TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

AL03 - A Modernized ANSYS-Based Finite Element Model for the

Thermal-Electrical Design of Aluminum Reduction Cells
Daniel Richard1, André Felipe Schneider2, Marc Dupuis3 and Stephan Broek4
1.Associate – Center of Excellence for Aluminum, Hatch Ltd., Saguenay, Canada
2. Numerical Analysis Specialist – Center of Excellence for Aluminum, Hatch Ltd., Montréal,
Canada
3. Consultant, GeniSIM, Inc., Saguenay, Canada
4. Associate – Center of Excellence for Aluminum, Hatch Ltd., Mississauga, Canada
Corresponding author: [email protected]

Abstract

Heat balance and magnetohydrodynamics are critical to the design of an aluminum reduction cell
since they largely determine its operational window. Furthermore, an inadequate lining design
generally leads to degraded cell performance and premature failures. The first task in lining design
is to determine the position of the frozen ledge and the cell superheat for a range of operational
parameters.

Although several different modeling approaches and computational domains have been proposed
to solve the Stefan problem, a widely accepted methodology, first proposed by Dupuis [1], is
based on the iterative repositioning of the ledge front in a thermoelectrical (TE) Finite Element
(FE) model. The algorithm involves successive displacements of the solidification front nodes
based on the calculated temperature field until the entire ledge-to-liquids interface reaches the
bath solidification temperature. The superheat is adjusted to minimize the difference between the
cell internal heat generation and the integrated heat losses over the control volume. Originally,
this approach was limited to two layers of first order elements across the ledge thickness moving
horizontally and did not include the liquids.

This paper presents a generalization as well as improvements to the original methodology,

enabling the prediction of the ledge profile using an arbitrary number of first or second order
elements through the ledge thickness while including the metal pad and the bath. The proposed
modeling framework has been implemented in ANSYS using the ANSYS Parametric Design
Language (APDL) scripting language and designed to minimize the computational cost of moving
the ledge. Another benefit is that the generic core macros developed also efficiently handle the
ledge front displacement in any orientation. Current technology ANSYS elements are used, in
such a manner that high-performance computing solvers can be leveraged.

The robustness of this improved methodology is illustrated in this article by comparing the results
obtained for a fictitious 300 kA cell technology against those computed by the standard approach.

Keywords: Aluminum reduction cells, heat balance, moving ledge profile, finite element
analysis.

1. Introduction

Hatch has long been involved in the assessment of aluminium reduction technologies for
greenfield and brownfield smelter projects and has recently been mandated to evaluate designs
for a brownfield retrofit project, and specifically to perform heat balance calculations. Numerical
modeling is considered the best tool to study the thermal-electrical behavior of the lining and the
methodology is mature and widely accepted.

563
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

An essential part of the modelling effort is to predict the ledge profile, which can be quite
challenging. Dupuis first developed a FORTRAN routine to reposition the ledge of a cathode slice
and solved the TE problem by means of a finite element analysis in the commercial software
ANSYS. The approach was then implemented in the ANSYS interpreted scripting language
APDL [1]. The model domain was progressively enlarged to include a full quarter cell [2] but did
not include the liquid phases at this point. The electrical boundary conditions were applied to the
immerged surface of the anodes and to the top surface of the cathode blocks as an electrical
equipotential.

Dupuis introduced a full cell slice (i.e. including both anode and cathode segments) with
combined convergence of cell superheat and the ledge profile in [3]. The liquid zone for the full
quarter was later introduced in [4] thus allowing one to obtain a representative current distribution
in the metal pad.

The same modeling approach to move the ledge continued to be employed by Dupuis from the
mid-1980’s up to his most recent work in 2019 [5]. During that period, the ANSYS software
platform has evolved, such that some of the functionalities used by Dupuis became no longer
entirely compatible with the most recent versions of said software package.

Furthermore, the Authors agree that some of the drawbacks of the original Dupuis approach are
the following:
The construction of the geometry is from the bottom-up and requires experience with ANSYS
APDL
The discretization of the domain is limited to linear elements
In particular, the discretization of the ledge thickness is limited to 2 elements
The ledge mesh is attached to the cathode surface, which is limiting for the construction of the
corner geometry. It also limits the position of the ledge toe as it depends on the cathode block
assembly topology
The processing time for moving ledge and update boundary conditions is substantial
The models are based on legacy ANSYS elements
These elements do not support present-day Distributed Memory Parallel (DMP) processing,
which substantially increases computational performance of large models
Furthermore, thermal loads are, in some instances, applied directly to the faces of solid
elements, which is no longer supported by some of the current-technology ones (e.g.,
SOLID231).

From all of the above, it became quite clear that the original Dupuis approach could be improved
in order to obtain a modernized methodology that is fully compatible with ANSYS 2020 R1. This
was accomplished by rewriting critical routines and restructuring the modeling workflow.

2. Modernized Approach Workflow

The new workflow is summarized in Figure 1.

Figure 1. Modernized approach workflow.

564
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

2.1 External Geometry

Solid geometry of a cell can be developed in any software package compatible with the ANSYS
geometry engines and imported for further processing. Examples of valid geometry formats
include CATIA, Parasolid, STL, and STEP files.

2.2 Geometry Editing

The geometry that is tied to the model assembly is built with either ANSYS/SpaceClaim or
ANSYS/Design Modeler. Within the context of the present work, the Authors used SpaceClaim
to preprocess the geometry. External components can be imported and modified, while additional
entities (e.g., the ledge) can be added directly in this module. Bodies are regrouped into parts that
share a mesh and the interaction between different parts is handled by contact-target element pairs
when building the finite element model itself.

2.3 Finite Element Model Assembly

The geometry is linked into ANSYS/Workbench (WB) Mechanical environment, where the finite
element (FE) mesh is built and contact is established between parts that do not share a mesh.
Named components are also defined in this environment.

The ANSYS/WB Mechanical environment is shown in Figure 2. The colors represent the different
parts in the model that will share a FE mesh. Contact and target elements are used to transfer heat
and electricity across non-matching meshes from the different parts, which facilitate the
discretization of the domain.

Figure 2. ANSYS/WB Mechanical environment with a quarter cell model.

Named selections (or, simply, components) are used in APDL routines to:
Assign materials properties
Assign element types and physics (e.g., thermal shell, thermal-electrical solid, etc.);
Define contact and target surfaces
Apply boundary conditions, like surface convection or prescribed voltages
Define the entities required for the ledge repositioning logic
Define the heat balance control volume
Define the reference location for reporting the anodic and cathodic voltage drops
Define the boundaries of the cell control volume at the anodic and cathodic panels

565
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

Define the bodies included in the finite element solution but that are outside of the cell control
volume
Define the surfaces for the detailed heat losses output
Define model assemblies for different types of solution (e.g., half anode, quarter anode panel,
cathode slice, full cell slice, quarter cathode panel and full cell quarter).

Examples of named selections can be found in Figure 3. A typical FE mesh is shown in Figure 4,
with coloring corresponding to the element types.

Figure 3. Example of named selections (or components). Left: component identifying the
cathode blocks for the attribution of material properties. Right: named selection
identifying the lower part of the pot shell for the definition of convection and radiation
boundary conditions.

Figure 4. Initial FE linear mesh for a full quarter cell model.

It is important to highlight that, once the FE mesh is built, there is no need to return to any of the
ANSYS/WB instances (either ANSYS/SpaceClaim, ANSYS/Design Modeler or ANSYS/WB
Mechanical environments) while computing the solution as the prediction of both operating
temperature and superheat as well as the repositioning of the ledge-liquids interface and the
update of heat transfer boundary conditions are entirely handled by generic APDL macros, which
require no User intervention (refer to section 2.5).

566
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

2.4 Parameters

Once the mesh is built and all the components are defined, the User needs to update the input text
files required for the solution. Model parameters defined in text files will be read by
ANSYS/APDL routines. These User input parameters are organized into four (4) main files:
General parameters file. This includes, for example, the cell amperage, the bath chemistry, the
ambient temperature, the model domain and the type of solution.
Materials and physics file. This is where the material properties are assigned to named selections
and the appropriate physics is selected.
Boundary conditions file. This is where convection and radiation heat transfer surface loads are
defined by the User and assigned to components.
Heat losses output definition file. This file is used to list the named selections that will be used
for the detailed heat losses standardized output.

2.5 Solution

Generic APDL macros are used to process the ANSYS Workbench database together with the
four (4) input text files described above. This can be done either interactively within the
ANSYS/WB Mechanical module, or separately as a batch solution running in the background.
Note that the User does not need to edit any of the generic macros to perform an analysis.

The following steps occur during this phase:

Reading of the ANSYS Workbench database;
Reading of the general parameters;
Reselecting the appropriate model domain;
Creating and reading the material properties;
Assigning the material properties and the element types;
Creation of the required contact-target element pairs;
Creation of the auxiliary data structure for the ledge repositioning logic;
Creation of the heat transfer coefficients;
Application of the heat transfer loads and the electrical boundary conditions;
Reading of the surfaces for the heat losses detailed output;
Definition of the extrapolation factors required for the complete cell output heat losses
corresponding to the model domain;
Application of the settings for the nonlinear thermal-electrical solution;
Solving the problem by following one (1) of three (3) possible routes, namely:
Static solution for half anode and anode panel models (i.e., thermal-electrical problem solved
only once);
Ledge convergence only for cathode slice and quarter cathode models at fixed superheat, or;
Simultaneous superheat and ledge convergence for full cell slice and full cell quarter models.

2.6 Postprocessing

Postprocessing of results can either be programmed in APDL macros or done interactively in

ANSYS. A typical temperature distribution is shown in Figure 5.

567
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

Figure 5. Full quarter cell temperature distribution (C) at converged ledge and
superheat.

2.7 Standard Output

Once the global problem has converged, a standard output text file is created that summarizes the
results and details the heat losses on the previously defined surfaces. An example is shown in
Figure 6 for the ledge convergence solution of a quarter cathode model (i.e., no anode panel)
using a quadratic FE mesh. Note that the detailed heat losses are reported for the model domain,
which means that the User must multiply the heat flow by the appropriate scaling factor for the
cell – in the particular case of a quarter cell model, by four (4).
============================================================
CATHODE QUARTER Model
Solution type Ledge convergence only
============================================================
OPERATING CONDITIONS
------------------------------------------------------------
Cell Current [kA] 300.00
Operating Temperature [°C] 975.00
Liquidus [°C] 968.00
Superheat [°C] 7.00
Ambient temperature [°C] 20.00
============================================================
HEAT IN
------------------------------------------------------------
Bath to ledge [kW] 83.29
Metal to ledge [kW] 173.78
Metal to lining [kW] 53.87
Joule Heat in Cathode [kW] 91.92
Total Heat into Cathode [kW] 402.87
Cathode Voltage Drop [mV] 295.09
============================================================
GLOBAL HEAT LOST

568
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

------------------------------------------------------------
Cathodic Panel [kW] 402.826818
------------------------------------------------------------
DETAILED HEAT LOST OVER MODEL DOMAIN
------------------------------------------------------------
[W] [W/m2] [Component]
7171.42 5242.27 CollBars
3956.95 1373.25 CathodeFlexes
1502.86 1097.98 Deckplate_In
36662.16 6004.28 SW_AboveBars
2930.94 2461.33 SW_BarLevel
1435.80 561.74 SW_BelowBars
1357.28 70.57 SW_CradlesFlange
10802.30 336.71 SW_CradlesWeb
7843.77 1086.40 SW_DeckPlate_Out
6674.86 432.24 Shell_Floor
1252.41 928.74 EW_BelowBox
1928.73 675.70 EW_Box
108.84 331.82 EW_CradlesFlange
420.19 333.48 EW_CradlesWeb
2092.54 911.78 EW_Deckplate_Out
9900.20 5292.81 EW_InsideBox
4665.44 31003.70 Heat leaving cathodic CV
0.00 Joule Heat outside cathodic CV
============================================================
ERROR CALCULATION
------------------------------------------------------------
Ledge converged? [-] Yes
Ledge Max (infinite norm) Temperature error* [°C] 8.88
Ledge RMS Temperature error** [°C] 0.045
Ledge Max (infinite norm) Iter. Disp.*** [mm] 0.1
Number of global iterations [-] 35
FEA Heat imbalance [W] 2.83
FEA Heat imbalance [%] 2.81E-03
============================================================
* Refers to the maximum absolute computed difference with the bath melting temperature for an individual
node at the ledge-liquids interface.
** Refers to the root mean square difference computed with the bath melting temperature for all nodes at
the ledge-liquids interface.
*** Refers to the maximum absolute computed iterative displacement for an individual node at the ledge-
liquids interface.
Figure 6. Standardized output example.

3. Modernized Approach Features

The present work follows the recommendations of Arkhipov et al [6] and includes the bath and
the metal pad in the electrical problem to obtain a representative current distribution in the metal
pad and an accurate anode voltage drop. However, the liquids are assumed to be isothermal at the
operating temperature and are not part of the thermal domain. The liquids are split into a fixed
part and a moving part that follows the ledge.

The latest generation ANSYS elements are used for the calculations. These elements are
compatible with the high-performance computing solvers, including the Distributed Domain
sparse Solver (DDS) that was employed to solve the full quarter model. The finite element mesh
can be linear or quadratic, which enables a tetrahedral mesh to be used with excellent accuracy
for the entire calculation domain with the exception of the ledge and the moving liquids, given
that these need to be structured as required by the ledge repositioning logic.

569
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

3.1 Ledge Repositioning Logic

During preprocessing, auxiliary matrices and local coordinate systems are built and stored. This
data structure is reused for every ledge moving iteration thus considerably reducing the overall
time spent accessing the ANSYS database over the course of the global problem solution. The
data structure is built in such a way that it can be easily expanded. Several additional controls
were implemented to overcome convergence difficulties with a particularly challenging lining
design, including:

The introduction of different ledge displacement relaxation factors by ledge zone (sidewall/end
wall and lower/upper portions);
The usage of an absolute displacement cut-off during a moving iteration, defined by ledge zone
(sidewall/end wall and lower/upper portions).

The moving logic was generalized to handle either linear or quadratic elements as well as an
arbitrary number of elements through the ledge thickness. The construction of the auxiliary
matrices is based on the connectivity of the finite element mesh instead of the spatial position of
the nodes, which makes the routines generic to any geometry with a structured (but not necessarily
uniform) ledge mesh. This feature can easily handle dissimilar side and end lining profiles – refer
to Figure 7 – as well as arbitrarily shaped ledge profiles.

Figure 7. Structured, non-uniform ledge FE mesh originating from dissimilar side and end
lining profiles. Left: as seen from the inside of the cell. Right: as seen from the outside of
the cell looking up.

As the nodes within the ledge are repositioned to obtain the bath melting temperature at the
interface with the liquids, the nodes in the moving liquids are also repositioned to maintain
element aspect ratios within appropriate limits in order to obtain an adequate current distribution.
An example for a quadratic mesh is shown in Figure 8.

Figure 8. Initial and final 2nd order FE mesh for ledge and moving liquids. Left: initial
mesh. Right: converged ledge profile mesh.

570
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

The convergence of the ledge profile is based on the infinite norm of the temperature difference
with the bath melting temperature and/or on the infinite norm of the iterative ledge displacement.

As previously mentioned, an important aspect of this work is that the ledge, the cathode and the
metal pad have independent meshes that are reconnected by contact elements. This allows the
ledge to freely move on the cathode surface irrespectively of the cathode topology. The ledge
displacement outwards (i.e., towards the center of the cathode panel) is constrained only by the
external surface of the moving liquids zone and, inwards (i.e., towards the potshell), by the cell
lining geometry.

3.2 Heat Balance Solution

Dupuis’s approach [3] is used to converge the heat balance solution. This involves iteratively
solving the thermal-electrical problem until the heat imbalance, i.e. the difference between the
cell internal heat and the total heat losses, is less than the convergence criterion AND the ledge
profile has converged. This scheme assumes a fixed bath chemistry. Therefore, for each global
iteration, a new thermal-electrical solution must be found, the voltage drops are obtained from the
finite element solution, a new cell internal heat is calculated, the cell heat losses are integrated
over the control volume and the convergence metrics are computed. If the global problem has met
the aforementioned criteria, the solution is done and the postprocessing is performed. Otherwise,
a new operating temperature is calculated, the ledge profile is repositioned, heat transfer boundary
conditions are updated and a new global TE solution is computed.

The calculation of the cell internal heat is based on the APDL implementation of the steady-state
cell voltage and heat balance routines used in the dynamic cell simulators ARC/Dynamic and
Dyna/MARC [7].

A generic iterative heat balance convergence scheme was implemented to predict both the
superheat and the ledge profile at the same time, which proved to be slow but robust – refer to
Figure 9. A standard Newton-Raphson convergence method – typically applicable for well-
behaved problems – is also available. Nevertheless, it must be stressed that it showed a tendency
to diverge for challenging lining designs, such as the one previously mentioned in section 3.1.

Figure 9. Global heat balance and ledge profile convergence example. Left: superheat and
heat imbalance, Right: ledge convergence infinite norms.

3.3 Standardized Output

The detailed heat losses reported in the standard output file are independent of the heat losses
calculated for the complete cell used for converging the heat balance. The definition of these

571
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

surfaces is entirely up to the User, which is particularly useful when planning a measurement
campaign for model validation. The output surfaces can therefore be defined to match what can
be measured in practice and refined to the level of details appropriate for the task at hand.

In the standard output file previously shown in Figure 6, the [Component]column corresponds
to a named selection on which the heat flow is integrated, except for two lines:
Heat leaving cathodic CV
Refers to the heat flow leaving the defined cathodic control volume. This contribution is used
in the cell heat balance;
Joule Heat outside cathodic CV
Refers to the Joule heat generated in the elements outside the control volume. This heat input
is not considered in the cell heat balance.

In this work, the control volume boundary is at the end of the cathodic flexes and there are no
elements outside the cathodic control volume. The anode rod elements located outside of the
enclosure, on the other hand, are excluded from the control volume.

4. Model Results with the Standard Ledge Topology

Within the context of this work, the demonstration model previously used by Dupuis (refer, for
example, to [2]) was rebuilt in the modernized environment with the standard ledge topology,
shown in Figure 8. In this configuration, the ledge toe moves parallel to the cathode surface and,
in order to allow for a direct comparison between the two (2) distinct approaches, a linear mesh
was used.

The heat balance problem was solved with both platforms using the same material properties,
boundary conditions and key simulation parameters – the latter of which shown in Table 1.

Table 1. Comparison case main process parameters.

Parameter Unit Value
Cell amperage [kA] 300.0
Excess (dissolved) [AlF3] % 10.825
Dissolved [Al2O3] % 2.5
[CaF2] 3.0*
[LiF] 0.0
[MgF2] 0.0
External voltage mV 200.0
ACD cm 5.0
Anode cover thickness cm 16.0
* Even though not representative of most aluminum reduction cell technologies, the reported CaF 2 content
was kept as indicated in order to ensure consistency with previously published data.

572
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

The main heat balance and computational results are summarized in Table 2.

Table 2. Comparison case main heat balance results.

Result Original Hatch Unit
Dupuis Modernized
Operating temperature 974.72 974.84 ºC
Liquidus temperature 967.99 967.99 ºC
Superheat 6.73 6.84 ºC
Cell internal heat 639.5 642.4 kW
Anode heat losses 236.3 234.4 kW
Cathode heat losses 403.5 408.0 kW
Total cell heat losses 639.8 642.4 kW
Heat imbalance 0.037 0.000 %
Anode voltage drop 334.3 339.3 mV
Cathode voltage drop 287.2 290.4 mV
Ledge infinite temperature norm* 20.8 5.2 ºC
Ledge RMS temperature norm** 0.15 0.013 ºC
Ledge infinite displacement norm*** 1.0 0.1 mm
* Refers to the maximum absolute computed difference with the bath melting temperature for an individual
node at the ledge-liquids interface.
** Refers to the root mean square difference computed with the bath melting temperature for all nodes at
the ledge-liquids interface.
*** Refers to the maximum absolute computed iterative displacement for an individual node at the ledge-
liquids interface.

The converged ledge profiles are compared in Figure 10, with cyan being for the original Dupuis
methodology and purple for the modernized version. The profiles are almost identical at the
sidewall, but the ledge toe in the modernized version is longer in the corner.

Point “O”

Original Dupuis (2 linear elements across ledge thickness)

Modernized (5 linear elements across ledge thickness)

Figure 10. Ledge profiles for the comparison case. Left: sidewall. Right: corner section at
45.

573
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

It was found that the ledge toe is limited in the corner of the original version due to the
construction of the cathode block mesh. This is shown in Figure 11 (Left), where the green ledge
is constrained by the dotted white line and, as a consequence, the ledge is prevented from moving
beyond Point “O” – refer also to Figure 10 (Right). This constraint does not exist in the
modernized version shown in Figure 11 (Right) due to the usage of contact-target elements
between distinct parts.

Point “O”

Figure 11. Corner ledge profile meshes for the comparison case. Left: corner ledge
movement is limited by the cathode panel-to-ledge mesh connectivity of the original
Dupuis model. Right: ledge movement not hindered by the cathode panel mesh of the
modernized version.

For the modernized model, the ledge was compared at the same superheat for the linear mesh and
a quadratic mesh with only two (2) elements through the thickness. The quadratic mesh was
previously shown in Figure 8. The obtained profiles are essentially identical, as shown on Figure
12 (Left) for the sidewall, and on Figure 12 (Right) for a cut through the corner. This built-in
capability allows the User to easily perform mesh independence tests to determine the appropriate
discretization scheme for the task at hand.

Modernized (5 linear elements across ledge thickness)

Modernized (2 quadratic elements across ledge thickness)

Figure 12. Linear and quadratic ledge profiles for the comparison case. Left: sidewall,
Right: corner section at 45.

574
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

For reference, the machine used to solve the modernized models is a 28 core Intel(R) Xeon(R)
Gold 6240 CPU @ 2.60 GHz with 128 GB of RAM. The ANSYS/WB database processing was
done using 4 cores in Shared Memory Processing mode (SMP) while the solution was performed
using the ANSYS sparse Distributed Domain Solver (DDS) with all 28 cores. The maximum
memory used was 30 GB of RAM. The elapsed time is detailed in Table 3.

It can be seen that the modernized approach using auxiliary matrices is very efficient since most
of the elapsed time is spent computing finite element solutions. In fact, the total elapsed time
spent to reposition each and every node of the ledge (including the redefinition of heat transfer
boundary conditions) was less than ten (10) seconds per global solution iteration.

Table 3. Modernized full quarter heat balance and ledge convergence computational data.
Elapsed time Elapsed time
Task (min) (%)
Process ANSYS/WB database 3.9 2
Solve the global heat balance problem 172.5 98
Elapsed time spent moving ledge 3.5 2
Elapsed time spent preparing solution 11.4 6
Elapsed time spent computing solution 143.8 82
Elapsed time spent postprocessing solution 1.6 1
Elapsed time spent post-processing model 12.2 7
Total 176.4 100

Parameter Value (-)

Number of degrees of freedom 534 079
Number of global iterations 64

For the original methodology, which involves building the model from scratch, the elapsed time
breakdown in shown in Table 4. Two-thirds of the time is spent in the preprocessor building the
model, moving the ledge and updating the boundary conditions. The machine used to solve this
model has 2 cores Intel i5-6300U CPU @ 2.40 GHz processors used in Shared Memory Parallel
(SMP) processing.

Table 4. Original full quarter heat balance and ledge convergence computational data.
Elapsed time Elapsed time
Task (min) (%)
Elapsed time spent pre-processing model 58.9 65
Elapsed time spent solution - preprocessing 0.2 0
Elapsed time spent computing solution 29.7 33
Elapsed time spent solution - postprocessing 0.0 0
Elapsed time spent post-processing model 1.8 2
Total 90.7 100

Parameter Value (-)

Number of degrees of freedom 162 506
Number of global iterations 17

5. Novel Topology

The Authors encountered a challenging industrial lining design that would lead the standard ledge
repositioning algorithm to diverge. In this specific design, the ledge tends to grow far in the corner
and then melt back at the sidewall and at the end wall. Near the corner, the side ledge and the end
ledge move in zones that have very low thermal gradients, such that one node would move

575
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

towards the wall and the neighboring node would move towards the cell center, creating “ledge
spikes”. Unfortunately, the global solution diverges once said spikes were formed.

The cause of the problem is thought to be a combination of cold cathodes, a cold corner and a
very insulated end wall. After many trials and refinements to the controls of the allowable ledge
displacement (once again, refer to section 3.1), further analysis revealed that it would never be
possible to converge the ledge with the standard algorithm since there would always be at least
one ledge slice that has two possible solutions for the position of the ledge toe. This is
schematically shown in Figure 13.

Figure 13. Challenging corner design for the standard ledge repositioning algorithm. Left:
temperature isosurfaces in the cell corner, ranging from white (very cold) to blue (cold), to
hot (red), Right: ledge profile at slice A-B.

A new ledge repositioning topology was devised in order to allow the ledge to move orthogonally
to a set of arbitrary planes, as shown in Figure 14 (Right). The initial solution is performed with
the elements on the cathode surface as thermal-only frozen ledge– see Figure 14 (Left). Once the
ledge thickness on the cathode surface has been reduced to a prescribed minimum, the elements
are switched to liquid metal, therefore being now able to conduct electricity between the
remainder of the metal pad and the cathode blocks surface. If the element surface temperature is
under the bath melting temperature, the stack of underlying moving elements is turned back to
thermal-only ledge properties. The applied thermal boundary conditions are updated at every
ledge moving iteration depending on the element status, while the physics of the contact-target
element pairs used to connect the different parts is modified accordingly.

Figure 14. Initial FE mesh and ledge displacement direction for novel ledge topology. Left:
initial mesh, Right: ledge displacement direction.

576
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

The construction of the ledge repositioning auxiliary matrices is identical to the standard topology
since the data acquisition is based on the finite element connectivity of the structured ledge and
moving liquids. Additional information was added in the existing data structure to track the
specific information required to process the ledge elements on the cathode surface.

The fictitious lining design used for demonstration purposes was modified with adjustments to
material properties and boundary conditions to obtain issues similar to those encountered with the
real-life stubborn lining design. The modifications made to the original lining design were devised
to obtain a cold cathode and a hot end wall. For the sake of simplicity, the problem was solved
only for the ledge profile at a constant superheat of 6.5 ºC.

The obtained ledge convergence curves are shown in Figure 15. The standard ledge topology does
not converge and oscillates up to the prescribed maximum number of global iterations, while the
novel topology converges within 33 iterations.

Figure 15. Problematic lining ledge convergence curves. Left: standard topology, Right:
novel topology.

The resulting ledge profiles are compared in Figure 16. Even though the whole ledge should have
disappeared at the end wall – see Figure 16 (Right) – a “ledge spike” appears for the standard
topology in a region of low thermal gradient – Figure 16 (Left). This is due to the fact that the
current flow into the center of the block is low as a result of the split bar design with castable in
the middle of the slot. Also, the thermal flux extracted by the collector bars is the lowest near the
symmetry plane.

Figure 16. Problematic lining ledge profiles obtained with both standard and novel ledge
topologies. Left: non-converged ledge profile for the standard topology, Right: converged
ledge profile for the novel topology.

577
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

6. Discussion

The novel topology allows solving problems that cannot be converged by the standard approach,
particularly when the lining design results in zones of low thermal gradient on the cathode blocks.
The ledge toe profile calculated by both methods differs – it is clear from Figure 17 that the
standard topology (in purple) produces elongated elements at the ledge toe while the novel
topology (in cyan) results in a smooth and thin layer of ledge at the same location.

Novel topology (5 linear elements

across ledge thickness)

Standard topology (5 linear

elements across ledge thickness)

Figure 17. Comparison of ledge profile elevation with standard and novel ledge topologies:
ledge toe behavior.

While the ledge profile solution appears better in elevation, one of the drawbacks of the novel
topology is the “checkerboard” pattern of the ledge toe introduced by the fixed horizontal
discretization of the ledge or, in other words, due to the fact that the ledge is only allowed to move
vertically in this region. This is shown in Figure 18 for the standard lining design. Note that this
checkerboard pattern effect can be reduced (but not entirely avoided) by refining the FE mesh of
the ledge over the cathode panel and by using a different meshing scheme in the corner.

Novel topology (5 linear elements across ledge thickness)

Standard topology (5 linear elements across ledge thickness)

Figure 18. Comparison of ledge profile with standard and novel ledge topologies:
checkerboard pattern effect for novel ledge topology. Left: isometric view, Right: plan
view.

578
 TRAVAUX 49, Proceedings of the 38th International ICSOBA Conference, 16 – 18 November 2020

7. Conclusions

A modernized approach to thermal-electrical finite element modeling of the cell heat balance
problem has been developed in the commercial software package ANSYS 2020 R1. The new
workflow is based on a combination of an interactive ANSYS/Workbench session, User-edited
input text files, and generic APDL scripts running in the background. The main limitations of the
original approach were eliminated and the use of the most recent ANSYS element technology
enables leveraging the high-performance parallel computing solvers, including the sparse
Distributed Domain Solver (DDS).

The modernized model was compared to the original approach for a complete quarter cell heat
balance and ledge profile solution and results were found to be very similar. Convergence of the
ledge profile was obtained for a linear and a quadratic mesh with (virtually) identical results.

A novel ledge topology was developed based on the displacement of the ledge nodes orthogonally
to any arbitrary set of planes. The new topology allows the computation of otherwise impossible
to converge problems.

Further work includes experimental validation of the heat balance model and assessment of the
ability of the novel ledge topology to accurately predict the position of the ledge toe.

8. Acknowledgements

The authors would like to genuinely express their gratitude to Jack Chapman from Hatch, whose
invaluable contributions made this work possible.

9. References

1. Marc Dupuis & Imad Tabsh, Thermo-electric coupled field analysis of aluminium reduction
cells using the ANSYS parametric design language, Proceedings of the ANSYS Fifth
International Conference, Vol 3, 1991, 1780-1792.
2. Marc Dupuis, Thermo-electric design of a 400 kA cell using mathematical models: a tutorial,
Light Metals 2000, 297-302.
3. Marc Dupuis, Computation of aluminium reduction cell energy balance using ANSYS® finite
element models, Light Metals 1998, 409-417.
4. Marc Dupuis, Computation of accurate horizontal current density in metal pad using a full
quarter cell thermo-electric model, Proceedings of CIM 2001, 3-11.
5. Marc Dupuis, How to limit the heat loss of anode stubs and cathode collector bars in order to
reduce cell energy consumption, Light Metals 2019, 521-531.
6. Alexander Arkhipov et al, Review of thermal and electrical modelling and validation
approaches for anode design in aluminium reduction cells, Proceedings of the 36th Conference
and Exhibition of ICSOBA, Belém, Brazil, Brazil, 29 October - 1 November, 2018, Travaux
47, 589-605.
7. Imad Tabsh, Marc Dupuis and Alexandre Gomes, Process simulation of aluminum reduction
cells, Light Metals 1996, 451-457.

579