Powder Technology 124 (2002) 112 – 118

www.elsevier.com/locate/powtec

The equivalence between different residence time distribution models

in ball milling
Heechan Cho a, Leonard G. Austin b,*
a
School of Civil, Urban and Geosystem Engineering, Seoul National University, Seoul 151-742, South Korea
b
College of Earth and Mineral Sciences, The Pennsylvania State University, University Park, PA 16802, USA
Received 17 September 2000; received in revised form 30 August 2001; accepted 16 October 2001

Abstract

A number of different models have been used to describe the residence time distributions (RTDs) measured for dry and wet ball milling by
various investigators in various countries. It is useful to be able to compare the results between different workers and this requires a
knowledge of the equivalence of the models. A method is presented here that permits this comparison for the usual two parameter residence
time distribution models, where one parameter is the mean residence time and the other is a characteristic parameter that defines the shape of
the residence time distribution. The method defines the best correlation as that which produces the closest agreement of the predicted mill
product size distributions over a wide range of values of mean residence time. The correlations are given in graphical form and in equation
form. In some circumstances simulation results are not sensitive to the RTD model chosen, provided that the correct match of characteristic
parameters is used. D 2002 Elsevier Science B.V. All rights reserved.

Keywords: Ball milling; Residence time distribution models

1. Introduction pared by their effect on mill product size distribution not by

Austin et al. [1] have reviewed the experimental results
of measurements of residence time distributions (RTDs) in
ball mills. Such results are often given as the characteristic
parameters of a model fitted to the data, treating the data as
equivalent to that expected from a plug flow section
followed by a fully mixed section, or m equal fully mixed
sections in series, or the axial mixing factor in the Mori et al.
[2] model, etc. These models are two-parameter models,
where the mean residence time s is one parameter and the
other characteristic parameter can be the number of sections,
the relative size of sections, the dimensionless mixing
factor, etc. The results can be compared only by establishing
the relations between these characteristic parameters.
Austin et al. [1] compared the RTD models by showing
the plots of variance of the distributions (in dimensionless
form) versus the characteristic parameters. However, there is
no proof that two residence time distribution models with
the same s and variance will produce the same degree of
grinding. In the method given here, the models are com-
*
Corresponding author. Tel.: +1-814-863-0749; fax: +1-814-865-3248.
E-mail address: lxa4@psu.edu (L.G. Austin).
the shape of the RTDs, since it is the equivalent grinding
behavior that is of interest.
2. Theory
First, it must be understood that the two residence time
distributions that have only s as the determinant (that is, the
plug flow case and the fully mixed case) cannot be used to
describe accurately the results actually found with tumbling
media mills. On the other hand, it is usually found that two-
parameter models are sufficient, bearing in mind the inher-
ent variability involved in mill tests, therefore the use of
three- or four-parameter models is not justified. Second, the
data of Concha and Pearcy [3] quoted by Austin and Concha
[4] showed excellent fits of the Mori et al. model [2] to
residence time distribution data from a 3.5
4.88 m (diam-
eter by length) industrial ball mill and a good fit to that from
a 3
4.25 m industrial rod mill, so it is assumed here that
this semi-theoretical model can be used as the basis for
comparison.
It is also assumed that the kinetics of breakage model
used in a mill simulator is strictly first-order, since it is

0032-5910/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 0 3 2 - 5 9 1 0 ( 0 1 ) 0 0 4 8 0 - 6
 H. Cho, L.G. Austin / Powder Technology 124 (2002) 112–118 113

where p1(s) is the fraction of the mill product that is in the

top size interval. Substituting for w1(t) and /(t)

Z " !#
l
1 ð1  t*Þ2
p1 ðsÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffi exp  S1 st* þ dt*
0 2 pDt *3 4Dt*
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1 þ 4DS1 s  1
¼ exp  ð4Þ
2D

where S1 can be taken as 1 (time  1). For the RTD resulting

from m equal fully mixed sections in series (where m is the
characteristic RTD parameter in this case), the same proce-
dure gives

1
p1 ðsÞ ¼
m ð5Þ
S1 s
1þ
m

Fig. 1. Comparison of the disappearance of the material from the top size
for the Mori et al. RTD model (D = 0.15) and the m equal fully mixed
sections in series RTD model (m = 4.08).
implicit in the use of the residence time distribution concept
to solve the grinding equations without integrating along the
mill axis [5,6]. Of course, in comparing the equivalence of
one residence time distribution to another the comparison is
made with all other parameters (such as the specific rates of
breakage and primary breakage distribution functions) the
same. With these assumptions, the structure of the grinding
equations is such that if two different RTDs can be found
that give a close agreement between the rates of breakage of
any one size over the complete range from little breakage to
almost complete breakage, then all the rest of the simula-
tions will also be in close agreement (see Appendix).
This means that the comparison can be made by finding
the pair of characteristic RTD parameters that give the best
match between the two RTDs in the simulation of the
disappearance of the top feed size, for any arbitrarily chosen
specific rate of breakage of this size.
For example, the Mori et al. differential RTD is [1,2]
It can be noted that although this equation is developed
for m equal sections in series, it will compute for m not an
integer. The correspondence between D and m is that which
gives the best fit of p1(s) from Eq. (4) with that from Eq. (5).
Fig. 1 shows the best fit (absolute least squares) for a
range of S1s values that takes p1(s) from 1 to about 0.01 for
D = 0.15, that is, 99% of size 1 is broken. The maximum
error in using m = 4.10 as equivalent to D = 0.15 is 0.005,
that is, 0.5%. However, this increases to 1.2% for D = 0.30.
This is to be expected because the Mori et al. RTD cannot be
used to approach a single fully mixed section, since it is
based on a solution of the one dimensional axial diffusion
equation [1] that uses a boundary condition at the mill exit
that cannot reduce to a fully mixed mill even as D becomes
large. The values of D reported by Austin and Concha [4]

1  
/ðtÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffi exp  ð1  t*Þ2 =4Dt* ð1Þ
2s pDt *3

where t* = t/s. In this equation, the characteristic parameter

is D (the dimensionless axial mixing factor or diffusion
coefficient), t is time in the mill and s is the mean residence
time. Let the fraction of feed of size 1 left after grinding for
time t in the mill be w1(t), where

w1 ðtÞ ¼ exp½S1 st* ð2Þ

Integrating over all possible times in the mill

Z l
p1 ðsÞ ¼ w1 /ðtÞdt ð3Þ
0 Fig. 2. Correlation between D and m.
114 H. Cho, L.G. Austin / Powder Technology 124 (2002) 112–118

for a typical ball mill and a typical rod mill were 0.18 and
0.16, respectively. Fig. 2 shows the correlation between D
and m over the range D = 0.05 to 0.50. The correlation
equation is

m ¼ 0:8632D0:8304 ð6Þ

D ¼ 0:8376m1:2042

and the maximum fractional error is

e ¼ 1:7
103 þ 0:0446D

3. Correlations

Equivalent equations to Eqs. (4) and (5) for a model

consisting of a plug flow section of relative length h and a
fully mixed section of length 1  h in series, and a model
consisting of a large fully mixed section of relative length /
in series with two smaller each of length (1  /)/2 are
readily derived [1]. Table 1 gives the matching results for
the following residence time distribution cases:
1. Mori et al. versus m equal fully mixed
2. Mori et al. versus plug flow + fully mixed
3. Mori et al. versus one large– two small
4. m equal fully mixed versus plug flow + fully mixed.
The correlations are shown graphically in Fig. 3, and the
corresponding maximum errors in Fig. 4.
Fig. 3. Graphical correlations between the four RTD models.
Although the Mori et al. RTD is usually appropriate for
normal ball mills, and it would be good for a system tending
to plug flow, it is not expected to be useful for systems
approaching fully mixed, such as the pancake type of semi-
autogenous grinding (SAG) mill. The one large–two small
model may be good for normal ball mills and SAG mills,
but cannot approach plug flow systems.
The m equal fully mixed series is limited in scope if m is
taken as integer, but if m is allowed to be noninteger then it
is not possible to do a mass balance on a fractional section
which has no physical reality. This means that the computa-

Table 1
Correlation equations (h = relative length plug-flow section, / = relative length large section)
Case Correlation equation Maximum error
 0.83
Mori /m equal m = 0.86D  0.002 + 0.045D
m equal / Mori D = 0.84m  1.20  0.002 + 0.037m  1.204
Mori /plug + f.m. h = 0.76exp[  3.89D]  0.002  0.031D + 0.108 D 2
plug + f.m./ Mori D =  0.26lnh  0.07 0.001 + 0.009lnh + 0.007(lnh)2
0:86
Mori /1L + 2S /¼ 0.0142  0.0545D + 0.1004D 2
1 þ expð14:13D þ 3:07Þ
 2
0:86  / 0:86  /
1L + 2S/Mori D ¼ 0:07 ln 0:0005 ln þ 0:38 þ0:007
21:50/ 21:50/

m equal / plug + f.m. h = 0.76exp[  3.24m  1.20] 0.031[1  exp(  0.538m)]  0.034
8 0 " #0:83 19
< lnh0:26 =
plug + f.m. /m equal m = 0.86[  0.26lnh  0.07]  0.83 @
0:031 1  exp 0:464 A  0:034
: 0:93 ;
 
0:86  / 0:83
1L + 2S /m equal m ¼ 0:86 0:07 ln 4.31  27.15/ + 63.50/2  65.23/3 + 24.84/4
21:50/
0:86 1:058
m equal /1L + 2S /¼ þ ðm  1:035Þ0:958
1 þ expð11:84 m1:20 þ 3:07Þ m  0:985
 
0:86  / 0:27 6
1L + 2S /plug + f.m. h ¼ 0:76  1.23
10  4(/1.65
10  exp/72.70)
21:50/
0:86
plug + f.m./1L + 2S /¼ 0.086  1.177h + 5.831h2  8.733h 3
1 þ expð3:65lnh þ 4:11Þ
 H. Cho, L.G. Austin / Powder Technology 124 (2002) 112–118 115

with equal partial sections, as can be seen by comparing the

two:

1 1

p

, f p 1,0 ð7Þ
S1 s m S1 s n S1 sf
1þ 1þ 1þ
m m m

where f = m  n, 0 < f < 1.0, and m, n are integers. The

correspondence between the two then depends on m and f,
not just on m, as seen in Table 2. The bi-variant correlation
equation is

2
x1:3335
6 ð115:7  114:7xÞ0:1954 , 1:0Vmð¼ 1 þ xÞV2:0
6
6 x
6 , 2:0Vmð¼ 2 þ xÞV3:0
6 ð42:5  41:5xÞ0:2659
6
f ¼6 ð8Þ
6 x
6 ð43:23  42:23xÞ0:2336 , 3:0Vmð¼ 3 þ xÞV4:0
6
4 x
, 4:0Vmð¼ 4 þ xÞV5:0
ð43:89  42:28xÞ0:2203
Fig. 4. The maximum differences in the cumulative product size distribu-
tions produced by replacing the Mori et al. axial mixing model with the
other RTD models, using the parameter correlation equations.
4. Test case

In order to confirm the agreement between the different

tional method proposed by Austin and Cho [7] that elimi-
nates the use of the Reid Solution [5] cannot be applied to
this case. The case of n equal sections plus 1 small section
can be treated by this method, but does not give an identity
two-parameter models, an open circuit mill simulation was
performed to compare the results using the Mori et al. RTD
to those using the m equal and the plug + f.m. RTD. The
parameters chosen for the specific rates of breakage Si and
the primary breakage distribution Bi, j were

2
a
Table 2 xi 1
Correlations of m equal (partial) sections (Eq. (5)) with n equal sections 6A x , n > iz 1
plus 1 smaller (see Eq. (8))
Si ¼ 4 o 1 þ ðx=lÞK ð9Þ
m x n f Maximum error 0 , i¼n
(fraction)
1.0 0 1 0 0 2

1.2 0.2 1 0.0548 0.006

xi c xi b
6 U þð1  UÞ , nz i > j
1.4 0.4 1 0.1293 0.006 Bi; j ¼4 xjþ1 xjþ1 ð10Þ
1.6 0.6 1 0.2334 0.004
1.8 0.8 1 0.3964 0.002 1 , iV j
2.0 1.0 1 1.0000 0
2.0 0 2 0.0000 0
where A = 1.36 min  1, a = 0.8, l = 4.0 mm, K = 2.22,
2.2 0.2 2 0.0711 0.001
2.4 0.4 2 0.1609 0.001 U = 0.65, c = 1.1, b = 5.8, xi is the upper sieve size of
2.6 0.6 2 0.2787 0.002 the M2 sieve interval indexed by i, and xo is a standard
2.8 0.8 2 0.4516 0.001 size (1 mm). The milling conditions are given in Table 3,
3.0 1.0 2 1.0000 0 and the simulations were run to give a coarse product
3.0 0 3 0 0
3.2 0.2 3 0.0800 0.0005
3.4 0.4 3 0.1774 0.0006
3.6 0.6 3 0.3014 0.0006 Table 3
3.8 0.8 3 0.4778 0.0004 Mill chosen for simulations
4.0 1.0 3 1.0000 0 Mill diameter, m 3.5
4.0 0.0 4 0.0000 0 Mill length, m 4.88
4.2 0.2 4 0.0855 0.0002 Ball diameter, mm 12.5 – 50
4.4 0.4 4 0.1874 0.0003 Volume fraction of balls, J 0.4
4.6 0.6 4 0.3151 0.0003 Ball density, tons/m3 7.9
4.8 0.8 4 0.4934 0.0002 Interstitial filling of balls, U 1.0
5.0 1.0 4 1.0000 0 Fraction of critical speed 0.75
 116
Table 4
Typical open circuit simulation results using the Mori RTD (D = 0.15), the m equal RTD (m = 4.08) and the Plug + f.m. RTD (h = 0.408)
Interval Size, Am S value, min  1 Feed, % < size Product size distribution, % < size
Coarse Medium Fine
Mori m equal Plug + f.m. Mori m equal Plug + f.m. Mori m equal Plug + f.m.

H. Cho, L.G. Austin / Powder Technology 124 (2002) 112–118

1 8000 1.18 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
2 5657 1.81 95.0 99.9 99.9 99.9 100.0 100.0 100.0 100.0 100.0 100.0
3 4000 2.12 87.0 99.8 99.7 99.8 100.0 100.0 100.0 100.0 100.0 100.0
4 2828 2.08 75.0 99.6 99.5 99.7 100.0 100.0 100.0 100.0 100.0 100.0
5 2000 1.88 60.0 99.3 99.2 99.5 100.0 99.9 100.0 100.0 100.0 100.0
6 1414 1.62 46.0 98.8 98.6 99.1 100.0 99.9 100.0 100.0 100.0 100.0
7 1000 1.34 34.0 97.8 97.5 98.2 99.9 99.8 100.0 100.0 100.0 100.0
8 707 1.05 24.0 96.0 95.6 96.4 99.8 99.5 99.9 100.0 100.0 100.0
9 500 0.81 17.0 92.7 92.4 93.1 99.4 99.1 99.7 100.0 99.9 100.0
10 354 0.61 12.0 87.5 87.4 87.7 98.6 98.1 99.0 100.0 99.9 100.0
11 250 0.47 8.0 80.0 80.0 80.0 96.7 96.2 97.3 99.9 99.7 100.0
12 177 0.35 5.5 70.5 70.6 70.4 93.3 92.9 93.8 99.6 99.2 99.8
13 125 0.27 3.8 59.7 59.8 59.6 87.8 87.6 88.1 98.7 98.2 99.2
14 88 0.20 2.5 48.7 48.7 48.7 80.0 80.0 80.0 96.9 96.3 97.5
15 62 0.15 1.6 38.3 38.2 38.4 70.3 70.4 70.1 93.4 92.9 93.9
16 44 0.12 1.1 29.3 29.0 29.5 59.4 59.5 59.2 87.8 87.6 88.1
17 31 0.09 0.7 21.8 21.5 22.0 48.4 48.4 48.3 80.0 80.0 80.0
18 22 0.07 0.5 15.9 15.6 16.1 38.1 37.9 38.1 70.2 70.4 70.1
19 16 0.05 0.3 11.4 11.2 11.6 29.1 28.9 29.2 59.4 59.6 59.2
20 11 0.04 0.2 8.1 7.9 8.2 21.6 21.4 21.8 48.4 48.5 48.3
21 8 0.00 0.0 5.6 5.4 5.7 15.7 15.5 15.9 38.0 38.0 38.1

Mean residence time, min 5.2 5.1 5.3 12.2 12.0 12.3 28.3 27.9 28.7
Mill capacity, tph 137.5 139.7 135.4 58.8 59.8 58.0 25.4 25.7 25.0
Percent difference in capacity 0 1.7  1.5 0 1.6  1.3 0 1.4  1.4
 H. Cho, L.G. Austin / Powder Technology 124 (2002) 112–118 117

(80% < 177 Am ), a medium product (80% < 88 Am) and a
fine product (80% < 31 Am).
Table 4 gives the results. Since the RTD forms are not
mathematical identities, the product size distributions are
not exactly the same nor are the mean residence times
required to meet the product specifications. In producing
the correlations given above (see Correlations), by absolute
Appendix A. Proof of equivalence matching
The treatments given by Gardner and Sukarijnajtee [9]
and Gardner and Verghese [10] can be joined with the
transfer function model given by Luckie and Austin [11]
to give the following simulation equation [12]
X i

least squares fits, the differences (referred to as ‘‘errors’’) are pi ¼ di j f j , n z i z j z1 ðA1Þ

j¼1
not randomly distributed as would be the case with exper-
imental errors. Therefore, there will be consistent differ- where
8
ences between the simulated product size distributions, as > ej , i¼j
>
<
can be seen. However, in a real mill simulation there are
dij ¼ X
i1 ðA2Þ
always experimental and model errors, and the characteristic >
> cik cjk ðek  ei Þ, i>j
:
breakage parameters (A, a, l, K, U, c, b ) are never known k¼j
with accuracy, so the match between the results from the with
different forms of RTD can be considered satisfactory if the Z l
differences are of the same order of magnitude as the ej ¼ eSj t /ðtÞdt ðA3Þ
experimental reproducibility of actual test results. 0
and
8
>
> X
j1
5. Conclusions >
>  cik cjk , i<j
>
>
>
>
>
<
k¼i
The correlations between the different characteristic 1, i¼j
cij ¼ ðA4Þ
parameters of two-parameter residence time distribution >
>
>
>
> 1 X
models that give virtually the same simulations of product j1
>
>
size distributions and mill capacities are given both graphi- >
> Sk bik ckj , i<j
: Si  Sj
cally and in equation form. These are based not on the k¼j

comparison of the shapes of the residence time distributions
but on the comparison of their effect on the results of mill
simulation result over a wide range of mean residence times.
The Mori et al. residence time distribution is appropriate
for normal ball mills but is not valid for mills where the
RTD approaches that of a fully mixed system. In addition,
allowance for different mean velocities of flow of different
particle sizes cannot be used with this RTD model. For that
case, m equal sections in series, plug flow plus fully mixed,
or the one large– two small model are more appropriate.
However, the m equal sections in series where m is not
integer can only be directly employed with a program that
uses the Reid solution in the simulation algorithms, and
again this prevents the application of the Austin technique
[8] that allows for different mean velocities as a function of
particle size. This case (m not an integer) can be converted
to one of the sections in series models and the Austin
technique is then applicable. The one large – two small
model is particularly convenient for this case.
Although three parameter models obviously enable a
better fit between model predictions and an experimentally
determined RTD, experience shows that the use of three-
parameter models is not generally justified by a worthwhile
improvement in the accuracy of mill simulations over a
range of operating conditions. The method given here could
in principle be applied to the correlation of the two adjust-
able parameters of a three-parameter RTD model with the
two of another but the correlations would be complex since
four variables are involved.
In these equations pi is the mill product particle size
vector, fi is the mill feed and dij is the transfer matrix (the
weight fraction of mill feed size j that arrives in mill product
size i). The differential residence time distribution occurs
only in the calculation of the vector ej, Eq. (A3).
Eq. (A3) can be put in dimensionless form as
Z l
ej ¼ expðsj *t*Þ/*ðt*Þdt* ðA5Þ
t *¼0
where si* = Sjs, t* = t/s, and /*(t*) is the dimensionless
residence time distribution for reduced time t/s, s being the
mean residence time. In comparing the product size distri-
butions (produced from a given feed size distribution at a
given s value in a given mill) for two different assumed
residence time distributions, the best fit occurs when the two
ej vectors agree most closely, as all other factors (Si, bij, etc.)
are the same. However, the best agreement over all relevant
values of s is when all ej values over the range s* = 0 to l
agree closely between the two compared RTDs. The ej
values are no longer dependent on j and Sj but only on
s*, where the match is required over the same range of s*,
that is
Z l
ej ¼ eðs*Þ ¼ expðs*t*Þ/*ðt*Þdt*,
t *¼0
0Vs*Vl ðA6Þ
There is only one curve of e(s*) versus s* for a chosen
(dimensionless) RTD model, and two different RTD models
118 H. Cho, L.G. Austin / Powder Technology 124 (2002) 112–118
are best matched over all times when the two curves of Eq.
(A6) agree most closely.
Since this match is independent of the particle size and S
values considered, it is convenient to choose size 1, set
S1 = 1, which makes s* = s, and use least squares minimi-
zation of the differences between the two sets of p1(s) values
to get the best characteristic model parameter for the second
model corresponding to the chosen values of the character-
istic parameter of the first model.
This treatment also shows that deducing the correlation
(between the characteristic parameters of two different RTD
models) by equating the variances of the two RTDs [13] is
not valid.
References
[1] L.G. Austin, R.R. Klimpel, P.T. Luckie, Process Engineering of Size
Reduction, SME-AIME, New York, NY, 1984, pp. 343 – 371.
[2] Y. Mori, G. Jimbo, M. Yamazaki, Kagaku Kogaku 28 (1964) 204 –
213.
[3] F. Concha, F. Pearcy, Fundamentos Dinamicos de la Mineralurgia,
Curso Panamericano de Metalurgia Extractiva, Universidad de Con-
cepcion, Concepcion, Chile, 1985 (see Ref. [4]).
[4] L.G. Austin, F. Concha, Diseno y Simulacion de Circuitos de Molien-
da y Clasificacion, Lamas Cia, Concepcion, Chile, 1994, p. 171.
[5] K.J. Reid, Chemical Engineering Science 20 (1965) 953 – 963.
[6] L.G. Austin, R.R. Klimpel, P.T. Luckie, Process Engineering of Size
Reduction, SME-AIME, New York, NY, 1984, pp. 435 – 451.
[7] L.G. Austin, H. Cho, An alternative method for programming mill
models, to appear in Powder Technology (2001).
[8] L.G. Austin, Minerals and Metallurgical Processing 16 (1999) 29 –
36.
[9] R.P. Gardner, K. Sukarijnajtee, Powder Technology 7 (1973) 169 –
179.
[10] R.P. Gardner, K. Verghese, Powder Technonolgy 11 (1975) 87 – 88.
[11] P.T. Luckie, L.G. Austin, Minerals Science and Engineering 4 (1972)
24 – 51.
[12] L.G. Austin, R.R. Klimpel, P.T. Luckie, Process Engineering of Size
Reduction, SME-AIME, New York, NY, 1984, p. 441.
[13] L.G. Austin, R.R. Klimpel, P.T. Luckie, Process Engineering of Size
Reduction, SME-AIME, New York, NY, 1984, p. 370.