Page 2

Page 1

underlying wear mechanism is not completely understood. In addition,

component wear in industrial mills is not a well characterised process,
and appropriate and accurate data to calibrate and validate numerical
models is sparse. In spite of these complications, the outputs of the
DEM have been coupled with relatively simplistic wear models with
promising results.

SAG
2006
DEPARTMENT OF MINING ENGINEERING

This paper addresses two major issues: the influence that the lifter
profile has on the charge behaviour within an industrial mill and the
development of a structured methodology to evolve lifter profiles within
the framework of the DEM.

UNIVERSITY OF BRITISH COLUMBIA

A STRUCTURED APPROACH TO MODELLING SAG MILL LINER

WEAR NUMERICAL MODELLING OF LINER EVOLUTION
Andrew McBride and Malcolm Powell
Mineral Processing Research Unit, University of Cape
Rondebosch, 7700, South Africa, [email protected].

Town,

ABSTRACT
A numerical framework to predict the three-dimensional evolution of the
liner profile within a rotating industrial SAG mill is presented. The need to
understand the manner in which the lifter profile develops is emphasised
by comparing two discrete element method (DEM) simulations with
measured new (i.e. unworn) and worn lifter profiles. The DEM informs
the subsequent wear model as to the loading conditions on the shell of
the mill. The wear model relates the loading conditions on the shell to
material loss. The lifter profile is then evolved and forms the input for the
next DEM simulation step in the wear analysis cycle.
INTRODUCTION
Predicting the optimal spacing between the lifters and the evolution of
the lifter profile during the course of the lifters useful lifespan will greatly
assist in the design of industrial mill. Numerical methods, such as the
discrete element method (DEM) (Cundall & Strack, 1979), coupled with
powerful modern computers allow the dynamic behaviour of the mill
charge to be approximated subject to various assumptions regarding the
shape of the charge particles (usually assumed spherical) and the
manner in which they interact. Predicting the evolution of the lifter profile
is a far more complex task than predicting the charge profile as the

The influence of the lifter profile on the motion of the charge has been
investigated by several authors (see e.g. Cleary, 2001a). The current
authors and co-workers compared the trajectory of a single particle
within the bulk charge of an experimental mill with DEM predictions for a
wide range of lifter profiles at varying mill speeds (Govender, 2005).
The small experimental scale of this investigation renders it
questionable to extrapolate all the results to a full scale industrial mill.
Cleary (2001a) investigated the influence of lifter shape on charge
motion for a 5m diameter SAG mill. The current work uses a well
characterised 8m diameter industrial SAG mill, treating competent
silicate ore, as the base case for the investigation. Considerable lifter
profile and charge size distribution data has been gathered over the
useful lifespan of the lifter. The details of the mill and charge are given
in Table 1. An in-house sampler was used to collect a 1000kg
representative sample of the mill charge. This measured size
distribution provides the charge size distribution for the DEM simulation.
DEM SIMULATIONS
Due to the large computational overhead (i.e. the amount of processor
time it takes) of DEM simulations, it is necessary to truncate the size
distribution to reduce the particles to a manageable number whilst still
attempting to capture the salient features of the bulk motion. As the
small particle sizes are orders of magnitude less in size than the larger
ones, they are probably too small to contribute in any significant manner
to the forces giving rise to liner wear, and they are removed from the
size distribution. In addition, the stable time-step size (the theoretical
maximum time-step required to accurately integrate the individual
particle's motion) is directly related to the smallest particle size.
Truncating the size distribution at 22mm was seen as acceptable as the
packing of the particles between the lifters, and the cataracting stream
is still accurately represented. In Table 1 the total number of particles
per one m3 is converted to a number per 0.25m slice of the mill length.
Truncating the size distribution reduces the number of particles from

Page 3

Page 4

over million to less than 40 000, thereby significantly reducing the

computational overhead. To put this in perspective, a 3.6GHz processor
with 2Gb of RAM takes approximately 3 weeks to conduct a full
simulation (4 revolutions) of this truncated charge. It is thus critical to
truncate the mill load but without adversely affecting the charge kinetics
and energy distribution profiles significantly. Assessing the influence of
charge size truncation is part of the ongoing scope of this study of liner
wear prediction.

The profiles were obtained using the UCT profiler gauge,

Chandramohan & Powell (2006). An image of the mechanical gauge in
use and the measured profiles over the life of a lifter are shown in
Figure 2 (see next page).

Table 1 Mill parameters.

mill diameter, m
shell
8.00 inside liner
7.801
speed, rpm
rpm
11.4 % crit
75
mill filling, %
total,ave.
40 balls
14.9
number of rows of lifters
52
feedrate, tph
230 (average)
gold, quartzite
ore
sg
2.7
size distribution, % retained
number of balls
number of rocks
3
3
size, mm
balls
ore
exp./m
sim.
exp./m
sim.
212
0.00
180
0.10
5
9
122
0.10
5
9
90
39.9
0.46
12
121
4
7
63
47.3
1.46
39
247
31
60
45
9.4
6.06
22
328
370
708
31.5
2.6
15.17
17
290
2623
5014
22.4
0.8
21.22
10446 19972
16
21.87
29740
0
11.2
15.66
60223
0
8
6.98
75911
0
5.6
3.52
114453
0
total numbers
90
986 293812 25779

Two measured lifter profiles are simulated. Figure 1a and b depict the
new lifter profile while Figure 1c and d shows the measured profile after
a period of 211 days. The resulting DEM representation of the mill was
directly generated from the CAD model of the mill that took as its input
the measured lifter data. It was seen as important to capture the
significant detail of the measured profile in the DEM model, as is evident
in Figure 1.

Figure 1 a) Mill with new lifters. c) Mill with worn lifters. b&d) Close
up of DEM model of new lifter profile. d) Close up of DEM model of
worn lifter profile.
A 0.25m slice of the mill containing approximately 38 000 particles was
simulated using the commercial DEM package EDEM from DEM
Solutions. Periodic boundary conditions in the longitudinal direction (i.e.
parallel to the axis of mill rotation) were implemented. These boundaries
allow the particles that cross through the boundary to reappear at the
opposite boundary. Thus, the natural three-dimensional packing, as
opposed to the altered packing that would occur if a rigid boundary was
used, and the particle population is preserved. Utilising a periodic
boundary, the 0.25m slice can accommodate the largest particles, but
should ideally be wider than this. A wider slice, with the resulting larger
number of particles, will be simulated using a parallel processing
environment in the near future. A three-dimensional slice has been
shown to be a far better approximation to the full three-dimensional

Page 5

Page 6

behaviour and in particular the particle packing than the associated twodimensional model (Cleary, 2001a).

Outputs from the DEM simulations for the new and worn lifter profiles
are shown in Figure 3 (see next page). Figure 3a and d show timeexposure (streak) images of the charge motion, coloured by the
magnitude of the velocity. Streak images give a far more detailed
overview of the charge motion than conventional snapshot images of
the charge at an instant in time (see Figure 3b and e). The new, more
aggressive lifters increase the angular position of the shoulder causing
more particles to enter the cataracting stream. The velocities attained
by these particles are higher than for similar particles in the reduced
cataracting stream produced using the worn lifters. Figure 3a clearly
indicates however that many of these high velocity particles in the
cataracting stream impact with the shell of the mill in the toe region.

300

New liner
250
17-Jul-05

height, mm

200

15-Feb-06
01-Mar-06

Simulated worn liner

150

06-Apr-06
11-Apr-06

100

50

0
0

50

100

150

200

250

300

350

400

450

500

horizontal position, mm

Figure 2 Liner profiles a) the gauge in situ b) progressive profiles.

The primary DEM input parameters are listed in Table 2. The widely
used simplified Hertz-Mindlin contact model is employed to approximate
the contact between the spherical DEM particles (see e.g. the review
article by Mishra, 2003). The magnitude of the shear modulus for both
the ball and ore particles has been decreased by a factor of 100 to
increase the minimum critical time-step size, whilst attempting not to
invalidate the assumptions of the contact model. The time-step duration
of approximately 6.2e-5s (30% of the Rayleigh time-step duration) is
fixed for the duration of the analysis. The correct number of particles to
generate the required filling corresponding to the measured particle size
distribution is placed in the mill and two complete revolutions simulated
in order to mix the charge and allow the mill to enter a pseudo-steady
state. Thereafter, another two revolutions are simulated in order to
assess the effect of lifter profile on charge behaviour.
Table 2 DEM simulation parameters.
Steel
7.752x108
0.29

Steel-Rock
Gr
r

Shear modulus, N/m2

Poissons ratio

Gs
s

Coefficient of rolling friction

Coefficient of sliding friction

r ss 0.001
ss 0.3

r sr
sr

0.002 r rr
0.5
rr

Coefficient of restitution

ss

sr

0.5

0.3

rr

Rock
7.550x107
0.2
0.005
0.8
0.5

INFLUENCE OF LIFTER PROFILE ON CHARGE BEHAVIOUR

These high velocity impacts with the shell damage the liner of the mill,
resulting in peening, cracking, and sometimes fracturing of the liners.
The charge profile attained using the worn lifter is less damaging as no
particle impacts occur directly with the shell in the toe region. The
milling action has changed considerably as the liner wears. Not only has
the impact point dropped down onto the charge, but there is clearly a
smaller fraction of the charge undergoing cataracting and impacting.
The consequence of this is a shift from impact breakage to abrasion
grinding. This will result in a lowered throughput, but finer product. To
recover the throughput this variable speed mill can be sped up to 80%
of critical speed as the liner wears to this profile. In order to link this
information into the liner design, a milling model that utilises DEM
outputs is required. At the University of Cape Town a Unified
Comminution Model (UCM) is being developed that does just this,
Powell (2006).
An ability to predict the manner in which the lifter profile evolves with
time is clearly required for the proper design of lifters. Measurements,
such as those given in Figure 2 and Table 3, indicate that the new
lifters wear at a lower rate initially and then the wear rate accelerates as
they wear below a critical height that allows excessive slip of the
charge. Thus, it is critical to understand the manner and rate at which
the lifters evolve in order to optimise the mill design process.
Figure 3c and f show close up images of the particle packing structure
against the lifter. A component of this research is to investigate the
effect of lifter spacing on mill performance. The spacing between lifters
is too close in the current design resulting in the charge packing into
position between the lifters, as shown by all the balls between the lifters
in Figure 3c all having the same velocity corresponding to the mills
rotational velocity.

Page 8

Page 7
Table 3 Liner wear rates of test mill.

Date
days
17-Jul-05
0
15-Feb-06
213
01-Mar-06
227
06-Apr-06
263
11-May-06
298

lifter wear, lifter wear,

height
mm/kt
mm/day
215
88
0.106
0.595
84
0.103
0.579
56
0.108
0.606
30
0.111
0.622

liner plate
wear,
mm/kt
0.070
0.101
0.106
0.111

This forms a dead zone of zero grinding action. For the worn lifters it is
evident that there is relative motion right down to the shell plate
between the lifters, so this remains an active zone. A considerable
portion of the grinding action takes place in the shear zones between
layers of charge as the mill draws the charge upwards, so locking up
18% of the ascending charge will have a considerable influence on the
grinding rate of the mill. In related work conducted on the mill that is
simulated in the current work, van der Westhuizen and Powell (2006), it
was found that the mill throughput was severely limited when new lifters
were installed in the mill.
This worn lifter simulation is at the point at which the liner wear begins
to accelerate, see Figure 2 for the profiles. Below this there should be
considerable sliding action extending onto the liner plate, and leading to
accelerated lifter and liner plate wear.
A series of simulations for different lifter spacing ratios is currently
underway and results will be presented in a forthcoming publication.
This will present the collision energy spectra and power draw of the
various mill configurations so as to provide the first stage of the link
between liner evolution and mill performance.
PREDICTING THE EVOLVED LIFTER PROFILE

Figure 3 Images of DEM trajectory data. (ac): the outputs for new
lifters and (d-f) for worn lifters. (a & b): - Streak images. (b & e):
Snapshot images of velocity profile. (c & f): Close-up of particle
packing between lifters.

A method to predict the evolution of the lifter profile for the full scale
industrial mill described in the previous section is presented. The
prediction of lifter evolution has been considered by several authors.
The following serves as a summary of several key contributions; others
have been omitted for the sake of brevity. The essential components to
consider when assessing the relative merits of a wear methodology are:
the manner in which the lifter profile is discretised and whether this
approach is valid in three dimensions, the wear model applied, and the
smoothing algorithm used to average the wear across the lifter to
prevent the formation of local depressions.

Page 9
Cleary (1998) discretised the profile of the liner into equal sized divisions,
henceforth referred to as bins. A two-dimensional DEM mill model was
considered. The kinetic energy of the particles impacting with the mill
shell was accumulated in the bins. The Finnie wear model (Finnie, 1972),
a variant of the widely used Archards model (Archard, 1980), was used
to relate the bin data to liner wear (i.e. material loss). The wear across all
lifters was then accumulated to a single representative lifter and the
amount of wear in each bin averaged using a cubic spline smoothing
algorithm.
The later work of Cleary (2001b) extended his earlier work (1998) to
three-dimensions. The surface of the lifter was tessellated into triangular
elements (i.e. decomposed into a set of non-overlapping triangular
facets). The data accumulated in each of the elements was then
averaged over the length of the mill and smoothed. The elements were
constructed to be smaller in diameter than the smallest particle in the
charge. The amount of wear was now assumed to be due to impact and
abrasion. Impact wear was estimated using two measures; namely the
energy dissipated in the normal direction at the contact and a measure of
the excess kinetic energy. The abrasive wear was likewise estimated
using two measures; namely the energy dissipated by the tangential (to
the contact surface) and the kinetic energy of each particle involved in
the collision. As such, the Cleary (2001b) model is the most flexible (and
complex) model in the current literature. Other notable contributions
include the work of Qiu et al. (2001) and Glover & de Beer (1997).
Unfortunately, a complete understanding of the wear mechanisms in a
mill that can conveniently utilise the outputs of the DEM in a relatively
simple manner and that is parameterised using well constructed
experiments, is not available at present.
The approach taken in the current work resembles that of Cleary (1998).
Each collision event with the mill shell is recorded during the simulation
process. The resulting data file is post-processed to map all the collision
events in the mill to a representative master lifter. The master lifter is
then decomposed into a series of rectangular elements (bins) that extend
the length of the mill. The width of the elements is selected to be
significantly smaller than that of the smallest particle in the mill. The
discretised profile of the master lifter is shown in Figure 4 (see page after
next), wherein 100 bins have been used. It should be noted that in this
and subsequent images of the lifter, the horizontal and vertical scale
shown are not equal. The units are in m, but are only shown to represent
scale and do not correspond to a specific position in the mill. The curved
shell of the mill between the lifters is in addition represented as a straight
line segment. Any possible contact data can be accumulated in the bins.
The normalised variation in the number of contact events and the normal

Page 10
energy dissipated across the profile of the lifter are shown in Figure 5
(see next page) and Figure 6 (see page after next) . Archards law is
used to relate the contact data to material loss in the direction normal to
the bin. The relatively short duration of the DEM simulation
(approximately 0.5s wherein 40 000 collision events occurred) and the
finite size of the bins produces data less smooth than one would expect
in an actual mill wherein the data would be accumulated (if this were
possible) over a period of weeks to months. The procedure to smooth
the data is as follows. A cubic spline function is first used to smooth the
data profile prior to applying Archard's law. Archard's law is then applied
and the updated lifter profile obtained. A least squares smoothing spline
is then used to smooth the resulting worn lifter profile. The worn lifter
profile so obtained then forms the basis for further wear simulations.
The profile is exported in a CAD compliant format and forms the input
for the next DEM simulation.
The iterative nature of the simulation process required to predict the
wear process, coupled with the computational expense of the individual
simulations, necessitated the development of a procedure to maximise
the amount of wear predicted by each simulation without introducing
significant errors. The approach taken in this work is as follows. The
amount of wear as a percentage of the lifter volume is determined. If
this amount is below a defined threshold, the updated lifter profile is
determined but the simulation not restarted. The loading on the previous
geometry is now applied normal to the updated lifter profile and a
successive updated lifter profile determined. This process repeats until
the amount of wear exceeds the specified threshold.
CONCLUSIONS
The effect of the lifter profile on the behaviour of a full scale industrial
mill has been investigated using the DEM. Considerable changes in the
response of the mill are predicted as the lifter profiles moves from an
unworn to a worn one. It is obvious that the liner should be designed to
maximise output over the working lifespan of the liners and that a
knowledge of the worn lifter profiles would be of considerable value.
The liner wear procedure developed in this work resembles that
presented by other researchers. The primary difference in the current
approach is that the system is not binned a priori and the manner in
which the collision data and worn profiles are smoothed.
The focus of future work will be on calibrating the wear model to the
measured lifter profiles, this is being conducted in related work

Page 11

Page 12

(Chandramohan & Powell, 2006), and determining a means to critically

assess the performance of the numerical model. In addition, procedures
to expedite the simulation of the wear process need to be investigated.
The longer term goal is to use laboratory wear data, such as that being
developed by Radzisewski (2001), to fully predict the wear and evolving
profile for a greenfields application.

Figure 6 Normalised distribution of the normal energy dissipated

during collision events with the shell. Both smoothed and
unsmoothed collision data is shown. The smoothing is performed
using a least squares smoothing spline algorithm (scale in m).

Figure 4 Discretised new master lifter profile (scale in m).

Figure 7 Updated original and worn master lifter profiles (scale,

m).
REFERENCES
Figure 5 Normalised distribution of the number of contact events
occurring with the master lifter (scale in m).

1) Archard, J.F., 1980, Wear Theory and Mechanisms, Wear Control

Handbook, American Society of Mechanical Engineers, pp. 35-80.

Page 13
2) Chandramohan, R. and Powell, M.S., 2006, A structured approach to
modelling SAG mill liner wear monitoring wear, Proceedings
International Autogenous and Semiautogenous Grinding
Technology 2006, Sep. 24-27, Ed. Muller et al., Published CIM.
3) Cleary, P.W., 1998, Predicting charge motion, power draw,
segregation and wear in ball mills using discrete element methods,
Minerals Engineering, Vol. 11, Issue 11, Nov., pp 1061-1080.
4) Cleary, P.W., 2001a, Charge behaviour and power consumption in
ball mills: sensitivity to mill operating conditions, liner geometry and
charge composition, International Journal of Minerals Processing,
Vol. 63, pp 79-114.
5) Cleary, P.W., 2001b, Recent advances in the DEM modelling of
tumbling mills, Minerals Engineering, Vol. 14, pp. 1295-1319.
6) Cundall, P.A. and Strack, O.D.L., 1979, A discrete numerical model
for granular assemblies, Gotechnique, Vol. 29, pp. 47-65.
7) Finnie, I., 1972, Some observations on the erosion of ductile metals,
Wear, Vol. 19, pp. 81-90.
8) Glover, G. and de Beer, J.C.K., 1997, The application of the discrete
element method to the optimization of mill liner systems, XX IMPC,
Aachen, Sep, pp. 219-228.
9) Govender, I., 2005, X-Ray motion analysis of charge particles in a
laboratory mill. PhD Thesis, University of Cape Town.
10) Mishra, B.K., 2003, A review of computer simulation of tumbling
mills by the discrete element method: Part Icontact mechanics,
International Journal of Mineral Processing, Vol., 71, No. 1-4 , 22,
pp. 73-93.
11) Powell, M.S., 2006, The Unified Comminution Model - a
conceptually new model, Proceedings of the IMPC 2006
conference, 3-8 Sep. Istanbul, Turkey.
12) Qiu, X, Potapov A, Song, M, and Nordell, L., 2001, Prediction of
wear of mill lifters using discrete element methods, Proceedings
International Autogenous and Semiautogenous Grinding
Technology 2001, Sep. 30 - Oct. 3, Ed. Barratt et al., Published CIM.
13) Radziszewski, P., 2001,
Determining Impact, abrasive, and
corrosive contributions to total media wear, In Proceedings
International
Autogenous
and
Semiautogenous
Grinding
Technology 2001, Sep. 30 - Oct. 3, Ed. Barratt et al., Published CIM,
Vol. IV, pp. 252-259.
14) van der Westhuizen, A.P. and Powell, M.S., 2006. Milling curves as
a tool for characterising sag mill performance, Proceedings
International Autogenous and Semiautogenous Grinding
Technology 2006, Sep. 24-27, Ed. Muller et al., Published CIM.