n
incidence angle, uref is a constant reference velocity and n a constant ex- er C BH F s V p F 49
ponent. The shape coefcient Fs is reported to take the value 1 for angu-
lar particles, 0.53 for semi-rounded particles and 0.2 for fully rounded
2 3 4 5
particles. F 5:410:11 10:93 6:33 1:42 50
The angle function f () is split into two ranges. Below the user-
specied transition angle 0 is a polynomial in, the incidence angle where C and n are empirical constants, equal to 2.17 107 and 2.41,
in radians. Above the transition angle, f () follows a trigonometric respectively. BH is Brinnell hardness of the eroded material, Fs is the par-
relationship ticle shape coefcient, in this case 0.2 for perfectly round particles, Vp is
the particle impact velocity and is the particle incidence angle. Like
2 2
f x cos sinw y sin z: 43 the Oka correlation, this is a robust model due to the fact that it depends
exclusively of ow information and the eroded material properties.

The constants w, x and y are 1, 0.147 and 0.745 respectively, 3.4. Coefcients of restitution
whereas z is calculated internally by requiring that f () is continuous
at 0. In order to obtain an accurate prediction of the particle trajectories, it
The correlation proposed by Neilson and Gilchrist [23] is: is necessary to select a particle restitution model. Conversely, knowl-
edge on how particles behave after collisions with walls is needed.
er erC erD 44 Upon collision, the particle loses energy, and the rebound velocity is
G.C. Pereira et al. / Powder Technology 261 (2014) 105117 111

lower than the particle incident velocity. This effect is taken into account location, the four-feet duct upstream of the bend was also included in
through the coefcients of restitution. In this work, three different the calculation domain. The simulation was then run attempting to re-
models were used, all of them being derived from experimental studies. produce the same conditions as those under which the experiment
The model proposed by Forder et al. [12] for the normal and parallel was done, shown in Table 1.
coefcients of restitution is given, respectively, by: The mesh used in the simulation was created using only hexahedra,
which guarantees more stability and generates less diffusivity in the
2 3 4
e 0:9880:78 0:19 0:024 0:0027 51 simulation. Another benet of using this sort of element is the resulting
higher quality of the mesh, with fewer highly distorted elements than
one would obtain with tetrahedra, for example. This element was also
2 3 4 5 chosen because it is possible to rene the mesh close to the walls,
epar 10:78 0:84 0:21 0:028 0:022 52
where great velocity gradients and boundary layer do exist, without se-
where is the particle incidence angle. verely increasing the cell count and therefore the computational cost.
Grant and Tabakoff [13] proposed the model after treating the post The mesh resolution used in all simulations is approximately
collisional particle movement dynamics in a statistical approach. 700,000 elements, and is shown in detail in Fig. 3.
Based on experimental data on aluminum and sand, they proposed
Eqs. (53) and (54) for the coefcients: 5. Results and discussion

2 3 5.1. Comparison among erosion models


e 0:9931:76 1:56 0:49 53

In order to provide an initial picture of the erosion contours and save


epar 0:9981:66 2:11 0:67 :
2 3
54 computational effort in the subsequent analyses, a rst approach was
conducted comparing the four erosion models implemented. For this
test, 100,000 computational particles were injected in the domain and
Sommerfeld and Huber [28] proposed a model for the normal coef-
the coefcient of restitution model chosen was the one proposed by
cient of restitution only, regarding the parallel component equal to
Grant and Tabakoff [13]. Indeed, this is the only restitution model
one. The reason for that is the low contribution of the parallel compo-
which actually relates sand particles and aluminum. Also, the roughness
nent on the reection of particles after collision. The correlation for
parameter was kept at zero, that is, the domain walls are assumed to be
the normal restitution coefcient is given by:
perfectly smooth, and the coefcients of friction were assigned 0.5. The
e max10:013; 0:7: 55 inuence of these variables on the erosion rate will be demonstrated
further in this work. The penetration ratio contours for the four erosion
models tested, as viewed from the opposite side of the domain inlet, are
3.5. Coefcients of friction shown in Fig. 4.
It can be seen that all the correlations generate similar erosion pat-
Friction is another important effect to be accounted for in particle terns, qualitatively speaking, although the penetration ratio magnitudes
wall interactions. Depending on the static and dynamic coefcients, par- are very different. Actually, in most correlations, the main parameters
ticles can lose energy and velocity, directly affecting the erosion. In affecting the erosion are the particle velocity and angle of impact.
UNSCYFL3D, the standard coefcient can be used ( = 0.5), as well as Since the ow eld is the same for all correlations, the general behavior
an empirical model proposed by Sommerfeld and Huber [28] described is expected to be similar. The shape of the penetration contours can be
below: explained based on the particle motion. Along the inlet pipe, the parti-
cles develop a parabolic prole as shown in Fig. 5a. The highest penetra-
max0:50:175; 0:15: 56 tion ratio can be seen to occur as the faster particles in the center of the
pipe collide directly with the outer wall of the bend with high impact
It is worth noticing that the static and dynamic coefcients of friction angles. Downstream of the highest penetration region the bend, it can
were assumed to be equal. No considerable difference was detected by be seen that the particles tend to ll up the pipe diameter and move to-
prescribing the dynamic coefcient lower than the static one. wards the inner part of the bend, as depicted in Fig. 5b. As a conse-
quence of these sliding collisions at lower velocities, the isocontours of
4. Test case description the penetration ratio display an inverted V shape, visible from Fig. 4.
Based on the reference experimental data, it was also possible to
The investigation of erosion models performed in this work had as a quantitatively compare the erosion prole in the outer wall of the
database an experiment conducted by Chen et al. [5], in which a test elbow as a function of the curvature angle. The origin (0) is set at the
piece (elbow) was xed at the end of an air line. Sand particles were elbow ow inlet, and 90 corresponds to the elbow ow outlet. Fig. 6
injected in the line at about 4 ft upstream of the test piece. The test shows the penetration ratio prole for the four simulated models, as
piece was a ninety-degree-elbow with a diameter of 1 in and a curva- well as the experimental data by Chen et al. [5].
ture radius of 1.5 in. In order to better represent the ow at the specimen It can be noticed that the Ahlert correlation resulted in much higher
erosion when compared to the experimental results, whereas the other
Table 1
three correlations displayed values closer to the experimental data. Ac-
Flow condition summary. tually, Chen et al. [5] simulations with the same model overpredict the
experimental ratio by a factor of 10. The correlations proposed by
Temperature. This is the rst line of the table, and shoud be formatted 298 K
Zhang et al. [32] and Oka et al. [25] were the ones that better matched
like the ones below.
the experimental reference. This shows the potential of these two
Fluid Air
models, more specically the model proposed by Oka et al. [25], since
Fluid velocity 45.72 m/s
Particle diameter 150 m the eroded material hardness is taken into account. It thus enables the
Particle mass ow rate 2.08 x safe prediction of the effect of different eroded materials. The correla-
10-4kg/s tion proposed by Neilson and Gilchrist [23] showed very low values
Particle volume concentration 0.0042% for the penetration ratio when compared with the experimental results.
Test piece material Aluminum
This might be owing to the fact that it is an older model, and the
112 G.C. Pereira et al. / Powder Technology 261 (2014) 105117

Fig. 3. Computational mesh used in the simulations.

correlation was obtained for a specic pair of materials. Thus, it number of parameters whose values are usually not available. In any
may not be suitable for further analyses with different pairs of event, taking into account the complexity of the phenomenon, the qual-
materials. ity of the results with the Oka and Zhang models is remarkable.
Another important conclusion is that all the model predictions are In the next sections, the effect of the number of computational par-
slightly shifted regarding the peak penetration ratio. Because it is seen ticles, surface roughness, coefcients of friction and restitution will be
to occur with all the models, this effect can be attributed to the gaspar- assessed. Since the Oka model yielded the best results so far, it was se-
ticle ow prediction. Naturally, the simulation of this ow depends on a lected for such investigations.

Fig. 4. Penetration ratio contours for: Ahlert (a), NeilsonGilchrist (b), Oka (c), Zhang (d).
G.C. Pereira et al. / Powder Technology 261 (2014) 105117 113

Fig. 5. Particle ow in the inlet pipe (a) and at the bend (opposed from the inlet) (b).

5.2. Inuence of the number of computational particles It can be concluded that there is no signicant difference in the pen-
etration ratio proles between the results obtained with 50,000 parti-
Because the number of actual particles inside the domain even at cles and with 100,000 particles. This suggests that a relatively small
low volume concentrations is extremely high, they are statistically rep- number of computational particles can be used, with a reduced compu-
resented by computational particles. Obviously, the particle mass ow tational cost, without compromising the quality of the predictions. Nev-
rate simulated must be the same as that of the experiment, and there- ertheless, when comparing the results obtained with 5000 particles, a
fore it is necessary to assess how many computational particles are nec- greater difference is noticed in the penetration ratio magnitude at
essary to statistically reproduce the experimental behavior. Since the some points. This is expected when dealing with particle methods,
particle trajectories are calculated stochastically, the increase in compu- since fewer particles bring about a noisier prole.
tational particles tends to increase the accuracy of the simulation, as
well as reduce random errors originated from a reduced number of 5.3. Inuence of the coefcient of restitution
samples in the simulation.
Based on the authors' experience, there exists a limit to the increased The analysis of the inuence of the coefcient of restitution model
accuracy resulting from the continuous increase in the number of com- was performed using the three models described in Section 4. As previ-
putational particles, particularly when RANS turbulence models are ously stated, the models predict the particle behavior after the collision
employed. Besides, this increase in the number of particles might with the walls, that is, the value of the post-collision translation velocity.
prove unaffordable, since the computational cost increases as well. For heavy particles, such as those in this work, the particle motion is
In order to investigate the inuence of the number of particles in the highly dependent on the collisions with walls, and therefore the coef-
erosion rate, three cases were tested, using only one erosion correlation cient of restitution is expected to play an important role. The resulting
[24] and one model for the coefcient of restitution [13]. Three simula- penetration ratio contours are presented in Fig. 8.
tions, using 5000, 50,000 and 100,000 computational particles, respec- Fig. 9 shows the plot of the penetration ratio as a function of the cur-
tively, were then run to assess their effect in the penetration ratio. vature angle for the three restitution models.
Fig. 7 displays the penetration ratio as a function of the curvature An analysis of Fig. 9 reveals that the model proposed by Grant and
angle for the three cases studied. Tabakoff [13] produced very accurate results in comparison with the ex-
perimental results, although a second peak in the penetration ratio can
be noticed. The model proposed by Forder et al. [12] provided the

Fig. 6. Comparison of numerical and experimental penetration ratio versus bend curvature
angle for the four erosion models. Fig. 7. Inuence of the number of computational particles in the penetration ratio.
114 G.C. Pereira et al. / Powder Technology 261 (2014) 105117

Fig. 8. Erosion contours for different restitution models: (a) Forder, (b) Grant and Tabakoff, (c) Sommerfeld.

lowest values among the three correlations, and the model proposed by
Sommerfeld and Huber [28] lies between them. This can be justied
based on the pair of materials used in the latter two models. For exam-
ple, the model proposed by Sommerfeld provided the coefcient of res-
titution for glass particles colliding against steel, whereas the model
proposed by Forder was adjusted for sand particles colliding against
steel. On the other hand, the model proposed by Grant and Tabakoff re-
lates sand particles and aluminum, which is actually the most suitable in
this analysis. Nevertheless, the correlation results in a smaller magni-
tude, second peak in the penetration ratio, which is neither observed
in the experiment nor generated by the other two correlations.
A lag between the simulation and the experimental penetration ra-
tios can also be distinguished, as it was observed previously with all
the correlations for the erosion. Apparently, the simulated particles
reach the bend outer wall in a lower position than they do in the exper-
iment. Since the ow in the bend is upward, this deviation was initially
attributed to a lift force over the particles. In order to check this possibil-
ity, an additional simulation was run, in which all forces, except for drag
Fig. 9. Inuence of the models for the coefcient of restitution in the penetration ratio. and weight, were excluded in the model. No difference was observed,
G.C. Pereira et al. / Powder Technology 261 (2014) 105117 115

leading to the conclusion that these forces do not cause any deviation in
the particle ow. Another possible explanation for such discrepancy is
related to the fact that, due to the erosion, the bend geometry actually
experiences modications during the experiment. Such process is not
accounted for in the current simulations. This subject is under investiga-
tion by the authors.

5.4. Inuence of the wall roughness

The wall roughness plays a very important role in the particle mo-
tion, as demonstrated by Lan and Sommerfeld [19], and therefore in
the erosion prediction. As described in Section 3, the stochastic rough-
ness model utilizes a normal probability distribution to change the im-
pact angle due to the roughness contribution. In UNSCYFL3D, it is
possible to change the wall roughness parameter, which actually repre-
sents the standard deviation in the probability distribution function for
the impact angle, from 0 to 10, with 0 representing a perfectly
smooth and 10 a very rough surface. For this investigation, three values
for the roughness parameter were tested, namely 0, 2 and 4, with the
Fig. 11. Inuence of wall roughness in the penetration ratio.
purpose of simulating different wall nishings. Fig. 10 shows the
resulting penetration ratio contours.

Fig. 10. Erosion contours for different wall roughness = (a) 0; (b) 2; (c) 4.
116 G.C. Pereira et al. / Powder Technology 261 (2014) 105117

By analyzing Fig. 11, it can be concluded that, rough walls cause the proposed by Sommerfeld, in which it depends on the impact angle.
magnitude of erosion to decrease up to a certain limit. This nding can Fig. 12 shows the resulting penetration ratio contours for the three
be explained by the increased particlewall collision frequency caused cases.
by the surface roughness. These particlewall collisions, in turn, remove Fig. 13 suggests that the penetration ratio does not strongly depend
energy from the particles when compared to a perfectly smooth surface, on the coefcient of friction, as long as it is constant, since the results for
which then hit elbow walls at a lower velocity. This occurs because the 0.5 and 0.8 are slightly different from each other. Additionally, they are
particle elbow diameter is relatively small compared to the particle di- closer to the experimental results, despite the shift in the curve. It ap-
ameter, as demonstrated by Lan and Sommerfeld [18,19]. It is interest- pears that the standard value for the coefcient of friction is valid for
ing to observe that the increase in roughness does not necessarily this case. However, when using the model proposed by Sommerfeld,
promote a decrease in the penetration ratio. we notice a signicant increase in the erosion magnitude. Again, this
might be attributed to the fact that the model was developed for a
5.5. Inuence of friction pair of materials different from the ones simulated.

As stated in Section 3, friction is a very important parameter in the 6. Conclusions


particle motion. It is thus likely to affect the erosion prediction, due to
its effects on the particle rebounce from the walls. Depending on the By using accurate CFD models for the gasparticle ow in a 90 bend,
friction, the particle may slide or not over the wall surface. Consequent- it was possible to assess the erosion pattern and predict its magnitude
ly, it is expected that the coefcient of friction inuences the erosion. In and location as a function of model parameters such as the coefcients
this work, three cases were run, considering the coefcient of friction of restitution, coefcients of friction and wall roughness. Reynolds Aver-
equal to 0.5 (standard) and 0.8 and calculated according the model aged NavierStokes turbulence modeling was used due to its good

Fig. 12. Erosion contours for different coefcients of friction (a) 0.5; (b) 0.8; (c) Sommerfeld model.
G.C. Pereira et al. / Powder Technology 261 (2014) 105117 117

[3] M. Benson, T. Tanaka, J.K. Eaton, Effects of wall roughness on particle velocities in a
turbulent channel ow, Trans. ASME J. Fluids Eng. 127 (2005) 250256.
[4] M. Breuer, M. Alletto, Langfeldt, Sandgrain roughness model for rough wall within
EulerianLagrangian predictions of turbulent ows, Int. J. Multiphase Flow 43
(2012) 157175.
[5] X. Chen, B.S. McLaury, S. Shirazi, Application and Experimental Validation of a Com-
putational Fluid Dynamics (CFD) Based Erosion Prediction Model in Elbows and
Plugged Tees, University of Tulsa, 2004. (2004).
[6] C.T. Crowe, Multiphase Flow Handbook, CRC Press Taylor & Francis Group, Boca
Raton, USA, 2006.
[7] C.T. Crowe, M. Sommerfeld, Y. Tsuji, Fundamentals of GasParticle and GasDroplet
Flows, CRC Press, 1998.
[8] S.C.R. Dennis, S.N. Singh, D.B. Ingham, The steady ow due to a rotating sphere at
low and moderate Reynolds numbers, J. Fluid Mech. 101 (1980) 257279.
[9] S. Elghobashi, On predicting particle-laden turbulent ows, Appl. Sci. Res. 52 (1994)
309329.
[10] J.H. Ferziger, M. Peric, Computational Methods for Fluid Dynamics, Springer, 2002.
[11] I. Finnie, Erosion of Surfaces by Solid Particles, WearElsevier, 1960.
[12] A. Forder, M. Thew, D. Harrison, Numerical investigation of solid particle erosion ex-
perienced within oileld control valves, Wear 216 (1998) 184193.
[13] T. Grant, W. Tabakoff, Erosion prediction in turbomachinery resulting from environ-
mental solid particles, J. Aircr. 12 (1975) 471547.
[14] A. Haselbacher, F.M. Najjar, J.P. Ferry, An efcient and robust particle-localization al-
gorithm for unstructured grids, J. Comput. Phys. 225 (2007) 21982213.
[15] J.A.C. Humphrey, Fundamentals of uid motion in erosion by solid particle impact,
Fig. 13. Inuence of the coefcients of friction on the erosion behavior. Int. J. Heat Fluid Flow 11 (3) (1990) 173195.
[16] I.M. Hutchings, Tribology, Friction and Wear of Engineering Materials, 1st ed.
Elsevier Limited, 1992.
[17] S.R. Mathur, J.Y. Murthy, A pressure-based method for unstructured meshes, Numer.
performance in the ow investigated and low computational cost, as Heat Transfer, Part B 31 (1997) 195215.
well as relatively small mesh requirement. [18] S. Lan, M. Sommerfeld, Characterisation of pneumatic conveying systems using the
Based on the simulation results, it can be concluded that the combi- Euler/Lagrange approach, Powder Technol. 235 (2013) 764782.
[19] S. Lan, M. Sommerfeld, Numerical calculation of pneumatic conveying in horizontal
nation of the models proposed by Oka et al. [24] for the erosion ratio and channels and pipes: detailed analysis of conveying behaviour, Int. J. Multiphase Flow
Grant and Tabakoff [13] for the coefcient of restitution is the most suit- 39 (2012) 105120.
able for the erosion prediction for this case, since it provided the best [20] S. Lan, M. Sommerfeld, J. Kussin, Experimental studies and modelling of four-way
coupling in particle-laden horizontal channel ow, Int. J. Heat Fluid Flow 23
agreement with the experimental penetration ratio in terms of magni- (2002) 647656.
tude and position. It can also be concluded that wall roughness inu- [21] C.K.K. Lun, H.S. Liu, Numerical simulation of dilute turbulent gassolid ows in hor-
ences directly the erosion, attenuating its magnitude. Also, the specic izontal channels, Int. J. Multiphase Flow 23 (No. 3) (1997) 575605.
[22] R. Mei, An approximate expression for the shear lift force on a spherical particle at
value for the friction coefcient appears not to be relevant for the ero- nite Reynolds number, Int. J. Multiphase Flow 18 (1992) 145147.
sion calculation. [23] J.H. Neilson, A. Gilchrist, Erosion by a stream of solid particles, Wear 11 (1968)
Based on the main ndings of this work, it is expected that the ero- 111122.
[24] Y.I. Oka, K. Okamura, T. Yoshida, Practical estimation of erosion damage caused by
sion rate in other geometries can be predicted accurately with the Oka solid particle impact, Wear 269 (2005) 102109.
model, therefore depending only on measurable parameters of the [25] S.I. Rubinow, J.B. Keller, The transverse force on a spinning sphere moving in a vis-
eroded material instead of very specically adjusted models. cous liquid, J. Fluid Mech. 11 (1961) 447459.
[26] P.G. Saffman, The lift on a small sphere in a shear ow, J. Fluid Mech. 22 (1965)
385400.
Acknowledgment [27] L. Schiller, A. Naumann, A drag coefcient correlation, Z. Ver. Dtsch. Ing. 77318
(1935).
The authors would like to acknowledge the nancial support from [28] M. Sommerfeld, N. Huber, Experimental analysis and modelling of particlewall col-
lision, Int. J. Multiphase Flow 25 (1999) 14571489.
Petrleo Brasileiro S. A. (PETROBRAS) (050.0072928.12.9), CAPES and [29] F.J. Souza, R.V. Salvo, D.M.M. Martins, Large Eddy simulation of the gasparticle ow
National Counsel of Technological and Scientic Development (CNPq). in cyclone separators, Sep. Purif. Technol. 94 (2012) 6170.
[30] M.M. Stack, G.H. Abdulrahman, Mapping ErosionCorrosion of Carbon Steel in Oil
Exploration Conditions: Some New Approaches to Characterizing Mechanisms and
References Synergies, Ph.D. thesis Strathclyde University, 2004.
[31] Y. Zhang, E.P. Reuterfors, B.S. McLaury, S.A. Shirazi, E.R. Rybicki, Comparison of com-
[1] AhlertK. , Effects of Particle Impingement Angle and Surface Wetting on Solid Parti- puted and measured particle velocities and erosion in water and air ows, Wear 263
cle Erosion of AISI 1018 Steel, MS Thesis University of Tulsa, USA, 1994. (2007) 330338.
[2] J.E. Bardina, P.G. Huang, T.J. Coakley, Turbulence Modeling, Validation, Testing and
Development, NASA TM 110446, 1997.

You might also like