Professional Documents
Culture Documents
pubs.acs.org/IECR
Department of Chemical Engineering, Worcester Polytechnic Institute, 100 Institute Road, Worcester, Massachusetts 01609-2280,
United States
Johnson Matthey, P.O. Box 1, Belasis Avenue, Billingham, Cleveland TS23 1LB, U.K.
ABSTRACT: A new velocity-based approach to xed bed radial heat transfer is presented. Axial and radial velocity components
were averaged from detailed 3D computational uid dynamics (CFD) xed bed simulations of computer-generated beds of
spheres and used to model radial thermal convection. The convection terms were coupled with a radially varying stagnant bed
thermal conductivity in a 2D pseudocontinuum xed-bed heat transfer model. The usual eective radial thermal conductivity kr
and apparent wall heat transfer coecient hw were not used, and there were no adjustable parameters. The radial and axial
temperature variation predicted by the velocity-based model agreed well with the angular-averaged temperatures from the
detailed 3D CFD simulations over the range 80 Re 1900 and for N = 3.96, 5.96, and 7.99.
1. INTRODUCTION
Heat transfer in xed bed tubes is an important topic in the
chemical industry because xed beds are extensively used in
applications with heat eects, such as reactors, thermal storage
units, and adsorption or desorption plants. In particular,
multitubular xed bed reactors with low tube-to-particle
diameter ratio (N) are used for extremely exothermic or
endothermic reactions such as partial oxidations and steam
reforming of methane, respectively. Heat must be rapidly
transferred into or out of a narrow reactor tube, in which the
tube wall has a strong inuence on heat transfer and ow of
reactants around the catalyst particles. These in turn aect
catalyst activity, selectivity, and deactivation.
Current reactor models for heterogeneous gassolid reactors
have been based on fairly radical simplifying assumptions, such
as pseudohomogeneity, eective transport parameters, and
uniform catalyst pellet surroundings. Despite the realization
that local ow structures are critically important in determining
the global behavior of a ow or transport system,1 in many
cases the hydrodynamic modeling of reactors is still based on
unidirectional axial plug ow. All mechanisms for radial heat
transport are lumped into an eective radial thermal
conductivity kr, which is taken as constant and used to describe
heat transfer up to the wall. The observed increase in resistance
to heat transfer near the containing wall has been a continuing
source of diculty. The classical approach to modeling this
increased resistance near the wall is to idealize it to occur at the
wall, and lump all the mechanisms into a wall heat transfer
coecient, hw. Thus the near-wall resistance is misplaced, and
the temperature of the near-wall particles is under-predicted
(for wall heating) along with the associated reaction rate. For
narrow tubes this can be a major problem.
A review has recently been presented of the present state of
research and understanding of radial heat transfer in xed beds.2
The classical eective parameter kr hw model was extensively
described and problems with typical approaches to obtaining and
analyzing experimental heat transfer data to get kr and hw were
2013 American Chemical Society
Article
that kr varied with tube radius, especially near the tube wall, it
was usually taken as constant to simplify both parameter
estimation and solution of the model equations. The strong
decrease in kr as the tube wall was approached was idealized to
be a heat transfer resistance located at the wall, and was
represented by the wall heat transfer coecient, hw, along with a
temperature jump at the wall. The parameters kr and hw each
reect the eects of several dierent heat transfer mechanisms, and
have proved dicult to determine over the years, especially at low
N, while reaction models based on them have been criticized as
being oversimplied. Several papers and reviews have addressed
these concerns, many of which were summarized recently.2
One of the main perceived failings of the PF model has been
the use of a radially uniform axial velocity u0 to represent ow in
the tube. A comprehensive review of uid ow in packed tubes up
to 1987 was given by Ziolkowska and Ziolkowski9 which
demonstrated that the prevailing opinion was that the constant
u0 should be replaced by vz(r) and the wall heat transfer coecient
hw should not be used. Several research groups have developed
various approaches to obtain vz(r), including extended capillary
models,10 the extended BrinkmanDarcyForchheimer equation
from either particle-based11,12 or porous-media based1315 methods,
a combination of these two models,16 and various models derived
from the volume-averaged NavierStokes equations.1720 Several of
these ow models found it necessary to introduce an eective
viscosity into the equations,12,17 which introduces another
parameter that requires estimation and to some degree negates
the advantage of dispensing with the wall coecient. In addition,
the eective radial thermal conductivity is retained, but must now
be re-estimated for use in the altered model.21,22
In parallel with the problems associated with heat transfer
modeling in packed tubes, several authors have expressed
dissatisfaction with the standard dispersion model (SDM) which
uses eective diusion to represent axial and radial dispersion23
and has drawbacks including innite speed of propagation and
overestimation of back-mixing. Some dierent approaches to this
problem have included the cross-ow model,24 the alternating ow
model,25 and the wave model, rst put forward by Stewart26 and
more recently strongly championed by Kronberg and his
colleagues.27,28 The application of the wave model to xed bed
heat transfer was demonstrated by Kronberg and Westerterp27
whose work showed that this model also results in parameters that
must be determined from experimental data.
Kronberg and Westerterp27 in particular presented a strong
argument for the need for a new approach to modeling transport
in xed beds. The crux of their argument was that for many years
we have used eective diusion and conduction models to
represent heat and mass transfer phenomena that are essentially
uid mechanical in nature. The reason for this has largely been
computational convenience, a constraint that is rapidly being eased
by the development of faster, larger computers and improved
numerical methods. It should be possible to move toward models
that more realistically represent the ow eld in a xed bed. Some
of these points of view have more recently been echoed by
Schnitzlein29 who pointed out that eective dispersion coecients
are commonly used with gradients in concentration and that a
large contribution to dispersion is uid mechanical which is solely
driven by the packing structure and not by any concentration
gradient.
One alternative approach to radial heat transfer has been to
consider two- or three-dimensional ow elds, that is, to
include velocity components transverse to the main direction of
ow. Early attempts to obtain such ow elds were made by
T
r
1 T
T
r
= kr
r r r
z
T
r
= hw (T |r = R Tw )
r=R
(1)
(2)
=0
r=0
T |z = 0 = Tin(r )
(3)
(4)
Article
Figure 2. Simulated packed beds for CFD analysis: (a) N = 3.96, (b) N = 5.96, (c) N = 7.99.
T |r = R = Tw
(expt)
(computer model)
% deviation
3.96
5.96
7.99
0.476
0.451
0.431
0.466
0.450
0.432
2.1
0.2
0.2
and others included axial terms. Most neglected the radial mass
ux Gr(r,z), for example Stanek and Vychodil35 concluded that
radial ow terms were less than 1% of the velocity magnitude,
while Eigenberger36 and Froment37,38 and their co-workers
stated that strong radial ow components were found only in
the rst particle layer or a short entrance region.
The conclusions of the previous paragraph seem to
contradict the well-established idea that the main contributor
to radial heat transfer, at least as Re increases, is radial
displacement of uid around the particles, that is, convective
dispersion. One explanation is that the uid mechanics models
that were used to obtain the velocity components were all
based on averaged measures of bed structure, usually voidage.
Some used (r,z) and others only (r); however, all involved a
degree of smoothing of the bed structure. We suggest here that
the higher values of vr along the entire bed that would be
needed to account for the observed rates of radial heat transfer
are suppressed by the use of smoothed voidage elds in the
NavierStokes equations. In fact, the void fraction at a point
can have only values of zero or one, and it changes abruptly at
the local level as the radial coordinate passes from particle to
uid and back many times. It is these abrupt changes in void
fraction that give rise to the redirection of ow, giving strong
local variations in p and vr, which in turn result in the observed
radial heat transfer rates. These local variations can be averaged
out in smoothed or global approaches to xed bed structure
and uid ow.
To avoid uncertainties in the existing models of ow in xed
beds and the desire to avoid premature smoothing of the
velocity components at the local level, a dierent approach is to
use simulations of velocity elds directly in heat transfer
models. For example, Dixon et al.5 put forward a model of
T
T
r
c pGz(r , z)
+ Gr(r , z) = k r
r r r
z
r
kr
T
r
T
r
= hw (T |r = R Tw )
r=R
(5)
(6)
=0
r=0
T |z = 0 = Tin(r )
(9)
(7)
(8)
15246
Article
Article
of the sphere positions for the entire bed in each case using the
formulas developed by Mueller.49,50
The voidage proles for N = 3.96 are shown in Figure 3. The
prole shows two minima corresponding to the two layers of
spheres along the radial coordinate. Overall, the features and
magnitude of the experimental () are well-reproduced. The
high void fraction at the tube center ( 2) is caused by the
hole down the tube center due to the packing structure. The
voidage prole for N = 5.96 is shown in Figure 4; some slight
shift of the prole toward the tube wall may be attributed to the
soft-sphere algorithm which produces a more compacted
packing. This feature was also observed in similar algorithms
previously.45 The downturn for values of 3 is due to
anomalies at the center of the bed, where it is dicult to dene
small enough surfaces to obtain accurate values. Nevertheless,
this region is very small and of lesser importance compared
to the near-wall region where excellent agreement is found.
The voidage prole of the N = 7.99 bed is shown in Figure 5
and shows similar features and good agreement. The
magnitudes of the maxima and minima are especially accurately
found by the algorithm. The low voidage at 4 (tube center)
for the computer-generated packing is not an anomaly; for
this particular packing the spheres lined up along the
center-line.
The general good agreement shown in the graphs
demonstrates that the computer-generated sphere pack
reproduces the essential features of experimental measurements. The locations of maxima and minima are correctly
reproduced, as well as their magnitudes. This nding as well as
the results for overall voidage gives us condence in our
computer-generated models for the simulation of the velocity
and temperatures in a xed bed.
(ui)
+
= Sm
t
xi
(10)
Article
set at the tube outlet. For the energy balance, the tube wall
temperature, Tw = 368.15 K and the temperature of the inlet
ow Tin = 298.15 K were specied. At the particle soliduid
interfaces continuity of temperature and heat ux was enforced.
CFD simulations were carried out to obtain velocity and
temperature elds in full beds of spheres for the three cases,
N = 3.96, N = 5.96, and N = 7.99. The nominal particle
diameter was 1 in. (0.0254 m) in all columns and the nominal
tube diameters were 3.96 in. (0.1009 m), 5.96 in. (0.151384 m),
and 7.99 in. (0.202946 m). The models had a length of 0.0254 m
of empty tube before the bed inlet and a length of 0.0508 m of
empty tube after the bed to be able to place the inlet and outlet
boundary conditions away from the packing. The packed
lengths of the columns were as given in Figure 2. Simulations
were run over a range of ow rates to give Re in the range
802000.
The uid for the CFD simulations was taken as air with
constant properties corresponding to a bed average temperature of 333.15 K. These were density = 1.059545 kg/m3,
viscosity = 2.0291 105 kg/ms, specic heat cp = 1800 J/kgK,
and thermal conductivity kf = 0.0287 W/mK. The particles were
taken to be alumina with properties as density s = 1947 kg/m3,
specic heat cps = 1000 J/kgK, and thermal conductivity ks =
1.0 W/mK.
The model geometries and the mesh were constructed using
the commercial software GAMBIT 2.4.6, with the help of
journal les to carry out the repetitive creation and placement
of the spheres. To obtain a ne enough near-wall mesh for the
k- method we used boundary layer prism cells at outside
particle surfaces and at the tube walls; tetrahedral cells were
used in the main uid volume and inside the particles. The
unstructured tetrahedral mesh cell size was 1.524 103 m
(dp/16.7) and the boundary layer mesh thickness was 2.54
105 m (dp/1000) with three layers on the tube wall and a
single layer on the particle surfaces. The N = 3.96, 5.96, and
(11)
(12)
j hjJj
u
Dp
T
+
+ (ik) i + S h
xi xi
xi
Dt
xk
(13)
In this equation h is the enthalpy and for the present study the
user-dened volumetric heat source term Sh was zero. Radiation
was not included in the CFD simulation model due to the
relatively low laboratory-level temperatures simulated. The
above equations were solved in their original form for laminar
ows; for turbulent ows the Reynolds-averaged Navier
Stokes (RANS) models were used, with the original equations
being ensemble-averaged.
With RANS models all turbulence length scales are modeled.
The solution variables are decomposed into mean, ui and
uctuating, ui components and integrated over an interval of
time that is large compared to the small-scale uctuations.
When this is applied to the standard NavierStokes equations,
the result is
uj
ui uj
( ui )
p
ui
= +
+
+
xi
xj
xi
xj xj
t
2 u
( uiuj )
l +
xj
3 xl
(14)
Figure 7. Verication of nite element solution of 2D pseudocontinuum model with constant coecients against analytical solution at
dierent bed depths.
15249
Article
Figure 8. Contours of (a) axial velocity (m/s), (b) radial velocity (m/s), and (c) temperature (K), in the x = 0 plane of the N = 3.96 xed bed and
for Re = 240. Dotted boxes indicate regions used for close-up velocity vectors shown in Figure 9.
Figure 9. Close-up analysis of boxed regions from Figure 8 with N = 3.96 and Re = 240: (a) velocity vectors colored by axial velocity (m/s),
(b) velocity vectors colored by radial velocity (m/s), and (c) temperature contours (K).
7.99 total mesh sizes were 8.904 million, 15.579 million, and
28.47 million cells, respectively.
To remove the problem of meshing around the contact points
between the particles and between the tube wall and the particles,
the technique of shrinking the diameters of the particles to 99% of
the original diameter was used, so that the particles had an actual
diameter 0.025146 m. To provide the same size gaps for the
particlewall contact points the tube diameters were all increased
by 2.54 104 m also. This decision implied that heat transfer by
particleparticle or particle-wall area contacts was not represented
in this model. Other approaches to the problem of meshing
around contact points have been developed, and these were
recently reviewed and compared.51 Although the use of gaps
between particles does aect heat transfer uxes and temperature
distributions, in this study the same simplication was made in
applying the formula for the eective stagnant thermal
15250
Article
Figure 10. Cross-section of N = 7.99 xed bed column showing radial surfaces and angular planes used for averaging and sampling of the CFD
results.
Figure 11. Averages over angular surfaces in the N = 3.96 bed, Re = 240: (a) void fraction, (b) temperature (K), (c) radial velocity (m/s), (d) axial
velocity (m/s).
Article
1 +
T |r = R = Tw
(18)
T
r
(19)
=0
r=0
(20)
2
NM
kp 1
kp
B
B+1
B 1
ln
2
kp
B
N M
2
(N M )
(15)
1 10/9
B = 1.25
(17)
T |z = 0 = Tin(r )
ke0(r ) = k f 1
T
T 1 0 T
+ vr(r , z) =
c pvz(r )
rke (r )
z
r r r
r
(16)
Article
Figure 12. Supercial axial velocity proles for the three dierent
values of N, Re = 240.
Figure 13. Average of radial velocity contours in xed bed columns of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 80.
15253
Article
Figure 14. Average of radial velocity contours in xed bed columns of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 240.
15254
Article
Figure 15. Average of radial velocity contours in xed bed columns of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 950.
15255
Article
Figure 16. Average of radial velocity contours in xed bed columns of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 1900.
Article
Figure 17. Comparison of CFD 3D discrete particle model temperature contours and 2D vzvr pseudocontinuum model temperature contours in
xed bed column of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 80.
Figure 18. Comparison of CFD 3D discrete particle model temperature contours and 2D vzvr pseudocontinuum model temperature contours in
xed bed column of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 240.
15257
Article
Figure 19. Comparison of CFD 3D discrete particle model temperature contours and 2D vzvr pseudocontinuum model temperature contours in
xed bed column of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 950.
Figure 20. Comparison of CFD 3D discrete particle model temperature contours and 2D vzvr pseudocontinuum model temperature contours in
xed bed column of (a) N = 3.96, (b) N = 5.96, and (c) N = 7.99 for Re = 1900.
15258
Article
For N = 5.96 the two temperature elds are very similar, with
the CFD temperature level slightly lower at the outlet, but
otherwise the comparison is good. For N = 7.99 the two
temperature elds are in excellent agreement, with even smaller
details the same. This demonstrates that the new model can
predict the convection and conduction heat transfer locally with
the pseudocontinuum energy equation very well in a 2D model
for low Reynolds numbers for which conduction and
convection both play signicant roles.
The temperature contours for Re = 240 for all N are shown in
Figure 18. It is seen that at this higher ow rate there is less
penetration of the thermal front into the bed than for the lower
ow rate. The temperature contours between CFD and the vzvr
model were in generally good agreement. However, close to the
center of the bed, the CFD temperature developed more than in
the vzvr model; this was especially so for N = 3.96 and somewhat
the case for N = 5.96. Since the center of the bed was taken as a
symmetric boundary condition in the vzvr model the velocity
prole close to the center had to satisfy this boundary condition
limitation and could not develop in the same way as for the CFD
model. For N = 7.99 the two temperature elds were again in
excellent agreement.
The temperature contours at the much higher ow rate Re =
950 for all three N are shown in Figure 19. The simulations for the
near-turbulent regime Re = 950 could be run as either laminar or
turbulent models in the CFD; in this work Re = 950 was assumed
as laminar ow and the ow from CFD showed reasonable results.
The results are more sensitive to the averaging method for higher
Reynolds number compared to the low Reynolds numbers since
the radial velocities have lower negative and higher positive values,
so it was necessary to average more angular surfaces to avoid low
or high radial velocities at dierent points, which would then
cancel with the averaging method. Averaging of more angular
surfaces extracted better radial velocities to be used in the
pseudocontinuum vzvr model. This may account for the stronger
appearance of discrete temperature features in the N = 3.96 vzvr
model results, although generally the near-wall comparisons were
good. For N = 5.96 the vzvr temperature is a little low, and also for
N = 7.99. The extent of the temperature contours and the general
shape of the developing contours are both good in all three cases.
The temperature contours for turbulent ow at Re = 1900 for
all N are shown in Figure 20. The CFD and pseudocontinuum
model temperature comparison showed some slight dierences
in the results. This was due to the dimensional reduction of 3D
to 2D. There is a wider wall region of high temperature for the
vzvr model than in the CFD simulations, but comparisons show
very good results for all three N for the temperature levels
across the tube radius, in terms of both extent and magnitude.
Overall, the 2D pseudocontinuum heat transfer model based
on the velocity elds from CFD produced very reasonable
results compared to the 3D CFD temperature simulations
without the need to introduce any adjustable parameters such
as kr/kf and hw or an eective viscosity.
AUTHOR INFORMATION
Corresponding Author
*E-mail: agdixon@wpi.edu.
Present Address
ACKNOWLEDGMENTS
This material is based upon work supported by the National
Science Foundation under Grant No. CTS-0625693.
8. CONCLUSIONS
The main object of this work was to demonstrate the feasibility
of modeling radial temperature proles in xed beds of spheres
without using any adjustable parameters such as kr/kf and hw,
and without using eective heat conduction approaches for
uid mechanical phenomena. Instead, radial heat transfer was
to be predicted using local position-dependent components of
axial and radial velocity to represent heat transfer by uid
motion and its decrease near the tube wall, and a local eective
15259
NOMENCLATURE
cp = uid specic heat, J/(kgK)
cps = solid specic heat, J/(kgK)
B = shape parameter for ZehnerSchlunder formula
Bi = wall Biot number, hwR/kr
dp = particle diameter, m
dt = tube diameter, m
Fi = external body force per unit volume, kg/(m2s)
dx.doi.org/10.1021/ie4000568 | Ind. Eng. Chem. Res. 2013, 52, 1524415261
Article
Greek Letters
= bed voidage
0 = initial voidage for bed generation
= eective thermal conductivity of the uid (molecular and
turbulent), W/(mK)
= angular coordinate, radians
= uid viscosity, kg/(ms)
= dimensionless distance from tube wall, (Rr)/dp
= uid density, kg/m3
s = solid density, kg/m3
ij = viscous ux of j-momentum in the i-direction, kg/ms2
= specic dissipation rate, s1
REFERENCES
Article
15261