Energy, Mines and Énergie, Mines et

Resources Canada Ressources Canada

CANMET
Canada Centre Centre canadien
for Mineral de la technologie
and Energy des minéraux
Technology et de l'énergie

2- 1

SIPOC
C9127

Simulated
Processing of
Ore and Coal

Chapter 3 Material Balance

Energy, Mines and Énergie, Mines et
Resources Canada Ressources Canada

CANMET
Canada Centre Centre canadien
for Mineral de la technologie
and Energy des minéraux
Technology et de l'énergie

The

SPOC Manual
Chapter 3 Material Balance
Material Balance Computation for
Process Evaluation and Modelling

D. Laguitton and D. Hodouin

Editor: D. Laguitton

Mineral Processing Plant Simulation

Minerals Research Program
Mineral Sciences laboratories

SP85-1 /3E
 Minister of Supply and Services Canada 1985
Available in Canada through
Authorized Book Agents
and other bookstores
or by mail from
Canadian Government Publishing Centre
Supply and Services Canada
Ottawa, Canada K IA 0S9
Catalogue No.: M38-16/3-1985E Canada: $7.00
ISBN 0-660-11859-9 Other Countries: $8.40
Price subject to change without notice
Disponible en français
 THE SPOC MANUAL

The SPOC* manual consists of eighteen chapters, published separately. Their numbers and short titles are as follows:

1. Summary 5. Unit Models: Part A

2. Sampling Methodology 5.1 Unit Models: Part B
2.1 SAMBA Computer Program 5.2 Unit Models: Part C
2.2 Grinding Circuit Sampling 6. Flowsheet Simulators
3. Material Balance 7. Model Calibration
3.1 BILMAT Computer Program 7.1 STAMP Computer Program
3.2 MATBAL Computer Program 7.2 FINDBS Computer Program
4. Modelling and Simulation 7.3 RTD and MIXERS Computer Programs
4.1 Industrial Ball Mill Modelling 8. Miscellaneous Computer Programs

These chapters are available from: CANMET, Energy, Mines and Resources Canada
Technology Information Division
555 Booth Street
Ottawa, Ontario

*Simulated Processing of Ore and Cal

i
 THE SPOC MANUAL

The SPOC* manual consists of eighteen chapters, published separately. Their numbers and short titles are as follows:

1. Summary 5. Unit Models: Part A

2. Sampling Methodology 5.1 Unit Models: Part B
2.1 SAMBA Computer Program 5.2 Unit Models: Part C
2.2 Grinding Circuit Sampling 6. Flowsheet Simulators
3. Material Balance 7. Model Calibration
3.1 BILMAT Computer Program 7.1 STAMP Computer Program
3.2 MATBAL Computer Program 7.2 FINDBS Computer Program
4. Modelling and Simulation 7.3 RTD and MIXERS Computer Programs
4.1 Industrial Ball Mill Modelling 8. Miscellaneous Computer Programs

These chapters are available from: CANMET, Energy, Mines and Resources Canada
Technology Information Division
555 Booth Street
Ottawa, Ontario

*Simulated Processing of Ore and Cal

i
 ABSTRACT
This chapter of the SPOC manual has been developed for the participants to a workshop on material balance
computation. The various methods of material balance calculation are reviewed in order of increasing complexity. The
two packages BILMAT and MATBAL are largely documented in Chapters 3.1 and 3.2 to which the reader is referred.
The justification of methods requiring computers is presented on the basis of a search for best estimators,
i.e., estimators that take into consideration as much experimental information as possible in order to provide "well
informed" estimates. Each family of methods is illustrated by numerical exercises.

RÉSUMÉ
Ce chapitre du manuel SPOC a été écrit pour les participants à un atelier sur les bilans matière. Les diverses
méthodes d'equilibrage de bilans matière y sont passées en revue par ordre de complexité croissante. Les deux
programmes BILMAT et MATBAL sont amplement décrits dans les chapitres 3.1 et 3.2, auxquels le lecteur est
renvoyé. Les méthodes requérant un ordinateur sont justifiées par la recherche d'estimateurs améliorés qui tiennent
compte du maximum d'information possible. Des exercices numériques illustrent chaque groupe de méthodes.

ACKNOWLEDGEMENTS

The SPOC project has benefited from such a wide range of contributions throughout the industry, the university, and
the government sectors that a nominal acknowledgement would be bound to make unfair omissions. The main groups
that contributed are: the various contractors who completed project elements; the Industrial Steering Committee
members who met seven times to provide advice to the project leader; the various users of project documents and
software who provided feedback on their experience; the CANMET Mineral Sciences Laboratories staff members who
handled the considerable in-house task of software development, maintenance, and documentation; the EMR
Computer Science Centre staff who were instrumental in some software development; and the CANMET Publications
Section. Inasmuch as in a snow storm, every flake is responsible, their contributions are acknowledged.

v
 CONTENTS
THE SPOC MANUAL i
FOREWORD iii
AVANT-PROPOS iii
ABSTRACT v
RÉSUMÉ v
ACKNOWLEDGEMENTS v

1. INTRODUCTION 1
2. SYMBOLS AND DEFINITIONS 2
2.1 Symbols 2
2.2 Definitions 2
3. TRADITIONAL METHODS 3
3.1 Two-Product Formula 3
3.2 Three-Product Formula 3
3.3 n-Product Formula 3
3.4 Limitation of the n-Product Formula 4
4. METHODS OF IMPROVING ESTIMATES 7
4.1 Regressions 7
4.1.1 Orthogonal regression 7
4.1.2 Other regressions 7
4.1.3 Regression forced through the origin when i('
,Y 0,0 8
4.2 Node Imbalance Minimization 8
4.3 Generalized Least-Squares Method 9
4.3.1 Lagrange multipliers 9
4.3.2 Weighting factors 10
5. ERROR STRUCTURE IN THE EXPERIMENTAL DATA 11
5.1 Components of the Total Error for an Assay Value 11
5.1.1 Fundamental error 11
5.1.2 Errors due to time and space variations in the lot 11
5.1.3 Errors due to sampler design and operation 11
5.1.4 Errors due to sample preparation and analysis 11
6. SENSITIVITY ANALYSIS 12
6.1 Introduction 12
6.2 The Analytical Method of Sensitivity Analysis 12
6.2.1 Principle 12
6.2.2 Factors influencing the variance estimates 12
6.2.3 Practical use of the estimate variances 13
6.2.3.1 Confidence interval 13
6.2.3.2 Residuals analysis 13
6.3 The Monte-Carlo Analysis 13
7. MATERIAL BALANCE PROGRAMS 14
7.1 MATBAL: Iterative Lagrange Solution to the GLS Problem 14
7.2 BILMAT: Hierarchical Solution to the GLS Problem 14
7.2.1 Flowsheet variables 14
7.2.2 Flowsheet analysis 15
7.2.3 Mass conservation equations 15
7.2.4 Independent variables 15
7.2.5 Data redundancy 15
7.2.6 BILMAT algorithm 16
7.3 Node Imbalance + GLS Combination Methods 16
7.4 Miscellaneous 16

vii
 8. EXERCISES ON MATERIAL BALANCE COMPUTATION 17
8.1 Exercise 1: n-Product Formula 17
8.2 Exercise2: Regression 17
8.3 Exercise3 18
8.4 Exercise4 18
8.5 Exercise5 19
8.6 Exercise6 19
8.7 Exercise7: Sensitivity Analysis by the Monte-Carlo Method 19
8.8 Exercise8: Sensitivity Analysis by the Analytical Method 19
9. ANSWERS TO EXERCISES 20
9.1 Exercise 1 20
9.2 Exercise 2 20
9.3 Exercise 3 20
9.4 Exercise 4 20
9.5 Exercise 5 20
9.6 Exercise 6 20
9.7 Exercise 7 21
9.8 Exercise 8 21
10. REFERENCES 22

TABLES
1. Symbols for material balance 2
2A. Hydrocyclone experimental data (non-cumulative) 5
2B. Hydrocyclone experimental data (cumulative passing) 6

FIGURES
1. One product 1
2. Two products 3
3. Three products 3
4. Four products 4
5. Hydrocyclone. CIR = - 5
0
6. Raw data 4" hydrocyclone (non-cumulative) 7
7. Regressions on hydrocyclone non-cumulative data 8
8. Regressions on hydrocyclone cumulative data 8
9. Principle of the two-error estimators 12
10. Various levels of the mass balance variables 14
11. Sample flowsheet for data redundancy test 15

viii
 1. INTRODUCTION
The material balance problem in a process flowsheet species 1 in the solid of stream j, the following mass
arises from the necessity for the process engineer to conservation equations are valid:
make periodical inventories of the distribution of mate-
=
rial through the various process streams. It may be
dictated by bookkeeping requirements (e.g., concen- Wprii =
trate inventories, daily, monthly, yearly), by trouble where the asterisk denotes the true value of the vari-
shooting or process control requirements (e.g., circulat- ables.
ing load calculations), by environmental requirements
(e.g., tailing flow rates and assays) or by other require- In practice, one can only measure an experimental esti-
ments. This problem is also designated as a "mass mate of W1 and X and it is essential to keep in mind that
balance", "metallurgical inventory", or "process audit". its use as a true value is a decision that bears con-
sequences on the confidence level associated with any
Ideally, if one could measure the true value of every figure resulting from such an assumption. Intuitively it is
assay and flow rate in a flowsheet, the material balance obvious that the better the experimental precision, the
problem would not exist. However, the reality of process more valid the assumption that it is the best estimate. In
instrumentation is such that 1) only a few process vari- practice, we will see that better estimates than W1 and Xj
ables can be measured and these are mostly assays, can be calculated. They will be called Wi and Rj, the
2) the experimental estimates of these variables are maximum likelihood estimates.
subject to errors from various sources and do not satisfy
mass balance principles. The material balance calculations can be divided into
two families; one where the experimental estimates are
The mass balance principle is that in our usual environ- taken as best estimates, and the other where improved
ment, the total mass of material undergoing a physical estimates are calculated. The following discussion will
or chemical transformation remains constant. This cover successively the traditional methods correspond-
stops being true in relativistic mechanics where mass ing to the former case, and more recent methods corre-
can be converted into energy and vice versa. sponding to the latter.
In all mineral processes therefore, the mass of con-
sumed elements is equal to that of produced elements.
For instance, the mass of copper or chalcopyrite enter-
ing a mill or a cyclone or a flotation cell per unit of time is WI W2
equal to that leaving the same unit per unit of time
provided the steady-state conditions have been XI
reached.
X2
In a single unit such as shown in Figure 1, if Wi repre-
sents the solid flow rate of stream i, and Xj an assay for Fig. 1 - One product

1
 2. SYMBOLS AND DEFINITIONS
2.1 SYMBOLS 2.2 DEFINITIONS
Symbols are to equations what words are to a text. In all Chapters 3.1 and 3.2 contain their own glossary of
likelihood two different authors will select different sym- terms. For the purpose of this chapter the following are
bols to represent the same theory unless a general useful.
convention has been established through years of prac- Random or stochastic variable
tice in a field of science. A variable which can take different values from one
Such conventions do not exist in material balance com- measurement to another, depending on the random
putation. The two packages MATBAL and BILMAT have error.
evolved from two distinct groups and suffer from this Deterministic, or error-free, variable
heterogeneity. For instance, the symbol for flow rates in A variable which can be determined without random
MATBAL is W (for Weight) while it is D in BILMAT (for error (e.g., number of increments in a composite
Débit, French for Flow Rate). The MATBAL symbols sample).
were kept as in the original papers by Wiegel et al. in
order to maintain a coherence within a family of pro- True-value of a random variable
grams and papers. Only through years of practice will That value which cannot be exactly determined.
these differences vanish. Estimate
The following Table 1 is an attempt to define uniform The value of a variable obtained via an estimator.
series of symbols for a material balance context and will Estimator
be adhered to in this chapter. A method to try to access the true value of a variable.
Maximum likelihood estimates
Table 1 — Symbols for material balance In the case of a material balance problem based on
normal error models, these are the minimum variance
general symbol for flow rate estimates of the process variables. They are the best
wp pulp flow rate available estimates.
sw solid flow rate Error
Ww water flow rate The difference between the true value and the estimate.
feed flow rate of a cyclone Residual
0 overflow flow rate of a cyclone The difference between the estimate and the experi-
mental value.
underflow flow rate of a cyclone
fi , oi , u1 assay i of the feed, overflow or Probability distribution
underflow A graph or table giving the probability that an estimate of
a variable takes a given value.
II imbalance for element i across node j
Re RI? Rr, RI,
Normal probability distribution (or gaussian, or bell
residual in stream j for water,
pulp, solids, assay i or assay k shaped)
Rik (or AXI AX) A distribution characterized by two parameters; its
in class i.
mean and its standard deviation. It is usually followed by
Er', Ep, Er, error in stream j for water, pulp,
etc.
random errors resulting from several contributing errors.
El, Elk
Biased estimator
xik assay for element k in class i in An estimator is said to be biased when the mean value
branch j of the estimate it produces does not approach the true
k*I true value (any symbol with asterisk) value as the number of estimations increases.
Rik estimate (any symbol with hat)
ci standard deviation, e.g., of
estimate of CF, e.g., sr
9 variance (or 0-2)
V estimate of V (or S2)
Si , , sL solid weight fraction in pulp, in liquid
Lp L liquid weight fraction in pulp, in solid
100 S P per cent solid

Note: To avoid confusion with exponents, assay XI is repre-

sented by Xi for i = 1, Xffor = 2, Xi'for i = 3, etc. In prac-
tical examples, one can use Cui , Zni , Pbi , Fei , etc.

2
 3. TRADITIONAL METHODS
These are known under the name of n-product formulae. Note that for a two-product unit, all streams can be
They consist in calculating n-product variables around a determined if one assay is known in each stream.
process unit where experimental variables can be mea-
sured. Usually the unknown variables are the flow rates 3.2 THREE-PRODUCT FORMULA
and the known variables are the assays. The symbol, n,
refers to the number of streams in which the flow rate In a three-product situation, as in Figure 3, the valid
must be calculated. In the following, W represents the equations are:
flow rate of interest and could apply to pulp, solid, or wi = W2 + ± W4
liquid depending on which is analyzed.
wixi = W2X2 W3X3 W4X4
The simplest case is the one-product-formula valid for a
simple unit, such as represented in Figure 1. The valid If the three-product streams' flow rates are to be calcu-
equations are: lated, one must measure a second assay Xi in all
streams in order to obtain a third equation:
wi = W2
WiX; = W2; ± WA + W4X4'
W1X1 = W2 X12 The same elimination technique used above gives the
The variables are: W 1 , W2, XII,X12 following results:
In order to calculate, for example W2, it is only neces- W4 = (XY — X2' )(X3 — X2) + (X1 — X2)(X2 — X3' )
sary to know W1 . No assay XI is required. Similarly, Xi2 is
Wi (X4' — ;)(X3 — X2) + (X4 — X2)(X2 — X'3 )
known if only XII is known. In both cases, the value of the
calculated variable has the same error as that of the and others by permutation: e.g.,
measured variable. W3 = (X1' — X2)(X4 — X2) + (X 1 — X2)(X2' — X4' )
Wi (X3' — ;)(X4 — X2) + (X3 — X2)(; — X;4)
3.1 TWO-PRODUCT FORMULA
Note that for a three-product unit, all streams can be
A two-product problem corresponds to the illustration in determined if two assays are known in each stream.
Figure 2. The valid equations are:
wi = w2 +
3.3 n-PRODUCT FORMULA
wixil = W2XI2 W3X6 As a generalization of the above cases, an n-product
These two equations allow us to calculate a maximum of problem can be solved if n-1 assays are known on
two unknowns. W2 and W3 are the usual variables to be each stream. As n grows bigger, the substitution method
calculated. By simple elimination one can easily show of solving the equation becomes cumbersome and the
that: determinant method can advantageously be used.
vv3 = Xi — 2 ■■■ , 1 and W2 -= X 1 — X3 WI This method simplifies the solution of a system of linear
w
X3 — X2 X2 — X3 equations to a calculation of determinants.
It is important to note that all conservation equations Consider the system:
remain valid to a scale factor, i.e., if the total mass of ai xi + b 1 x2 + C1X3 = di
material in the circuit is, for example, doubled, the above
a2x1 + b2x2 + c2x3 = d2
formulae apply and it is therefore customary to express
all flow rates as a ratio to a feed of 1 or 100 flow units. 3x1 + b3x2 + c3x3 = d3 a

2
WI 3
X2
Xi W3 4
1X3
Fig. 2 - 71,vo products Fig. 3- Three products

3
It can also be written in matrix notation: using the determinant method:
MX= D 1 1 11
with (al b1 7.1 1 21
5.7 40 2 2
= a2 b 2 C2
a3 b3 C3 2.26 1 10 0.1 2216.07
W2 - 0.1
(Xi 1 1 11 22160.7
X X2) 60 1 21
X3 140 22
1 1 10 0.1
c
D d)
2
1 1 11
d3 60 7.1 21
15.7 22
By definition, the determinant of matrix M is: 1 2.26 10 0.1 2216.02
W3 - 0.1
22160.7 22160.7
al ID1 cl
det M = a2 b2 C2 = a1 (b2c3 - b3c2)
and similarly: W4 = 0.2, W5 = 0.60
a3 b3 C3 - b1 (a2c3 - a3c2)
+ c1 (a2b3 - a3b2)

Any determinant can be calculated along any row or 3.4 LIMITATION OF THE n-PRODUCT
column by carefully alternating the +1 and -1 factors in FORMULA
front of the various terms. The solution of the system of The limitation of the n-product formula comes from the
equations M X = D is given by: fact that the calculated variables are influenced by the
_
- M* precision of the obse rved variables in a non-quantified
way. Consequently, when more than n-1 assay types are
known in each stream each subset of an n-1 assay will
where Mi*designates the determinant of matrix M after give a different answer which can discredit the results
column i has been replaced by the values in D. completely to the plant engineer's eyes. Let's, for
This can be illustrated by an example from Taggard's instance, consider a two-product circuit where the fol-
handbook of mineral dressing, (pages 12 - 123), for a lowing data are known:
four-product circuit. Cu Zn % Fe %
The following assays are known: 1. Feed 0.163 3.93 11.57
Pb % Zn % Cu % 2. Conc 0.657 52.07 14.67
Feed 7.1 5.7 2.26 3. Tail 0.140 0.49 13.09
Pb conc 60 1 1 Although three assay types have been measured in
Zn conc 1 40 1 each stream, only one is required to calculate the con-
Cu conc 2 2 10 centrate and tailing flow rates; W2 and W3 by the two-
Tail 1 2 0.1 product formula.
For a feed of 100 t/h, the results corresponding to each
The valid equations, according to Figure 4, are: mode of calculation are:
Wi = W2 + W3 + W4 + W5 with Cu data: W2 = 4.5 W3 = 95.5
7.1W1 = 60 W2 + W3 + 2 W4 + W5 with Zn data: W2 = 6.7 W3 = 93.3
5.7W1 = W2 + 40 W3 + 2 W4 + 2 W5 with Fe data: W2 = -96.2 W3 = 196.2
2.26 Wi = W2 + W3 + 10W4 + 0.1W5 These three sets of results are quite di fferent due to the
large variations in assay values and precisions.
A conventional error calculation performed on the two-
product formula gives the following result:
AW2 < AW1 AX1 + AX3 AX2 + AX3
3 W2 X1 -X3 X2 - X3
W1
4 The relative error on a calculated flow rate depends,
• therefore, not only on the magnitude of the errors on all
assays but also on the di fferences, X1 - X3 and X2 - X3,
which are proportional to the quality of the separation
Fig. 4- Four products being performed.

4
For an ideal separation, one has:
o
X3 = 0 and '^'W2 :5; '^^Wi + ^Xi + ^X2
W2 Wi Xi X2
If all assays are known with a 6.0% accuracy in the
example above, the maximum relative error on W2 is
found to be 19.2% with Cu assays, 13.1% with Zn, 156%
with Fe.
A more realistic estimate of the sensitivity of the solution
to assay errors can be obtained by a technique called
Monte-Carlo Simulation which requires a computer. In
this method, a perturbation similar to the suspected
random error present in all assays is imposed on the
measured value in order to simulate a set of synthetic Fig. 5 - Hydrocyclone. CIR = ô
assay values. These are used for an n-product solution
and the process is repeated enough times to accumu-
late a number of possible solutions. The mean and In most cases, the variable of interest is the circulating
standard deviation of these possible solutions can be load ratio U/O, i.e., the ratio of solid flow rates in the
calculated. underfiow and the overflow respectively. Using the two-
This method, applied to the problem above, gave the product formula, with f,, u,, and o,, the non-cumulative
following results: per cent retained on size interval i, one has:

,^W2 = 72% with Cu assays, 8% with Zn, CIR, = U = f, - o,

W2 1000% with Fe O u, - f,
The problem of selecting the best subset of assays in each interval i gives a different value of CIR, as shown in
order to calculate the unknown variables becomes Table 2.
especially acute when the assays consist of size dis- Having so many possible estimates of CIR, one can
tributions with up to a dozen intervals or more. Process start thinking of a method to determine a best estimate.
evaluation around a hydrocyclone is typical of this situa- Several methods of obtaining such an estimate are sub-
tion as shown in Figure 5. sequently described.

Table 2A - Hydrocyclone experimental data (non-cumulative)

F U O F-U O-F (F-U)(F-U) (O-F)(O-F) (F - U)(O - F) CIR

3.97 5.95 0.00 -1.98 -3.97 3.92 15.76 7.86 2.01
15.25 21.45 0.00 -6.20 -15.25 38.44 232.56 94.55 2.46
14.80 20.70 0.00 -5.90 -14.80 34.81 219.04 87.32 2.51
9.71 13.05 0.00 -3.34 -9.71 11.16 94.28 32.43 2.91
8.27 11.05 0.80 -2.78 -7.47 7.73 55.80 20.77 2.69
4.70 6.03 0.00 -1.33 -4.70 1.77 22.09 6.25 3.53
10.06 9.44 0.00 0.62 -10.06 0.38 101.20 -6.24 -16.23
4.01 4.53 1.82 -0.52 -2.19 0.27 4.80 1.14 4.21
2.25 1.56 6.61 0.69 4.36 0.48 19.01 3.01 6.32
1.92 1.23 16.83 0.69 14.91 0.48 222.31 10.29 21.61
2.26 1.04 22.32 1.22 20.06 1.49 402.40 24.47 16.44
2.56 0.50 20.70 2.06 18.14 4.24 329.06 37.37 8.81
2.67 0.99 16.04 1.68 13.37 2.82 178.76 22.46 7.96
3.07 0.65 9.85 2.42 6.78 5.86 45.97 16.41 2.80
3.38 0.31 3.21 3.07 -0.17 9.42 0.03 -0.52 -0.06
3.63 0.56 1.39 3.07 -2.24 9.42 5.02 -6.88 -0.73
2.29 0.23 0.42 2.06 -1.87 4.24 3.50 -3.85 -0.91
5.20 0.73 0.01 4.47 -5.19 19.98 26.94 -23.20 -1.16
SUMS
100.00 100.00 100.00 0.00 0.00 156.92 1978.52 323.64 65.17

MEANS
5.56 5.56 5.56 0.00 0.00 8.72 109.92 17.98 3.62

5
 Table 2B - Hydrocyclone experimental data (cumulative passing)

F U, 0 F-U 0- F (F-U)(F- U) (0- F)(0 - F) (F- U)(0 - F) CIR

96.03 94.05 100.00 1.98 3.97 3.92 15.76 7.86 2.01
80.78 72.60 100.00 8.18 19.22 66.91 369.41 157.22 2.35
65.98 51.90 100.00 14.08 34.02 198.25 1157.36 479.00 2.42
56.27 38.85 100.00 17.42 43.73 303.46 1912.31 761.78 2.51
48.00 27.80 99.20 20.20 51.20 408.04 2621.44 1034.24 2.53
43.30 21.77 99.20 21.53 55.90 463.54 3124.81 1203.53 2.60
33.24 12.33 99.20 20.91 65.96 437.23 4350.72 1379.22 3.15
29.23 7.80 97.38 21.43 68.15 459.24 4644.42 1460.45 3.18
26.98 6.24 90.77 20.74 63.79 430.15 4069.16 1323.00 3.08
25.06 5.01 73.94 20.05 48.88 402.00 2389.25 980.04 2.44
22.80 3.97 51.62 18.83 28.82 354.57 830.59 542.68 1.53
20.24 3.47 30.92 16.77 10.68 281.23 114.06 179.10 0.64
17.57 2.48 14.88 15.09 -2.69 227.71 7.24 -40.59 -0.18
14.50 1.83 5.03 12.67 -9.47 160.53 89.68 -119.98 -0.75
11.12 1.52 1.82 9.60 -9.30 92.16 86.49 -89.28 -0.97
7.49 0.96 0.43 6.53 -7.06 42.64 49.84 -46.10 -1.08
5.20 0.73 0.01 4.47 -5.19 19.98 26.94 -23.20 -1.16
SUMS •
603.79 353.31 1064.40 250.48 460.61 4351.56 25859.50 9188.98 24.29

MEANS
35.52 20.78 62.61 14.73 27.09 255.97 1521.15 540.53 1.43

6
 4. METHODS OF IMPROVING ESTIMATES
Confronted with a situation where a method provides Another method of extracting a best estimate of CIR is to
several estimates for some process variables, two dif- plot the values of fi —0 against those of ui — f1 . The slope
ferent avenues are open to produce unique best esti- of the best fit curve to those values can be taken as a
mates. One consists in taking the means as the most best estimate of CIR. This can be done graphically as
efficient estimates of the variables; the other in using a shown in Figure 6.
different method of estimation which would give a If non-cumulative values are used, the graph neces-
unique estimate of each variable and could be justified sarily goes through the origin since the mean values of
as giving the best estimate. fi — oi and u i are zero, and a best fit curve always goes
through the mean coordinates of its points.
4.1 REGRESSIONS The calculation of the best fit curve can also be done
As discussed above, a typical case with several esti- by regression, using a computer program such as
mates available is the calculation of the circulating load STAMP (1).
ratio of a hydrocyclone when the size distribution of its The regression consists in calculating the best esti-
feed and product streams is known. The two-product mates 9, of variables yi = f1 — 01 , and/or 5 1 = u i — f1 , as
formula gives: those which minimize a figure of merit or minimization
CIR1 — U — criterion which can be one of the following:
0 u— f1 1. Regression of y on x: 1(9 — 9) 2
The mean of all values obtained for the n size intervals is 2. Regression of x on y: — x) 2
often used as a best estimate:
3. Regression of x and y:
Ma+ bx—y)2 (p2 + 1)
CIR = ClR (p — b) 2

4. Regression of orthogonal: 1h 2 = Ma+ bx—y) 2

oI)
or CIR — b2 + 1
(u1 — fl )
hi = distance from point i to regression
It must be noted that the size distributions used in
curve
conjunction with these formulae can be cumulative or
non-cumulative. The non-cumulative is preferred p = split coefficient
because of a less complex error correlation among the = b slope of regression line
various sizes. Table 2 gives the results obtained with
both formulae. The regression coefficients for a linear regression
Y = a+ bX are obtained by the following formulae:

0-F 4.1.1 Orthogonal Regression

-8.00 0.00 8.00 16.00 24.00 32.00

o 1e2 g2 e
Slope = b —
o
where
e = ( 1 Y1)2 — (/X1) 2 — n (/Y —

g = 2(nX 1Y1 — 1,X11Yi)

o

o
/Y — b/X1
'

o intercept = a — 1

7'4 O
o 4.1.2 Other Regressions (Y on X, X on Y,
o on X and Y)
00
• — na if /X1 # 0
Slope = b =
o

b = (nee n/Yr + np/X1Y1 if /Xi = 0

Fig. 6 — Raw data 4" hydrocyclone (non-cumulative) 1Y1 np/XF—

7
where The method can be refined further by weighting each
intercept = a variable in the regression inversely to its assumed vari-
a = X 1 Y — YX1Y1 + p(IYIZXF — ance so that it contributes to the regression curve only in
proportion to its precision.
— nXiYi + p(nUF — IXIEX1)

split coefficient = p = slope of residuals 4.2 NODE IMBALANCE MINIMIZATION

p = 0 for regressions of X on Y
The mass conservation equations satisfied by the flow
p = 00 for regressions of Y on X rates and assays in a two-product circuit can be written:
pI = —2 for regressions on Y and X — W2 — W3
El — W2X2 — W3X3 =
Ei ---- relative error on variable i where Is and Ix are the node imbalances, or departure of
the mass conservation equation from balance, or
p = — 1 for orthogonal regression residuals. Several equations with lx residuals can be
written if several assay types are known.
The node imbalance minimization method consists in
saying that rather than choosing the minimum subset of
4.1.3 Regression Forced Through the the assays that allows the use of a n-product formula,
Origin When X,Y 0,0 one uses as best estimates of Wi the values that mini-
mize the sum of the squares of all imbalances:
a =0 b = Ç'/)—< F = mir >2 I(l r )2 = minimum
Intuitively one can indeed accept that since there are no
The results obtained for several types of linear regres- Wi 's that balance exactly all equations, those which
sion applied to non-cumulative and cumulative size dis- come closest to an overall balance qualify as best esti-
tributions of Table 2 are given in Figures 7 and 8. The mates. The calculation of these Wi's requires the deriva-
poor correlation coefficient obtained in particular with tives of F with respect to all unknown Wis. The solution
non-cumulative data confirms that some points are defi- values are such that all derivatives are zero at the same
nite outliers and might as well be remeasured or deleted. time.

20 o

ORTHOGONAL REG
Gr, 70 ORTHOGONAL :_-
18 y.-53.6 +5.5x
y" 5.8 x e-'
16
o
q; 60
14 o
12 0
I0 50
X on Y i‘-1-
8
:k*
6 Y on X 40
X on Y through origin
4
Yon X
17: 2 6 30
R,7 ,71
0 o 25
-2 o 20
CUMULATIVE DATA
-4
4" Hydrocyclone
-6 10

-8
- 10 NON - CUMULATIVE DATA
Hydrocyclone
-12
- 14 o
- 10
- 16
- 6 - 5 -4 -3 -2 -I 0 I 2 7 12 16 20
FI U1 Ft - Ul

Fig. 7— Regressions on hydrocyclone non-cumulative data. Fig. 8 —Regressions on hydrocylone cumulative data. Correlation
Correlation coefficient: 0.58 coefficient: 0.81

8
Therefore: These equations are of the type:
8[I(Is) 2 + Z(I)o) 2] 0 Wi – — W3 = 0
8w2
W1 5‘(1 *25'(2 ‘Â/35(3 = 0
8[1(p)2 /012]
SW3 The second type of equation can be written:
F = I(Is) 2 + I(lx) 2 = – + ‘fq + Wî (W1 – 131)(X1 – – (W2 – R5)(X2 –
2W1W2 2W1W3 2W2W3 VW1 IX?! – (W3 – Rg)(X3 – 1:1) = 0
+ We 1Xe wg - 2w1w2 )(1)(2 using the relation
2W1W3 X1X3 2W2W3 1,X2X3 X = X* + Ex
i.e., experimental value = true value + experimental
SF = 2 (W2 – W1 + W3 + WgXe — W1IX 1 X2 error which is approximated by X = X + Rx. (Experi-
W31X3X2) = 0 mental value = estimate + residual.)
The number of unknown Rs and Rx in a general mass
By symmetry: balance problem is usually much larger than the
SF number of mass convervation equations. This implies
= 2 (W3 – W1 + W2 + W3 1,Xe — W1 1X1X3 that an infinite number of solutions exist that satisfy the
3 8W mass balance equations. The objective of a mass bal-
+ W2 1X3X2) = 0
ance method is, therefore, to produce the maximum-
By standard elimination technique: likelihood solution.
W3 = (1 + EXiX3)(1 +1Xe) — (1 +IX 1 X2)(1 +1X2X3) As stated above, if the assay and flow rate populations
can be considered as normally distributed as well as
W1 (1 +X)(1 +IXe) — (1 +1X2X3) 2
randomly sampled (i.e., unbiased and uncorrelated), the
W2 maximum-likelihood estimates are such that:
— is
and obtained by symmetry.
W1 – the estimates variance is minimum;
– the mass balance equations are satis-
If one has only one assay type, the sign can be fied.
removed and the equation simplifies to:
The reader of Chapters 3.1 and 3.2 (2,3) will notice that
W3 = X1 — X2 the notion of data redundancy is sometimes very subtle
W1 X3 — X2 and hard to grasp at first. In particular the BILMAT
which is the two-product formula corresponding to the presentation is based on the concept that the only true
special case (1s) 2 + (1 ) 2 0: the equations have an variables to be estimated are the flow rates and that the
exact solution. number of equations is therefore usually much larger
than the number of unknowns. The solution to this prob-
lem also happens to produce better estimates of the
4.3 GENERALIZED LEAST-SQUARES assays. The MATBAL presentation, from the onset,
METHOD classifies assays and flowrates as problem variables for
which better estimates are sought. In that case the
All methods of material balance computation reviewed number of equations is smaller than the number of
above assume that the experimental values of the variables. Nevertheless, both programs solve essen-
assays are their best estimate. In the following, a tially the same problem. (See also Section 7.2.5.)
method is discussed in which the experimental values
are also adjusted by an amount which is such that the
sum of squares of all adjustments is minimum, whence 4.3.1 Lagrange Multipliers
the name "least squares" method. The problem configuration described above is called a
If best estimates and W are associated to each assay constrained minimization problem which has the gen-
X and flow rate W, the maximum-likelihood estimators eral form:
for a normal population of X and W are such that: Min {F(x)} minimum criterion
– 5Q 2 1(W — W) 2 minimum G 1 (x) = 0 constraints
This estimator is also called an estimator of minimum
variance. It can be shown that the values of x that are solutions to
this problem are also solutions to:
In a process where X and W have been measured, one
may want to calculate the best estimates SC and W. Stat {L(x 1 ,X1)} Stat = Stationary point
These can be obtained by solving al valid mass balance
equations in the flowsheet.

9
where 60 x 4 + 60 x 2 + 90 x 1
= 64.3
L(xiXi) = F(x) + Z Xp i(x) (4 + 2 + 1)
XI is called a Lagrange coefficient or multiplier; More weight has been given to a low score; therefore,
L is called a Lagrangian. the mean has been lowered.
In the minimization of the Lagrange function, it is appro-
The material balance problem can therefore be solved priate to weight each adjustment of a variable by the
by calculating the stationary point of the corresponding inverse of its variance. The larger the variance (i.e., the
Lagrangian. less precise the assay), the lower the weighting factor
+ /(W — W) 2 + GI (5'(,W)
The stationary-point coordinates will not be as affected
where G I (5C,W) are the mass conservation constraints by an imprecise assay as by a precise one. This adds to
and X1 the Lagrange unknown coefficients or multipliers. the efficiency of the estimator.
As discussed later on, various methods can be used The most efficient Lagrangian is, therefore:
to compute this stationary point. 1 (x _ jso2 oN _ W%2
2
+ Z X.G.,
(SC W)
0-2
X Crw
4.3.2 Weighting Factors The necessity for the user to provide a realistic estimate
One further refinement can be introduced in the method of the variances of all measurements is an intrinsic part
by weighting each variable adjustment in proportion to of the mass balance problem. This is referred to as
its precision. providing an error model for the experimental values.
The analogy with a weighted mean for several scores in This error model is characterized by assumptions: nor-
school can illustrate this point. For scores of 60, 60, and mal distribution; unbiased, uncorrelated variables; and
90 in three subjects equally weighted, the mean score is by data: variable variances; or standard deviations.
70. If the scores are weighted 4, 2, and 1 respectively,
the weighted mean is:

10
 5. ERROR STRUCTURE IN THE EXPERIMENTAL DATA
The problem of error structure in the experimental data Inversely, for a concentrate, X L --- 1, and ΠFE can remain
is the subject of a separate chapter in the "SPOC Man- small when smaller samples are collected:
ual", Chapter 2 (4). The main facts about error models £— for slurry sampling should be given a
for material balance computation are summarized in this value of 1;
section.
f— for most concentrated solids should be
given a value of 0.5;
5.1 COMPONENTS OF THE TOTAL
ERROR FOR AN ASSAY VALUE* g- for material containing several sizes of
particles, should be 0.25 for single size
The total error is a combination of: material, g = 1.
1. the sampling fundamental error;
2. the errors due to time and space variations in 5.1.2 Errors Due to Time and Space
the lot; Variations in the Lot
3. the errors due to sampler design and opera- These errors can follow various patterns, such as long-
tion; range and short-range variations. The periodicity of
4. the errors due to sample preparation and these variations, however, can introduce another source
analysis. of error.
Factors 1 and 4 cannot be eliminated; 2 and 3 can be The only way to reduce these errors is to collect com-
minimized. posite samples. Systematic sampling (at regular inter-
vals) is preferable to random sampling unless a cyclic
variation exists in the lot.
5.1.1 Fundamental Error (FE)
P.Gy (5) shows the variance of FE to be: 5.1.3 Errors Due to Sampler Design and
02FE= Cef g d3 [
1 1 Operation
]
M S— ML These errors occur when all particles do not have the
where same probability to enter the sampler.
M5 = sample mass (g) The basic rules to minimize them are:
ML = lot mass (g)
— The cutter edges must be parallel.
C= composition factor (g/cm3) — The cutter must move perpendicular
to the axis of the stream, at a constant
e = liberation factor (0 or 1) speed (-. 0.6 mis).
f= particle shape factor (0.5) — The cutter width must be at least
g = size distribution factor 10 mm and greater than three-times
d = maximum particle size (cm). the largest particle.

C, the most variable coefficient is given by: 5.1.4 Errors Due to Sample Preparation
,____ 1 — XL
and Analysis
C [(1 — XL) pc + XLpgj
XL The sources of such errors are numerous; e.g., con-
tamination, loss, alteration, sub-sampling, analysis
where method. They normally represent a minor fraction of the
XL = critical concentration total error.
PC = specific gravity of critical component In the solution of material balance problems, the stan-
pg = specific gravity of non-critical compo- dard deviation of the experimental data is an essential
nent (g/cm 3). part of the input data. It is used to distribute the data
adjustment in proportion to the data accuracy and also
One easily sees that for a trace element, Xi.<< 1. to calculate the resulting error on the variable estimates.
Therefore, C is large and, consequently, cre E is large In a first approximation, all errors are assumed to be
unless M s itself is very large (large sample). normal, unbiased, and uncorrelated.

11
 6. SENSITIVITY ANALYSIS
6.1 INTRODUCTION 6.2 THE ANALYTICAL METHOD OF
No material balance problem is complete without a SENSITIVITY ANALYSIS
value being calculated for the precision of the flow rate
estimates and, ideally, for the adjusted assays. In other 6.2.1 Principle
words, the confidence interval of the material balance
The principle of the method is illustrated by the calcula-
results must be estimated. The benefits of a sensitivity
tion of the variance of a linear combination of two ran-
analysis are, therefore:
dom variables, X, and X2 .
— assessment of the error on the esti-
If Y = aX„ + bX2, then (4 = a2cyli + b2(42
mates;
—2abux„ x2.
— reflexion on the experimental results
and the error model used to perform If X1 and X2 are independent, the covariance term,
the computation. CrX/X2, is zero.
A few basic principles of statistics must be reviewed The variance of the dependent variable, Y, is therefore a
before the sensitivity analysis can be understood: function of the functions of the independent variables.
— For a normally distributed variable, In a material balance problem, the mass conservation
the mean value and the standard equations are non-linear and the variance relationships
deviation of a series of measure- are more complex. They must be linearized around a
ments (sample) are the maximum- consistent set of values of the variables in order to derive
likelihood estimates of the distribution the suitable relationships.
parameters. They are estimates of
Ideally, the linearization should be done around the true
minimum variance.
values of the variables. In practice, it is done around a
— In a material balance problem, the set of estimates generated by the material balance pro-
experimental variables, the esti- gram.
mates, the residuals, and the errors
are assumed to be normally distrib- If W and k represent these estimates, one can define
uted. The estimates being of mini- the following variations:
mum variance, the variance of the cal-
culated variables is smaller than, or = W —W
equal to, that of the experimental vari-
ables. Similarly, the variance of the SX = X —
residuals is smaller than that of the
errors. =
The sensitivity analysis consists in studying the rela-
tionship between the error model of the experimental
85'c 5'( —
data and that of the calculated estimates. It can be done When Wand X are
considered as true values, the first
by two different approaches; the analytical method, and two variations approximate the errors on flow rates and
the Monte Carlo method illustrated in Figure 9. class fractions. Therefore, the variances of SW and SX
are the variances of the measured variables W and X.
Similarly, the variances of SW and SX are the variances
of the variable estimates.
I ! New The BILMAT User's Guide shows how the values of
.
-- Experimental 0-ei and d can be calculated by variance algebra as a
X 's function of ot and cr?<. The resulting variances are con-
Simulated X servative; i.e., larger than their true values, but very
often accurate.
•
6.2.2 Factors Influencing the Variance
X crx , new W Estimates
The following factors contribute to the validity of the
.
variance estimates obtained by the analytical method:
— The measurement errors follow a nor-
mal distribution.
Fig. 9 — Principle of the two-error estimators — They are uncorrelated.

12
 - The assumed error variances are rea- Case 1 is more likely if only one Xi is out of range.
sonable. Case 2 is more likely if several Xi's are out of range in the
- The estimate used as a true value is same stream j for different species i.
accurate. Case 3 is more likely if several )(I's are out of the range
The following cause-to-effect relationships must be for the same species i in several streams j.
underlined: Repeated positive (or negative) values in the residuals
- The better the data, the smaller the obtained for several material balances of the same cir-
estimate variances. cuit can also help to detect bias in the experimental
- The more redundant the data, the data. Usually, residuals with values outside a ± 3 o-x
smaller the estimate variances. range are considered as outliers and should lead to
either data rejection or error model modification.
- The more efficient the process with Repeated abnormally-small residuals should also sug-
respect to the problem variables (e.g., gest a reassessment of the assumed error variance.
good concentration ratio), the smaller The overall contribution of some data to the material
the estimate variances. balance solution can also be estimated and suggest
- The fewer recycle streams that exist, new sampling strategies.
the smaller the estimate variances
are. 6.3 THE MONTE-CARLO ANALYSIS
A digital computer can generate random numbers
6.2.3 Practical Use of the Estimate belonging to a given probability distribution. For
Variances instance, given a mean and a standard deviation, a
program can generate as many samples of a normal
6.2.3.1 Confidence interval population as required.
If 0-2 = 10 -4 and 5( = 5% = 0.05, In a material balance problem, each measurement - X
then crR = 10 -2 = 0.01. and W - is characterized by:
This means that there is a 68% chance that X* is - its standard deviation (from the error
between 5( - 0-5z and 5( + 0-5 : i.e., between 0.04 and model);
0.06.
- its best estimate 5( and W (from the
The chances of X* being within 5(.± 20-i< (i.e., 0.03 and result).
0.07) are 95%.
It is, therefore, possible to generate pseudo-experimen-
tal values belonging to the same population. In other
6.2.3.2 Residuals analysis words, we can simulate assay values that are equivalent
The residuals (e.g., W - W) are random variables with to those that could be obtained in repeated sampling.
their own distribution. Their variances are given by: This can be done only if the assumed error model is
valid.
w w and x
After one mass balance calculation has been com-
Using the example abcive where ŒR = 0.01 and pleted, all the experimental values can be renewed by
iX = 0.02, the 95% confidence interval of the residual
( simulation and solved for a new series of calculated
X - X is: variables. The accumulation of several such series
allows the computation of their means and standard
±2 — cre = -± 2 1410-4 - 10-4 deviations. These quantify the sensitivity of the calcu-
= ± 3.56 10-2 lated flow rates to the natural variations in the observed
variables.
If the calculated residual 5( - X is outside this interval The larger the number of repeats, the more reliable the
(- 3.56 10-2 + 3.56 10 -2), the possibilities are: errors estimate.
1. The experimental data X have been The potential of this method to check the sensitivity of
accidentally altered. the results to the error model must be emphasized. If the
2. The sampling procedure has intro- error model used for the Monte-Carlo simulation is dif-
duced a significant bias, or the sam- ferent from that used in the mass balance solution, the
pling variance is underestimated. magnitude of the variations in the estimates and their
standard deviation are an indicator of the importance of
3. The assay procedure (sample prepa- using the proper error model in a particular problem.
ration and analysis) has introduced a
significant bias.

13
 7. MATERIAL BALANCE PROGRAMS
An exhaustive survey of all published material balance 7.2 BILMAT: HIERARCHICAL SOLUTION
programs is beyond the scope of this text and the reader TO THE GLS PROBLEM
is referred to such a survey by Reid, Smith, Voiler and
Cross (6). 7.2.1 Flowsheet Variables
Material balance programs can be classified into four BILMAT has been designed to simultaneously balance
broad families depending on the method of solution of data describing various subdivisions of a pulp stream.
the material balance equations: The first level of division consists of two phases; usually
1. programs using an iterative Lagrange one solid and one liquid. Each phase can itself be
solution, e.g., MATBAL; described by classes (e.g., particles in a size interval).
2. programs using a hierarchical solu- Each class can be divided into sub-classes (e.g., parti-
tion, e.g., BILMAT; cles in a specific gravity range within a size interval).
Each class or subclass can be described by charac-
3. programs using a two-step node teristics (e.g., chemical assay, mineralogical assay,
imbalance + GLS solution; etc.).
4. miscellaneous. This is summarized in Figure 10. The upper two levels of
The following discussion covers types 1 and 2 more the stream description (e.g., pulp, solid and/or liquid) are
extensively. called macroscopic. The lower two levels are called
microscopic. Specific variables can describe each level
7.1 MATBAL: ITERATIVE and they are referred to as macroscopic and micro-
scopic variables.
LAGRANGE SOLUTION TO THE
Level one is described by flow rate measurements.
GLS PROBLEM Level two is described by flow rate and phase concen-
In the MATBAL approach, the stationary point of the tration in the pulp (e.g., liquid fraction in the pulp, or solid
Lagrangian is obtained by a classical calculus method: fraction in the pulp). Level three is described by mass
the cancellation of all derivatives with respect to the fractions of each class in one of the level-two phases
variables. (e.g., mass fractions of a particle size distribution). Level
four can either be described by mass fractions of sub-
81_ = 0
— classes within a level-three class, or assays of level-
three classes.
8L = 0 The scale of four levels can be translated by considering
Sw one phase as level one, classes of that phase as level
8L two, sub-classes as level three, and characteristics as
= 0
level four. In comparison, MATBAL handles only data
Ôxi

The above system of equations has necessarily as

many unknowns as equations.
All equations can be derived by hand in order to bring
the system to the form:
MX=C
where X is a vector of variables )k, W, xi , and M and C are
matrices of coefficients. !SOLIDS • SOLIDS
COMPO- SUB-COMPONENTS
These equations have been coded in FORTRAN and I
LVEA/
can be solved automatically by the program. The user MACROSCOPIC MICROSCOPIC
has to input a flowsheet description that is suitable, VARIABLES VARIABLES
using MATBAL nomenclature. LEVEL I LEVEL 2 LEVEL 3 LEVEL 4

As one can imagine, the systems to be solved are large •FO-F FLOW RATE .S0,.IDS AND WATER•CHEMICAL •CHEMICAL -CONCEN-
FLOW RATES CONCEN- TRATIONS IN
and the applicability is strictly restricted to the type of : TRATIONS PARTICLE SIZE
- .PARTICLE . INTERVALS
mass constraints that have been considered in the SIZE DISTRI.,
derivatives calculation. In this family, several other pro- BUTIONS

grams have been published especially by Cutting (7). RESTRICTEO DOmAIN

OF VARIABLES FOR
For full narrative, mathematical, and programming THE SENSITIVITY
ANALYSIS
details on the MATBAL program, the reader is referred to
Chapter 3.2 of the "SPOC Manual" (3) which is a user's Fig. 10 — Various levels of the mass balance variables
guide.

14
relative to two adjacent levels; e.g., (pulp + solid and 7.2.5 Data Redundancy
water), or (solid + assay), or (size interval + assay) etc. This concept is a consequence of the variable indepen-
The BILMAT User's Guide gives examples of flowsheet dence described above. If a set of independent vari-
variables. ables has been measured, all variables can be
calculated; the problem is said to be exactly determined,
or to correspond to a minimal experimental design. If
7.2.2 Flowsheet Analysis fewer variables have been measured, no complete solu-
Flowsheet analysis in terms of networks is a prerequisite tion can be calculated and the problem is called under-
to using the BILMAT program. A network is a graphical determined. If extra variables have been measured,
representation of the mass conservation across process they constitute redundant data.
units. A problem can be redundant with respect to some vari-
Each level of subdivision of the process streams must ables, and underdefined - or exactly defined - with
be considered for network definition. A total of eight respect to others. The checking of redundancy is an
different networks can be described: one at level one, important and sometimes delicate prerequisite to using
one at level two, four at level three, and two at level four. BILMAT. This is illustrated by the following example.
Details on network definition are given in the BILMAT Consider a two-stage flotation circuit (as illustrated in
User's Guide. For computation purposes, networks are Figure 11). The following four data sets can be analyzed
expressed as matrices according to the definitions of the for redundancy:
User's Guide. A: W1, W2, X2, X3, X4, X?, X3
B: Wÿ, W2, X2, X3, X4, X5, X?, X3, X4
7.2.3 Mass Conservation Equations C: Wi, W2, X2, X3, X4, X5, X;, X2, X3
The mass conservation equations corresponding to the D: W,, W2, Ws, X2, X3, X4, X5, X6, X2,
existing networks are described in Sections 2.4.2 to
X2, X3, X4, X5
2.4.5 of the BILMAT User's Guide and in Chapter 3.1 of
the "SPOC Manual" (2). The data describe one phase (solid), and two classes X1
and X2. The number of independent variables is:
7.2.4 Independent Variables (6 + 12) - (3 + 6) = 9 (assuming that Xj + Xj :1/- 1).
Independent flowsheet variables are those that allow Case A is underdefined.
the others to be calculated by the material balance
equations. For instance, in a solid network consisting of Case B is exactly defined.
q nodes and p branches, there are q equations of con- Case C is overdefined for all variables except Xj
servation, p flow rates and, consequently, p- q indepen- (j = 4,5,6) which are not calculable.
dent flow rates. There may be several sets of indepen- Case D is overdefined.
dent flow rates, but no set of p- q values is a set of
independent flow rates. The program can produce meaningless results when
This concept of independence can be applied to vari- underdefined data are used. If only one flow rate is
ables of several levels in the global material balance measured (or fixed), no adjustment of its value can be
context. done, the measured value is the best estimate and the
For instance, if we consider solid flow rates Wj number of independent variables should be reduced by
(j = 1, ..., p) and mass fractions X] in classes one (p-q-1) or n(p-q)-q-1.
i= 1, ..., n, the valid material balance equations are:
q - solid conservation equations
nq - class conservation equations
The number of independent variables is:
(p+np) - (q+nq) - p = n(p-q) - q
If the set of classes is complete (i.e., the total mass
fraction is one), the number of independent class mass
fractions must also be reduced by one per stream.
Therefore, the actual number of independent variables
in the above problem is:
(p+np-1) - (q+nq) = (n+1)(p-q) - 1 Fig. 11 - Sample flowsheet for data redundancy test

15
It can also be written in matrix notation: using the determinant method:
MX= D 1 1 11
with (al b1 7.1 1 21
5.7 40 2 2
= a2 b 2 C2
a3 b3 C3 2.26 1 10 0.1 2216.07
W2 - 0.1
(Xi 1 1 11 22160.7
X X2) 60 1 21
X3 140 22
1 1 10 0.1
c
D d)
2
1 1 11
d3 60 7.1 21
15.7 22
By definition, the determinant of matrix M is: 1 2.26 10 0.1 2216.02
W3 - 0.1
22160.7 22160.7
al ID1 cl
det M = a2 b2 C2 = a1 (b2c3 - b3c2)
and similarly: W4 = 0.2, W5 = 0.60
a3 b3 C3 - b1 (a2c3 - a3c2)
+ c1 (a2b3 - a3b2)

Any determinant can be calculated along any row or 3.4 LIMITATION OF THE n-PRODUCT
column by carefully alternating the +1 and -1 factors in FORMULA
front of the various terms. The solution of the system of The limitation of the n-product formula comes from the
equations M X = D is given by: fact that the calculated variables are influenced by the
_
- M* precision of the obse rved variables in a non-quantified
way. Consequently, when more than n-1 assay types are
known in each stream each subset of an n-1 assay will
where Mi*designates the determinant of matrix M after give a different answer which can discredit the results
column i has been replaced by the values in D. completely to the plant engineer's eyes. Let's, for
This can be illustrated by an example from Taggard's instance, consider a two-product circuit where the fol-
handbook of mineral dressing, (pages 12 - 123), for a lowing data are known:
four-product circuit. Cu Zn % Fe %
The following assays are known: 1. Feed 0.163 3.93 11.57
Pb % Zn % Cu % 2. Conc 0.657 52.07 14.67
Feed 7.1 5.7 2.26 3. Tail 0.140 0.49 13.09
Pb conc 60 1 1 Although three assay types have been measured in
Zn conc 1 40 1 each stream, only one is required to calculate the con-
Cu conc 2 2 10 centrate and tailing flow rates; W2 and W3 by the two-
Tail 1 2 0.1 product formula.
For a feed of 100 t/h, the results corresponding to each
The valid equations, according to Figure 4, are: mode of calculation are:
Wi = W2 + W3 + W4 + W5 with Cu data: W2 = 4.5 W3 = 95.5
7.1W1 = 60 W2 + W3 + 2 W4 + W5 with Zn data: W2 = 6.7 W3 = 93.3
5.7W1 = W2 + 40 W3 + 2 W4 + 2 W5 with Fe data: W2 = -96.2 W3 = 196.2
2.26 Wi = W2 + W3 + 10W4 + 0.1W5 These three sets of results are quite di fferent due to the
large variations in assay values and precisions.
A conventional error calculation performed on the two-
product formula gives the following result:
AW2 < AW1 AX1 + AX3 AX2 + AX3
3 W2 X1 -X3 X2 - X3
W1
4 The relative error on a calculated flow rate depends,
• therefore, not only on the magnitude of the errors on all
assays but also on the di fferences, X1 - X3 and X2 - X3,
which are proportional to the quality of the separation
Fig. 4- Four products being performed.

4
 8. EXERCISES ON MATERIAL BALANCE COMPUTATION
8.1 EXERCISE 1: n-PRODUCT FORMULA
1. Derive the two-product formula. b) Calculate several of them using the
2. Consider the four-product circuit illustrated here- NPRD* program.
after, in which each stream has been assayed for six c) If you were responsible for calculating
species. inventories on this circuit, what would
a) How many solutions can be obtained by you do?
the n-product formula?

FEED 26.03/48.84/0.34/0.39/0.105/4,248
1

31 \4
72.5/5.6/0.007/0.02/0.011/7.73 15.9/64.2/1.63/1.82/0,44/4.19

0,71/70.2/0,013/0,024/0,014/0,35 4.44/66.3/0.049/0.073/0,05/4.93

SIX ASSAYS IN A FOUR PRODUCT CIRCUIT

*NPRD is one of the miscellaneous utility programs documented in Chapter 8 of the SPOC Manual (11).

8.2 EXERCISE 2: REGRESSION

Consider the attached data collected around a hydro- 3. Use the STAMP* program to determine CIR by
cyclone. F is the feed, U the underflow; 0 the overflow. linear regression.
These data are the size distribution of the solids. XF Xu XC, XF - Xu X0 - XF
1. Using the two-product formula, calculate the ratio 3.97 5.95 0.00 - 1.98 -3.97
U/O for several size intervals. What conclusions do 15.25 21.45 0.00 -6.20 -15.25
you draw? 14.80 20.70 0.00 - 5.90 -14.80
2. The circulating load is given by: 9.71 13.05 0.00 -3.34 -9.71
8.27 11.05 0.80 -2.78 -7.47
X0 - XF = CIR 4.70 6.03 0.00 -1.33 -4.70
XF - Xu 10.06 9.44 0.00 0.62 -10.06
4.01 4.53 1.82 - 0.52 -2.19
Calculate the mean circulating load. 2.25 1.56 6.61 0.69 4.36
Plot X0 - XF VS XF - X. What is the best value of CIR 1.92 1.23 16.83 0.69 14.91
2.26 1.04 22.32 1.22 20.06
that you can get from your plot?
2.56 0.50 20.70 2.06 18.14
2.67 0.99 16.04 1.68 13.37
3.07 0.65 9.85 2.42 6.78
3.38 0.31 3.21 3.07 -0.17
3.63 0.56 1.39 3.07 -2.24
2.29 0.23 0.42 2.06 -1.87
5.20 0.73 0.01 4.47 -5.19
100.00 100.00 100.00 0.00 0.00

*See SPOC Manual, Chapter 8 (11).

17
8.3 EXERCISE 3
Consider the attached flowsheet. All eight streams have
been assayed for Cu, Zn, Fe with results as given on the
flowsheet.
1. For a feed of 100 t/h of solids, calculate flow rates
W4 and W6 by the two-product formula.

2. Calculate the precision, AW6, when all assays

W6
are known with a 6% relative precision and
0.1
W1
3. Calculate W4 and W6 , using the node imbalance ASSAYS
formula. 1 2 3
4. Use the SAMBA program to calculate the general 1 STREAM 1 0.163 3.93 11.57
node imbalance minimization.* 2 STREAM 2 0.447 11.63 12.79
3 STREAM 3 0.466 8.11 14.35
4 STREAM 4 0.140 0.49 13.09
5 STREAM 5 0.960 34.64 14.37
6 STREAM 6 0.657 52.07 14.67
7 STREAM 7 1.020 42.73 14.66
8 STREAM 8 0.440 11.86 13.75

8.4 EXERCISE 4 1. For a feed of 100 t/h, calculate the three concen-
trates and the final tailings flow rates.
Consider the attached flowsheet. Ten numbered sam-
pling points have bee assayed for four species, except 2. Prepare the MATBAL input file for this problem. The
2, 3, 7 and 8 which have been assayed for three species. Table of assay precisions is given below.*
3. What can you say about sampling point No. 10?
Assays are as follows: Table of assay precisions
1. Mill Heads .87 1.27 3.69 1.61 VG G VG F
2. Blk Scav Tls .27 .52 3.54 F F G F
3. Cu-Pb Conc 10.45 13.06 4.05 F F F F
4. Pb Conc 1.50 16.70 3.62 11.22 VG F F F
5. Cu Conc 25.29 4.22 4.63 11.51 F F VG F
6. Zn Conc .87 1.37 48.4 2.39 F F G F
7. Zn Scav Tls .22 .49 .80 F F F F
8. Zn Cl Scav Tls .37 .99 1.63 G G G F
9. Final Tails .21 .51 0.89 0.67 F F F F
10. Str. Cntrl .90 .95 30.0

*See "SPOC Manual", Chapter 2.1 (1).

MILL
PEED

9
MILL
TAILS

Zn CLNR
SCAV

Cu RGHR Zn
2nd,3rd,4th
CLNR
LEAD
CCNC.

ZINC
CONC.
Cu CLNR

COPPER
18 DONC.
For an ideal separation, one has:
o
X3 = 0 and '^'W2 :5; '^^Wi + ^Xi + ^X2
W2 Wi Xi X2
If all assays are known with a 6.0% accuracy in the
example above, the maximum relative error on W2 is
found to be 19.2% with Cu assays, 13.1% with Zn, 156%
with Fe.
A more realistic estimate of the sensitivity of the solution
to assay errors can be obtained by a technique called
Monte-Carlo Simulation which requires a computer. In
this method, a perturbation similar to the suspected
random error present in all assays is imposed on the
measured value in order to simulate a set of synthetic Fig. 5 - Hydrocyclone. CIR = ô
assay values. These are used for an n-product solution
and the process is repeated enough times to accumu-
late a number of possible solutions. The mean and In most cases, the variable of interest is the circulating
standard deviation of these possible solutions can be load ratio U/O, i.e., the ratio of solid flow rates in the
calculated. underfiow and the overflow respectively. Using the two-
This method, applied to the problem above, gave the product formula, with f,, u,, and o,, the non-cumulative
following results: per cent retained on size interval i, one has:

,^W2 = 72% with Cu assays, 8% with Zn, CIR, = U = f, - o,

W2 1000% with Fe O u, - f,
The problem of selecting the best subset of assays in each interval i gives a different value of CIR, as shown in
order to calculate the unknown variables becomes Table 2.
especially acute when the assays consist of size dis- Having so many possible estimates of CIR, one can
tributions with up to a dozen intervals or more. Process start thinking of a method to determine a best estimate.
evaluation around a hydrocyclone is typical of this situa- Several methods of obtaining such an estimate are sub-
tion as shown in Figure 5. sequently described.

Table 2A - Hydrocyclone experimental data (non-cumulative)

F U O F-U O-F (F-U)(F-U) (O-F)(O-F) (F - U)(O - F) CIR

3.97 5.95 0.00 -1.98 -3.97 3.92 15.76 7.86 2.01
15.25 21.45 0.00 -6.20 -15.25 38.44 232.56 94.55 2.46
14.80 20.70 0.00 -5.90 -14.80 34.81 219.04 87.32 2.51
9.71 13.05 0.00 -3.34 -9.71 11.16 94.28 32.43 2.91
8.27 11.05 0.80 -2.78 -7.47 7.73 55.80 20.77 2.69
4.70 6.03 0.00 -1.33 -4.70 1.77 22.09 6.25 3.53
10.06 9.44 0.00 0.62 -10.06 0.38 101.20 -6.24 -16.23
4.01 4.53 1.82 -0.52 -2.19 0.27 4.80 1.14 4.21
2.25 1.56 6.61 0.69 4.36 0.48 19.01 3.01 6.32
1.92 1.23 16.83 0.69 14.91 0.48 222.31 10.29 21.61
2.26 1.04 22.32 1.22 20.06 1.49 402.40 24.47 16.44
2.56 0.50 20.70 2.06 18.14 4.24 329.06 37.37 8.81
2.67 0.99 16.04 1.68 13.37 2.82 178.76 22.46 7.96
3.07 0.65 9.85 2.42 6.78 5.86 45.97 16.41 2.80
3.38 0.31 3.21 3.07 -0.17 9.42 0.03 -0.52 -0.06
3.63 0.56 1.39 3.07 -2.24 9.42 5.02 -6.88 -0.73
2.29 0.23 0.42 2.06 -1.87 4.24 3.50 -3.85 -0.91
5.20 0.73 0.01 4.47 -5.19 19.98 26.94 -23.20 -1.16
SUMS
100.00 100.00 100.00 0.00 0.00 156.92 1978.52 323.64 65.17

MEANS
5.56 5.56 5.56 0.00 0.00 8.72 109.92 17.98 3.62

5
 9. ANSWERS TO EXERCISES
9.1 EXERCISE 1 2. The MATBAL4 input file for this problem is:
1. See Section 3.1. DATA FOR EXERCISE 4 CHAPTER 3, SPOC MANUAL
9 4 1 3 0 9 1 1 10
2a. 20 solutions (there are 20 combinations of six 0 0 0 0 1 0 0 8 0 0 0 99 0.1 0,1
assays, 3 by 3). 1 MILL HEADS
2 BLK SCAV TLS
2b. See CIM Bull. Vol. 75, No = 840, p 166-170. 3 CU-PB CONC
2c. Use MATBAL or BILMAT. 4 PB CONC
5 CU CONC
G ZN CONC
9.2 EXERCISE 2 7 ZN SCAV TLS
8 ZN CL-SCAV TLS
1. See Table 2A. 9 FINAL TAILS
1 .87 1.27 3.69 1.81
2. 3.62 (see Table 2A and Fig. 6). 2 .27 .52 3.54 0
3. See the STAMP program user's guide, Chapter 7.1, 3 10.45 13,06 4,05 0
4 1.5 18.7 3.32 11.22
SPOC Manual. 5 25.29 4.22 4.83 11.51
G .87 1.37 48.4 2.39
9.3 EXERCISE 3 7 .22 .49 .80 0
8 .37 .99 1.33 0
1. Using assay type 1, W4 = 4.5 we = 95.5 9 .21 .51 .89 .87
1 ASSAY 1
Using assay type 2, W4 = 6.7 W6 = 93.3 2 ASSAY 2
Using assay type 3, W4 = -96.2 W6 = 196.2 3 ASSAY 3
4 ASSAY 4
VG G VG F
2. Using assay type 1, AW6 = 0.72 FFGF
FFFF
W6 FFFF
Using assay type 2, AW6= 0.08 VG F F F
F F VG F
W6 FFGF
FFFF
Using assay type 3, AW6= 10.0 GGGF
W6 1 5 1 -4 -5 -8 -9
1 3 1 -2 -3
3. Using the formula given in Section 4.2: 2 3 3 -4 -5
3 3 2 -8 -9
a ,„ (1+ 374.47)(1+171.6)-(1+ 153.4)(1+ 217.6) x 100 1 100 .01
" 6= 1 3 1 2 3
(1 +2926.9)(1 +171.6)-(1 +217.6) 2
2 3 1 2 3
6.81; W4 = 93.19 3 3 1 2 3
4 1 2 3 4
4. Solution given by SAMBA:
W1 to We = (100.0, 236.6, 207.5, 93.6, 29.0, 6.4, 3. Sampling point No. 10 is not useful in calculating a
22.66, 113.9) material balance. It was probably sampled for con-
trol purposes.
9.4 EXERCISE 4
1. By the four-product formula (the NPRD program) 9.5 EXERCISE 5
Using assays 1, 2, 3: W1 = 100, W4 = 3.88, See "SPOC Manuar, Chapter 3.2 (3).
We = 2.29, W6 = 5.49, W9 = 88.34
Using assays 2, 3, 4: W1 = 100, W4 = 3.37 9.6 EXERCISE 6
W5 = 4.55 We = 5.34 W9 = 86.74 See "SPOC Manual", Chapter 3.1 (2).

20
9.7 EXERCISE 7 2. Cleaner-5 concentrate: 95% confidence interval.
1. Using the NPRD program and data for Exercise 3, Measured Estimated
the following results were obtained for 20 repeats of
Relative ore flow
the Monte-Carlo option, using assay types 1, 2, 3,
rate .025 ± .004
successively.
Solid flow rate .295 ± .176 .337 ± .066
Nb assay 36.27 ± 14.51 33.48 ± 6.17
Assay type 1 5% pre- 10% pre- 2% pre-
cision cision cision The Nb assay reliability is improved by a large factor
95.55 ± 1.8 ± 4.23 ± 0.78 greater than two. Its standard deviation is still around
W4 10% but could be lowered if Nb were assayed with a
We 5.29 ± 1.8 ± 4.23 ± 0.78
better accuracy (see Section 5.2.5 of Chapter 3.1 [2]).
Assay type 2
93.33 ± 0.43 ± 1.05 ± 0.19 3. Using the results of the mass balance calculation,
W4 redo the sensitivity analysis without using the Nb
W6 6.67 ± 0.43 ± 1.05 ± 0.19
assay. Results and discussion are in Section 5.2.5,
Assay type 3 Chapter 3.1.
W4 196.2 ± 103.4 ± 555.8 ± 21.5
-96.2 ± 103.4 ± 555.8 ± 21.5 4. Use the sensitivity analysis editor to change the
W6 error model of the cleaner-4 concentrate assay and
to mask the solid flow rate on that stream. Then run
The estimate of the precision of the calculated flow the sensitivity analysis program.
rates is more reliable when the experimental data
The results are:
are more reliable.
Relative standard deviation (°/0)
2. By masking successively the various assays, it is Nb 35.86
found that assay type 2 is the only data that seem Mg 72.11
acceptable in this data set. The solution obtained Si 41.33
when this assay alone is used is less sensitive to 21.25
experimental errors: Ca 13.81
Fe 14.49
W1 = 100.0 8.54
W2 = 339.8 ± 84
An error model of 100% was used on that stream.
W3 = 297.3 ± 78
After computation, the same values have an acceptable
W4 = 93.26 ± 0.7
accuracy. But this has destroyed the reliability of the
W5 = 42.5 ± 15 assay of the cleaner-5 output streams. Therefore,
We = 6.74 ± 0.7 cleaner-4 concentrate should be assayed with accuracy.
By performing the same experiment with cleaner-1 con-
W7 = 35.78 ± 15.6 centrate, one would see that this stream is well known
We = 204.0 ± 78. after computation and that it has no significant effect on
other streams.
See also the treatment of this problem by the
5. Use the sensitivity analysis editor to mask Mg assay
SAMBA program in Chapter 2 of the "SPOC Man-
and then P assay. Observation of the relative ore
ual", example 5.3 (4).
flow rate accuracies leads to the conclusion that
deleting P assays does not significantly affect the
reliability of the results (see Section 5.2.5 of Chap-
ter 3.1 [2]). Therefore, P could not be assayed in a
9.8 EXERCISE 8 future sampling campaign.
1. The value of the relative standard deviations of a 6. Use the mass balance editor to modify the Fe assay
relative ore flow rate depends mostly on the value of in the rougher tailing. Run BILMAT and then the
the flow rate. A low flow rate generally has a high sensitivity analysis program. The mass balance
relative standard deviation. results are affected by that modification and, at first
Absolute standard deviations are aff ected by the glance, nothing wrong can be detected. However,
error model defined by the user and by the general the sensitivity analysis results show that the outlier
structure of the plant network. flag is equal to 3 for that assay in that stream.

21
 10. REFERENCES
1. Seguin, F; Leung, J.; Flament, E; and Laguitton, D.; • 6. Reid, K.J.; Smith, K.A.; Voiler, V.R.; and Cross, M. "A
"SPOC Manual, Chapter 7.1: STAMP computer pro- survey of material balance computer packages in
gram for statistical analysis and modelling"; Report the mineral processing industry"; 17th APCOM con-
SP85-1/7.1; CANM ET, Energy, Mines and ference; Society of Mining Engineers; AIME; pp.
Resources Canada; 1985. 41-62; New York; 1982.

2. Hodouin, D.; Flament, E "SPOC Manual, Chapter 7. Cutting, G.W. "Estimation of interlocking mass-bal-
3.1: Material balance computation for process eval- ances on complex mineral beneficiation plants"; Int
uation and modelling — BILMAT computer pro- J Miner Process 3:207; 1976.
gram"; Report SP85-1/3.1; CANMET, Energy,
Mines and Resources Canada; 1985. 8. Smith, H.W.; Burrough, D.; and Laguitton, D.;
"SPOC Manual, Chapter 2.1: SAMBA computer
3. Laguitton, D. "SPOC Manual, Chapter 3.2: MAT- program for sensitivity analysis of sampling data
BAL computer program for material balance data used for material balance computation"; Report
adjustment"; Report SP85-1/3.2; CANMET, Energy, SP85-1/2.1; CANMET, Energy, Mines and
Mines and Resources Canada; 1985. Resources Canada; 1985.

4. Smith, H.W. "SPOC Manual, Chapter 2: Sampling 9. White, J.W. et al. "A useful technique for metal-
methodology for ore and coal process evaluation lurgical mass balances applications in grinding"; lnt
and modelling"; Report SP85-1/2; CANMET, J Miner Process 4:39; 1977.
Energy, Mines and Resources Canada; 1985.
10. Mular, A.L. "Data adjustment procedures for mass
5 Gy, P. The Sampling of Particulate Materials: The- balances"; Computer Methods for the 80's in the
ory and Practice; Amsterdam; Elsevier Scientific Minera! Industry, edited by A. Weiss; pp. 843-849;
Publishing; 1979. Soc Min Eng AIME, Pet Eng; New York; 1979.

11. Laguitton, D., and Leung, J.; "SPOC Manual, Chap-

ter 8: Utility programs", Report SP85-1/85; CAN-
MET, Energy, Mines and Resources Canada; 1985.

22
 Table 2B - Hydrocyclone experimental data (cumulative passing)

F U, 0 F-U 0- F (F-U)(F- U) (0- F)(0 - F) (F- U)(0 - F) CIR

96.03 94.05 100.00 1.98 3.97 3.92 15.76 7.86 2.01
80.78 72.60 100.00 8.18 19.22 66.91 369.41 157.22 2.35
65.98 51.90 100.00 14.08 34.02 198.25 1157.36 479.00 2.42
56.27 38.85 100.00 17.42 43.73 303.46 1912.31 761.78 2.51
48.00 27.80 99.20 20.20 51.20 408.04 2621.44 1034.24 2.53
43.30 21.77 99.20 21.53 55.90 463.54 3124.81 1203.53 2.60
33.24 12.33 99.20 20.91 65.96 437.23 4350.72 1379.22 3.15
29.23 7.80 97.38 21.43 68.15 459.24 4644.42 1460.45 3.18
26.98 6.24 90.77 20.74 63.79 430.15 4069.16 1323.00 3.08
25.06 5.01 73.94 20.05 48.88 402.00 2389.25 980.04 2.44
22.80 3.97 51.62 18.83 28.82 354.57 830.59 542.68 1.53
20.24 3.47 30.92 16.77 10.68 281.23 114.06 179.10 0.64
17.57 2.48 14.88 15.09 -2.69 227.71 7.24 -40.59 -0.18
14.50 1.83 5.03 12.67 -9.47 160.53 89.68 -119.98 -0.75
11.12 1.52 1.82 9.60 -9.30 92.16 86.49 -89.28 -0.97
7.49 0.96 0.43 6.53 -7.06 42.64 49.84 -46.10 -1.08
5.20 0.73 0.01 4.47 -5.19 19.98 26.94 -23.20 -1.16
SUMS •
603.79 353.31 1064.40 250.48 460.61 4351.56 25859.50 9188.98 24.29

MEANS
35.52 20.78 62.61 14.73 27.09 255.97 1521.15 540.53 1.43