Light Metals 2004 Edited by Alton T.

Tabereaux TMS (The Minerals, Metals & Materials Society), 2004

DEMONSTRATION THERMO-ELECTRIC AND MHD MATHEMATICAL MODELS

OF A 500 KA ALUMINUM ELECTROLYSIS CELL: PART 2
Marc ~ u p u i s ' Valdis
, ~ojarevics'and Janis ~ r e i b e r g s ~

' ~ e n i ~ iInc.
m
3 1 1 1 Alger St.,Jonquiere. Quebec. Canada. G7S 2M9
%
.!-
l#l;,L
.- -..,A{.-.lj:3L?k<L--32.:.LL7g
. "

'university of Greenwich, School of Computing and Mathematics

'University of Latx.ia. Institute of Physics

32 Miera S

Keywords: Thermo-electric models, MHD iilodels. .4luminum electrolysis cell

Abstract
In the plesent study, a full 3D cell-cathode thenno-electr~cmodel
of a 500 kA demonstrdtion cell has been deleloped
In parallel, a non-lmeai tianslent \za\e MHD model foi the same
500 kA demonstiat~oncell. takmg Into account the s h ~ e l d ~ neffect
g coeffic~entsIn the thermo-electric model
of the optnn~zedgeonietly of the potshell, has been deleloped
Prellminary iesults of the Impact of the Interactions b e t w e n the
cell thcrmo-electi~cand MHD models ale prescnted
Introduction
111 Part 1 of this colldboratne vork. the authors developed both a
3D quartel-cell thelmo-electrlc model and a n o n - h e a r wa\e MHD
niodel of a 500 ItA denionstration cell des~gn[ I ] The quni-te~-cell
modcl comprises one quartel of the cell wlth cathode and anode
linked together The othei three quarters of the cell are prescnt by
~ i r t u eof symmetry For detalls on the design of thls 500 kA cell,
see the artlcle published In the magazme Alumm~umlast year [2]
The present collaboratire n o i k 1s also a direct follolz-up of the
n o l k presented in 2002 on the tentat~vedevelopment of a 3D full-
cell and exte~nalbus-bai thenno-electric model [3] At that bme. a
niodel of a 300 ItA denionstratlon cell deslgn wds b u ~ l tbut could
not be soh ed on the PI11 compute1 then a\ a11'1ble
The need to lmk togcthe~a 3D full-cell thermo-electilc model and
the non-lineal \za\e MHD model Into a coupled thermo-elect~~c-
MHD model of an aluni~nuiiielectrolys~scell has already been
evplamed 111 detalls pre\ ~ously[4] The ultmate goal of the cullent
collaborat~rework I S to d e ~ e l o psuch a coupled thenno-electr~c-
MHD model vhich would be onc of thc steps towa~ds the
development of a complete multl-physlcs model of an alummum
electrolys~scell
In t h ~ spaper. the ~csultsof a 3D full-cathode and extemal bus-bar
theimo-electric model ale p~esented Attempts to calculate the~mo-
electr~cfields In a full-cell and a half-cell model ale also d~scussed
A full-cell model IS essentially a full-cathode model with the anode
coupled to it A half-cell model 1s one half of the full-cell model,
the other half is plesent by v~rtueof symmetry
453
The n o n - h e a r nave MHD model IS evtended to take Into account
the shleldmg effect of the deta~ledand optimized geometly of the
potshell F~nally.the velocity fields calculated in the MHD model
are used to define the metallledge and bath ledge local heat t~ansfer
3D full-cathode and external bus-bar thermo-electric model
As demonstrated In [5], m the filst approximation, ~t is not
necessary to modcl In deta~lthe anode panel 111 older to calculate
the cunent dens~tyIn the metal pad when all the anodes are
carlying nedily e q i d current The positlon of the ledge, the
geometly of the block collector bdi connect~onm d the external
bus-ba~ des~gnall hale mole mfluence on the metal pad current
dens~tythan small res~stance discrepancy in the anode netnoili
Hence the 3D 5111-cathode and external bus-bar thernlo-electiic
model I S sufficient to compute the deta~ledcunent dens~tyIn the
metal
Figure 1 shows the mesh of the 3D full-cathode and extemal bus-
bar thenno-electric model. For the 500 kA demonstration cell, the
mesh is made of 3541 16 finite elements. Figure 2 shows the
distribution of the local heat transfer coefficients at the metaliledge
and bath;ledge interface for the preliminary run. This distribution
is not based on the velocity fields calculated by the MHD model or
any other velocity field. It is rather arbitrarily generated by a
sinusoidal function, which gives a variation along the walls, similar
to what four flow pools would give.
A P4 3 2 GHz computer was used to compute the solutlon It took
17 CPU hours to s o h e the model 111order to perfonn 3 loops of the
ledge shape convergence scheme The conbergence vAeme u a s
stopped before rcach~ngcomergence, as this mas not requlrcd for
thls prelmimary run
F~gure3 shows the obtained thennal solution, mh~leFigures 4 and
5 show the mctal pad voltage solutlon and the metal pad curlent
dcns~ty,respect~vely It can be obscrlcd that the rcsultmg ledge
profile th~cknessvanes along the cell permeter. thc ledge IS thicker
\\here the local heat transfer coeffic~entmas set lom and th~nner
where the local heat transfer coefficient mas set h ~ g h As a result.
the mtens~tyof the hor~zontal current In the metal pad I S not
un~fornialong the cell but rather harle5 from cathode block to
cathode block depending partlc~llarly on the ledge toe xarlatlon
along the cell pernneter

Figure 1. Mesh of the 3D full-cathode and external bus-bar

thermo-electric model.

Figure 4. Voltage solution In the metal pad of the 3D full -

cathode and external bus-bar thermo-electric model.

Figure 2. Setup of the local heat transfer coefficients pattern at the

h q u ~ d sledge ~nterface.

Figure 5. Current densty In the metal pad of the 3D full- cathode

and external bus-bar thermo-electric model.

3D full-cell and external bus-bar thermo-electric model

F~gure6 shows the mesh of the 3D full-cell and external bus-bar

thenno-electric model. That mesh 1s made of 585016 finite
Figure 3. Thermal solution of the 3D full-cell cathode and elcments Once agam [3]. that n~odclcould not be solxcd even on
external bus-bar the~mo-electricmodel. the more pouerful P4 3 2 GHz computer nlth 2 GB of RAM
Unfortunately, that model I S usmg constramt equations to connect
the anode panel mesh to the cathode mesh, and ~t seems that those
constranit cquatlous ale plex entlng the AYSYS solve1 to conxerge
Honchei, it 1s wolth notlng that the authors are st111 usmg a 10
454
yeals old Teislon of AYSYS and ~t IS hoped that nnpio\ements
made to the ANSYS s o l ~ esince
~ then nlll ensuie conxelgence
Figure 6. Mesh of the 3D full-cell and external bus-bar thermo-
electric model.
3D half-cell thermo-electric model
Figure 7 presents the mesh of a 3D half-cell model. This mesh is
made of 290410 finite elements. That model is the biggest model
with constraint equations that could bc solved. The P4 took 100
CPU hours to converge with the assumed ledge profile. Figure 8
shows the thermal solution obtained. Of course. uith that much
CPU time required for an assumed ledge profile, it was not
practical to repeat this 5 - 10 times. the number of iterations
needed for the calculation of the ledge profile!
I
Figure 7. Mesh of the 3D half-cell model.
455
Figure 8. Temperature solution of the 3D half-cell model.
MHD model refinement
Magnetic field in an aluminium cell is created by the currents in the
cell itself and from the complex bus-bar asrangement around the
cell, in the neighboring cells and the return line, and by the effect
of cell construction steel magnetization. The magnitude and
distribution of the magnetic field is of prime importance for
maintaining a desirable gentle mixing in metal and bath and to
avoid unstable wave growth with the intense flow and associated
wall erosion. The complexity of any practically usable
magnetohydrodynamic (MHD) model of the cell arises from the
coupling of the \wious physical effects: fluid dynamics, electric
current distribution, magnetic field and thcrnlal field.
The MHD model presented here accounts for the time dependent
coupling of the current and magnetic fields Lvith the bath-metal
interface movement [h]. Ledge profile is assumed given for now.
Coupling with the thermal model. which will supply ledge profile
to the MHD model, will be made in near ft~ture. This model
calculates non-linear waves and turbulent flows in the two liquid
layers. The fluid flow and wave model is transient and effectively
three-dimensional. using for the metal and bath flow the shallow-
layer approximation. The model couples the waves and the
electromagnetic field distribution: both, electric and magnetic fields
are recalculated as the wave shape changes. Although it has been
validated against analytical solutions for nonlinear gravity waves
and against physical models with mercury in the past, it has never
been tested with real cells because of high cost of measurements
and because involvement of an aluminium producer would be
neccssaly. We wish that in the future, such validation could be
done.
The first calculation step needed for an MHD model is the electric
current distribution in the busbars. This is calculated by coupling
thc clectric current in the fluid zone to a resistance network
rcprescnting individual anodes and cathode collector bars as well as
the wholc bus-bar circuit between two adjacent cells. The
Kirchhoff equations are generated automatically and solved at each
time step in order to simulate the effect of waves on electrical
current redistribution in the whole electrical circuit. The software
pcrmits adding connections. easily changing bus locations and
cross-sections, giving the freedom to experiment and optimize the
bus-bar network.
In the 500 kA test cell, all electr~callymdependent cathode bus- contribution to the solution and can not be simply discarded:
bars ale connected In 12 sections, each of nhich takes tlie cunent instead the analytical singularity elimination is used to give smooth
from 4 collector bars, and then the upstream and dounstream results.
counterparts are connected by 6 anode rlsers to the anode bus-bars
The asyrnmetr~cbus arrangement for t h ~ scell p a ~ t l ycompensates Once the magnet~zat~on of steel is knovn. the magnetic flux dens~ty
the ~ e t u n ihne on the left at x = - 60 m T h e ~ eIS almost a 50 - 50% B = p,, H f o ~tlie f l u ~ dzones can be calculated from the equation (2)
upstream - do\>nst~eanicurrent d n ISIOII o%mg to carefully adjusted u m g the obser\at~onlocation r on the conlputational grid wheie
bus-ba~cross-sect~ons the electromagnetic force distribution 1s needed This method used
for magnetic field calculation \\as validated against mercury
The second step in the MHD model is to calculate the magnetic phys~calmodels ~51thand n ~ t h o u tsteel palts Some validat~onwas
field B (called more precisely, magnetic induction). which is also camed out on c o m m e ~ c ~cells al
necessary to determine the electromagnetic force distribution
.~cithinthe liquid zone. f = j x B. The magnetic field B is the sum
of two contributions: B = BT+ B,I ; BTis generated by currents and
Bll by ferromagnetic steel material. The inagnetic field B1 from the
currents in the full bus-bar network is recalculated at each time step
during the dynamic simulation using the Biot-Savart law. A very
similar technique is used on the 3D grid within the cell fluid layers
where a special analytical technique is applied to deal with the
singularity in the Biot-Savart Law in order to obtain a smooth and
converging solution when the field calculation position coincides to
the electric current.

The calculation of the magnetic field BXIfrom steel requires much

more effort and cannot be done at each time step: we calculate it
only once with the steady state current distribution. The difficulty
arises because the steel parts of the cell are made of ferromagnetic
material whose magnetization 31 (H) depends non-linearly on the Figure 9. The experimental 500 kA bus-bar: four neighboring
local magnetic field intensity H in the magnetic material. The local cells used in magnetic field calculation.
magnetic induction B in the ferromagnetic material is orders of
magnitude higher than in the non-magnetic material. like air. liquid
aluminium, bath etc. Equation ( I ) gives the relationship between
~iiagneticinduction. magnetization and magnetic field intensity.

where p,, is the permeability of vacuum. equal to 4x x 10.' (H.:I~)in

the Interilational System of Units. In the non-magnetic material :M
= 0 and B = p,, H. In the magnetic material the unknown magnetic
field intensity H is related to the magnetization M (H) by the
material properties of a particular material (depending also on the
temperature, carbon content in steel. previous magnetization, etc.).
The typical curves used for the aluminium electrolysis cells are
discussed in [ 7 ] . In order to find the unknown magnetic field Figure 10. The ferromagnetic elements sho~vnfor the test cell
intensity, we need to solve the integral equation:
For the magnetic field computation, the busbar network was
extended to include four neighboring cells as shown in Figure 9.
,
,I
M(rf) 3(r - r ' ) (2) The steel is included as shoun in Figure 10. The steel was divided
--
[-- (r - r') , M(r')]dVf, into approximately 30000 nonlinear elements. Figure 11 shows the
'IT r - rfi7 lr - rfli magnetic ficld at the liquid metal top level without the effect of
steel and Figure 12 - with the effect of steel. The difference
uliele the magnetlc field H, IS g n e n by the Blot-Sabart lau from between the tw'o is considerable (see the legend in the Figures).
all the external electric currents. the co-ordinate locat~onr IS for the
field calculat~onposltlon (observation po~nt),thc integ~at~on pomt The MHD model uses a ~ e l a t n e l ylough mesh of 2 5 x 4 5 ~ 2elements
positlon r IS 111 thc element holume dV s-unnmg through all thc In older to be able to rc-conipute the current d ~ s t r ~ b u t ~ato neach
ferromagnetic mateiial In I;,, The Iteratne solutlon procedure for trmc step In a ~easonabletune Nevertheless, the solution IS
the equation (2) starts n ~ t han assumed distrlbut~onof R I in the suffic~ently smooth because of the global pseudo-spectlal
ferioinagnetic elements, calculates H for these elements, then uses approxmiation used for the space discretisat~onperniltting much
VI (H) matenal property to obtam the updated magnetlzatlon The h ~ g h c raccuracy In comparlson to fin~teelemcnt or fimte \olunie
p~ocedurc cnds when conlergence 1s a c h ~ e ~ c d The obv~ous a p p ~ o x ~ m a t ~on
o n the
s s ~ i n ~ lgrid
ar s ~ z c
compl~cat~on IS due to the s ~ n g u l a ~ In
~ t y(2) ~ + h c nthc ~ n t e g ~ a t ~ o n
pomt conmdes with the observat~onpoint T h ~ 15 s a bery ~mportant
456
The alumln~u~n-electrolyte interface defonnat~onmakes the anode Interface, dH(m) ##fg
cunents unequal because of the local ACD change The model
~ncludes an opt1011 to account for the tlme aberage gradual
consumption of the anode bottom to conforn~to the ACD change
An art~fic~ally
accelelated anode burn-out 1s permtted In order to
ach~ehethe result In a reasonable coniputatlonal tlme mtenal

Figure 13. The metal-bath interface at the last time steps of the
computation.

+
vsc 20 0.10 mls

Figure 11. The magnetlc field calculated without the

ferromagnetic elements.

Figure 14. The velocity and the effective turbulent viscosity in the
liquid metal at the last time step of the computation, t=1000 s.

The coupling between pressure and velocities has some special

effects on the metal-bath interface deformation. A large scale
Figure 12. The magnetic field calculated with the ferromagnetic horizontal vortex will create a considerable fiee surface dip in its
elements. center. This is because the circulation pattern in the metal layer is
different fioni the one in electrolyte; this creates pressure
A transient MHD simulation is started with the flat metal-bath differences across the interface to which the interface adjusts
interface and zero velocity. Then the usual magneto-hydrostatic vertically.
approximation is used to compute the interface defo~~nation as it
would be if the electric currents are unifonn. This interface Interface oscillations, (a) - DHcornerl
DHcOrnerNA
Pot drop
position is used as the actual initial state. A time step of 0.2 s was 0 005 -
used in the simulation shonm. The waves take about 200 - 300 s to
build up to a constant amplitude as shown in Figure 15. In a stable 0-
cell. this would be the time needed to achieve a stationary interface
position and a corresponding cuirent distribution. A snapshot of
the computed metal-bath interface defo~mationafter 1000 s of
simulation is shown in Figure 13. Predominantly horizontal re- -0 01 - -3 56
circulating flow also grows gradually with time. A snapshot of the
velocity patterns and the effective turbulent viscosity distribution at 250 500
I000 s is shown in Figure 14.
L / Fourier spectrum (b)

The flow is turbulent for the typical condtt~onsof the electrolys~s

cell The MHD model presented here uses 2 equation 'k-w', t m e -F DHcornerl
dependent turbulence model [6] nhich IS solled s~m~~ltancously - F DHcornerNA
Pot drop
xbith other ~ . a r ~ a b l eats each lteratlon The hlghly non-un~forn~ _ Gravbty frequenc~es

disti~bution of the effectne turbulent \iscosity I S shoun by the

color-filled contour lines 111 Flgure 14 The reglons of high
turbulent \ iscos~tyare asociated wlth enhanced turbulent heat and
mass transport, w h ~ c hale important for plcdlcting heat loss (ledge
th~ckncss)and thc cell wall eloslon
0 0 025 0 05 0 075 f (Hz)

Figure 15. The computed metal-bath mtcrface oscdlat~onsat two

d~agonallyopposlte comers (under anodc N 1 and anode N40) and
thc voltage osc~llat~on
(a), and the correspond~ngFouriel spectra
shoun a t t h the typcal grablty frequencies (tnangles) for
comparison (b)
457
The other important effect included in the MHD model is the flow-
induced electric current. resulting from the term o Y x B (o is the
electrical conductivity, which is very high for liquid metal. v -
metal velocity). This effect usually stabilizes the cell fluid
dynamics because, according to the Lenz's law, the induced field
always acts against the mechanical action causing it.
As ~eference,base on t h e ~ revpe~~mental . a z e r ~and a1 [9]
w o ~ kN
are proposlny a slope o f 3360 (W m2K) (nz s) Unfortunately,
experimental data, not a\ allable to the authors, would be requ~red
to establ~sha rel~ableconelat~onbetween the l e l o c ~ t yfields In the
xlclnlty of the cell ledge perlmeter and the local heat t~ansfer
coeffic~entsat the hquids/ledge Interface

The osc~llat~on patterns beneath the first comer anode and the The calculat~onof t h ~ smodel on the P4 computer took 32 CPU
d~agonallyoppos~tefar coiner anode are shown In the top part of hours to build and s o h e 1 1 lteratlons of the ledge shape
the F~gure15. The respect11 e anode currents fluctuate In the same conxelgence scheme. The results are shoun In F~gures17, 18 and
manner In spite of fairly nice and symmetric circulat~onpatterns, 19
shoun 111 Fig~ire14 for the metal (and a different 4 koitex pattern in
the bath. not shoun here). the nave is not damped as uould be
expected in a stable cell; the oscillations remain at about 1 cm
amplitude above and below the initial flat surface. This also gives
rise to the cell 1-oltage oscillations of a ~nagnitudeabout 3 mV
which are show'n in Figure 15. They are the result of the overall
cell resistance change, caused by the non-uniform ACD changes
during the wave motion. The voltage noise is typically observed in
all commercial cells. This particular 500 ItA demonstration cell,
evidently. will need a further busbar optimization to make the cell
inore stable. Reducing the nlagnetlc field inagn~tude at the
downstream left corner would certainly help We are not
part~cularlyconcerned about t h ~ shere, because the purpose of t h ~ s
article is to present the model and not to design a con~mercialbus-
bar arrangenient. The Fourier spectrum for this oscillation Figure 16. Distr~butionof the local heat transfer coefficients at
indicates a dominant kequency between the third and fourth the metallledge and bathiledge interface.
gra~itat~onal frequencies in the full spectrum of the normal grak~ty
u a \ e s As explained m [a], the MHD lnteractlon 1s respons~blefor
the s h ~ ffrom
t the purely hydrodynam~cosc~llat~on frequenc~es

First interaction trial between models

Up to now. the themno-electnc model and the MHD model w ele

used completely nidependently In the follow~ngIUII. for the first
t m e . the I cloc~tyfields calculated by the MHD model 111 both the
bath and the metal hake been used to set up the local heat tiansfer
coeffic~entsat the I~quids/ledge~nteiface111 the thermo-elcctr~cfull-
cell-cathode model (see F~gure16)

Due to time constraints in preparing this paper, however, the metal

velocity field used is not the one presented in Figure 14, but rather
a similar one from a previous run. The equations used to compute
the local coefficient in this first interaction trial are:

uhele h 1s 111 W/ni7 "C and c Is In n i s As In the MHD model the

\eloclty ( I ) is set to zelo at the ledge surface, the heloc~tyused 111
Equations (3) and (4) is the bulk f l u ~ dl e l o c ~ t yin the immediate
~ i c i n l t yof the ledge surfacc The idea of h e a r relationship w ~ t h Figure 17. The~inalsolut~onof the 3D full-cathode and external
\eloc~ty, used 1x1 Equat~ons (3) and (4), is taken fiom the bus-bar thermo-electr~cmodcl.
expermental w o k in [9], but the coefficlents are d~fferent The
coeffic~ents I84 1 5 and 1286 0 \I ere \elected In order to get
average lalues ~ d e n t ~ c atol the constant values uscd befo~e. as
typical ~ a l u c sof the heat transfei coefficlcnts ale estnnatcd to bc Future 11 ork
aiound 2000 W/rn2K when u s ~ n gthe cell l ~ q u ~ d usuperheats The
choice of the slope 15 qulte a ~ b ~ t r a rat y t h ~ sstage but 5000 The obv~ousnext step would be to use the ledge profile calculated
(~lm'K)i(m,s) IS barely large enough to really sign~ficantlyaffect by the full-cell cathode the~io-electncmodel and feed ~t into the
the ledge th~cluiessalong the p e r ~ m e t eof ~ the ccll c o n s ~ d e ~ i nthe
g MHD modcl. Work on t h ~ is s In progless.
rclatlvcly Ion ~ e l o c ~ t i In
e s thc v ~ c ~ n of
~ tthat
y cell ledge permetel
458
459