Minerals Engineering 86 (2016) 75–95

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Comminution mechanisms, particle shape evolution and collision energy

partitioning in tumbling mills
P.W. Cleary a,⇑, R.D. Morrison b
a
CSIRO Digital Productivity and Manufacturing Flagships, 71 Normanby Road, Clayton, Victoria 3168, Australia
b
Julius Kruttschnitt Mineral Research Centre, The University of Queensland, 40 Isles Road, Indooroopilly, Qld 4068, Australia

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

Article history: A computational model for use with DEM (Discrete Element Method) is proposed for the five mechanisms
Received 30 July 2015 applying to particle comminution in AG, SAG and coarse feed ball mills. Chipping and rounding are mech-
Revised 4 October 2015 anisms that lead to preferential mass loss from corners and edges of particles and which produce shape
Accepted 3 December 2015
change for the particles as well as size reduction. These are controlled by the energy dissipation at a con-
Available online 14 December 2015
tact and the location of the contact on the surfaces of the particles. These mechanisms lead to rounding or
conditioning of angular feed particles. Single impact body breakage is a very weak contributor to overall
Keywords:
size reduction with most body breakage occurring via damage accumulation over many contacts. This
Tumbling mills
Breakage
incremental damage mechanism is inherently less inefficient but leads to size reduction from the many
Chipping weak collisions experienced by particles within the mill. Finally, size reduction from abrasion is well rep-
Rounding resented by the shear energy absorption of the particles. Particle size and shape evolution due to chip-
DEM ping, rounding and attrition is demonstrated using a well characterised pilot mill for which detailed
particle mass data is available. The relative contributions of the five mechanisms and a quantification
of the wasted energy for both AG and SAG charges and for new and substantially comminuted material
are reported. Changes in the energy spectra with the decreasing particle sizes over time are described
with increases in the fraction of collisions above the elastic energy threshold leading to faster damage
accumulation and reduced energy wastage. Finally, it is shown that the attribution of dissipated energy
between particles in a collision is material dependent with significantly more energy absorbed by rock
particles than balls. This needs to be accounted for in DEM models, particularly when attempting to
explicitly predict particle size reduction.
Crown Copyright Ó 2015 Published by Elsevier Ltd. All rights reserved.

1. Introduction example the comminution special issue edited by Cleary and

Morrison, 2008). A recent review article by Weerasekara et al.
The Discrete Element Method (DEM) is a computational scheme (2013) describes the current status of DEM modelling and the chal-
for predicting the flow of particulates. Since milling is predomi- lenges related to its use in comminution.
nately a process involving collisions between particles and One of the most important aspects of mill performance that has
between particles and the milling machine, it has been used exten- been least studied using DEM is that of particle breakage. Although
sively to investigate different aspects of mill performance. The ear- this is the purpose of milling, the use of DEM for the prediction of
liest use for modelling in SAG mills was by Rajamani and Mishra size reduction is challenging because several different modes of
(1996) who used two dimensional DEM to help design lifters by breakage are active and because the particle change (size and
predicting media overthrow. The most common use has been in shape) needs to be explicitly included in the DEM model. Com-
understanding the charge distribution, material flow patterns, monly DEM simulation is used to provide average process data
energy utilisation and power draw (Mishra and Rajamani, 1992, from which selection functions can be determined for use in tradi-
1994; Cleary, 1998, 2001a, 2001b, 2001c; Herbst and Nordell, tional population balance models (Tuzcu and Rajamani, 2011;
2001; Cleary, 2004, 2009; Morrison and Cleary, 2004, 2008; Tavares and de Carvalho, 2009). The accuracy of such models is
Cleary et al., 2008 and many authors since then including for limited by the ability to adequately characterise the collisional
environment and the limited physics captured by the population
⇑ Corresponding author. balance models. To be able to accurately predict, from first princi-
E-mail address: [email protected] (P.W. Cleary). ples, breakage, product size distribution and resident particle size

http://dx.doi.org/10.1016/j.mineng.2015.12.006
0892-6875/Crown Copyright Ó 2015 Published by Elsevier Ltd. All rights reserved.
76 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

distribution, particle breakage needs to be included directly in the as is the variation of these rates with time as the particles become
DEM model. The breakage model and the mechanisms that they rounder. The decrease in particle size over time will be shown to
implement also need to be sufficiently correct to make accurate change the collisional environment in the mill. It will be further
predictions. This avoids the averaging over the different flow con- shown that as the particles shrink, their specific collision energies
ditions within different parts of the mill that is inherent in increase leading to more collisions being above on the elastic
approaches that seek to predict particle populations using aver- threshold E0 which will lead to increasing efficiency of the incre-
aged selection functions. An approach for direct inclusion of break- mental breakage mechanism. Finally, the attribution of collision
age in DEM was first proposed by Cleary (2001b). This used a energy absorption between colliding entities will also be shown
replacement strategy where compressive or impact breakage frac- to be important and bounds on the splits required to match exper-
ture of a parent particle into finer progeny was represented iment will be established.
directly in the DEM model. This model was limited by the use of
spherical parent and progeny and the limited ability to include
2. Breakage mechanisms in mills
characterised breakage properties. Recently, a more sophisticated
variant that can represent breakage of non-round particles into
The first few minutes of operation with a fresh charge causes
non-round progeny was proposed by Delaney et al. (2010a). This
rapid rounding of initially angular particles and a relatively coarse
was specifically developed to be able to model incremental break-
product. The production rate of progeny then decreases to a more
age due to damage accumulation of particles over many collisions
or less steady state production rate of much finer progeny. The rate
and uses breakage characterisation data from tests such as the
of particle wear can be modelled based on inter-particle friction
JKDWT (JK Drop Weight Test) or JKRBT (JK Rotary Breakage Tester).
and shear energy absorption. Under suitable conditions, the num-
At SAG 2006, we reported (Morrison et al., 2006a) the simulated
ber of impacts which exceed the elastic threshold E0 allows the ore
and measured outcomes of treating a well characterised ore in a
particles to begin to accumulate incremental damage.
1.2 m diameter mill. This well instrumented, pilot scale mill at
Previously (Morrison et al., 2011), postulated five mechanisms
the University of KwaZulu-Natal had been combined with some
occurring in AG, SAG mills and coarser feed ball mills:
new approaches to ore testing to allow different modes of breakage
to be tested. It showed that autogenous mill loads of various sizes

Body breakage (single impact breakage through the particle –
and shapes could be reasonably predicted by abrasive mass loss
the traditionally conceived mechanism occurring in tumbling
proportional to the estimated frictional energy experienced by
mills but in reality infrequent in a large SAG mill and com-
each particle. However, this approach was inadequate for SAG
pletely absent in the pilot mill used here).
operation where incremental damage produces more non-trivial

Incremental damage (body breakage due to accumulated dam-
body breakage and quite different progeny size distributions.
age or fatigue from many weak collisions).
Another drawback of this initial investigation was that the parti-

Attrition or abrasion (mass loss at the surface of rounded rocks
cles were spherical in the DEM models of both the abrasion mill
as other particles slide over them or they slide over the liner).
and the pilot mill.

Rounding (preferential and higher abrasive mass loss from slid-
At SAG 2011, we extended this work (Morrison et al., 2011;
ing at the corners and edges of blocky particles).
Delaney et al., 2013) by exploring the accuracy of the breakage pre-

Chipping (loss of corners, edges and larger asperities from small
diction when using just an incremental damage mechanism. This
scale body breakage for irregular shaped or non-round
used the elastic energy of contacts to estimate the incremental
particles).
damage that controls the probability of a particle breaking and
the progeny produced. Defining E0 to be the elastic threshold
All five mechanisms have been implemented in the DEM code
energy at which damage starts to occur, it was found that a range
described in Cleary (2004, 2009). For the results reported in this
from 3.6 to 5.4 J/kg was needed to produce appropriate amounts of
paper, only the last three surface erosion mechanisms are active.
breakage. The model was able to reproduce with good accuracy
The final stage of this work where all the mechanisms are used
both AG and SAG product size distributions. In using only the
together to provide a complete representation of the breakage
incremental damage, the E0 values were found to be around an
environment experienced by the rock particles will be reported
order of magnitude lower that typically measured (Morrison
in a following paper. This should, in principle, allow direct predic-
et al., 2007; Whyte, 2005). Much of the size reduction was found
tion of the complete rock size distribution over time.
to occur from very weak collisions that removed very small mass
increments from the particles rather than by substantive body
breakage. Essentially, the incremental damage model was also 3. Implementation of size reduction in DEM
directly predicting attrition of the particles, although little correla-
tion was found between the predicted and measured fine progeny. 3.1. Determination of energy absorbed by a particle from DEM
In this paper, we identify five key breakage mechanisms and
provide details of the computational method for implementing The contact force is best represented in a collision frame with a
them. We then focus on the surface damage mechanisms by explic- normal direction and a tangent plane. This is defined locally at the
itly including the chipping and rounding of non-spherical parent contact point between the two colliding surfaces. In this frame,
particles in addition to the abrasion mechanism that has been pre- using a linear spring–dashpot contact model, the normal force Fn
viously reported by Morrison et al. (2006a). This allows us to consists of a linear spring to provide the repulsive force and a dash-
explicitly predict the size and shape evolution of the particles in pot to dissipate a specified proportion of the relative kinetic
the mill resulting from surface damage accumulated over time energy:
from their collisions. The surface mass loss rate as a function of
F n ¼ kn Dx þ C n v n : ð1Þ
energy input is calibrated from one identified rock in the first
experiment and this is then used in simulating the evolution of The magnitude of the overlaps Dx between particles is deter-
all the particles in three test cases (with differing test conditions). mined by the stiffness kn of the spring in the normal direction (in
The model predictions using only surface damage form an upper N/m). The damping term is proportional to the normal component
limit on the size of each particle at each time. The nature of the of the contact velocity vn. The normal damping coefficient Cn is
change of shape of the particles during milling is explored in detail chosen to give the required coefficient of restitution e (defined as
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 77

the ratio of the post-collisional to pre-collisional normal compo- tangential energy dissipation. This is based on the normal energy
nent of the relative velocity), and is given in Schwager and being directed into the body of the particles while the tangential
Pöschel (2007) in a more DEM friendly implementation form in energy is directed in the shear direction and will affect primarily
Thornton et al. (2013). In many DEM application simulations the asperities that contribute to the generation of friction.
including in the ones presented here, the stiffness is used as a The best test of the correctness of the energy dissipation calcu-
numerical rather than a physical parameter and is chosen to be lar- lations is to integrate these over all collisions to estimate the total
ger than the real material stiffness so as to maximise the timestep energy dissipation in the granular system over the simulated time.
and therefore minimise the simulation time. This is generally pos- For mill simulations, the power draw is accurately given by the
sible as long as the numerical contact deformations remain small product of the torque applied to the mill by all the particle contacts
and the contact force networks are not altered by the moderate by the mill rotational speed. The total energy supplied by the mill
softening of the springs. Typically, average overlaps of 0.1–0.5% is just the instantaneous power draw integrated over time. For long
are adequate to ensure a balance between computing speed and duration mill simulations that have reached steady state operation,
realistic predictions. the energy dissipated and the energy supplied by the mill should
The tangential force is given by: be equal. In practice, balances within a few percent indicate an
n X o accurate DEM simulation. When using the energy dissipated in col-
F t ¼ min lF n ; k t v t Dt þ C t v t ; ð2Þ lisions to predict energy spectra which are used as proxies for
understanding breakage or for explicit breakage prediction (as
where the vector force Ft and tangent velocity vt are defined in the we will do in this paper) it is critically important that these ener-
plane tangent to the surface at the contact point. kt is the tangential gies are accurate.
spring stiffness (N/m), Ct is the tangential dashpot coefficient and Dt The energy dissipated in a collision is the energy that is dissi-
is the integration timestep. The summation term represents an pated between the colliding surfaces of the two bodies. The final
incremental spring that stores energy from the relative tangential step in this energy calculation process is to attribute the energy
motion and models the elastic tangential deformation of the con- to each of the bodies involved in the collision. In the absence of
tacting surfaces, while the dashpot dissipates energy from the tan- better information to the contrary general practice in DEM appears
gential motion and models the tangential plastic deformation of the to be to attribute the energy equally to each body. For collisions of
contact. The total tangential force Ft is limited by the Coulomb fric- similar materials (such as rock particles with rock particles) this is
tional limit lFn, at which point the surface contact shears and the not necessarily an unreasonable assumption. For rock particles
particles begin to slide over each other. Here l is the dynamic fric- impacting the liner or with steel balls, this assumption has no par-
tion coefficient of the contacting surfaces. The tangential stiffness ticular foundation and will be a question that we start to address in
used is half the normal stiffness. this paper. This question of energy dissipation between bodies is
In terms of choice of contact model Thornton et al. (2013) critical for DEM modelling that includes breakage since it controls
demonstrated that all inelastic contact models behaved broadly the evolution of the particles. When breakage is not explicitly
as two classes depending on whether the inelasticity was plastic included then this split affects only the energy spectra and any
or viscous in nature. Both classes produce similar types of single conclusions that are drawn from these.
particle rebound kinetics at higher coefficients of restitution but
increasing different behaviour at moderate to low values. It is an 3.2. Attributing energy between mechanisms and particle shape
open question as to which class of contact model best represents attributes
collisions of different materials, but it appears less likely that
elastoplastic models will be appropriate for representing rock col- The super-quadric particle shape representation is described in
lisions which are generally not plastic in nature. For brittle rocks, Cleary (2004, 2009) which contain examples of use in industrial
much of the dissipation arises from small scale fracture at contact applications and an explanation of the importance of including
points and from spalling. More fundamental investigations of the shape within DEM simulation). A super-quadric is defined in a
contact forces for elastic and elastoplastic particles can be per- canonical frame with Cartesian coordinates (x, y, z) as:
formed using the FEM (Wu et al., 2003), but these methods are xm  y m zm
not suitable for direct body breakage predictions where large þ þ ¼1 ð3Þ
deformations, fragmentation and fragment collisions all occur. a b c
For these problems the SPH methods formulated for brittle fracture The power m is called the blockiness and controls how sharp
(see Das and Cleary, 2010 for such examples) are better suited for the corners and edges of the particle are. The semi-major axes in
determining how brittle particles break. the principle directions are a, b and c and their ratios determine
Energy dissipation during a collision can be calculated as the the aspect ratios of the particles with Axy = b/a and Axz = c/a. This
integral of the dot product of the dissipative force components shape description is particularly well suited to applications where
(in the collision frame) with the relative contact velocity vector the shape of the particles evolves over time. Since the shape repre-
(also in the collision frame). For the linear spring–dashpot model sentation is continuous, the shape parameters (m, a, b and c) can be
used in this work, the normal component of the dissipative force varied incrementally and continuously with the degree of surface
is the normal dashpot in Eq. (1). The tangential contribution is erosion closely replicating the real physical behaviour without
more complex as the calculation depends on whether the contact any artefacts arising from the nature of the numerical shape repre-
is sliding or not. If the contact is not sliding then the dissipative sentation. In contrast, alternative approaches using either polyhe-
component is the tangential dashpot in Eq. (2) while for the sliding dral shapes which have sharp or simply rounded edges or bonded
case it is the entire tangential force Ft. The normal component of spheres where the curvature of edges and corners is limited by the
the energy dissipation is the integral of normal dashpot force with radius of the sub-particles (which cannot be easily be varied), the
the normal velocity over the collision and the tangential compo- super-quadric shape easily supports continuous shape change.
nent of energy dissipation is the integral of the tangential dissipa- Some examples of equi-axed (that is ones with aspect ratios of
tive force with the tangential velocity. The total energy dissipation unity) super-quadric particles are shown in Fig. 1. For m = 2 the
in a collision is then the sum of these two components. It is com- particles are spherical (as used for the steel balls in SAG case 3).
mon practice to consider that body breakage is controlled by the For m = 3 the particles are mid-way between spherical and blocky
normal energy dissipation while the abrasion is controlled by the with quite rounded edges and corners. In contrast, for m = 5 the
78 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

The weight vector w is then normalised to give the fractions of

energy used to change each of the shape parameters of the particle
from the energy dissipated in that collision.
Since the conditioning process is strongly preferential and abra-
sion is relatively small, we have used a = 1 and b = 10 for the sim-
ulations reported in this paper. We use c = 5 so that the sum of the
energy used in the two aspect ratios is comparable to the energy
Fig. 1. Super-quadric particle shapes, (a) spherical (m = 2), (b) mid-way from used in the blockiness change. If a particle is spherical then
spherical to blocky (m = 3), and (c) moderately rounded cubical particle (m = 5). w2 = w3 = w4 = 0 and all the attrition energy is attributed to abra-
sion (and therefore goes into producing radial shrinkage of the par-
ticle). For a particle with the most extreme shape used in this paper
particles are reasonably blocky. For higher values of m the edges having m = 6 and aspect ratios of 1.6 for both intermediate and
become increasingly sharp with m = 10 giving very cubical small axes, the energy attribution weight vector is w =
particles. (0.05, 0.31, 0.31, 0.31). This means that all three shape attributes
Fig. 2 shows a schematic of the chipping and body breakage absorb the same fraction of the energy and the abrasion mecha-
energy calculation using information from the collision such as nism uses around 1/6 of this level. The energy used for changing
the contact normal N and the moment vector D from the contact each particle attribute is then converted to an explicit mass loss
point to the particle centre. These are used to apportion the normal for the particle for that collision using a calibration constant. In this
energy dissipation between chipping and body breakage (by either work, we use the same calibration value for each of the shape attri-
high energy single impact collisions such as in a vertical shaft bute calibration constants.
impact crusher or incremental damage such as in SAG mills). The Once experimental data becomes available on the detailed
fraction of the normal energy attributed to body breakage (single shape variation of particles in some form of test mill then this
impact or incremental damage) is simply the dot product D  N. data can be used to calculate suitable values for the breakage
The fraction assigned to chipping is then (1.0  D  N). model parameters a, b and c which are the calibration constants
The mass loss rate due to abrasion and rounding of each particle for each shape component. Here we use the indicative values
in the simulation is proportional to the shear energy dissipation. noted above so as to demonstrate how the model works and
Note that there is no minimum threshold energy for collisions to the quantitative predictions of particle size and shape change that
generate surface damage from these mechanisms. The mass loss it enables. These mathematical forms used in Eqs. (5)–(7) are the
from the abrasion is used to reduce the diameter of the particle, simplest ones that capture the basic relationship between the
while the energy for rounding is used to reduce the blockiness energy absorbed and the geometry of the particle. When detailed
and aspect ratios of the particles. experimental data becomes available for comparison this will
The fraction of the attrition energy attributed to each of the allow refinement of these forms as well as calibration of the
shape change components with w1 being for diameter, w2 for model parameters.
blockiness, w3 for intermediate aspect ratio and w4 for small axis The mass loss rate due to chipping is chosen to be proportional
aspect ratio is chosen according to the metric: to the fraction of the normal energy dissipation in the particle that
w1 ¼ a ð4Þ is not directed towards the centre of the particle (as described in
the last sub-section). It is controlled by the shape of the particle
  and the location of the contact. This energy component is the
10
w2 ¼ b 1 ð5Þ energy that determines the extent of the chipping of the corners
12  m
of particles. This mechanism attempts to capture the type of break-
age where an oblique impact near a sharp corner or edge is able to
w3 ¼ cðAxy  1Þ ð6Þ
produce modest body breakage that removes the corner or edge in
one or a small number of moderate sized fragments leading to a
w4 ¼ cðAxz  1Þ ð7Þ
more rounded particle. Note that there is no minimum threshold
where a, b and c are user selectable weightings, m is the blockiness energy applied for collisions to generate chipping damage. The
and Axy and Axz are the two aspect ratios of the particle. In Eq. (5) the chipping energy is attributed to the change of the blockiness and
form is chosen so that w2 is 0.0 for spherical particles which have aspect ratios in the same way as was done for the rounding, except
m = 2 and increases as the particle blockiness increases. In Eqs. (6) that the first component w1 = 0 (as there is no diameter change
and (7) the weights are chosen to be zero when the particles have produced by chipping).
aspect ratios of unity and to increase with increasing aspect ratio. Although this model is able, in principle, to predict mass loss
from the parent particles due to chipping, rounding and attrition,
it currently does not allow prediction of the fine progeny size dis-
tribution that results.

3.3. Modelling body breakage by incremental impact damage

The basic hypothesis of incremental breakage is that:

At less than some specific energy (E0), no body damage occurs

as all deformation is elastic even though the particle surface
may be abraded to some extent (Larbi-Bram et al., 2010).

At higher energies, the particle accumulates incremental dam-
age proportional to (Ei–E0) and the probability of survival
Fig. 2. Chipping and body breakage energy calculation using the contact normal N
decreases with the level of accumulated damage until it breaks
and the moment vector D from the contact point to the particle centre to apportion (where Ei is the energy absorption by the i-th collision of the
energy dissipation. particle).
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 79

The incremental damage model then has three components: The commonly used JKMRC relationship for severity of breakage
can be modified in the same way following the strategy devised by
1. An estimate of the probability of breakage for each impact. Shi and Kojovic (2007):
2. A rule for accumulation/severity of breakage. !!
X 0
3. A progeny distribution model (or breakage function) that allows t10 ¼ A 1  exp b ½Ei  fE0  ð12Þ
the generation of a family of progeny for each breakage event. i

where t10 is the fraction of the mass of the original particle which
Morrison et al. (2007) reported on various approaches to mod- will pass through an aperture of 1/10 of the original particle size
elling incremental damage for this ore type. A satisfactory model – after the impact event. The analysis by Shi and Kojovic (2007)
for the probability of breakage, degree of breakage and likely pro- did not contain the factor f as almost all of their data was for single
geny size distribution was developed based on the standard JKMRC impact tests at much higher energies. This factor will only be
impact breakage model (Napier-Munn et al., 1996) and the work of important for multiple, low energy impacts. For high energy
Vogel and Peukert (2004) with modifications suggested by Shi and impacts, it will be absorbed into the fitted parameters of Eq. (10).
Kojovic (2007). The same parameter b0 is used in both Eqs. (11) and (12) for this
In this model, the probability of survival after a series of k equal model. The degree to which this is correct is not currently fully
impacts is predicted by clear. There can also be variation in b0 with particle size. This was
S ¼ 1  expff mat  x  kðW m;kin  W m;min Þg ð8Þ addressed at some length also by Shi and Kojovic (2007) but not
all ores exhibit this variation. See Morrison et al. (2007) for further
where discussion of these issues.
S = breakage fraction, Eq. (12) provides an estimate of the severity of breakage in
Wm,kin = single impact energy (J), terms of the t10 which will occur after multiple impacts which
Wm,min = no damage energy (J), cumulatively lead to body breakage. It does imply a simplification
x = diameter of particle (m), as minor damage will probably also occur during each such event
k = number of (equal) impacts, and some of this will be accounted for by the surface damage mod-
fmat = material parameter. els. However, when the probability of breakage exceeds a suitable
trigger energy level, a particle is assumed to break. This is of course
The likely severity of breakage can be estimated in terms of the a random distribution which matches the experimental distribu-
net energy absorbed by the body of the particle. tion. The net accumulated energy provides an estimate of the
The Vogel and Peukert (2004) equation provides a useful model actual t10 at which breakage occurred.
to test. In slightly modified form, the probability of selection S for In a single impact event, t10 is predicted as a function of Ecs
breakage after i events (that is, E–E0). It seems reasonable to assume that the sum of the
!
X 0
net crack propagation energy should generate a similar degree of
S ¼ 1  exp b ½Ei  E0  ð9Þ new surface area. Therefore in the multi-event case, we expect
i each event which causes damage to continue to propagate cracks
where E0 is the threshold energy per unit of particle mass (in J/kg) until some fraction – say f – of E0 is reached. However, this raises
below which the particle does not accumulate any impact damage. the question as to whether the probability of actual breakage
Note that [Ei–E0] may not be less than zero and b0 is a material should also increase in proportion to the degree of damage which
parameter which may contain a particle size component. For more has been acquired to that time. Hence, it seems likely that the E0
detail, see Morrison et al. (2006a). which we infer from multiple events may already contain this fac-
The probability of survival after i events is simply (1  S). tor. Decoupling these effects would require an experiment which
! measures the actual stress as it is applied as well as measuring
X 0 multiple particles to estimate the probability of actual breakage.
PðSurvival=iÞ ¼ exp b ½Ei  E0  ð10Þ
i
This might require two parameters but it would be reasonable
for the final stress to be a fraction f of the stress required to trigger
Whyte (2005) carried out a series of preliminary tests on crack extension. The authors are not aware of any published
16  19 mm and 22  26 mm size fractions of a broad rock size reports of such an experiment. However, this line of reasoning does
distribution at four energy levels. These provide a manageable justify the use of Eq. (12) with f set to unity to estimate the degree
sub-set to test the survival model. of breakage even though a proportionally greater stress than that
Breakage testing in this range of energies is unusual because producing E0 may be required to trigger the damage event.
the results of single impact testing appear to become erratic. This incremental damage model is used in the DEM simulation
For the early experiments, some preliminary testing was car- to give predictions of body breakage due to damage accumulation.
ried out to determine the energy level which would almost The energy Ei is amount of normal collision energy from the DEM
certainly cause breakage in a single impact. Once this level collision model that is directed into the body of the particle (as
was established, separate sets of particles were subjected to described in Section 3.2 and recalling that energy is also directed
successive impacts at fractions of this energy level. However, into the surface erosion mechanisms as well). The probability for
the progeny where more than a few impacts were required breakage to occur in the simulation is then determined from Eq.
for breakage tend to be clustered around a t10 value of 5–6. (9) and the resulting progeny from such a breakage event are pre-
Hence, it may well be that after the requirement for damage dicted using the JKMRC breakage map (Narayanan and Whiten,
is exceeded; the cracks continue to propagate until internal 1988).
friction equals the driving force suggesting a modified form The size distribution of the progeny particles in the DEM simu-
for Eq. (10) where f is the fraction of E0 at which the cracks lation is calculated using a spline fitted version of Eq. (12):
will stop propagating. !!
X X
0
PðSurvival=iÞ ¼ expðb ½Ei  fE0 Þ ð11Þ tðkÞ ¼ AðkÞ 1  exp bðkÞ ½Ei  fE0  ð13Þ
i i
80 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

where A(k) and b(k) are coefficients for a set of k = 6 size classes et al., 2013) and double roll crusher (Weerasekara et al., 2013),
from t2 to t75, which have been determined from drop weight tests. in a tumbling mill (Cleary, 2001a) and for a broad range of crushers
The method of Narayanan and Whiten (1988) can then be used to in Cleary and Sinnott (2015) and Sinnott and Cleary (2015).
generate a full size distribution for the progeny. The method uses A more sophisticated variant that can represent breakage of
a matrix which relates other degrees of breakage in a particular non-round particles into non-round progeny was proposed by
event to the t10 parameter and then links those into a cumulative Delaney et al. (2010a). This was specifically developed to be able
size distribution using a cubic spline function. It is well proven to model incremental breakage due to damage accumulation of
and in common use in comminution models (Napier-Munn et al., particles over many collisions and uses breakage characterisation
1996). In terms of implementation in DEM, when a particle break- data from tests such as the JKDWT (JK Drop Weight Test) or JKRBT
age event occurs, first the reduced size of the parent particle is cal- (JK Rotary Breakage Tester). The resulting set of progeny particles
culated by extrapolation from the predicted t(1) and t(2) values. The (from Eq. (13)) are packed into the volume of the parent particle.
model then iteratively determines the size of the next progeny par- Details of packing of super-quadric particles and methods for den-
ticle by using the left over mass to calculate t(next) which is then sely packing these can be found in Delaney and Cleary (2010b). The
substituted into the size distribution curve. The progeny are there- progeny generation, packing and replacement of the parent are
fore estimated as a sequence of particles which exactly duplicates performed in the DEM simulation as an instantaneous event since
the splined, cumulative weight distribution (given in Eq. (13)) sub- the timescale for a particle to break is substantially shorter than
ject to an estimate of the largest particle in the progeny list based any of the flow related timescales. The fraction of the mass of the
on t(1) and t(2). fracturing parent particle that is in the progeny generated is prin-
The progeny model has been calibrated for breakage energies cipally controlled by the choice of the minimum fragment size.
E0 > 0.3 kW h/t. Use at lower breakage energies can cause the dis- Packing fractions as high as 97% can be achieved when the mini-
tribution curve to give physically unrealistic progeny sizes that mum resolution size is very small compared to the size of the frac-
are larger than the total remaining mass that is available for this turing particle. More commonly a larger fragment size limit is used
and smaller progeny. In such cases, an estimate of the reduced par- and resolved mass fractions are then more typically in the range of
ent size can still be determined from interpolation from t(1) and 70–90%. It should be noted that this is not a restriction on the DEM
the original parent size. breakage model but rather a reflection of the usual DEM require-
Eq. (13) is a general form of breakage behaviour that applies ment to restrict the overall number of particles in a simulation
broadly to mineral rocks. The ore specific behaviour is determined so that the compute time is reasonable. Progeny that are smaller
by the parameters A(k), b(k), and E0 which need to be measured or than the resolved size limit are independently accounted for which
estimated for each ore type. Note, that in the DEM simulations it is provides a prediction of the finer size fractions of the product
usual to specify a minimum size of progeny fragment which will be material. Each breakage event, its progeny and the packing into
resolved within the simulation. However, since the full empirical the parent are performed on the fly during the DEM simulation.
size distribution for each breakage is known, the finer unresolved This allows good statistical sampling of the progeny size distribu-
progeny is known and can also be accumulated. The summation tions. Use of pre-computed progeny and packing would lead to dis-
of these unresolved fragment distributions then gives a direct pre- cretisation artefacts in the resolved progeny size distribution that
diction from the DEM of the finer breakage products generated by would degrade the predictive capability of the breakage model.
the mill. In this way, the full size distribution for both the particles This approach has been used recently for predicting breakage in
which are resolved in the simulation, and also the unresolved fine the Loveday mill (Delaney et al., 2013) and performance of a cone
material can be predicted. crusher (Delaney et al., 2015).
An important question about incremental damage that is not
fully resolved is about what fraction of the incremental damage 4. Experimental mill, ore characterisation and DEM setup
done to a particle leading up to its body breakage remains as dam-
age inherited by its progeny. There is some evidence that a sub- A substantial sample of a uniform, fine grained ore was col-
stantial fraction of the damage is used up in the fracture (Whyte, lected in conjunction with detailed surveys of an industrial SAG
2005). Whyte’s thesis implicitly assumes that each particle’s initial operation. The sample was characterised using JKDWT abrasion
condition is ‘‘undamaged” and does not depend on its history. If it and Bond ball work indices. Further repetitive DW test work was
last passed through a crusher and was broken in a single event, carried out at much lower energies to investigate incremental
then this is reasonable. A sample from a SAG mill load might how- breakage of this ore (Morrison et al., 2007; Whyte, 2005). Bbosa
ever carry some history. On this basis, we set the initial damage (2007) found that E0 ranged from 0.001 to 0.0015 kW h/t (3.6 to
energy for a new progeny particle to be zero in the simulations 5.4 J/kg) for smoothed particles or about one tenth of the Whyte
reported in this paper. It is quite possible though that there is some estimate of 0.008 kW h/t for unsmoothed particles.
residual damage that should be quantified and inherited and it A sample of the ore was also tested in a 1.12 m by 0.31 m long
would be far from surprising if this was material dependent. semi-batch pilot mill. Key operating parameters for this mill are
provided in Table 1. This mill was developed by Loveday and
3.4. Replacement methodology in DEM for body breakage Hinde (2002) to calibrate and test an abrasion based AG/SAG model
(Loveday and Whiten, 2004). This mill has been used because it
The original body breakage method was proposed in Cleary provides for rapid removal of progeny as they are generated via
(2001a). This used a replacement strategy where compressive or slots at the base of the front of each lifter. This allows fine product
impact breakage fracture of a parent particle into finer progeny to be rapidly removed minimising re-breakage and giving an attri-
was represented directly in the DEM model. The size of the pro- tion mass loss of fines per unit time. Additionally, the experiments
geny and their packing arrangement into the body of the parent were performed in a way that the mass of each particle in the
was determined by building an Apollonian packing of the parent. charge was carefully tracked through time. This is ideal for detailed
This model was limited by the use of spherical parent and progeny comparison with DEM predictions of the comminution process
and the limited ability to include characterised breakage proper- where the evolution behaviour of each particle can be explored.
ties. Despite these limitations it has been able to provide useful In addition to the mass of surviving particles, the progeny resulting
predictions of crusher performance such as the VSI (vertical shaft from all the modes of comminution were collected, weighed and
impactor) and jaw crusher (Cleary, 2009), cone crusher (Cleary sized. This provides a mass balance around the mill.
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 81

Table 1 Thornton et al. (2011, 2013) for elastic and inelastic collisions.
Key operating parameters of the pilot mill. Many examples of simulation by the code can be found in Cleary
Mill diameter 1.12 m (2004, 2009).
Mill length 0.31 m
Lifters 14 – 40  40 mm
Mill speed 30 rpm 5. Size reduction from attrition, chipping and rounding
Test 1 30% autogenous load
Test 2 20% autogenous load
In the previous work (Morrison et al., 2006a), the standard
Test 3 10% – 10% SAG rock – ball
Rock charge 0.5–5.0 kg per rock at an SG of 2.65 JKMRC abrasion test mill (300 mm in diameter and 300 mm in
Ball charge 80 mm nominal at an SG of 7.85 length) was used for the calibration of the surface attrition rate
constant. The abrasion mill was expected to be dominated by
shear, so the shear energy absorption by the particles was used
Tests were performed for 12 min for three different charges: to calibrate the degree of abrasive wear expected in the 1.2 m
diameter mill in terms of mass lost per unit of energy dissipated.
1. 30% AG charge, In this paper, we take an alternative approach of using one specific
2. 20% AG charge, particle from Test 1 to calibrate the surface mass loss rates against
3. SAG charge with 10% rocks and 10% balls. the DEM rate of shear energy absorption by the particles. In the
experiments some specific particles were painted and each was
DEM simulations were run exactly matching these conditions. carefully re-measured every 2 min so the specific mass loss rate
The simulated mill charge contains 53 rock particles for Test 1, of this and selected other particles is known to good accuracy.
46 rock particles for Test 2 and 24 rocks and 45 steel balls for Test Most importantly, the correct matching of these particles at start
3. In each test case, the exact starting mass of each rock is known and end could be made so that a reliable mass loss rate was exper-
and these were used in the DEM setup to create particles that imentally known for these particles. The largest of these identified
exactly matched these. The rock particles used in the simulation particles was used here to calculate the surface erosion calibration
were non-round which is required since the real rocks were not constants. The resulting mass loss constant was 0.07 g/J (which
round. DEM should only be used to model particles as spheres if corresponds to an energy input of 4 kW h/tonne over the 10 min
the particles are genuinely close to spherical (Cleary, 2004, of the experiment, see Morrison et al., 2006b). This is ratio of the
2009). To achieve this, the DEM particles were represented as mass lost by the largest particle in the experiment with the
super-quadrics (see Cleary, 2004 for details). The shapes of the par- amount of shear energy that it was predicted to absorb in the
ticles in the experiments were not measured so these could not be DEM simulation. This calibration constant was then used for the
used to specify the DEM particle shape. Instead, representative dis- rock erosion rates in the simulations for all three tests (since the
tributions of the shape factors corresponding to a hard ROM (run of same rock was used in each of the experiments).
mine) ore were used. The blockiness of the particles used (from Eq. The same erosion calibration constant is used for all three sur-
(3)) was from m = 3.0 to 6.0 which produces particles from reason- face erosion mechanisms since we have no current data available
ably rounded to quite angular. For Tests 1 and 2 the rocks had an to enable a different choice to be made for the different mecha-
aspect ratio range from 1.0 to 1.4 for the intermediate axis (to nisms. The faster early rate of conditioning (rounding) of the rock
major axis) and 1.1 to 1.65 for the minor axis (to major axis) aspect particles observed in the experimental work in this pilot mill
ratios. For Test 3 a more platey rock was used with intermediate (Yahyaei et al., 2014) suggests that the rounding and chipping
aspect ratios of 1.0 and a minor axis aspect ratio of 1.0–2.0. Exam- mechanisms do remove larger masses of rock per unit energy input
ples of such super-quadric particles are given in Fig. 1. The steel than is found later for the abrasion. The corners and edges of the
balls were almost spherical and so were represented as particles particles are intrinsically weaker than the middle of the particle
that were nearly spherical. The collision properties and numerical so this type of differentiation would also seem intuitively reason-
parameters used in the DEM model are given in Table 2. able. However establishing this properly would require access to
Particles are divided into size classes with classes 1–10 being more information on the masses of the rock particles at intermedi-
for balls and classes 11–20 being for rocks. The ball classes are only ate times throughout the milling process rather than just initial
present in SAG case 3. The ball classes are uniformly distributed and final values that are used in this study. This question will be
74.65 mm and 81.44 mm. Since this range of variation is not large considered in future work.
we will not report in detail about the specific ball size classes. The
rock classes 11–20 are linearly distributed. The minimum size for 5.1. Test 1: 30% AG load
the smallest rock class is 40 mm for all three test cases. The upper
limit is 190 mm for Test 1, 169 mm for Test 2 and 183 mm for Test Fig. 3 shows the size and shape evolution of the rock particles
3. due to surface erosion by the abrasion, chipping and rounding
The collision model used in this work is a linear spring–dashpot mechanisms for AG Test 1. Initially, the rocks are quite angular
model. Details of this and other collision models can be found in and have reasonable aspect ratios. Collisions occur preferentially
with corners and frictional contacts also preferentially occur at cor-
ners and edges. This concentrates the erosion at these more
Table 2
exposed locations leading to preferential rounding of corners and
Parameters used for the DEM simulations. edges. As a result the particles become increasing round and
increasingly equi-axed as well as smaller in diameter. In Fig. 3
Normal spring stiffness 500,000 N/m
Shear spring stiffness 250,000 N/m
the particles are coloured by their blockiness (the power m in Eq.
Density of rock 2650 kg/m3 (3)). The progress of the rounding can be seen as the particles
Density of ball 7850 kg/m3 become increasingly green and then blue. By 10 min of grinding,
Coefficient of restitution (rock–rock) 0.3 they are well conditioned and well rounded. Note that as they wear
Coefficient of restitution (rock–steel) 0.5
the particles better engage with the lifters and are thrown more
Coefficient of restitution (steel–steel) 0.8
Friction (rock–rock, rock–steel and steel–steel) 0.5 efficiently. So by 10 min the centre of the charge region has few
particles. Rather the particles are transported along the shell by
82 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

Fig. 3. Rounding and size reduction of rock particles in the pilot mill for a 30% AG charge for Test 1. The colours show the blockiness of the particles with red being quite
angular and blue quite rounded with green being intermediate. The frames show the charge at 3, 6, 9 and 12 min. (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

the lifters in the lower left part of the mill (as seen in this figure) 5.2. Test 2: 20% AG load
and then thrown as a relatively coherent stream from the shoulder
on parabolic trajectories back down to the liner at the impact toe. Fig. 5 shows the size and shape evolution of the rock particles
The predictions of the final masses of the charge particles for due to surface erosion by the abrasion, chipping and rounding
the autogenous Test 1 and the comparison to the experimental mechanisms for AG Test 2. The flow pattern and the particle size
results are shown in Fig. 4. Only a few of the particles have reduction predicted are similar to that found for Test 1 but with
not had some form of body breakage. These measured masses the charge being a bit more dilute because of the lower fill level
of these three particles are close to the line of particles masses used. The nature of the rounding of the particle shapes and the
predicted from the DEM. Recalling that the incremental breakage shrinkage of the particles due to the surface erosion mechanisms
is not yet active in this simulation, the requirement is that the is similar to the previous case, indicating that the particle evolution
DEM predictions match or are larger than the experimental dynamics are not particularly sensitive to the fill level at least in
masses – which is achieved by the model in this simulation. this pilot mill.
For many of the particles this leaves around 40–60% of the total The predictions of the final masses of the charge particles for
mass loss to be performed by the incremental damage model. the autogenous Test 2 and comparison to the experimental results
This contrasts with the last increment of this work (Morrison are shown in Fig. 6. Again, only a few of the particles have not had
et al., 2011) where all the damage was generated by the incre- some form of body breakage. The line of final simulation masses
mental damage mechanism but at the cost of using a lower again forms an upper limit on the experimental masses with two
value of E0. In this new work, the reduced reliance on the incre- particles matching extremely closely. There is generally closer
mental damage should enable use of larger and more physically agreement for the masses for many of the particles for this case
reasonable E0 values. than there was for Test 1. This indicates that the incremental
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 83

6000 Fig. 9 shows the change in average values of the particle shape
attributes for AG case 1 using the abrasion and rounding mecha-
nisms. Fig. 9a shows the variation in blockiness over 12 min of
5000 milling. The rate of change in blockiness is higher in the early
Experiment stages of milling and then slows between 4 and 8 min. It then
DEM decreases more slowly but linearly for the remaining milling time.
4000 The rate of change of blockiness in the later stages is 40% lower
than in the early stages. This reflects the changing distribution of
energy dissipation between mechanisms as the particles become
Mass at end (g)

rounder. Recalling that the energy used for chipping and rounding
3000
are the energy components directed away from the centre of the
particle this means that as the particles become rounder the frac-
tion of the energy being attributed to these mechanisms declines
2000
and the rate of chipping and rounding declines. This behaviour is
to be expected and the ability of the model to predict such changes
is encouraging. Fig. 9b and c shows the change in the average inter-
1000 mediate and minor axis aspect ratio. Both vary in a close to linear
manner and unlike for blockiness there is no change in the rates of
change over time. This suggests that the blockiness of the particles
0 changes faster in response to these mechanisms than do the aspect
0 1000 2000 3000 4000 5000 6000
ratios.
Inial mass (g)
Recalling that the rocks are represented as super-quadrics with
Fig. 4. Comparison of the rock size distribution for the experiment and the their blockiness uniformly distributed in the range 3 < m < 6, the
simulation for Test 1 with 30% AG charge after 12 min of grinding. intermediate aspect ratio in the range 1.0 < Aint < 1.4 and the minor
aspect ratio in the range 1.1 < Aminor < 1.65. We choose to use a
similar approach for plotting the shape attributes as used for the
damage mechanism is less significant for the lower fill level with it size distribution, that is, the cumulative probability of passing a
needing to contribute only about half of the mass loss. given value of the shape. We do this both because it is a familiar
approach for representing size distributions and because the fre-
5.3. Test 3: 10% rock and 10% ball SAG load quency distributions of the shape attributes are intrinsically noisy
because of the very small numbers of particles in these test cases.
Fig. 7 shows the size and shape evolution of the rock particles Fig. 10 shows the cumulative distribution of the blockiness for AG
due to surface erosion by the abrasion, chipping and rounding case 1 at the start of the simulation and after 12 min of shape evo-
mechanisms for SAG Test 3. The round grey particles are the steel lution caused by the surface erosion mechanisms. The initial distri-
balls. The balls tend to move on higher trajectories than the rocks bution is shown in blue with diamond glyphs at the data points.
because their smaller size and regular shape allows them to better There are no particles initially with m < 3.0 so the cumulative dis-
engage with the lifters and so be thrown more efficiently than the tribution starts here and then rising roughly linearly reaching 100%
non-round rock which tends to more occupy the middle region of at m = 5.75 which is consistent with the specified range of the
the charge. The rate of mass loss is increased by the presence of the blockiness. The probability distribution used to generate the speci-
balls. The high circular arcing flow, particularly later in the simula- fic blockiness values for the particles is uniform – which is the rea-
tion is more strongly established than for the AG charges with the son for the broadly linear trend in the distribution. The bumpiness
balls being typically on the highest trajectories on the top side of arises from the variation resulting from the small sample size of
the cataracting stream. only 53 particles. After 12 min of milling the blockiness curve
The predictions of the final masses of the charge particles for (shown by the red1 curve with square glyphs at the data points)
the SAG Test 3 and comparison to the experimental results are has moved to the left meaning that the particles have become
shown in Fig. 8. There is very good agreement for the larger rocks, rounder. The amount of movement is small for smaller values of m
which are too large for the impacts they experience to exceed their as there particles are already reasonably rounded and so less of
E0 values and so they are not able to accumulate incremental dam- the collision energy is directed into the rounding and chipping
age. Consequently, their size reduction arises purely from the sur- mechanisms. The change at the large value end of the distribution
face erosion mechanisms and the DEM model (using only the three is substantial with the most blocky particle have a blockiness of only
erosion mechanisms) is therefore able to quite accurately predict 4.5. The strong change in gradient of the distribution demonstrates
the mass reductions of these larger particles. The surface interac- the strong shape dependence of the chipping and rounding mecha-
tions with the steel balls are likely to be at a higher average energy nisms. It is also worth noting that the distribution is now quite flat
producing slightly better agreement for the SAG case than for the with little sign of variability arising from the small sample size.
previous AG cases. The milling has reduced the initial variability of the particle shape.
For particles of initial mass of around 2.5 kg and below, the This movement of the blockiness distribution is similar to the beha-
incremental damage produced by the steel balls in the experiment viour of the particle size distribution as the particles become finer
is significant and there is a strong and systematic deviation of the with the distribution moving to the left – but in this case the move-
experimental masses from the simulation mass line. ment reflects rounding rather than size change.
Fig. 11 shows the matching cumulative distribution for each of
6. Predicted particle shape change from chipping and rounding the particle aspect ratios for AG case 1, again at the start of the sim-
ulation and after 12 min of shape evolution. The initial distribution
It is instructive to explore the nature of the shape change aris-
ing from the chipping and rounding mechanisms. Figs. 3, 5 and 7
give a qualitative idea of how the particles change but with DEM 1
For interpretation of colour in Figs. 10 and 11, the reader is referred to the web
we are also able to explicitly quantify the nature of the changes. version of this article.
84 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

Fig. 5. Rounding and size reduction of rock particles in the pilot mill for a 20% AG charge (Test 2). The colours show the blockiness of the particles with red being quite angular
and blue quite rounded with green being intermediate. The frames show the charge at 3, 6, 9 and 12 min. (For interpretation of the references to colour in this figure legend,
the reader is referred to the web version of this article.)

of the intermediate axis aspect ratio (Fig. 11a) is quite smooth and with milling observed is very similar in nature to that observed
starts with some particles having values of unity (meaning they are for case 1 but perhaps slightly weaker in the total amount of
around the same width in the major and intermediate directions). change. This indicates that the shape change arising from chipping
All particles initially have intermediate aspect ratios of less than and rounding is not particularly sensitive to fill level, at least for
1.4. After milling for 12 min the aspect ratios have all been reduced this scale of pilot mill. The behaviour of the aspect ratios is similar
(as shown by the red curve being uniformly to the left of the orig- to that for AG case 1 (Fig. 11) and so is not shown here.
inal red curve) but the degree of change is much smaller than the Fig. 13 shows the cumulative distribution of the blockiness for
blockiness. The gradient of the distribution has only increased SAG case 3 at the start of the simulation and after 12 min of shape
slightly with only weaker preferential reduction in this aspect ratio evolution caused by the surface erosion mechanisms. In this case
being observed for larger initial values. The effect is significantly there are round balls which do not evolve in shape and initially
weaker than for the blockiness. The minor axis aspect ratio (shown platey super-quadric rocks with blockiness in the range 3 < m < 6
in Fig. 11b) is similar in its behaviour. The initial distribution starts and minor axis aspect ratio in the range 1.0 < Aminor < 2.0. The
are value 1.1 and is below 1.65. It is again fairly smooth and linear. change in the blockiness distribution is stronger than for the AG
The degree of minor aspect ratio reduction from the chipping and cases with the final distribution being steeper than previously.
rounding is larger for large aspect ratios and quite small for smaller There is also much more change in the moderately rounded sizes
values. The overall change in this aspect ratio distribution is argu- (around m = 3). The final distribution is noisier than for the earlier
ably smaller than for the intermediate aspect ratio and is more cases because there are only 24 rocks in this case. The behaviour of
concentrated in the larger part of the distribution but still much the aspect ratio is again similar to that for AG case 1 (Fig. 11) and so
less so than for the blockiness. is also not shown here.
Fig. 12 shows the cumulative distribution of the blockiness for Summarising the effect of chipping and rounding mechanisms
AG case 2 at the start of the simulation and after 12 min of shape on the shape distributions of the particle attributes for all the
evolution caused by the surface erosion mechanisms. The change cases, we find that the:
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 85

balls (86.0 kg) has not changed and they continue to dissipate sig-
4000
nificant amounts of energy throughout the milling process. More
Experiment quantitatively, there is a 31.7% reduction in power draw compared
3500 to a 24.7% decline in charge mass. This means that the specific
DEM
power has decreased from 6.4 W/kg for fresh charge down to
3000 5.82 W/kg at the end. Note that these values are similar to those
found for the AG case, but the specific power draw decreases for
the SAG case instead of the increase observed for the AG case. In
2500
broad terms though, the power draw is essentially proportional
Mass at end (g)

to the charge mass.

2000 The rock–rock type of collision initially consumes 53% of the
energy supplied by the mill for AG case 1. This declines to only
1500 38% at the end of milling. The remaining energy is used in rock–
liner collisions. This indicates that as the particles become smaller
they are less likely to collide with each other and more likely to fall
1000
completely through the charge and impact directly onto the liner.
In contrast, for the SAG case 3 there are five different combinations
500 of collisions possible. The energy used in each type of collision is
shown in Table 3. Initially the rock–rock collisions consume only
0 11% which is much lower than for the AG case. This is due to the
0 1000 2000 3000 4000 small number of rocks and the relatively larger number of balls
Inial mass (g) which dominate the collisions. As the rocks shrink, the direct
rock–rock collision energy consumption drops drastically to only
Fig. 6. Comparison of the rock size distribution for the experiment and the
simulation for Test 2 with 20% AG charge after 12 min of grinding.
1.9% of the energy consumption as the increasing small rocks able
to collide with sharply decreasing frequency. The rock–liner colli-
sions dissipate a larger fraction of the energy (around 18%) which

Blocky or angular particles (those with m > 4.5) lose their cor-
is much smaller than the 47% occurring for the AG mode. This pro-
ners and edges more rapidly compared to initially more
portion decreases by 25% as the particles shrink. Since the domi-
rounded particles
nant ball component of the charge does not change over time,

The cumulative blockiness distribution moves to smaller block-
the chance of the decreasing sized rocks hitting the liner directly
iness values and its gradient increases (meaning that the range
is not significantly affected. The rock-media collisions initially dis-
of blockiness is much narrower)
sipate a larger fraction of the energy at 21.7%. This declines by 38%

Blockiness varies much more rapidly than aspect ratio.
with grinding reflecting the smaller rock masses and their lower

The rate of aspect ratio change depends on aspect ratio with
chance of being hit by the media. The energy dissipated in
particles of 1.2 or less changing this aspect ratio only very
media-rock collisions is initially 1.96 times higher than for rock–
slowly but with faster reduction rates for higher aspect ratios.
rock collisions. This increases to 7.1 times for the smaller rocks

Changes in the average shape properties are almost linear –
at the end of the grinding process. This highlights the importance
essentially proportional to the mass loss rate. However, the
of the grinding media and how large a role they play in size reduc-
average blockiness decline is faster in the early stages but then
tion and how this importance increases as the ball to rock mass
decreases to be around 40% lower in the latter half of the milling
ratio increases. The remaining energy consumption does not con-
time.
tribute to comminuting the rock particles, but simply leads to
media and liner wear. The media-liner collisions initially consume
7. Change in energy utilisation as rock particles erode a massive 36.1% of the energy rising to 48% for the worn charge.
The direct media on media collisions initially consume 17.3% of
In the physical experiments the amount of rock mass loss by the energy which increases strongly to 24% as the decreasingly
attrition is moderate. In order to better demonstrate the effect of small rock particles are less frequently trapped between approach-
rock erosion on the pattern of energy utilisation, we increase the ing media. So the inclusion of media maintains a more constant
attrition rates by a factor of 10. This means that the mass losses grinding environment for the rock particles but at a cost of a signif-
are 10 times larger and the particles are significantly smaller at icant increase in energy consumption (all of which is wasted in via
the end of grinding. We then compare the energy usage patterns media and liner wear). This extra energy consumption explains the
for the first minute of operation (initial performance) and the final entire observed difference in power draw between case 1 and case
minute (highly eroded rocks) for both AG case 1 and SAG case 3 to 3 at the end of the grinding process.
see how the mass loss has affected the nature of the collisional It is also instructive to examine the distribution of energy used
environment and how it uses the energy supplied. for each of the size reduction mechanisms for each case at both the
The initial power draw for AG case 1 is 791.2 W (calculated over start and end of the grinding process. The start and end distribu-
the first minute of milling) for a charge mass of 137.6 kg. This tions of energy between mechanisms for AG case 1 are given in
declines to 380.7 W at the end of the milling process at 12 min Table 4. The energy dissipated in boundary wear is substantial at
due to the large reduction in the mass of the rocks (62.7 kg). This around 24% at the start rising above 31% at the end (due the
represents a 52% reduction in power draw for a 54% reduction in increased weighting of rock–liner vs. rock–rock collisions. Since
charge mass. This means that the specific power draw (ratio of the mass of rock is changing as is the fraction of energy absorbed
power draw and charge mass), changes only slightly from by the rocks it is also useful to consider the relative contribution
5.75 W/kg up to 6.07 W/kg for the heavily eroded rocks. In contrast of each mechanism as a percentage of the total damage energy
SAG case 3 has a similar initial power draw of 850.4 W but a much absorbed by just the rocks which is given in Table 5. This data is
larger (compared to the AG case) power draw of 580.9 W at the end obtained by re-normalising the first 5 rows of data from Table 4.
of grinding. Although the rock has been heavily eroded (from Initially, the dominant rock damage mechanism is rounding
46.6 kg at the start down to 13.8 kg at the end) the large mass of with around 36% of rock damage energy being used to remove
86 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

Fig. 7. Rounding and size reduction of blocky rock particles in the pilot mill for a 10% rock and 10% ball SAG charge (Test 3). The colours show the blockiness of the rock
particles with red being quite angular and blue quite rounded with green being intermediate. The steel balls are round and shaded grey. The frames show the charge 3, 6, 9
and 12 min. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

the corner material on the particles and thereby causing the strong is 51.4% of the total energy available. The mechanism with the
shape change observed. The chipping mechanism at 21.6% is also smallest damage energy utilisation is abrasion which consumes
quite strong and also contributes meaningfully to rock shape only 6.6% of the damage energy (and 5% of the total energy). This
change. The energy available for incremental breakage (which is is responsible for the radial shrinkage of particles due to frictional
the part of the normal component energy dissipated that is direc- sliding. An important reason for this being so small a fraction
ted towards the centre of the particle and which also exceeds the relates to the particles initially being quite angular and having
E0 value – set here at 1.0 J/kg) is only 2.6% of the rock damage higher aspect ratios which leads to very high corner and edge ero-
energy (and 2.0% of the total energy). This explains the low level sion from rounding and chipping. Unfortunately, the abrasion
of incremental damage observed in the experimental results for mechanism is the one that is responsible for the shrinkage of large
case 1 and the general difficulty of tumbling mills to break larger particles bringing them down to the size range at which the mill
particles. The wasted energy is defined here as the total energy can start to produce body breakage, so a slow rate of abrasion
from the normal component of collision energy which is directed means that it takes a long period for the particles to become small
into the particle and which is less than the elastic threshold E0 enough for this to happen. The incremental damage is the mecha-
and which therefore does not contribute to volumetric damage of nism then responsible for this body breakage and this is only 2.6%.
the particle. This is a substantial fraction of the rock damage So the fraction of the energy being used in the two most important
energy at 31.6%. This energy is dissipated as heat within the parti- mechanisms for significantly reducing the size of rock particles is
cles. This energy wastage is one of the key reasons that tumbling only 7%. Around 43% of the energy is used for removing corners
mill performance is inherently inefficient. and edges, a process often described as conditioning (Loveday
With the identification of these mechanisms and their methods and Naidoo, 1997). This does contribute to the creation of finer
of calculation, we can for the first time estimate from simulation rock particles but these mechanisms also serve to protect the
the fraction of energy wasted in a tumbling mill. For this case it integrity of the rocks by preventing energy being absorbed by
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 87

4000 (a) 4.7

4.6
3500 4.5
Experiment 4.4

Blockiness
3000
DEM 4.3
4.2
2500
4.1
Mass at end (g)

2000 4.0
3.9
0 2 4 6 8 10 12
1500
Time (min)

1000 1.195
(b)

Aspect rao (intermediate axis)

500 1.190

0 1.185
0 1000 2000 3000 4000
Inial mass (g)
1.180
Fig. 8. Comparison of the rock size distribution for the experiment and the
simulation for Test 3 with 30% SAG charge after 12 min of grinding.
1.175

the core of the particles (which is not at all what one would like to 1.170
0 2 4 6 8 10 12
happen in a grinding mill in terms of the larger particles while not
Time (min)
forgetting that production of finer particles is also important).
At the end of the grinding process (see the third column of
Tables 4 and 5) the distribution of energy between mechanisms (c) 1.360
has changed substantially. The fraction of energy in the liner wear 1.355
has increased strongly so the amount of energy being used for 1.350
Aspect rao (minor axis)

grinding is reduced. However, the mass of rock to grind has also 1.345
declined. The fairest comparison is to consider the relative distri-
1.340
bution of rock damage energy between the rock breakage mecha-
nisms, as given in Table 5. The particles having become much 1.335
more rounded causes more energy to be directed towards their 1.330
centres. This leads to an 82% increase in abrasion to a still relatively 1.325
small percentage of 12%. There is a sharp corresponding decrease
1.320
in the amount of rounding (due to the particles now being much
rounder). In contrast, the amount of damage energy used for chip- 1.315
0 2 4 6 8 10 12
ping has increased strongly by 57%. This change in the balance of Time (min)
rounding and chipping (which both depend on particle shape) is
a result of the changing balance of collisions being from rock–rock Fig. 9. Change of average shape attributes for AG case 1, (a) blockiness, (b)
(which has a larger sliding frictional component to their interac- intermediate axis aspect ratio, and (c) minor axis aspect ratio.

tion) to rock–liner (which is more predominantly in the normal

direction). The strong rounding of the particles at the end of the
grinding period means that a larger fraction of the increased 1.0
energy in the normal direction is directed into the core of the par-
ticles causing more collisions to exceed the elastic threshold E0. Inial
0.8
This leads to a strong 73% increase in the amount of incremental Aer 12 min
damage. For the wasted energy, the increased fraction of normal
Cumulave probability

energy is counter-balanced by the decreased fraction of energy 0.6

below the elastic threshold, which combined give a small decrease
of 2%. It is worth noting that even with these large changes the
incremental damage and abrasion mechanisms are still relatively 0.4
weak (only 11.4% of energy consumption and 16.5% of the rock
damage energy) compared to the other mechanisms.
0.2
Table 6 shows the distribution of energy use by the mechanisms
for SAG case 3. The quantities in the first five rows are calculated
for just the rock particles. An additional row gives the percentage 0.0
of energy consumed by the media. This is dominant at the start 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
at 41% and increases slightly to 42.9% at the end of grinding. The Blockiness

liner wear mechanism contribution for the SAG case is the second Fig. 10. Change in blockiness distribution as rock particle shape evolves for AG case
largest energy consumption mechanism, initially at 26.2% and 1.
88 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

1.0 1.0

Inial Inial
0.8 0.8
Aer 12 min Aer 12 min

Cumulave probability
Cumulave probability

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
1.0 1.1 1.2 1.3 1.4
2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
Intermediate axis rao
Blockiness

Fig. 13. Change in blockiness distribution as rock particle shape evolves for SAG
1.0
case 3.

Inial
0.8
Aer 12 min Table 3
Energy consumption for different collision types at the start and end of milling for
Cumulave probability

SAG case 3.
0.6
Collision type Initial percentage Final percentage % change
(first minute) (last minute) over time
0.4 Rock–rock 11.1 1.9 83
Rock–liner 19.2 12.4 25
Rock–media 21.7 13.5 38
0.2 Media–media 17.3 24.2 40
Media–liner 30.8 48.0 56

0.0
1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7
Minor axis rao
Table 4
Fig. 11. Change in aspect ratio distributions of rock particles as their shape evolves Distribution of energy dissipation between different mechanisms at the start and end
for AG case 1. of the milling process for AG case 1.

Dissipation Initial percentage Final percentage % change

mechanism (first minute) (last minute) over time
1.0 Abrasion 5.0 8.3 66
Rounding 27.3 12.9 53
Chipping 16.4 23.4 43
Inial
0.8 Incremental damage 2.0 3.1 55
Aer 12 min Wasted 27.4 21.4 22
Liner wear 24.0 31.1 30
Cumulave probability

0.6

0.4 Table 5
Relative contributions of the breakage mechanisms to the rock comminution at the
start and end of the milling process for AG case 1.
0.2
Dissipation Initial percentage Final percentage % change
mechanism (first minute) (last minute) over time

0.0 Abrasion 6.6 12.0 82

2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 Rounding 35.9 18.7 48
Blockiness Chipping 21.6 34.0 57
Incremental damage 2.6 4.5 73
Fig. 12. Change in blockiness distribution as rock particle shape evolves for AG case Wasted 31.6 31.0 2
2.

28.9% at the end. This is slightly larger than for the AG case at the mechanisms is very similar to that of the AG case. The largest dif-
start and then slightly lower at the end. Between these two energy ference is for the incremental damage which is relatively smaller at
consumption mechanisms, around 47–52% of the energy input is 1.7% compared to 2.6% for the AG case. At the end of the grinding
used purely for wearing the liner and the media. process, the abrasion fraction of damage energy is only modestly
The fraction of the damage energy used by each mechanism to increased, the rounding has declined by just 9%, the chipping by
comminute the rocks for SAG case 3 is given in Table 7. This is a only 2% whilst the incremental damage has increased sharply by
renormalization of the first five rows of the data in Table 6. At 88% and the wasted energy has increased slightly (by 8%). The
the start of the grinding process, the balance of damage changes observed over time for the SAG case are different in
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 89

Table 6 interact with. The accepted ball mill type grinding model of a layer
Distribution of energy dissipation between different mechanisms at the start and end of slurry on a particle offers much less mass again to absorb any
of the milling process for SAG case 3.
useful energy. By way of contrast, a glancing blow which shears
Dissipation Initial percentage Final percentage % change off an asperity only needs to generate a small quantity of new sur-
mechanism (first minute) (last minute) over time face area. A normal impact is more likely to load a particle and
Abrasion 2.1 2.1 0 begin to damage it through induced intra-particle tension at stress
Rounding 12.0 9.4 22 levels with elastic energy greater than E0. A gouging action is more
Chipping 7.0 5.9 16
Incremental damage 0.56 0.90 61
plausible for incremental surface damage – once a particle has
Wasted 10.7 10.0 6 been sufficiently smoothed. This is consistent with the effective
Media wear 41.0 42.9 5 friction model tested by Djordjevic at al. (2006) and the results
Liner wear 26.2 28.9 10 in this paper.
Musa and Morrison (2009) investigated tumbling mill efficiency
by comparing measured energy with that required to achieve a
similar product using a JK drop weight tester. SAG mill efficiency
Table 7 was found to be typically about 40%, that is, about two and a half
Relative contributions of the breakage mechanisms to the rock comminution at the
start and end of the milling process for SAG case 3.
times as much energy was required compared with single impact
breakage. The incremental breakage hypothesis is sensibly consis-
Dissipation Initial percentage Final percentage % change tent with this analysis.
mechanism (first minute) (last minute) over time
Abrasion 6.4 7.5 17
Rounding 36.6 33.2 9
Chipping 21.3 20.8 2
8. Effect of size reduction on energy spectra
Incremental damage 1.7 3.2 88
Wasted 32.6 35.3 8 Energy spectra are constructed by calculating the energy dissi-
pation in each separate collision and then building a probability
distribution of the different collision energy levels. These can be
constructed using either the collision rate or the energy dissipation
pattern and generally smaller in magnitude than for the AG case. rate as the dependent variable and the independent variable can be
The large increases in abrasion and chipping observed over time either collision energy (J) or specific collision energy (J/kg). These
for the AG case do not occur for the SAG case and the chipping frac- spectra are a way of characterising the collision environment in a
tion is essentially constant instead of increasing. The only mecha- mill and providing a quantitative framework for evaluating the
nism that increased more over time for the SAG case was the effects of changes in design and operating variables. Since the col-
incremental damage, increasing modestly more at 88% compared lision between particles generate the breakage of the rock, then
to 73% for the AG case. Note that while it has increased more the changes in the collision spectra link directly to changes in product.
fraction of incremental damage is still smaller for the SAG opera- The form of the energy spectra used here was introduced by Cleary
tion. The grinding media (which does not change size) appears to (2001b). It has been used for example to look at the effect of chang-
maintain a more constant grinding environment over time as the ing feed size to a coarse clinker grinding ball mill (Cleary et al.
rock particles are comminuted. This suggests that full size AG mill 2008) and for tower mills (Sinnott et al., 2006). Other authors have
performance could be somewhat more sensitive to the changing used variants of the spectra for similar purposes. For the current
size and shape distributions of the rock axially along the mill from paper we will use the form from Cleary (2001b) and will use colli-
feed to discharge. In contrast, the presence of the media in a SAG sion rate and specific collision energy as the dependent and inde-
mill appears likely to maintain a more constant grinding environ- pendent tabulation. Spectra are generated for each size class, for
ment along the mill with the performance perhaps less influenced each collision type and over all collisions. Here we will focus on
by the specific nature of the rock charge variation within the mill. the spectra for the specific size classes.
This is consistent with industrial experience that SAG mills are Figs. 14 and 15 show some energy spectra from AG case 1 and
easier to operate (Napier-Munn et al., 1996). Hence, SAG mills SAG case 3, respectively. Again the attrition rates have been
dominate the industry. When AG milling was in more common increased by a factor 100 so that there is a large change in charge
use, some designers provided for storage of coarse and fine feed particle size. A number of important questions relating to the
fractions to improve control. In practice, operators tended to use change in behaviour of the energy spectra with changing charge
up the faster milling coarse feed and leave the next shift with a attributes can be explored with these simulation cases.
problem and no autogenous grinding media. The first question is whether there is any difference between
It is also worth briefly comparing the actual SAG environment the spectra of the AG and SAG charge configurations. Comparing
with the test mill environment. While the test mill environment Fig. 14 (top) and Fig. 15 (bottom/left) for size class 16 at the start
is depleted of fines and smaller particles compared with an indus- of the grinding process we observe that the spectra for this com-
trial SAG mill, the implications for process (in)efficiency are still mon size class is similar in shape and magnitude. There is a central
reasonable. model maximum which is indicated by the vertical line. The colli-
An interaction between a 60 mm and a 6 mm particle can break sion frequency declines slowly for both higher and lower energy.
the smaller particle at a t10 of A (the largest degree of breakage pos- There is a long tail with energies recorded out to 108 J but these
sible in a single impact) without much effect on the stress level of are of no importance since they do not contribute to breakage
the larger particle because it masses about 1000 times more. Sim- and are not shown. At higher energies the collision frequency
ilarly, the destruction of the 6 mm particle does not produce much declines at an increasing rate dropping sharply at energies above
product compared with breakage of the larger particle. Earlier DEM 1 J with peak energy observed of around 5 J. The collision frequen-
modelling of a 1.8 m  0.6 m industrial pilot SAG mill (Cleary and cies are slightly lower for the SAG case because of the smaller num-
Morrison, 2004), suggested that most interactions between parti- ber of particles in the charge to generate the collisions. The peak
cles occurred within the same size fraction or one above or below energy is similar at around 5 J and the modal peak is modestly
for a combination of geometric access and probability density rea- higher at around 0.028 J/kg. It is important to note that the vast
sons. That is, there have to be enough particles of similar sizes to majority of the collisions are very weak. The E0 value divides this
90 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

1. There is little difference in these spectra with the shape being

similar, the modal peak being little changed and even the very sen-
sitive peak collision energy varying only slightly. The energy
absorbed per unit mass is similar for similar mass particles despite
long milling and shape change. There is a subtle change in the
shape of the spectra at collision energies of 1–2 J with secondary
inflection having developed. There is a hint of this inflection point
in the initial spectra but it has become more clearly defined at the
end of the grinding process. This modest increase in collision fre-
quency for this narrow range of energies does not have any clear
cause nor does it have any obvious consequence. Note that the
actual particles that are recorded as class 16 are not the same par-
ticles as those recorded in that class at the start. The original class
16 particles for AG case 1 have become class 14 in AG case 1
because of their size reduction from surface erosion. The particles
in class 16 at the end of 12 min of grinding are typically those ini-
tially found in class 19.
The final (third) question is around how does the energy spectra
vary with size? Here we use the spectra from SAG case 3. Some of
the spectra (even numbered size classes) are shown in Fig. 15. The
spectra for classes 12 and 14 (which are product size classes) are
from the end of the simulation whereas classes 16 and 18 are feed
classes and are taken from the start of the simulation. The spectra
for class 18 (a larger feed class) is quite similar to that of the pre-
viously discussed class 16. For class 14, the upper half of the spec-
trum is similar but there is an increase in frequency of lower
energy collisions. For class 12, the changes are similar in nature
but more extreme leading to an almost constant collision fre-
quency for collisions below 0.2 J. This indicates that the particles
experience increasing numbers of weak collisions as they become
smaller.
Table 8 shows the peak collision energy intensity for different
Fig. 14. Top end of the collision energy spectra for size class 16 (finest feed class) at rock size classes for AG case 1 and SAG case 3. Peak energies are
start of milling process and after 12 min for AG case 1. taken to be the largest energy which has collisions with a fre-
quency of 0.01 Hz or more. The spectra used are for the full
12 min of mill test operation so there are rock particles in both feed
spectra into a small range of collisions with collision energies and product size fractions. The threshold frequency of 0.01 corre-
above this level that contribute to incremental damage and ones sponds to around one collision per particle over the 12 min dura-
below this which do no damage to the particles and so contribute tion. Less frequent collisions can over no effect on the rock
only to the wasted energy. However, if the particle is coated with charge over this duration. There is a general trend of increase in
slurry, the fine particles in the slurry may be comminuted within the peak (most frequently observed) collision energy per unit mass
the contact zone as the energies required are most strongly depen- as the particle size becomes smaller for both AG and SAG type
dant on mass. A 0.5 mm particle has a mass which is only charges. This means that as a particle shrinks the proportion of
1/125,000 that of a 25 mm particle. Many of these fine particles its energy spectra that exceeds the E0 elastic threshold increases.
may be broken by an impact which has negligible effect on the lar- Therefore, an increasing fraction of the available collision energy
ger particle. is able to manifest as incremental damage. So as a particle shrinks
It is further important to note that the collision frequencies of in size from abrasion, chipping and rounding its rate of accumula-
the higher energy collisions are quite small (between 0.1 and tion of incremental damage increases and its chances of undergo-
0.01 Hz). This means that collisions able to do damage to rock par- ing catastrophic body breakage increases. This highlights the
ticles in this mill occur on average between once every 10 and complex interaction between the different breakage mechanisms
100 s. Over the duration of the grinding process this is only 7–70 that control the evolution life cycle of particles within a tumbling
collisions spread across the particles in the class, so around 1–10 mill.
collisions per particle. Only these small numbers of collisions are
able to create any form of body damage to the rock particles. The
small differences between the AG and SAG spectra cause one to 9. Why are ball collisions more effective at breakage than rock
ask the question as to how can there be such a large difference collisions?
in the observed final size distribution for the SAG case (in Fig. 8)
compared to the AG case (in Fig. 4)? This question will be explored A critical concern with the energy spectra results shown earlier
in the next section. is that there is little meaningful difference between those for AG
The second question is about whether the energy spectra vary and SAG modes. The peak energies are both around 3.9–4.6 J. On
with time (i.e. with the degree of particle shrinkage) for the same this basis, one would expect relatively similar amounts of incre-
size class? That is, does the shrinkage and shape change of the par- mental damage for cases 1 and 3. However, Figs. 2 and 6 show that
ticles in the charge change the nature of the collision environment there is clearly a significant contribution to size reduction for SAG
that a particle of the same size experiences. Fig. 14 shows the case 3 for the smaller and intermediate size particles. This
upper end of the collision energy spectra for size class 16 (finest observed behaviour cannot be explained except if the SAG case
feed class) at start of milling process and after 12 min for AG case energy spectra were to have a clearly higher peak value for its
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 91

Fig. 15. Upper end of the energy spectra for SAG case 3 for different rock size classes (12 is the smallest and 18 is the highest class shown).

Table 8 mill is not able to create collisions with total energy dissipation
Peak collision energy intensity (J/kg) for different rock size classes for AG case 1 and above about 9 J. So the only way that the breakage observed for
SAG case 3. Peak energies are determined as the largest energy with a frequency of
0.01 Hz or more over the entire 12 min grinding duration.
SAG case 3 can therefore occur is for the rock particles to absorb
a much greater fraction of the energy dissipated in each collision.
Class AG case 1 SAG case 3 As a demonstration of the important effect of the attribution of
12 – 4.5 energy dissipation in collisions to the participating bodies, the AG
13 – 4.6 case 1 and SAG case 3 simulations were run again, this time
14 3.9 4.6
attributing 90% of the collision energy dissipation to the rock par-
15 3.6 4.3
16 3.8 3.6 ticles. Fig. 16 shows the energy spectra for rock size class 18 (fairly
17 3.1 3.4 coarse) for (top row) AG case 1 and (bottom row) SAG case 3. The
18 3.2 3.0 left column shows the spectra for a 50–50 split of energy between
19 2.1 1.9
balls and rocks while the right column shows the spectra for a 10–
90 split in favour of the rock. For the AG case, there is only a small
difference which results from the greater energy transfer to the
energy spectra. This would cause many more collisions to exceed rock particles when they collide with the steel liner. For the SAG
the elastic threshold and to exceed it by a larger degree leading case, there is a much stronger effect as the steel ball–rock collisions
to much greater incremental damage energy to be accumulated also now attribute 90% of their collision energy to the rock parti-
and therefore significant body breakage. cles. The shapes of the spectra are very similar but the maximum
The maximum fall height for a particle in this mill is around energies recorded have increased from around 3.7 J to 6.0 J.
1.2–1.4 m and the majority of particles have masses of 2 kg or less Finally, we compare the maximum collision energy intensities
so the maximum kinetic energy of a particle falling in the cataract- observed for the two different levels for the attribution of energy
ing stream on impact at the toe is 24–28 J. The lowest coefficient of dissipation to the rock for each of the size classes. In AG case 1
restitution is for rock–steel collisions with a value of 0.5 being (see Table 9) the increase from considering the reduced steel
used. This means that around 75% of the initial kinetic energy absorption of energy varies from 15% (for class 16) up to 68%
can be dissipated which is 18–21 J. It is rare for a particle to fall (for class 19). The mean change across all the classes is a modest
the full distance without some intermediate collisions so the peak 28%. The change is not larger because all the dominant rock–rock
energy dissipations observed per particle pair are around 8–9 J collisions still split their energy dissipation 50–50 between the
which is around half the theoretical limit just calculated. The peak rock particles. Only the collisions with the steel liner increase the
collision energies observed are around double the peaks for the collision energy intensity strongly, but these are a minority of
spectra for each size (see Table 8) since we have assumed a 50– the collisions. The mean peak energy intensity for the AG mode
50 split of energy dissipation between the colliding bodies. The is 4.8 J/kg. If E0 were instead to be 4.0 J/kg then only a small num-
92 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

Energy split 50% ball -50% rock Energy split 10% ball - 90% rock

Fig. 16. Energy spectra for rock size class 18 (fairly coarse) for (top row) AG case 1 and (bottom row) SAG case 3. The left column shows the spectra for a 50–50 split of energy
between balls and rocks while the right column shows the spectra for a 10–90 split in favour of the rock.

Table 10
Table 9 Peak collision energy intensity (J/kg) for different rock size classes for the SAG case 3
Peak collision energy intensity (J/kg) for different rock size classes for the AG case 1 for different levels of energy absorption by the rock. These are calculated over the first
for different levels of energy absorption by the rock. These are calculated over the first minute of grinding operation.
minute of grinding operation.
Class 50% absorption 90% absorption % change
Class 50% absorption 90% absorption % change 16 3.3 6.3 91
16 5.9 5.0 15 17 4.0 6.8 70
17 4.0 5.8 45 18 3.8 6.0 58
18 4.0 4.5 13 19 2.5 4.0 60
19 2.2 3.7 68

ber of collisions would exceed this level and the incremental dam- The attribution of a much larger fraction of the collision energy
age to the rock particles would be minimal. For SAG case 3 (see to the rock particles is physically quite reasonable. Rock is an
Table 10) the increase from the reduced steel absorption of energy intrinsically much more brittle material and is much more able
varies from 58% (for class 18) to 91% (for class 16). The mean to crack during collision. The stress waves within the steel are
change in the maximum collision energy intensity is 70%, which transmitted without damaging the steel and are reflected from
is quite significant. For SAG operation, the collisions with both the rear of the steel structure back into the rock where it can create
the steel balls and liner transfer much more energy to the rock par- fracture damage. The choice of a 90–10 split was made here purely
ticles so the effect is much more significant than for the AG mode. to demonstrate the effect and to show that it can explain the
The mean peak energy intensity for the SAG case is 5.8 J/kg. So if E0 observed experimental behaviour.
were to be 4.0 J/kg then a reasonable number of collisions would In analysis relating to an Ultra Fast Load Cell, Tavares and King
exceed this level and the rate of incremental damage would be (1998) assuming a perfectly elastic non-frictional contact between
much larger than for AG mode or for the SAG mode with a 50–50 a rock and impactor, gave the relationship for the fraction of the
attribution of collision energy. elastic energy of the contact that was stored in the rock particle as:
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 93

1
Blocky or angular particles (those with m > 4.5) lose their cor-
u ¼ kp ; ð14Þ
þ1 ners and edges more rapidly compared to initially more
ks
rounded particles.
where kp is the stiffness of the rock particle and ks is the stiffness of
The cumulative distribution of blockiness moves to smaller
the steel ball. With typical values of ks = 200 GPa and kp = 1–50 GPa, blockiness values and its gradient increases (meaning that the
this gives a range for u of 0.8–0.995. This means that considering range of blockiness becomes narrower) as the particles are
only purely elastic effects, the energy split deduced from matching ground.
the DEM simulations to the experimental data is the middle of this
Blockiness changes much more rapidly than the aspect ratios of
theoretical range. It should also be noted that for collisions between the particles.
particles of the same material that Eq. (14) predicts equal elastic
The rate of aspect ratio change is dependent on the aspect ratio,
energy being stored in each particle which is consistent with histor- with particles having aspect ratios of 1.2 or less changing only
ical DEM practice of using a 50–50 split of energy. slowly but with faster reduction rates for higher aspect ratios.
However, for a brittle frictional material that fractures during
Changes in the average shape properties for the population of
the contact, it is not clear what material properties control this particles in these batch mills is almost linear with time – which
energy split, whether it is Young’s modulus (as is the case for means that they essentially proportional to the mass loss rate
the elastic non-frictional material), the brittleness of the material, arising from the chipping and rounding mechanisms. However,
the restitution coefficient or some combination of other proper- the rate of decline of the average blockiness is faster in the early
ties. It is important to appreciate that the energy being absorbed stages but then decreases by around 40% in the latter half of the
by the particles which controls breakage is the energy dissipated milling period.
within the particles rather than the elastic energy stored (which
is at least partially restored to the particle as kinetic energy after For the dilute batch grinding setup used in this work, the ini-
unloading of the contact). This energy dissipated is, however, tially dominant rock damage mechanism is rounding with more
likely to be at least partially dependent on the peak elastic energy than a third of the rock damage energy being used to remove the
or elastic strain. This effect is captured to some degree by the corner material from the particles thereby causing the strong
coefficient of restitution which is used in setting the dashpot shape change. The chipping mechanism consuming more than
strength in the DEM contact model (Eqs. (1) and (2)). The under- 20% of the energy is also quite strong and contributes meaningfully
lying mechanisms controlling this energy split and methods for to rock shape and size change. The energy used for incremental
characterising the controlling material properties need further breakage is quite small (less than 3% of the rock damage energy
investigation now that its importance in breakage prediction for this milling configuration). Wasted energy (which is defined
using DEM is understood. as the energy directed into the particle and which is less than
the elastic threshold E0) and does not contribute to particle damage
and can also be estimated using this model. This represents a sub-
10. Conclusions stantial fraction of the energy consumed at more than 30%. This
energy wastage is one of the key reasons for tumbling mill perfor-
A computational method for attributing the energy dissipated mance being inherently inefficient. For SAG mills with conven-
in particle collisions to five mechanisms identified as causing size tional charge size distributions some of this energy will result in
and/or shape change for particles in AG, SAG and coarse feed ball the breakage of finer particles (which are absent from this test
mills has been presented. Chipping and rounding are shape depen- mill) which could (slightly) reduce the level of energy wastage
dent mechanisms that lead to preferential mass loss from corners from the incremental damage mechanism. This possibility is con-
and edges of particles at much lower energy levels than required sidered further at the end of Section 6.
for body breakage and which are responsible for rounding or con- The rounding or conditioning of the particles (as they are
ditioning of particles as well as some size reduction. They are con- ground) causes more energy to be directed towards their centres.
trolled by the energy dissipation at a contact and the location of At the end of grinding for 120 min (which is equivalent to the
the contact on the surfaces of the particles (which defines both model prediction with the erosion rates scaled by a factor 10) with
the contact normal and the moment vector). Single impact body an AG charge, this leads to a near doubling of the abrasion with
breakage has been found to be a weak contributor to overall size sharp corresponding decreases in the amount of rounding. How-
reduction in AG, SAG and coarse feed ball mills. Most body break- ever, the amount of damage energy used for chipping was found
age occurs as a result of damage accumulation over many colli- to increase strongly (by around 50%). This change in the balance
sions. This damage is proportional to the collision energy above of rounding and chipping (which both depend on particle shape)
an elastic threshold with energy dissipation below this resulting is a result of the changing balance of collisions being from rock–
only in heat and sound. This incremental damage mechanism is rock (which has a larger sliding frictional component to their inter-
inherently less efficient than single body breakage as a much larger action) to rock–liner (which is more predominantly in the normal
fraction of the total input energy is below the elastic threshold. direction). The strong rounding of the particles at the end of the
Much of the size reduction is then the result of the accumulation grinding period means that a larger fraction of the increased
of the damage from many weak collisions experienced by particles energy in the normal direction is directed into the core of the par-
within mill. Finally, size reduction from abrasion is well repre- ticles causing more collisions to exceed the elastic threshold E0.
sented by the shear energy absorption of the particles. This leads to a near doubling in the total amount of incremental
The chipping and rounding mechanisms have a significant damage. For the wasted energy, the increased fraction of normal
effect on the distributions of the particle shape attributes. For the energy is counter-balanced by the decreased fraction of energy
three test conditions used in this study of a well characterised below the elastic threshold, which combined give a very small
dilute pilot AG/SAG mill, we find that the: decrease.
The changes observed over time for a SAG charge are different

Collisions occur preferentially with corners and frictional con- in pattern and generally smaller in magnitude than for an AG
tacts also preferentially occur at corners and edges. This con- charge. The large increases in abrasion and chipping observed over
centrates the erosion at these more exposed locations leading time for the AG case do not occur for the SAG case and the chipping
to preferential rounding of corners and edges. fraction is essentially constant instead of increasing. The only
94 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95

mechanism that increases more over time for the SAG case was the DEM simulation of comminution can no longer be ignored. How-
incremental damage. The grinding media (which does not change ever, it is not fully clear which material properties control this
size over this grinding period) appears to maintain a more constant energy split. The underlying mechanisms controlling this energy
grinding environment over time as the rock particles are commin- split and methods for characterising the material properties need
uted. This suggests that full size AG mill performance could be significant future investigation.
more sensitive to the changing size and shape distributions of
the rock within the mill. In contrast, the presence of the media in
Acknowledgement
a SAG mill appears likely to maintain a more constant grinding
environment along the mill with the performance perhaps less
The authors would like to thank Matt Sinnott for his assistance
influenced by the specific nature of the rock charge variation
with the rendering of the DEM related images.
within the mill.
Attrition is the mechanism that is responsible for the shrinkage
of large particles bringing them down to the size range at which References
the mill can start to produce body breakage. The incremental dam-
age is the mechanism then responsible for this body breakage and Bbosa, L., 2007. Measurement of Impact Breakage Properties of Ore Particles using a
this is found to be only around 2% for this mill and charge. So the Series of Devices. M.Sc. Thesis, University of Cape Town.
Cleary, P.W., 1998. Predicting charge motion, power draw, segregation, wear and
fraction of the energy being used in the two most important mech- particle breakage in ball mills using discrete element methods. Miner. Eng. 11,
anisms for significantly reducing the size of rock particles is only 1061–1080.
7% of the energy input. Around 43% of the energy is used for Cleary, P.W., 2001a. Charge behaviour and power consumption in ball mills:
sensitivity to mill operating conditions, liner geometry and charge composition.
removing corners and edges (conditioning). This does contribute Int. J. Min. Process. 63, 79–114.
to the creation of finer rock fragments but these mechanisms also Cleary, P.W., 2001b. Recent advances in DEM modelling of tumbling mills. Miner.
serve to protect the integrity of the rocks by preventing energy Eng. 14, 1295–1319.
Cleary, P.W., 2001c. Modelling comminution devices using DEM. Int. J. Numer. Anal.
being absorbed by the core of the particles (which is not at all what Methods Geomech. 25, 83–105.
one would like to happen in a grinding mill). Cleary, P.W., 2004. Large scale industrial DEM modelling. Eng. Comput. 21, 169–
The inclusion of the size and shape change of particles during 204.
Cleary, P.W., Morrison, R.D., 2004. Using DEM to model ore breakage within a pilot
grinding within the DEM model also provides information on scale SAG mill. Miner. Eng. 17, 1117–1124.
how these particle changes affect the energy spectra (which are Cleary, P.W., Morrison, R.D., 2008. Editorial. Miner. Eng. 21, 743.
used to characterise the collision environment). There is a general Cleary, P.W., Sinnott, M.D., Morrison, R.D., 2008. DEM prediction of particle flows in
grinding processes. Int. J. Numer. Methods Fluids 58, 319–353.
trend of increase in the most commonly generated collision energy
Cleary, P.W., 2009. Industrial particle flow modelling using DEM. Eng. Comput. 26
per unit mass as the particle size reduces for both AG and SAG (6), 698–743.
charges. This means that as a particle shrinks the proportion of Cleary, P.W., Cohen, R.C.Z., Sinnott, M.D., Prakash, M., Mead, S., 2013. Prediction of
its energy spectra that exceeds the E0 elastic threshold increases industrial, biophysical and extreme geophysical flows using particle methods.
Eng. Comput. 30, 157–196.
and an increasing fraction of the available collision energy is able Cleary, P.W., Sinnott, M., 2015. Simulation of particle flows and breakage in crushers
to contribute to the incremental damage. So as a particle shrinks using DEM: Part 1 – compression crushers. Miner. Eng. 74, 178–197.
in size due to abrasion, chipping and rounding its rate of accumu- Das, R., Cleary, P.W., 2010. Effect of rock shapes on brittle fracture using smoothed
particle hydrodynamics. J. Theor. Appl. Fract. Mech. 53, 47–60.
lation of incremental damage increases and its chances of undergo- Delaney, G.W., Cleary, P.W., Sinnott, M.D., Morrison, R.D., 2010a. Novel application
ing catastrophic body breakage increases proportionally. This of DEM to modelling comminution processes. In: IOP Conf. Ser., Mater. Sci. Eng.,
highlights the complex interaction between the different breakage vol. 10, p. 012099.
Delaney, G., Cleary, P.W., 2010b. The packing properties of superellipsoids.
mechanisms that control the evolution life cycle of particles within Europhys. Lett. 89, 34002.
a tumbling mill. Delaney, G.W., Cleary, P.W., Morrison, R.D., Cummins, S., Loveday, B., 2013.
The charge composition controls the balances of collision types Predicting breakage and the evolution of rock size and shape distributions in
AG and SAG mills using DEM. Miner. Eng. 50–51, 132–139.
and energy dissipation mechanisms which then control the pattern Delaney, G.W., Morrison, R.D., Sinnott, M.D., Cummins, S., Cleary, P.W., 2015. DEM
of size reduction and shape change of the rock charge. The evolving modelling of non-spherical particle breakage and flow in an industrial scale
charge constantly changes the balance of these mechanisms with cone crusher. Miner. Eng. 74, 112–122.
Djordjevic, N., Morrison, R.D., Loveday, B., Cleary, P., 2006. Modelling comminution
abrasion and chipping increasing but rounding decreasing (mech-
patterns within a pilot scale AG/SAG mill. Minerals Engineering 19, 1505–1516.
anisms controlled by the particle shape) and with an increase in Herbst, J.A., Nordell, L., 2001. Optimization of the design of sag mill internals using
the incremental damage energy and a matching reduction in high fidelity simulation. In: Barratt, D.J., Allan, M.J., Mular, A.L. (Eds.), Proc. of
wasted normal impact wasted energy (mechanisms controlled by the SAG Conference, Vancouver, B.C., vol. IV. University of British Columbia, pp.
150–164.
particle size). The fraction of energy dissipated by the liner is also Larbi-Bram, S., Shi, F., Kojovic, T., 2010. Experimental study of ore characterization
significant and increases as the rocks reduce in size and collisions methods for AG/SAG mills. In: XXV IMPC, Brisbane, Australia, pp. 929–941.
occur increasingly with the liner. Loveday, B.K., Hinde, H.A., 2002. Application of a rock abrasion model to pilot-plant
and plant data for fully and semi-autogenous grinding. Miner. Process.
The final issue considered by this paper is the effect of the Extractive Metall. 111, C39–C43.
choice of energy attribution between colliding particles. Histori- Loveday, B.K., Naidoo, D., 1997. Rock abrasion in autogenous milling. Miner. Eng. 10,
cally, DEM usage has attributed this equally to each particle. If this 603–612.
Loveday, B.K., Whiten, W.J., 2004. Application of a rock abrasion model to pilot-
was valid then we would observe similar breakage behaviour plant and plant data for fully and semi-autogenous grinding. Trans. IMMt, Sect.
between the AG and SAG charge cases in the experiments used C 111, C39–C43.
here which is not the case. The attribution of a much larger fraction Mishra, B.K., Rajamani, R.J., 1992. The discrete element method for the simulation of
ball mills. Appl. Math. Model. 16, 598–604.
of the collision energy to the rock particles is physically quite rea- Mishra, B.K., Rajamani, R.K., 1994. Simulation of charge motion in ball mills. Part 1:
sonable. Rock is an intrinsically much more brittle material and is Experimental verifications. Int. J. Miner. Process. 40, 171–186.
much more able to crack during collision. The stress waves within Morrison, R.D., Cleary, P.W., 2004. Using DEM to model ore breakage within a pilot
scale sag mill. Miner. Eng. 17, 1117–1124.
the steel are transmitted without damaging the steel and are
Morrison, R.D., Loveday, B., Powell, M., Djordjevic, N., Cleary, P.W., 2006a. Applying
reflected from the rear of the steel structure back into the rock discrete element modelling to different modes of breakage in AG and SAG mills.
where it can create fracture damage. A semi-arbitrary choice of a In: Mular et al. (Eds.), Proc. International autogenous and semiautogenous
90–10 split between rock and ball was made here and demon- grinding technology, SAG 2006, September 24–27, vol. I–III. CIM.
Morrison, R., Loveday, B., Djordjevic, N., Cleary P.W., Owen, P., 2006. Linking
strates that this does indeed explain the observed experimental Discrete Element Modelling to Breakage in a Pilot Scale AG/SAG Mill. SAG 2006,
behaviour. This means that the energy split between particles in vol. 3, pp. 407–420.
 P.W. Cleary, R.D. Morrison / Minerals Engineering 86 (2016) 75–95 95

Morrison, R.D., Shi, F., Whyte, R., 2007. Modelling of incremental rock breakage by Sinnott, M., Cleary, P.W., 2015. Simulation of particle flows and breakage in crushers
impact – for use in DEM models. Miner. Eng. 20, 303–309. using DEM: Part 2 – impact crushers. Miner. Eng. 74, 163–177.
Morrison, R.D., Cleary, P.W., 2008. Towards a virtual comminution machine. Miner. Tavares, L.M., King, R.P., 1998. Single-particle fracture under impact loading. Int. J.
Eng. 21, 770–781. Min. Process. 54, 1–28.
Morrison, R., Cleary, P.W., Loveday, B., Delaney, G.W., Cummins, S., 2011. Predicting Tavares, L.M., de Carvalho, R.M., 2009. Modeling breakage rates of coarse particles in
the evolution of rock size and shape distributions using DEM based on different ball mills. Miner. Eng. 22, 650–659.
modes of breakage in AG and SAG mills. In: Major, K., Flintoff, B., Klein, B., Thornton, C., Cummins, S., Cleary, P.W., 2011. An investigation of the comparative
McLeod, K. (Eds.), Int. Conf. Autogenous Grinding, Semiautogenous Grinding behaviour of alternative contact force models during elastic collisions. Powder
and High Pressure Grinding Roll Technology 2011 (SAG2011), Vancouver, BC, Technol. 210, 189–197.
Canada. Thornton, C., Cummins, S., Cleary, P.W., 2013. An investigation of the comparative
Musa, F., Morrison, R.D., 2009. A more sustainable approach to assessing behaviour of alternative contact force models during inelastic collisions.
comminution efficiency. Miner. Eng. 22, 593–601. Powder Technol. 233, 30–46.
Narayanan, S.S., Whiten, W.J., 1988. Determination of comminution characteristics Tuzcu, T., Rajamani, R.K., 2011. Modeling breakage rates in mills with impact energy
from single particle breakage tests and its application to ball mill scale-up. spectra and ultra fast load cell data. Miner. Eng. 24, 252–260.
Trans. Inst. Miner. Metall. 97, C115–C124. Vogel, L., Peukert, W., 2004. Determination of material properties relevant to
Napier-Munn, T.J., Morrell, S., Morrison, R.D., Kojovic, T., 1996. Mineral grinding by practicable labscale milling tests. Int. J. Miner. Process. 74S, S329–
Comminution Circuits: Their Operation and Optimisation. Julius Kruttschnitt S338.
Mineral Research Centre, Brisbane, pp. 49–92. Weerasekara, N.S., Powell, M.S., Cleary, P.W., Tavares, L.M., Evertsson, M., Morrison,
Rajamani, R.K., Mishra, B.K., 1996. Dynamics of ball and rock charge in sag mills. In: R.D., Quist, J., Carvalho, R.M., 2013. The contribution of DEM to the science of
Proc. SAG 1996. Department of Mining and Mineral Process Engineering, comminution. Powder Technol. 248, 3–24.
University of British Columbia. Whyte, R., 2005. Measuring Incremental Damage in Rock Breakage by Impact. BE
Schwager, T., Pöschel, T., 2007. Coefficient of restitution and linear-dashpot model (honours) Thesis, The University of Queensland (unpublished).
revisited. Granular Matter 9, 465–469. Wu, C.-Y., Thornton, C., Li, L.-Y., 2003. Coefficients of restitution for elastoplastic
Shi, F., Kojovic, T., 2007. Validation of a model for impact breakage incorporating oblique impacts. Adv. Powder Technol. 14, 435–448.
particle size effect. Int. J. Miner. Process. 82, 156–163. Yahyaei, M., Weerasekara, N.S., Powell, M.S., 2014. Impact of mill size on low-
Sinnott, M.D., Cleary, P.W., Morrison, R.D., 2006. Analysis of stirred mill energy surface damage. In: Proc. XXVII Int. Min. Proc. Congress (IMPC 2014),
performance using DEM simulation: Part 1 – Media motion, energy Santiago, Chile, pp. 1–10.
consumption and collisional environment. Miner. Eng. 19, 1537–1550.