Modelling Dense Media Separation
Modelling Dense Media Separation
Modelling Dense Media Separation
00
Printed in Great Britain © 1991 Pergamon Press pie
T.J. N A P I E R - M U N N
ABSTRACT
This paper reviews the literature on the modelling of both D M S bath and
cyclone separators, and identifies the strengths and weaknesses in the present
simulation capabilities. Functions for partition curves are considered in the
context of process modelling. Recent work at the JKMRC is summarised,
including the development o f models o f DM drums and cyclones, and a new
DMS computer simulation module, JKSimDM, is reported. It is concluded
that effective simulation models are now available or within reach, and the
paper concludes with a statement o f objectives for future research in this area.
Keywords
Dense medium separation; dense medium baths; dense m e d i u m cyclones;
viscosity; modelling; simulation; partition curves
INTRODUCTION
The use of computer simulation for mineral process design and optimisation is now a reality.
For many years the technique was really accessible only to those involved in the
development of the process models themselves. Most practising engineers had neither the
time nor the inclination to acquire the skills necessary to interpret and apply the models
which had been developed in academia. The experience base, so important in the evolution
of such technologies, was therefore small and likely to remain so.
This situation changed dramatically with the advent of two important recent innovations:
Several commercial simulators are now available, and the user base, though still small, is
growing rapidly. This has returned the spotlight to the essential ingredient of the simulator
- the process models themselves.
329
330 T.J. NAPlER-MUNN
McKee and Napier-Munn [1 ] have described the features of the J K M R C mineral processing
simulator, JKSimMet, and presented some case study summaries to demonstrate the
capability and application of such packages. JKSimMet has until recently incorporated only
comminution, screening and classification models, and the simulator stream structure (the
protocol for conveying stream information from one unit operation to the next) was
originally designed specifically for such models. The reality is that the great majority of
mineral processing simulation studies conducted in the industrial environment (ie by
operators, designers or consultants) has involved only comminution circuits. This reflects
the advanced status and utility of these models compared with those of concentration and
separation processes. Even flotation, which has been extensively modelled, cannot yet be
reliably simulated in a general sense because the models are not sufficiently comprehensive,
and the simulations therefore tend to be very site-specific. Until recently, the same could
also be said of gravity concentration and dense medium separation (DMS) processes, which
had enjoyed very little attention in the modelling literature.
DMS separations are conventionally represented by their partition curve, which is shown in
Figure 1. It defines the proportion of material of any given density in the feed which
reports to either the sink (conventional for mineral separations) or float product (often used
for coal separations) - the two options realise curves which are mirror images of each other.
This proportion, usually expressed as a percentage, is termed the partition number. The
location of the curve is described by the separating density or cut-point, P50, which is the
density of material dividing equally between sink and float products. The efficiency of the
process is determined by the departure of the curve from a perfect partition represented by
a vertical line at the cut-point. An empirical measure of inefficiency is the Ep Cecart
probable moyen," or probable error). Although it only describes half the curve (the central
portion) it is widely used and has therefore become a de facto standard measure. The error
area represents the material misplaced in the separation over the whole density range, but
it does not describe the curve itself.
Since particles of different size partition in different ways, a separate partition curve must
be drawn for each particle size interval. The methods of deriving the experimental partition
curve from heavy liquids fractionation of the separation products are well known [2] and
will not be repeated here. Suffice it to say that a knowledge of the size-by-size partition
curves of a given separation, together with the assays associated with each particle
size/density interval, permit the complete metallurgy of the separation to be determined,
including the yield and assay of the products. The purpose of any DMS model is therefore
to predict the partition curve for a given set of operating conditions.
Dense medium separation processes 331
Error Area
A
100
75
E
Z
50 t
I
c
0 I
I
I
m
25 t
I
I
I
The function must be capable of representing the wide range of curve shapes encountered
in practice whilst utilising as few parameters as possible. Stratford and Napier-Munn [4]
enumerated the desirable features of such a function:
Equation 1 has a fixed symmetry (defined by a) and cannot reflect independently varying
asymptotes, which arise as a consequence of short-circuiting and bypass flow, particularly
332 T.J. NApw_.s-Mtml,,r
with fines, evidenced by a failure of the observed curve to meet the 0% or 100% abscissas.
King and Juckes, in a study of fine coal beneficiation in a cyclone [8, 9], introduced two
additional parameters to control the tails of the curve, giving (for yield to float, the mirror
image of Figure 1):
Yc = B + (1 - a - B)f(x) (2)
where
. .
Y_ is the corrected partition number and c~ and 13 are the fractions of feed short-
. U
circuiting to the sink and float respectively (which can be estimated directly from the
experimental curve). Equation 2 therefore has four parameters (o~, 13, a and p50), which can
be fitted simultaneously b y non-linear procedures. The Ep can then be calculated directly
from the parameters of the function. King and Juckes [9] used the Imperfection I, defined
as
Ep = 1.099 pso/a
A number of other functions have been proposed to describe the partition curve [eg 15].
For example, Erasmus [14] proposed a 4-parameter function based on a conceptual model
of coal washing DM cyclones, which assumed that the basic separation was perfect (a
vertical partition line), but varied cyclically with time to yield the imperfect performance
observed in reality. The function is
C -- p50 for a symmetrical curve (Itll = It21). The curve is not asymptotic to Y = 0 and 1,
and t 1 and t 2 must therefore be constrained to ensure that the curve does not exceed these
Dense medium separation processes 333
bounds. Descriptors such as the Ep and error area can be calculated analytically from
Equation 9.
A typical partition curve function, then, may incorporate four parameters - one for location
(related to Ps0), one for scale (or inefficiency), and two describing the tails of the curve,
either in terms of short-circuit or bypass of fines, or to locate the end points of the curve
at Y = 0 and 1. The selection of a function will depend on its robustness, its effectiveness
in fitting particular datasets, statistical parameter estimation criteria, and no doubt personal
prejudice and comfort [4, 16].
However, once the parameters are known, perhaps from published values or from surveys
of the process in question, the characteristics of the product from the process can be
predicted by numerically applying the partition curve to the feed washability. The process
can be "simulated" by exploring the effect on product characteristics of selecting different
parameters (or even using look-up tables of partition numbers). The US Department of
Energy Coal Preparation Plant Simulator utilises this approach [17]. This can be helpful for
example in considering flowsheet options or providing guidance as to the performance (PS0
and Ep) required to achieve a given yield and product specification. It can also help m
broad equipment selection decisions (eg cyclones vs baths), if the necessary data are
available; the US DOE in particular has acquired large volumes of such data in coal
preparation applications.
However, it provides no guide as to how to achieve the required performance through the
selection of operating conditions (eg medium density or feedrate) or equipment size and
other details. This requires a process model. Progress in the development of such models
for DM baths and cyclones is now discussed.
The early dense medium processes employed bath-type separators. Most authorities identify
Bessemer's 1858 patent for a conical vessel using solutions of metal chloride salts as the first
reported DM bath process. However the first commercial plants were only built in the
1920s and these employed the Chance process, which was also the first to use the now
conventional unstable aqueous suspension of a finely divided solid (in this case, sand).
The modern DMS process involving higher density magnetic media had to await the
development of efficient magnetic separators for medium cleaning and regeneration, and
plants using magnetite in coal cleaning were operating in the 1940s [12]. The development
of 14-16 % corrosion-resistant ferrosilicon powders of specific gravity 6.6 - 7.0 in the late
1950s allowed the principles of DMS to be extended to high density separations, first in
diamond and iron ore beneficiation, and later with other minerals.
It is significant that medium properties have to some extent driven the evolution of the
technology. A particularly important property is the viscosity of the medium. The
influence of viscosity on the process, in a hydrodynamic sense, has been recognised from
the earliest studies, but very few models have been proposed in which viscosity is explicitly
incorporated [18]. A specific objective of the recent JKMRC work on the modelling of
DMS processes has therefore been to incorporate quantitatively the effects of medium
viscosity, and to decouple this property from the medium density, with which it is strongly
correlated.
For deep baths (such as cones) and coarse particles, a model is almost superfluous because
such separations tend to exhibit perfect partitioning with a cut-point only slightly displaced
from the medium density. For other conditions (finer particles and/or shallow baths which
are more susceptible to turbulence) this is not the case, and a predictive model would have
considerable value in design and optimisation.
There have been several studies of baths (mostly in coal preparation) which have established
important process trends and operating guidelines. Some have been reviewed by Napier-
334 T.J. N^pmR-Mtn~
Munn [18] and Baguley [19]. However there exist no published quantitative models of DM
baths which can be used to predict a partition curve for a particular separator and set of
operating conditions.
Scott and Lyman [20] illustrated some of these ideas in a simple diagram reproduced as
Figure 2, which assumes a bottom-fed bath such as a drum in which float particles must
rise in order to be separated. They used simple sedimentation theory to derive a general
expression for the cut-point:
a and b are exponents which depend upon the prevailing particle flow regime.
• ReJectIfloatl 1 (sPrl ~ t
Fig.2 Conceptual model of DM drum separator (after Scott and Lyman [20])
(Terminal velocities should be read as rising positive)
Based on theoretical calculations for tracers in the size range 15-60mm in an iron ore
application, Scott and Lyman concluded that most of the particles moved in the intermediate
regime and suggested values of 0.6 and 1.6 for parameters a and b respectively, though this
was not tested experimentally.
Dense medium separation processes 335
Equation 10 only considers the mechanism of sedimentation. Clearly, as has been stated,
other mechanisms also play a role and must be incorporated (empirically, if all else fails) in
any usable model. Napier-Munn [18] has considered the problem in more detail, in the
context of the literature. Some important process trends include:
Inefficiency (defined by Ep) increases with viscosity and particle fineness. Whitmore
[21] actually developed a model for Ep based on surveys of several production baths
in coal preparation.
Results on small separators do not necessarily scale to large units because of changes
in bath depth, degree of turbulence and other factors; machine variables are therefore
important.
The yield stress of the medium (characterised as a Bingham plastic) can play a role
in partitioning performance, under certain circumstances.
In 1986 the JKMRC set out to develop a realistic model of industrial DM drum separators,
based on a consideration of the hydrodynamics involved. Details of the model (which is
now available in JKSimDM) are given by Baguley [19], and were reported to a closed
industry conference by Baguley and Napier-Munn [22]; they will be published in the open
literature in due course.
The model assumes (as implied above) that there is a correlation between the partitioning
performance of particles of a given size and density, as observed in practice, and their
terminal velocity in the bath. The model therefore consists of two parts:
Empirical correlation of the calculated terminal velocity with the observed partition
number, permitting the prediction of the full partitioning performance. This
correlation is then specific to a particular unit.
The terminal velocity is calculated from the correlation published by Concha and Almendra
[23], with an adjustment for shape factor. The calculation is performed for the mean size
and mean density of each size/density interval describing the feed washability. Input data
includes medium density and viscosity. The viscosity is at present defined as an apparent
equivalent Newtonian viscosity as measured by a Debex on-line viscometer [24, 25]. The
model structure available in JKSimDM includes an empirical correlation for viscosity in
terms of medium solids concentration, C v (which can' be calculated from medium
composition and density), and temperature, t:
In (r/) -- C o + C I C v Z - C 2 t (12)
The value of the constants are specific to a particular medium, and will depend upon the
medium composition, size distribution, contamination by slimes, and extent of magnetisation
[26]. The relationship must be determined by experiment; Figure 3 shows the relationship
obtained from a particular plant using the on-line viscometer.
Surveys were made of Wemco drums in two plants, one treating iron ore and one treating
manganese ore, in order to generate experimental size-by-size partition curves. Plotting the
computed terminal velocity for each size/density interval against the observed partition
ME 4/3-.4"--1
336 T.J. NAplr.g-Muss
curves produced S-shaped curves whose characteristics varied with particle size, as shown
in Figure 4. These curves were modelled using the empirical expression:
120
°I
100
mll
8O
6O
4O
2O
0
2200 2400 2600 2800 3000 3200 3400 3600
'°° t
60
Particle Size
The model parameters A,B and V10 0 were either found to be constant for a particular unit,
or were empirically correlated with operating conditions such as medium density and
viscosity, and feed tonnage. Simulating a drum then consists of the following steps:
Dense medium separation proecsscs 337
. Calculate an array of terminal velocities based on the size and density intervals for
which feed washability information is available.
2. Select the operating conditions (medium density and viscosity, and feed tonnage).
. Select values of A,B and Vl00 to insert in Equation 13 (these will be specific to a
particular installation, and can be used in green field situations if the drum is
assumed to have the same characteristics, including feed conditions).
. Calculate a vector of Yi values from Equation 13. These constitute the predicted
partition curve, and can be used to determine the product characteristics, given the
feed washability.
As noted earlier, two sets of model parameters have so far been obtained from industrial
plants. The model has been used in a number of simulation applications, including:
The model has given plausible predictions in all these applications, although some have yet
to be tested. The next phase of development should include the acquisition of circuit-
specific parameters from more installations, and the validation of the model in coal
applications. In addition, it should be pointed out that the model structure is appropriate
for any DM bath separator. Units other than drums should therefore now be modelled.
The model structure should also be modified to limit the empirical nature of Equation 13.
Terms which should be specifically incorporated include the hole size of the drum lifters,
and a Reynolds Number term to represent turbulence in the drum.
The DM cyclone has been rather better served by the research community than the DM
bath. The process was developed from a fortuitous observation by the Dutch State Mines
in 1939 that fine coal concentrated preferentially in the overflow of a thickening cyclone
in a DM bath circuit. Patents were granted in 1942, and the DSM licensed this and
associated processes (such as the DSM screen, or sieve bend) through a subsidiary,
Stamicarbon. Stamicarbon licensees around the world promoted the process and a large
installed base developed, in both coal and minerals. As a consequence Stamicarbon
accumulated an extensive database of operating data, which was used to develop predictive
capabilities sufficiently accurate to enable Stamicarbon and its licensees to offer
performance guarantees to its clients. By any definition, this constituted a model, and a
powerful one, but it remained proprietary to Stamicarbon and was never placed in the
public domain.
The early published work on modelling DM cyclones often approached the unit as a
classifier in which the sizing function is suppressed and the density separation enhanced.
Lilge, for example, developed his Cone Force Equation [27] from a consideration of the
hydrodynamics of the hydrocyclone, which produced an expression for the classification
size, d50, in terms of cyclone geometry, fluid density and viscosity, particle density and
flowrate. Lilge and Plitt [28] then used the Cone Force Equation to develop a design
procedure for DM cyclones based on selecting a finer d50 for the dense component of the
ore than the finest size worth recovering in the underflow. The d50 was also constrained to
338 T.J. NxpmR-Mtmr~
be coarse enough to prevent excessive medium segregation. It is fair to say that some of
Lilge's ideas were contentious, and certainly the method has not enjoyed widespread
application. However it did emphasise the importance of considering the behaviour of the
medium separately from that of the density separation. Later research showed the medium
behaviour to be process determining, and modelling it is therefore essential.
Tar jan had also addressed the behaviour of the m e d i u m in classification terms [29] and used
an equilibrium orbit approach to define the conditions under which different radial medium
density profiles would occur in the cyclone; the quality of separation would then depend on
these profiles and the resulting differential between underflow and overflow medium
density.
Several workers have equated performance criteria with medium behaviour, expressed as
product medium densities. Davies et al. [30] equated the cut-point, p50, with the underflow
medium density, #., and also presented an expression for the Ep for a given size in terms
of the Ep prevailin"g for coarse (10 m m ) particles.
Collins et al. [31] also reported an equivalence of p50 and Pu, until at lower feed densities
and a correspondingly high differential ( Pu - Po), p50 dropped below Pu" They interpreted
the cyclone performance in terms of the medium behaviour, and in particular the stability
of the medium as reflected in the differential. They correlated their data for the 254mm
DSM cyclone used, with atomised ferrosilicon and magnetite mixtures as medium, as
follows:
Ep :(ajd) + a s (18)
(A regression equation for pressure drop factor was also presented). This model can
presumably only be used in a design situation if the material-specific constants are known,
which must be obtained from testwork. Also, no function was proposed for the partition
curve, so one would have to be assumed and the predicted Pso and Ep values inserted.
Napier-Munn [6], in studies of a constant geometry 610mm cyclone treating diamond iferons
ores, found that in nearly all cases P5o > Pu, which is at variance with the work reported
above. He developed an arbitrary regression model for the parameters of the Erasmus
function (Equation 9) and the product medium densities, in terms of medium and ore
characteristics, feed head and feedrate. The model has good predictive capability within
the conditions under which it was developed, but like any such model should be
extrapolated only with extreme care. In particular, it is valid only for the particular cyclone
configuration tested. Details of the model and the validity range are given in the Appendix.
Dense medium separation proczsscs 339
Later work by Napier-Munn [33] with a 100mm cyclone using density tracers also found
that Ps0 > Pu. The data were well correlated by the simple empirical expression
where a, b and c were constants which were found to be dependent on tracer size. Pu was
predicted from a bulk sedimentation model. Equation 19 was also used by Davis [12] and
Davis and Napier-Munn [34] to model both Ps0 and Ep for tracer separations in coal
washing, using Equation 1 to model the partition curve.
The studies of DM cyclones in coal washing by JKMRC workers in recent years were
encapsulated in the form of a regression model which was reported by Wood et al in 1987
[35]. Based on both pilot and full scale data, the model consists of a number of sub-models
describing all the features necessary to simulate the process for design or optimisation.
These include correlations for feed and product flowrates, product medium densities, cut-
point and Ep. Correlations for predicting the onset of particle retention (which decreases
separation efficiency) were also presented.
The model (which is currently undergoing revision with the advent of additional data)
incorporates some of the concepts outlined earlier, notably the view that the density
separation is strongly dependent on medium behaviour, and that the behaviour of coarse
particles is a characteristic from which fine particle behaviour can be deduced. One
expression given for the cut-point is not dissimilar to Equation 19:
An alternative expression was also given, as a function of vortex finder and apex diameters,
feed medium density, head and medium particle size. The cut-point for coarser particles
was found to be the same as that for -4 + 2mm. For finer particles of size d, the correlation
was proposed. The experience of most workers is that Ep is correlated with Ps0 and Wood
et al. quantified this correlation as follows:
King and Juckes [8, 9] used Equations 1, 2 and 5 to model partition curves for fine coal
beneficiation in a 150mm x 15 ° cyclone, and presented empirical correlations for the model
parameters:
where al, a2 are constants whose value depends on the presence or absence of" slimes.
( d in #m )
It should be remembered that this model was developed for fine coal (- 500 #m ). For
coarse particles., a = 13 = 0, and the partition curve function reduces to Equation I. This
model form provides a method of predicting the effect of the short-circuiting of fines
(leading to partition curves which are not asymptotic to Y = 0 and 1), and further
emphasises the strong dependency of cut-point and efficiency on particle size. However
the values of the constants are specific to the cyclone and medium utilised in the
experiments.
The model was derived from a force balance (centrifugal and radial fluid drag) with a
turbulence mechanism imposed, c~ can in practice be fitted directly to experimental
partition curves, and Equation 28 then used to simulate performance through plausible
selection of the values of the operating variables. Clarkson's model is significant in its
consideration of turbulence as a mechanism, one of the very few examples in the DM
cyclone literature.
Of the cyclone models discussed above, only the JKMRC empirical model developed by
Wood et al. [35] for coal preparation makes any attempt to accommodate all the design and
operating variables necessary to undertake useful simulation exercises. This model, and the
refinements currently being implemented, is therefore being made available within the
JKSimDM structure. However, it is not appropriate for high density (mineral) separations.
In particular, it does not incorporate medium viscosity as a separate term, which becomes
more important at the high operating densities [18]. The objective of the study by Scott
[13], therefore, was to develop a model which could decouple viscosity from the other
variables, and could also provide a comprehensive description of the process, particularly
at the higher operating densities. The work has been described in a closed industry
conference [37] and will be published in the open literature in due course.
Scott proceeded from the observation that size-by-size partition curves for DM cyclones
"pivot" about a point which is controlled by the behaviour of the m e d i u m in the cyclone.
In the case of pure liquids or stable (neutrally buoyant) suspensions, such a phenomenon
would be expected [38]; the 'pivot point' would be defined by the density of the medium
and the yield of medium to underflow, since particles of the same density as the medium
would experience no separating force and would therefore partition in the same proportion
as the medium. The effect is illustrated in Figure 5. A more surprising result inferred
from the literature [e.g.9] and Scott's own work was that the pivot phenomenon prevailed
even for cyclones operating with conventional unstable media, despite the segregation of the
medium which occurs. He therefore incorporated the phenomenon into the partition curve
function and developed empirical models to predict the resulting function parameters.
A consequence of the pivot phenomenon is that the P50 and Ep must be positively correlated
(for size-by=size curves), The position of the pivot point is defined by the pivot parameters
Yp, pp (Figure 5). Substituting these values into Equation 8 gives for the Ep
100
Siz
80'
40'
It is well known that Ep is an inverse function of particle size (see Equations 18 and 26),
and Scott's data confirmed this. He therefore proposed the relation
where k is a model parameter, and n is a hydrodynamic constant (1.3 for the pilot plant
work). Equations 30 and 31 therefore constitute a 4-parameter model of the process, the
parameters being Y,, p,, Ep and k. Scott carried out pilot and industrial experiments on
100, 200 and 400mrh USM cyclones with ferrosilicon media over the density range 1456 -
3025 kg m "3 to establish predictive correlations for these parameters. In the pilot plant
work, the medium viscosity was varied directly by adding molasses and in all eases the
viscosity was measured using an on-line viscometer.
The general form of the resulting linear regression equations for the 200ram pilot plant unit
was (in the order in which they must be computed for simulation):
Ep = z + k d "n (33)
n was set at unity, and z was a linear function of Ap and volumetric medium/ore ratio.
Other relevant correlations were:
In k -- f (In r/, In De )
Y-o -- 0.61 Rm
In~[(Vp/Vf) - 1] -- f [In Vf,ln (o/Q)] (34)
In both cases, an estimate is required of Ap, which is a dependent variable. Scott developed
two approaches to this problem. The first consisted of empirical correlations for Pu and Po
in terms of operating conditions (including viscosity). The second modelled the
classification of the medium in terms of the efficiency curve (Equation 1) with a constant
value of a, and cut-size and water recovery empirically correlated with feed medium
density and viscosity. Scott also reported a procedure for considering the influence of
cyclone diameter in terms of a prevailing bulk flow Reynolds number, which modified the
correlations for the model parameters.
In summary, the JKMRC has developed two models of DM cyclones, one specific to coal
and the other a more general model. Both have strong empirical elements, but they do
incorporate all the design and operating variables which would normally be considered in
a simulation exercise. An important innovation in Scott's model is the decoupling of
medium viscosity and medium density which allows the viscosity to be separately
manipulated in the model, although this does introduce the requirement that a value for
viscosity be known. Also, both models have some scale-up capability through the
incorporation of cyclone geometry in some form.
Neither model has yet been extensively tested in simulation, and both will undoubtedly
benefit from further refinement. However, they do provide an effective and useful
structure within which this refinement, which is now underway, can take place.
This paper has reviewed the modelling of dense medium separation processes in terms of
the needs of practical, effective and useful process simulation methodologies. In this
context, the modelling of dense medium processes is not yet as advanced as that of
comminution and classification processes, although significant progress has been made in
recent years. Historical studies of both baths and cyclones have resulted in the
eharacterisation of important process effects which must be taken into account in any
comprehensive model. These include -
Very little DM process simulation has been reported in the literature. This is because the
quantitative correlations derived by most workers have tended to be directed to exploring
Dense medium separation processes 343
one particular aspect of the process, or have excluded industrially important variables such
as feedrate. Consequently, no comprehensive predictions model of either baths or cyclones
has been reported. King and Juckes [8, 9] provided a useful structure for modelling the
performance of fine coal separations in cyclones, and indeed any DM cyclone separation in
which significant by-pass occurs, and Napier-Munn [6] presented a quantitative cyclone
model for one particular industrial application. Apart from these, however, the only models
which can claim both industrial utility and some general application is Baguley's model of
DM drums [19, 22], the cyclone model of Wood, Davis and Lyman [35] for coal applications,
and the general cyclone model of Scott [13, 37]. All these are, or shortly will be, available
in the JKMRC dense medium process simulator, JKSimDM.
Baguley's drum model has a phenomenological form which is probably suitable for any
bath-type separator, although this has not yet been tested for units other than drums.
Future work on the model will address the empirical nature of the size-dependency of the
correlation between the prevailing partition number and the calculated terminal velocity
(Equation 13). The utility of the model will be considerably enhanced by expressing this
relationship in terms of design and operating variables, such as a bulk flow Reynolds
number (describing turbulent mixing effects) and lifter dimensions (describing the tendency
to misplace fine heavies though the lifter drainage holes). Expansion of the operating
database is also necessary to develop a parameter library for particular machines, and to
calibrate the model for coal applications.
Further operating DM cyclone data are also required, for the same reasons. The objective
of the next phase of modelling work must be to integrate the various approaches, to
synthesise a phenomenological description of the process, possibly in terms of Scott's pivot
partition function with elements of Clarkson's model [36] to account for turbulence effects.
New surveys of industrial plants have just been conducted with this in mind, and a new
phase of model development is underway.
Much work remains to be done, but the ability to simulate industrial dense medium plants
effectively and with confidence is for the first time firmly within our grasp.
REFERENCES
36. Clarkson C.J. A model of dense medium cyclones. Dense Medium Operators Conf.,
Brisbane, 235 - 245 (Aus. Inst. Min. Met.), 1987.
37. Scott I.A. and Napier-Munn T.J. A dense m e d i u m cyclone model for simulation.
4th Samancor Symp. on Dense Media Separation, Cairns (Samancor Ltd), Feb. 1990.
38. Napier-Munn T.J. The influence of medium viscosity on the density separation o f
minerals in cyclones. 1st Int. Hydrocyclone Conf., Cambridge, 63 - 82 ( B H R A ) ,
1980.
APPENDIX
The model consists of four regression equations to predict the parameters of the Erasmus
partition curve function (Equation 9), and two equations to predict the product medium
densities. The independent variables are functions of ore feed characteristics, ferrosilicon
medium size distribution, medium density, head and feed rate. The Erasmus parameters
exhibit some degree of correlation, and the K and C parameters therefore themselves also
appear as independent variables. The equations were developed b y standard stepwise
multiple linear regression analysis procedures.
It should b e remembered that the Erasmus function is not asymptotic to Y = 0 and 100%,
and for simulation purposes the constraints
Y = 0 f o r ! p r e d -< 0
Y : 100 for Ypred -> 100
should be applied, where Ypred is the partition number predicted by Equation 9 using
parameter values predicted by Equations A l.l. - A l.4.
The medium and ore size distributions are expressed in terms of the parameters of the
R o s i n - R a m m l e r function, which were estimated for each dataset:
NOMENCLATURE
A location p a r a m e t e r of R o s i n - R a m m l e r distribution.
B -- scale p a r a m e t e r o f R o s i n - R a m m l e r distribution.
C -- location p a r a m e t e r of Erasmus funct i on.
d --- particle size (#m f or m e di um, m m f o r ore)
F -- feedrate ( t / hr ) .
GA --- A - parameter o f R o s i n - R a m m l e r fit to ore f e e d size distribution.
GB = B - par a m et er of R o s i n - R a m m l e r fit to ore f e e d size distrtibution.
Gs0 = Mean ore SG (determined as 50% point on cum ul at i ve distribution curve)
H -- static head (m), or equivalent head = P x 0 . 1 0 1 9 7 / p f .
K -- scale p a r a m e t e r of Erasmus function.
MA -- A - parameter of R o s i n - R a m m l e r fit to f e e d m e d i u m size distribution.
MB = B - p a r a m e t e r of R o s i n - R a m m l e r fit to f e e d m e d i u m size distribution.
P = pressure drop (k Pa)
t1 = scale p a r a m e t e r (upper tail) o f Erasmus f u n c t i o n .
t2 = scale p a r a m e t e r (lower tail) o f Erasmus f u n c t i o n .
pf = feed m e d i u m SG
Po = overflow m e d i u m SG
#u = u n d e r f l o w m e di um SG