You are on page 1of 14

Minerals Engineering 160 (2021) 106667

Contents lists available at ScienceDirect

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

Repurposing legacy metallurgical data part II: Case studies of plant


performance optimisation and process simulation
Yousef Ghorbani a, *, Glen T. Nwaila b, Steven E. Zhang c, Martyn P. Hay d
a
Department of Civil, Environmental and Natural Resources Engineering, Luleå University of Technology, SE-97187 Luleå, Sweden
b
School of Geosciences, University of the Witwatersrand, Private Bag 3, Wits 2050, South Africa
c
PG Techno Wox, 43 Patrys Avenue, Helikon Park, Randfontein 1759, South Africa
d
Eurus Mineral Consultants (EMC), Plettenberg Bay, South Africa

A R T I C L E I N F O A B S T R A C T

Keywords: The history of the metallurgical industry is rich with data. An enormous amount of data is generated from mining
Data analytics operations and industrial factories, and as deployment of new technologies such as on-line monitoring and in-situ
Dry laboratories instrumentation proliferate through the 4th industrial revolution, the quantity and quality of data will increase
Data bank
dramatically. The first paper (Part I), describes a range of promising technologies that integrate well with
Gold
Flotation
existing mineral processing plants and testing laboratories to demonstrate the enormous potential of a dry
Metallurgy laboratory. A dry lab is a type of laboratory that includes applied or computational mathematical analyses for an
extensive range of different applications. In both laboratories and mineral processing plants, integration of
timely, accurate and reliable data analytics is key to leveraging data to enable data-driven plant design, opti­
misation and monitoring. However and despite progresses in analytical technology and increasing availability of
data and sophisticated data analytics, legacy metallurgical plant and test work data are being underutilised.
Understanding the insights contained within legacy metallurgical plant data is critical to the transition into a
data- and analytics-driven industry. This paper (Part II) details two case studies that use legacy data to benefit
metallurgical processes. One case study focuses on operational data from a gold recovery plant and provide
indirect knowledge of the structure and/or composition of the feed sources, and insights to guide the optimi­
sation of the operation. The other case study focuses on laboratory flotation tests, and demonstrates the effec­
tiveness of aggregated data in establishing empirical guidelines that can guide the design and optimisation of
new and existing processing operations.

1. Introduction are properly retrieved, processed, and interpreted using data analytics
techniques, they can be of significant and timely value, particularly in
In our first paper (Part I of this study), we have presented a review of the design of new metallurgical plants and optimisation of those that are
recent technological advances in the metallurgy discipline covering the underperforming. We focus on showcasing insight extraction through
full-spectrum of the mineral processing value chain. The review also the application of a variety of data analytics, process diagnostics and
examined the potential for the integration of modern analytical tech­ process simulation on legacy data in two case studies. In order to ensure
nology with data shifting towards dry laboratory 1environments as a a traceable transition from wet to dry laboratories, we use a combination
future-oriented approach. This paper (Part II) demonstrates that if data of tools and techniques that currently exist, but perhaps underutilised in

* Corresponding author.
E-mail address: yousef.ghorbani@ltu.se (Y. Ghorbani).
1
A dry lab is a laboratory where metallurgical processes can be simulated in a controlled laboratory environment through the use of computational or applied
mathematical analyses performed on a computer-generated model. The applied mathematical analyses include advanced data analytics and machine learning al­
gorithms. It is a computer-based laboratory where data from various sources can be stored and processed using high-end computing techniques. The dry lab differs
from Laboratory Information Management Systems (LIMS), which are operating laboratories for analysing assays to establish product quality. The scope of a dry lab
is wider, as an all-encompassing laboratory that is capable of processing and simulating different metallurgical processes from data derived from varied sources such
as test work laboratories, operating plants, and specific mineral processing equipment. Database systems such as LIMS could potentially be extended to include
mineralogical analysis as well as metallurgical testing data. Thus, LIMS could also benefit significantly from the dry lab approach.

https://doi.org/10.1016/j.mineng.2020.106667
Received 28 June 2020; Received in revised form 29 September 2020; Accepted 11 October 2020
Available online 3 November 2020
0892-6875/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 1. Gold metallurgical processing flowsheet (Ghorbani et al., 2020).

the industry. In addition, we seek to illustrate the value of legacy lack of standardised data governance within the business, which reduces
metallurgical data. Our first case study focuses on gold plant- the efficacy of data analysis and the strength of the interpretation of
performance optimisation and the second case study provides an inno­ data-derived information, and therefore the efficacy of the business in­
vative approach for plant design using flotation process simulation. The telligence. However, as this closely mimics the current state of opera­
availability of diverse metallurgical data and the development of mod­ tional data within the mining industry, we seek to best utilise this data to
ern and often highly-automatable data analytics algorithms make a gain insights on plant performance and potential plant optimisation
compelling case for the industry to move towards a dry-laboratory era. strategies.

2. Database descriptions and methods 2.2. Case study 2: SUPASIM® flotation model

2.1. Case study 1: gold plant predictive analytics Given limited sample size, ore characterisation and flotation per­
formance has commonly been measured by conducting laboratory-scale
In this case study, we apply data analytics on legacy operational data test work. The Denver D12 cell has been largely adopted as the standard
that records both metallurgical and metal accounting information from laboratory-scale flotation test machine since about the 1950s. This 2.5 L
a gold processing plant. The data records characteristics of gold ores cell is convenient and ideal for treating one-kilogram milled samples, as
from its intake to output, such as feed grade, treated and residual mass of well as being a practical volume for testing pulp taken from an operating
ore mixtures per shift and feed ore tonnage. From the data, we derived plant. In the majority of cases a common set of operating conditions
plant recovery using a mass balance. The database contains 33 records, (impellor speed, air rate, pulp level) were used to generate recovery-
covering just over a month of continuous operation on a daily basis. time and recovery-grade profiles (i.e. a rate test) as well as final
These records represent time-weighted averages of the measured vari­ concentrate via open-circuit batch tests
ables over the day, totalised where applicable (such as throughput). The An approximation of plant performance could be obtained by con­
company’s primary objective is to recover gold from defunct gold tail­ ducting a locked-cycle test, however the correlation with production
ings. A total of four defunct tailings storage facilities (TSFs) are sourced plant performance was often poor (Agar and Kipkie, 1978; Agar et al.,
by the company’s gold processing plant. The TSFs were generated from 1982). The best conditions were determined by comparing sets of
processing primary Witwatersrand Basin (South Africa) auriferous recovery-time and recovery-grade profiles generated via a full-factorial
conglomerates, where gold recovery percentages from primary ores are experimental design process. Although this generates a result, modelling
estimated to have been between 55 and 79% (Janisch, 1986). To respect production plant performance from laboratory batch simulation has
the company’s confidentiality, we do not disclose the company name, remained challenging. Some best practices are invariably ore-type based
instead opt to use pseudonyms for the ore sources. Fig. 1 shows a high- and statistically applicable to similar ores. As with plant data, head
level process flowsheet. The data characterises the ore mixtures through grade, recovery, concentrate mass and grade and recovery-time and
a fixed process that includes crushing, milling, conditioning, thickening, recovery-grade profiles offer only limited insight into flotation re­
carbon-in-leach (CIL), elution columns and electrowinning circuit. sponses. Data points produced from laboratory-scale flotation results
The plant operates 24 h per day, 7 days per week on a three-shift-per- can be modelled by Kelsall’s two-component equation (Kelsall, 1961) to
day basis (morning, afternoon and night shifts). The data contains determine the flotation kinetics of metal, mineral and floatable gangue.
missing entries, cryptic metadata information, inconsistent and un­ These are the parameters that drive flotation response, recovery, mass
known numbers of significant digits, unknown accuracy and only pull and concentrate grade. Laboratory kinetics characterise the ore and
inferred precision. These are largely data quality issues that result from a are predictors of plant-scale metallurgical performance and circuit

2
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 2. Fundamental controls on flotation performance.

design (Fig. 2). In essence, flotation performance can be modelled • Fast floating fractions of economic metals and slow floating rates of
numerically, enabling flotation behaviour to be compared on an equal economic metals and gangue from rate tests on rougher feed, and
basis that is independent of ore type and head grade. Using flotation • Metal recovery of the laboratory rougher rate test.
kinetics as the main input, a flotation simulation program, SUPASIM®,
was developed by Eurus Mineral Consultants in the 1980s, in order to Excluded are elements generally found in silicates (e.g. Si, MgO, Fe,
predict production plant performance from laboratory batch simula­ Al) and metal contaminants (e.g. As, Sb, Mg, Hg) as both of these groups
tions. Details of Kelsall’s equation can be found in Supplementary Ma­ are part of floatable gangue and usually have kinetics similar to floatable
terial 1 and methodology and application of SUPASIM® in gangue. Chromite is a metal contaminant associated with UG-2 ores and
Supplementary Material 2. also excluded because its recovery is almost entirely by entrainment. In
In this case study, we use a database consisting of 178 simulations of addition, individual metals (Pt, Pd, Rh, Au) that make-up the 4E plat­
plant operations that were conducted since 1985, using flotation kinetics inum group metals assay are excluded, as their kinetics are sufficiently
derived from laboratory batch rate tests conducted with the same grind mutually similar and almost indistinguishable from that of the combined
and using the same reagents as the plant. These simulations were con­ 4E platinum group metals. Including these analyses would double the
ducted as part of consulting work to duplicate production plant per­ number of simulations without producing additional overall knowledge
formance and validate the SUPASIM® program. After program or change the outcome of data analysis.
validation, other reagents, plant designs and operating conditions were Plant circuits are limited to those of standard configuration consist­
routinely simulated but are not part of this case study. In this case study, ing of roughing and one to three stages of cleaning operating in closed
we present a consistent set of simulation conditions to focus on the circuit. Twenty two percent of circuits include regrinding of rougher
feasibility of similar techniques and highlight the usefulness of data, concentrate. As is typical of regrinding, concentrate grade is improved
rather than the validation of SUPASIM®. For the database used, the but not overall recovery. One exception of a Cu-Zn plant in north-east
quality of the data and metrology adhere to employed laboratories’ Africa is excluded, where regrinding of rougher concentrate in both
standards and covers the following ore types and recovery of individual circuits markedly improved recovery. Multiple circuits such as MF2,2
metals: MF33 and plants that consist of two to four flotation stages in sequence,
such as in a Cu-Pb-Zn plant and Si, CO3, Nb, S recovery stages in a
• Platinum group metals South African Merensky Reef and Upper niobium plant are treated individually. Recoveries quoted are stage re­
Group 2 (UG-2), Zimbabwean Great Dyke, Canadian Pd-rich ore (Lac coveries and not recovery relative to main plant feed. Other plant sim­
Des Iles), ulations involving changes in grind, reagents, temperature, residence
• Base Metals: Cu-Pb-Zn, Cu-Zn, Pb-Zn, Cu-Ni, Cu and Ni (including time, circuit configuration and ore type that were not implemented in
pyrite-rich ores), the plant (or where subsequent plant data was not made available due to
• Niobium, confidentiality restraints) are not included in the database.
• Phosphate,
• Furnace and Converter Slags: platinum group metals and Cu, 3. Case study 1: Gold plant performance diagnostics using data
• Tailings: platinum group metals, analytics
• Ore Type: mined ore and furnace slag, and Merensky Reef and UG2
ores treated through the same plant, 3.1. Optimisation of a gold plant using legacy data
• Economic Secondary Metals: Cu and Ni in platinum group metals
ores (excluding UG2), Co in Cu and Ni ores, In this case study, we focus on process and metal accounting data
• Economic Metal Contaminants: Cu in Ni ores, Ni in Cu ores, Cu in Pb gathered during processing of defunct low-grade gold tailings and how
and Zn ores, Pb in Cu and Zn ores, Zn in Cu and Pb ores, the data can be used to improve plant performance through the appli­
• Gangue Contaminants: S in Nb ores. cation of data analytics. Data analytics, even without advanced machine
learning and artificial intelligence components can optimise plant per­
For this study, data for analysis has been limited to: formance beyond what is possible with conventional control systems by

• Actual plant recovery,


• Simulated plant recovery at the same concentrate grade as the plant,
2
Secondary mill/float stage treating primary mill/float stage tailings.
3
Tertiary mill/float stage treating secondary mill/float stage tailings.

3
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 3. A scatter plot of the source grade versus the percentage recovery of gold through a fixed plant process. Several interesting correlations between the source
grades and the percentage recovery are observed. The parameters for the regression models are shown in the figure as overlays. Red points are missing data points
that are interpolated using means of the dataset (not used for modelling). (For interpretation of the references to color in this figure legend, the reader is referred to
the web version of this article.)

providing additional control dimensions, increased rate of feedback (e.g. tonnage needs; ore conditioning and thickening processes, where the
near real-time) or more quantitative feedback control loops. Plant per­ gold is extracted into a cyanide-based solution and adsorbed into acti­
formance can be optimised to a greater degree when the process is made vated carbon; extraction of dissolved gold in elution columns using
more stable through predictive analysis and quality feedback. For oxidation-fixation of gold in activated carbon masses and; extraction of
complex systems, simultaneous multivariable control may be required to solid gold from carbon using electrowinning cells.
achieve desirable control granularity of key outcomes. A successful plant
optimisation strategy considers various layers of control, their in­
teractions and optimisation to create a holistic, system-level approach. 3.3. Data analytics
Analytics expedite the decision-making process by uncovering insights
using highly established and automatable techniques that can naturally From a systems analysis perspective, plant recovery is a composite
coexist with existing practices and seamlessly integrate into the 4th in­ performance indicator that can be used to understand plant perfor­
dustrial revolution. A good understanding of the drivers of plant per­ mance. There are an unlimited variety of performance indicators, some
formance, and their links to profitability is required to identify and of which can be important, or key performance indicators (KPIs) that
prioritise opportunities that will have a significant impact on gold re­ measure performance, such as that of a process at a low level or of a
covery. The top-level drivers of performance can be characterised by plant at a high level. One of the most important perspectives of plant
conventional value modelling. The impact of the underlying parameters performance is the relationship between ore characteristics and feed
on these top-level drivers is typically less well-understood. Our approach grade, and their impacts on the plant’s performance. Although max­
combines analytics with fundamental process knowledge to highlight imisation of the plant recovery is important to value production, tuning
the viability of an evidence-based, plant performance optimisation using various processes and activities through established plant processes is
routine operational data. not a simple task, since metallurgical processes contain complex in­
teractions including nonlinear effects. Our approach here is data-driven
and we seek to understand the relationships that may exist between the
3.2. Repurposing of defunct gold tailings plant’s input (ore mixtures), output (recovered gold mass) and perfor­
mance (recovery percentage). An advantage of the data-driven approach
Ores from the defunct gold tailing storage facilities (TSFs) are is that discovery of insights is not predicated on metallurgical knowl­
labelled TSF1 to TSF4 and are generally processed as blends of random edge or scientific hypothesis-testing. Therefore, this approach is highly
proportions, although within the data, TSF1 appears to be very under- applicable to operational data. Indeed, in our first case study, we employ
utilised relative to the other sources. The general process of extraction strictly data analytics to discovery relationships within the dataset and
includes sourcing ore mixtures from the various TSFs depending on only employ knowledge of metallurgy and the plant’s process to perform

4
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 4. Scatter plots of the source total source mass versus the percentage recovery of gold through a fixed plant process. No obvious relationships are observed. Red
points are missing data points that are interpolated using means of the dataset. (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

post-hoc analysis. The relationships discovered can be used as a basis for scales, and the gold liberation processes are identical between the
scientific inquiry via hypothesis testing, plant process diagnostics or assay- and plant-processes, then with increasing assay gold grade, the
improvement, or statistical inferences. plant recovery should not decrease. Identification and, if possible,
remediation of the undesirable inverse relationships should yield in­
3.3.1. Effects of feed ore on plant performance creases in plant performance. Additionally, these observations are
Within the available data, there are identifiable relationships be­ crucial from a plant-optimisation perspective, as given sufficient data,
tween various quantitative measurements of the 4 TSFs and plant re­ fitted models can be used to predict the effectiveness of ore mixtures, or
covery (Figs. 3 and 4). As is expected of operational data, the data is conversely, optimising an extraction process for inflexible fixed blends.
noisy. No relationships can be observed between plant recovery and From the metal accounting perspective, there seems to be inadequate
TSF1. A noisy positive correlation can be seen for TSF4 between plant controls on metal in process as indicated by plant recoveries that are
recovery with feed grade. For TSF2 and TSF3, there are inverse re­ sometimes above 100%. The causes for > 100% plant recovery could be
lationships, implying an increase in ore grade results in decreased plant poor sampling efficiency or unrealised plant lockups that could lead to
recovery. These relationships are unlikely to be caused by the noisy delayed metal in process (MIP) or work in progress (WIP)
nature of the data, as the R-square value of two power law regressions accountability.
are sufficiently high (Figs. 3 and 4). The exact choice of the regression Weak correlations between the total mass all sources and the plant
functions is not critical in this application, as long as the functions recovery are observed (Fig. 3). The behaviour of TSF2 and TSF3 are
feature a minimum of free parameters and a model can be chosen based probably different than TSF4. TSF1 is the least-utilised source and over a
on its R-square value. A plausible explanation for the differences in substantial portion of the dataset, its utilisation is zero. Daily mixtures
behaviour between the sources is that the microstructure and compo­ used at the plant consistently contain greater amounts of TSF2 and TSF3
sition of the sources are different, which leads to differences in behav­ ores (>50%) relative to the other 2 ore sources. As higher TSF2 and TSF3
iour under a fixed gold-liberation process. For example, there may be grades translates into poorer gold recovery, over-reliance on these
some mineralogical phases or spatial arrangement of phases that are sources is a subtle plant efficiency issue. In this context, plant optimi­
preventing effective extraction of TSF2 and TSF3. Since the data does sation could focus on designing an optimal process for TSF2 and TSF3
not capture residue grades per source, because they are mixed there­ ores, if their predominant usage is unavoidable, or minimise the usage of
after, we cannot easily check this. The presence of inverse correlations as these ore sources if possible. Alternatively, experiments could be con­
observed for TSF2 and TSF3 indicates that there is substantial difference ducted to examine possible synergistic effects of ore mixtures under
between ore homogeneity and/or the assay and the plant processes. If fixed gold liberation processes (e.g. Nwaila et al., 2019).
the volumes of ore mixture are homogeneous at the assay- and plant-

5
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 5. Recovered gold versus run-of-mine gold


content (total raw gold). Roughly, two positively
correlated trends (not regression-fitted, but indi­
cated by dotted lines 1 and 2) are observed with
some outliers (circled). Trend 1 is more efficient
than trend 2, since for a given total raw gold, the
recovered gold is higher. Data records with missing
components are interpolated using mean values and
are shown as red circles. (For interpretation of the
references to color in this figure legend, the reader is
referred to the web version of this article.)

Fig. 6. Eluate gold concentration versus run-of-mine gold content in total raw gold (kg) (fuschia pink). Two roughly linear relationships and some outliers are
observed. Data records with missing components are interpolated using mean values and are shown as red circles. Substantial outliers are shown in black. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3.3.2. Process plant efficiency analysis single correlation between the input and output, and a more efficient
At a high level, the plant process can be conveniently visualised as a plant process should exhibit a transfer function that is closer to the 1:1
transfer function that maps some input (run-of-mine gold content) to the line. The data from our plant is best described by two linear transfer
output (recovered gold). A consistent plant process should produce a functions (Fig. 5), which indicates the existence of two different

6
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 7. Plant recovery versus eluate concentration.


A linear relationship is observed up to 100% re­
covery and is modelled by considering that at 0 mg/
L eluate concentration, the recovery is 0%. The pa­
rameters of the model are shown in overlay. Data
records with missing components are interpolated
using mean values and are shown as red circles. The
red dashed line at 100% indicate the proper limit of
Au recovery while points above 100% shows poor
metal accounting due to work in progress or unac­
counted ore accumulation in the plant. (For inter­
pretation of the references to color in this figure
legend, the reader is referred to the web version of
this article.)

processes with substantially different extraction efficiencies. The out­ chemical extraction of the gold from the thickener (Fig. 6). As the plant
liers (Fig. 6) are particularly concerning, as these high run-of-mine gold operates with 2 batches of thickeners per day and no other part of the
content ore mixtures are highly under-extracted. process is known to be conducted in two disparate pipelines, it is
This inconsistency is already present at the eluate stage after reasonable that the difference occurs in the methodology and/or

Fig. 8. Recovered gold (kg) versus plant recovery. A roughly linear but noisy relationship is observed and the relationship is modelled here assuming that at 0%
recovery, there is no recovered gold. Data records with missing components are interpolated using mean values and are shown as red circles. (For interpretation of
the references to color in this figure legend, the reader is referred to the web version of this article.)

7
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 9. Multiple-range plot of masses of ore mixtures before and after their passage through the gold liberation process as conducted by various shifts at the plant for
all days. The mean and its standard error (SE) are shown.

implementation of the preparation of the two thickeners from original between eluate gold concentration versus total recovered gold, not
ore mixtures. However, the usage of the two thickeners is untraceable in shown here). This also indicates a thorough extraction of gold during
the data past the eluate stage. Despite the inconsistency, there is a strong electrowinning as no saturation behaviour is observed (e.g. an asymp­
positive correlation between gold in eluate solution and the run-of-mine tote in Fig. 7). Hence, the information could be interpreted to suggest
gold content (Fig. 6). Most of the outliers (Fig. 5) are already present at that the process leading up to the formation of an eluate is a limiting and
the eluate stage and are recognisable (Fig. 6). This again indicates that highly variable factor in the plant’s overall performance. This infor­
the processes that preceded the formation of the eluate should be sub­ mation combined with the inverse relationships observed between the
jected to additional scrutiny to examine the causes of low extraction incorporation of TSF2 and TSF3 ores (Fig. 3) suggests that a sound
efficiency from very high run-of-mine gold content feeds. method to optimise plant performance is to mix ores from TSF2 and
The electrowinning process seems to be much more consistent. A TSF3, and the other two sources in two separate single thickener tanks
single linear relationship is observed between the eluate concentration respectively. The chemical process associated with each tank can be
and the recovery percentage (Fig. 7). This indicates that the amount of optimised to maximise gold extraction for each type of ore behaviour.
gold extracted from the eluate is directly proportional to the concen­
tration of gold in the eluate (and confirmed by the linear relationship

Fig. 10. Multiple-range plot of grades of ore mixtures before and after their passage through the gold liberation process as conducted by various shifts at the plant for
all days. The mean and its standard error (SE) are shown.

8
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Table 1
Statistics of masses and grades of ore mixtures for various shifts. Here MS, AS and NS denotes morning, afternoon and night shifts, respectively.
Shift Data Mean Standard Deviation Sum Minimum Median Maximum

Thickener Treated MS (t) 31.01 5.48 1023.35 12.82 32.51 42.69


Thickener Treated AS (t) 33.22 5.65 1096.29 6.45 33.58 43.48
Thickener Treated NS (t) 35.77 8.72 1180.35 20.46 34.54 78.53
Treated Mass MS (t) 4199.76 998.76 138592.00 741.00 4526.00 5269.00
Thickener Grade MS (t) 0.35 0.07 11.62 0.21 0.35 0.71
Treated Mass AS (t) 4501.39 929.26 148546.00 318.00 4699.00 5317.00
Thickener Grade AS (t) 0.31 0.06 10.38 0.04 0.33 0.37
Treated Mass NS (t) 4675.42 711.38 154289.00 1902.00 4887.00 5308.00
Thickener Grade NS (t) 0.33 0.05 11.00 0.25 0.33 0.54
Residue Mass MS (t) 4185.91 950.43 138135.00 1665.00 4256.00 5545.00
Residue Mass AS (t) 4313.33 752.89 142340.00 919.00 4432.00 5278.00
Residue Mass NS (t) 4449.52 688.03 146834.00 2252.00 4629.00 5434.00
Residue Grade MS (g/t) 0.13 0.04 4.23 0.04 0.12 0.27
Residue Grade AS (g/t) 0.13 0.04 4.13 0.05 0.12 0.28
Residue Grade NS (g/t) 0.13 0.04 4.27 0.06 0.13 0.26

3.3.3. Plant performance optimisation analysis (Figs. 9 and 10). Therefore, the performance variation across the three
From a business perspective, the total plant productivity is a product shifts could be caused by some other form of variability of the ore
of labour efficiency and operating hours. It can be characterised by the sources or of the process (e.g. the TSFs are not well-mixed prior to
total recovered gold, which is another business KPI that is also a function processing). Environmental factors that vary significantly for morning-
of the many plant inputs and processes. Under a linear approximation shifts include changes in ambient temperature, which affects slurry
and assuming that all other factors are fixed within a plant, such as viscosity and thus gold dissolution kinetics (e.g., significant drops in
operating hours and labour efficiency, a plant’s recovered gold should temperature at night usually occur between 2:00 and 5:00 AM). How­
be proportional to the plant recovery, which is indeed observed for the ever, these factors should not affect the raw mass of ore mixtures treated,
studied plant (Fig. 8). An ideal plant is fully consistent in each of its which is measured prior to chemical processes. Labour factors are more
processes, such as labour efficiency and ore mixture homogeneity. plausible, such as early-morning sleepiness. To further examine the
However, a physical plant deviates from this ideal model in many ways. plant performance from a labour perspective, more data is needed, as
Inconsistencies known at the studied plant such as variabilities in ore- inferences from a small dataset (n = 33) could be unreliable. Nonethe­
mixture grades affects recovery. In general, a more consistent plant less, the nature of the substantial outliers is interesting and the identi­
should feature a narrower scatter cloud on a recovered gold versus re­ fication and remediation of factors leading to such outliers may be
covery percentage plot. Hence, a primary target to improve the studied worthwhile to yield significant gains in overall plant productivity.
plant’s performance is to identify and remediate issues that result in the
greatest outliers within the data (Fig. 8). Although this is true and 4. Case study 2: Flotation characterisation and performance
perhaps intuitive for our studied plant, this may not be true in general.
For example, a plant that utilises several processes aimed to maximise 4.1. Test work and simulation of plant flotation performance
recovery of several types of ores may not exhibit a single, convenient
relationship between recovered gold and plant recovery. Based on the The second case study demonstrates possibilities brought by a
insights from the feed source behaviour and process consistency ana­ simulation technique (i.e. SUPASIM®) that in the context of modern­
lyses, numerous simple modifications and additional scrutiny for sig­ isation of measurement technologies and increasing deployment of data
nificant outliers should lead to increases in plant performance. analytics is novel. In the past, several successful applications using
SUPASIM®, have been demonstrated in flotation plant design and
3.3.4. Influence of work schedule shift on plant performance analysis methodology (Hay and Rule, 2003; Hay and Martin, 2004).
It is possible to analyse labour productivity by examining the shift- Although the original SUPASIM®, methodology was developed in the
specific data. The morning shift exhibits the highest standard devia­ mid-1980s to predict plant performance from standard laboratory
tion and the lowest mean of treated and residual mass of ore mixtures flotation tests, the technique has since been upgraded to integrate recent
(Fig. 9). This variability is unlikely to be related to the quality of the technological advances, such as new variables measured using new
feed, as the grade of the feed and residual mixtures are statistically technologies.
identical between all shifts (Fig. 10). A summary of the general statistics The database compiled since 1985, consists of 178 simulations of
of the various shifts is given in Table 1. The night shift appears to be the operating plants duplicating existing average performance using flota­
most consistent (Fig. 9, Table 1). The reduction in the morning shift’s tion kinetics derived from laboratory batch rate tests conducted at the
treated mass of ore mixtures is significant (~10% relative to the other same grind and using the same reagents as the plant. The process
shifts, Table 1) and three characteristics in the data are observed (Fig. 9, involved the following steps;
Table 1): (1) the mean of the morning shift treated mass is lower relative
to all other shifts; (2) the spread of the data points is greater and; (3) 1. Conduct rougher and cleaner stage rate tests.
there are three significant outliers below 3000 t. This suggests the pos­ 2. Determine flotation kinetics for economic metals assayed, mass and
sibility of both a systematic and random component responsible for the floatable gangue.
reduction of the morning shift’s performance. Since the plant operates 3. Set up the production plant flowsheet with cell numbers and nominal
on a 24-hour basis, plant factors such as transient phenomena associated volumes sizes in the SUPASIM® program.
with the start-stop of sub-processes involved in the plant are unlikely to 4. Input feed tonnage, % solids by mass, ore specific gravity and head
be responsible. If the feed and residual ore mixture-grade data (Fig. 10) grades of economic metals.
behave similar to the treated- and residual-mass data (Fig. 9), then the 5. Input flotation kinetics of fast and slow floating fractions and rates
variability could be traced to a heterogeneity of the ore sources (e.g. a for economic metals and floatable gangue.
significant variability in grain size and a portion of the morning shift’s 6. Input cell and bank aeration rates and water rate constants typical for
labour is spent on remediation processes). However, this is not observed the ore type being processed.

9
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 11. Simulated plant recovery (%) versus actual plant recovery (%). A good correlation and a nearly unity slope is observed.

Table 2
Characteristics of incremental recovery categories.
Laboratory Values

Incremental Recovery Actual Plant Metal Fast Slow Floating Slow Floating Slow Floating Rougher Conc Actual Plant % of
Categories (Average) Recovery (%) Floating Rate of Gangue Rate of Metal Ratio (SFR) ** Recovery (%) Recovery less IMF Data
Fraction as %* (%)*

>2.0 74,9 69,1 0,00262 0,04516 21,0 82,5 5,8 26,4


0.0 to 2.0 76,1 75,2 0,00250 0,03038 12,4 86,2 0,9 15,2
¡2.0 to 0.0 57,9 59,3 0,00256 0,01745 9,5 72,2 − 1,4 9,0
0.0 to ¡10.0 61,5 66,6 0,00241 0,01672 9,4 73,6 − 5,2 28,1
¡10.0 to ¡15.0 52,9 65,0 0,00838 0,02249 5,0 77,7 − 12,0 7,9
¡15.0 to ¡20.0 39,1 55,7 0,00913 0,02146 4,1 68,1 − 16,6 6,7
< ¡20.0 26,4 67,1 0,07895 0,29870 4,2 81,2 − 40,7 6,7
*
Metal fast floating fraction is abbreviated as IMF.
**
SFR = [slow floating rate of metal/slow floating rate of gangue].

7. Set the program to mass balance (values of scale-up factors for each mineralogically complex than they were in the last century. For
kinetic parameter in roughers and cleaners based on previous sim­ example, in copper flotation, head grade has declined from 1.5% in 1910
ulations of production plant conducted to validate SUPASIM® are to 0.4% in 2018 (Northey et al., 2014; Nwaila et al., 2019) and with-it
allowed to vary by up to 10%). recovery has also decreased. The difference between the 1910 and
8. Compare simulated with actual plant mass balances. 2018 ore bodies may not always be immediately apparent from the first
9. Fine tune the simulated mass balance by adjusting cell/bank aeration package of rougher batch and rate tests. Recovery for more- and less-
rates within practical limits. desirable ores in the roughing stage can often be similar (top and bot­
tom rows of Table 2), but plant recoveries can be very different and each
The floatabilities of mineral and gangue are determined from a would require quite different processing in the cleaning circuit. Without
laboratory-scale flotation rate test using Kelsall’s unmodified two- these guidelines presented here, a test program can proceed to an
component equation (Kelsall, 1961). Scale-up factors derived from advanced stage until issues have been properly identified. The guide­
laboratory to plant simulation case studies conducted since 1985 lines can provide prior indication early in a testing program and guide
translate laboratory open-circuit batch-scale performance into closed- cleaning tests to replicate performance problems and to produce useful
circuit plant performance (Fig. 11). An analysis of these data provides insights.
a link between laboratory kinetics and plant performance. In its totality This analysis is not primarily focussed on recovery of metal (where
these correlations are complex, but they can be condensed and simpli­ metal is referred to include minerals) but on the floatability of slow
fied to provide a means of interpreting laboratory rate test results to floating gangue, which, over a certain value has a detrimental effect on
identify positive and negative responses which need to be investigated recovery. Mineralogically complex ores tend to be more altered, which
further. This encapsulates the aim of the case study which is to better increases the floatability of slow floating gangue. Modern plant design
understand the results of an initial set of laboratory tests by using can exacerbate this situation. To compensate for lower head grades and
empirically-based guidelines developed from a large database of historic finer grain sizes, plant capacity has increased in conjunction with a finer
information. With better insight early in the test program, the oppor­ grind to maximise mineral liberation. Although mill sizes have increased
tunity exists to reduce the number of tests required together with cost significantly, classification efficiency has remained largely unchanged
and time. In addition, key information about plant performance and and finer grinding has increased the proportion of ultra-fine material
design may also be uncovered. feeding flotation. In flotation kinetics, fine and ultra-fine gangue (typi­
Ore bodies being mined today are lower grade and more cally < 15 and 5 µm respectively) is slow floating. It is not easily rejected

10
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Fig. 12. Plant recovery versus fast floating fraction (FFF) and slow floating ratio (SFR). Three categories of plants are identified and are shaded. The average SFR for
each category of plant is also shown. The plant recovery curve is color-coded as per the three categories of plants.

and is variably incorporated into final concentrates mostly by entrain­ it should in theory, be equal to plant recovery. In practice, 41% of
ment. If a kinetic mass balance is conducted around a flotation circuit, operating plants in the database recover the well-liberated fraction of
slow floating gangue typically constitutes upwards of 70% of all float­ the main economic metals in plant feed (top two rows of Table 2). This is
able material and can be as high as 92% (Hay, 2005). As such, it is the increased to 50% if a precision of 2% is assumed for sampling and
main source of dilution in concentrate and reduces final concentrate experimental errors, and operations included where recovery is up to 2%
grade; the effect being greater as slow floating gangue d80 µm size be­ less than the fast floating fraction (Fig. 12 and the top three rows of
comes finer (Çilek and Umucu, 2001). The result is that in order for a Table 2).
plant to achieve a required concentrate grade, the operation must sac­ The remainder of the plants recover ~ 12% less than the fast-floating
rifice recovery and operate at the lower end of the recovery-grade curve. fraction of metal or mineral. A portion of the well-liberated mineral is
The degree of fine gangue floatability has a considerable impact on unrecovered, because final concentrate has been diluted by slow floating
flotation behaviour. An understanding of this aspect at a laboratory scale gangue and this is represented by the slow floating ratio (SFR), which is
provides useful information about probable plant-scale recovery, the slow floating rate of metal divided by the slow floating rate of
concentrate grade, bank residence time and circuit configuration. Test gangue. Under these conditions in a plant environment, concentrate is
work on individual ores and using a database to interpret flotation not being generated at the required grade and recovery is consequently
performance enables project economics to be evaluated at an early stage reduced by moving down the recovery-grade curve. The lost material is a
in the project and future laboratory test work to be designed according portion of the fast-floating fraction of metal or mineral. An in-depth
to the characteristics of the ore and performance requirements. This can analysis of the data shows that this happens when the slow floating
save considerable cost and time compared to conducting a matrix-style ratio is<18 and slow floating rate of gangue is>0.0026 (associated slow
sets of tests aimed at characterising the ore in order to plan subsequent floating rate of metal or mineral is 0.0464). These values are a general
test sets. rule and vary a little depending on how the data is analysed.
Plants with recoveries greater than their fast-floating fraction of
metal or mineral recover slow floating metal. The amount recovered is
4.2. Interpreting flotation kinetics
referred to as incremental recovery. It was found that 26% of plants
simulated fall into this category with slow floating metal contributing an
Selected flotation kinetics from only the rate test on rougher feed
average of an additional 6% to plant recovery (Fig. 12). Slow floating
have been used as these are the smallest dimensionality of data that can
metal and gangue typically have a finer particle size distribution than
be effectively correlated with ore characteristics, general flotation
the fast-floating fraction of metal or mineral. It should be noted that data
behaviour and plant performance. The metal fast floating fraction (FFF)
in the figure are described by moving average trends, as this is the
is a measure of the degree of liberation and is the proportion of mineral
clearest way to present the relationships whilst at the same time being
in float plant feed occurring as free particles. These particles have the
detailed enough to capture the extent of the variation within the data. In
highest rate of flotation and are most easily recovered at a high
addition, the horizontal axis is ore type as defined in section 2.2 and
concentrate grade. They also tend to have a particle size within the ideal
sorted by incremental recovery.
range for flotation from 15 to 120 µm. Logic dictates that all well-
Plant recovery generally decreases in line with slow floating ratio
liberated particles should be recovered in an operating plant, thus if
(Table 2). Metal fast floating fraction does not follow the same general
metal fast floating fraction in rougher feed is expressed as a percentage,

11
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Table 3
Modelling of the influence of ore type in plant recovery.
Laboratory Values

Ore Type Actual Plant Metal Fast Slow Floating Slow Floating Slow Floating Rougher Conc Actual Plant
Recovery (%) Floating Fraction Rate of Gangue Rate of Metal Ratio (SFR) ** Recovery (%) Recovery less IMF
as %* (%)*

Av Pb 84,8 80,5 0,00260 0,07710 30,9 94,4 4,3


Av PGMs 72,9 72,1 0,00231 0,04108 20,7 86,0 0,9
Av Cu, Ni 70,6 70,0 0,00273 0,02615 12,1 79,9 0,6
Av Co 61,7 70,2 0,00242 0,01341 9,5 77,7 − 8,5
Av Nb, P2O5 63,7 91,0 0,12340 0,90195 8,2 96,5 − 27,3
Av Zn 78,7 90,1 0,01125 0,05224 7,5 93,0 − 11,4
Av All Tailings and Slag 26,6 35,8 0,00171 0,00718 4,3 45,3 − 9,2
* Metal fast floating fraction
is abbreviated as IMF
**
SFR = [slow floating rate
of metal/slow floating rate
of gangue]

trend as mineral characteristic and degree of liberation are fairly floating rate of gangue and slow floating rate of metal values of all ores.
consistent. The cause of a reduction in plant recovery is not entirely due Niobium-, phosphate- and zinc-bearing ores perform very well in the
to poorer liberation, nor is it due to an increase in slow floating rate of laboratory and exhibit a high fast floating fraction of metal, but have
gangue as the latter remains constant down to a 10% loss in metal fast high rates of slow floating gangue and consequently perform poorly in
floating fraction. Loss in plant recovery is due to a reduction in slow the plant relative to that of the laboratory environment.
floating rate of metal or mineral, which in turn reduces the slow floating Low laboratory recovery is an insufficient indicator of ore charac­
ratio. Slow floating gangue begins to adversely affect plant performance teristic. For example, some of the worst performing ores in a plant can
when its value exceeds 0.0032 and slow floating ratio falls below 9.0. generate high recoveries in the laboratory. In particular, the recovery-
Plant recoveries>10% below the metal fast floating fraction are char­ time profile of niobium and phosphate ores do not tend to a horizon­
acterised by slow floating rate of gangue>0.0080 and slow floating tal asymptote but maintain a steep slope in the second half of the rate
ratio<5.0, irrespective of the value of slow floating rate of metal or test. This should not be mistaken as being indicative of a high plant
mineral, even though it is greater than other ores that generated higher recovery due to a continuous flotation of metal. The mistake often made
recoveries. in plant design is that residence time is increased to take advantage of
The mineralogy of ore types varies considerably and has a corre­ this high rate of slow floating metal, whereas in reality the cleaners are
spondingly large influence on plant performance. As a generalisation, flooded with highly floatable fine-grained gangue that is problematic.
plants treating platinum group metals, copper and nickel ores tend to Only a kinetic analysis will identify how an ore will potentially behave
recover all metal fast floating fraction (Table 3). In the database, copper in an operating plant and what the appropriate circuit design should
and nickel ores consist almost exclusively of chalcopyrite and pent­ encompass.
landite that are known to liberate well and be good “floaters”, which is
also the case for the approximately 60 platinum-bearing minerals found 4.3. Implications for plant design and specifying additional laboratory
in platinum group metals ores. The gangue component of these ores are scale test work
usually fairly inert silicates that give rise to good liberation between
slow floating mineral and gangue with slow floating rate of gangue Prior to conducting any rate tests, initial batch test work should
below 0.0032 and slow floating rate of metal or mineral of a least 14. establish the optimum grind parameters and reagent suite and if
Galena in lead-bearing ore commonly has a high degree of liberation appropriate, other key conditions such as Eh (oxidation potential), pH
across both fast and slow floating fractions. Fast floating fraction of (the negative logarithm of the hydrogen ion concentration [H+]), water
metal and slow floating ratio are both high and recovery of slow floating quality and whether a pre-float aeration stage is necessary. The best
metal can be as much as 4–5%. All of these platinum group metals, Cu, conditions can be evaluated in more detail by comparing floatation ki­
Pb and Zn ores are characterised by a slow floating rate of gangue of netics from rate testing.
0.0027 or less and a slow floating ratio of 12 or more. In general, at The rougher rate test and derived flotation kinetics provide an initial
optimum conditions if a laboratory rougher has a slow floating ratio characterisation of the ore, which can indicate probable plant perfor­
of<12, the probability is that the production plant will have a recovery mance, aspects of flotation behaviour that need to be addressed and the
problem and laboratory test work should be focussed on this issue. type of tests that should be included in future work. Guidelines are
Although cobalt occurs with nickel, it is presented as a separate presented for four categories as combinations of slow floating rate of
category to illustrate the impact of low slow floating rates of metal. Fast gangue and slow floating ratio (Table 4). The kinetics and floatability of
floating fraction of metal is equal to that of copper and nickel ores and at metals are not directly considered because they will have been previ­
0.00242 slow floating rate of gangue is less. However, since the slow ously optimised during initial batch test work investigating grind pa­
floating ratio is 9.5, plant recovery is 8.5% below the fast-floating rameters and reagent suite, etc. The guidelines focus on slow floating
fraction of metal. Note that this large difference in plant performance gangue kinetics, which has a direct influence on metal floatability (via
between cobalt and copper-nickel ores is unapparent from their concentrate grade) and recovery in an operating plant. If an ore is
respective laboratory recoveries, which have similar values. This high­ mineralogically clean4 (category 1), slow floating metal floats (on
lights the common misconception in the industry sector that the dif­ average) 31 times faster than slow floating gangue and in the cleaners is
ference between plant and laboratory performance is due to lower sufficiently upgradeable for some slow floating metal to be recovered
mechanical efficiency in a plant. The kinetic characteristics of an ore into a final concentrate. Plant recovery increases as more of this material
have a far greater influence on flotation performance than any me­
chanical or hydrodynamic difference between laboratory and plant scale
cells. 4
Ores with a low degree of alteration, low gangue floatability and a high
Tailings and slag present an extreme example with the lowest slow degree of liberation.

12
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

is recovered in the rougher concentrate, therefore increasing rougher

for this type of


residence time beyond the norm can be justified. An example of this is

Not applicable
Unlikely to be
Worth testing
Concentrate
found in the Merensky Reef ore in the centre of the Western Limb,

successful
Process?
Bushveld Igneous Complex (South Africa). It is one of the plant opera­

Two
tions in the database, which has been simulated and optimised, but not

ore
published. This is an exceptional ore, as is the plant configuration. Its
slow floating rate of gangue is very low at 0.00113 with a slow floating

Highly likely
Unnecessary

Unnecessary
Operation?
ratio of 45.4 and a fast-floating fraction of metal of 80.5%. A single bank

necessary
Probably
Circuit
of roughers, without any cleaners, produces a final concentrate of 80–95
Open

g/t from a head grade of 4.3 g/t. Nominal residence time in the roughers
is 106 min and incremental recovery averages 6.5%. If the ore kinetics
1–2 (3 will had not been interpreted and the plant designed using a rougher resi­
Required
Stages of
Cleaning

dence time typical of platinum group metals ores of 45 min, plant


Nos of

harm)
do no

2–3

2–4

3–4
measurements showed that recovery would be 3.7% less (4.2% by
simulation) resulting in a revenue loss of US$ 32.3 M per annum (mid-
Unlikely to be

2020 platinum group metals basket and US dollar values).

Highly likely
Regrinding
in Cleaner

Not only are fast and slow floating metal fractions well liberated but
necessary

necessary
Probably
unlikely
Circuit

Highly

they also have distinct flotation behaviours and treating them separately
increases recovery. The cleaning circuit is split into two sections (the
two-concentrate process) to produce high- and low-grade final concen­
trate streams. This circuit can be conveniently tested in the laboratory.
Residence

Roughers

Increase

Increase
Time in

Ores with this characteristic respond well to the two-concentrate process


Normal

Normal

as changes to a UG-2 concentrator in 1994 demonstrated by an increase


in recovery of 4% and concentrate grade from 405 g/t to 660 g/t (Hay,
2005).
Liberation in

Mineralogically poor ores5 in categories 3 and 4 may benefit from


Very Poor
Fraction
Floating

regrinding in the cleaner circuit. All ores yield slow floating metal in the
Metal

Good
Slow

High

Poor

second half of the rougher concentrate process. In category 3 ores, slow


floating ratio is low and this fraction is consequentially rejected in the
cleaners. Fine grinding of cleaner tailings generates new fractions of fast
Recovery

80% + ≈

80% + ≈

≪ 80%
< 80%

and slow floating metal, which have different flotation characteristics to


Plant

4%

2%

rougher concentrate and as such should be treated in a dedicated sec­


ondary cleaner circuit to produce its own concentrate. By regrinding,
each single parameter of slow floating rate of metal and slow floating
rate of gangue in the rougher concentrate translates to a new fast
Data
% of

20,2

46,1

30,3
3,4

floating fraction and slow floating rate of metal in the reground cleaner
tailings stream. This is the basis behind the introduction and success of
the IsaMill™ by Anglo American into many of its platinum group metals
Recovery
less IMF

concentrators (Rule et al., 2008).


Actual

− 12,3
Plant

(%)*

− 3,4

In category 4, compared to other categories, a much greater portion


2,6

2,3

of slow floating gangue tends to occur as free particles with a high value
of slow floating rate and tends to concentrate in cleaner tailings. Its
Floating
Rate of

value may be increased further through regrinding (by overgrinding and


Metal
Slow

30,8

22,3

sliming) and recirculating it back to the head of the rougher concentrate


7,1

5,3
Categories of kinetic characteristics of ores and processing requirements.

process decreases the slow floating ratio in total rougher feed. Taking
into account circulating load and its increased gangue concentration,
Database
Averages

Floating

slow floating rate of gangue can increase by as much as 4x resulting in a


Gangue
Rate of

0,0014

0,0063

0,0019

0,0238
From

Slow

decrease in final concentrate grade when the cleaners are operated in


closed circuit. The historical rule of thumb for plant designs was that to
maximise float yield for slow floating ores, add scavengers and recir­
Metal fast floating fraction is abbreviated as IMF.

culate scavenger concentrate and cleaner tailings to the head of the


rougher process. Plants in the bottom row of Table 2 and to the left in
Floating
Rate of
Metal

Fig. 12 were designed in this manner. The indication is to process ores


Slow

> 14

> 14

< 14

< 14

with this characteristic in open circuit and the impact of this should be
tested in the laboratory.
< 0.0032

> 0.0032

< 0.0032

> 0.0032
Floating

Gangue
Rate of

5. Summary and outlook


Slow

Our case study on gold plant legacy data indicates that routine
Metal Fast

operational data contains valuable information regarding relationships


Fraction
Floating

as %*

between common feed characteristics and performance indicators.


80

80

80

80
Category

5
Ores, which have been geologically, altered containing minerals with
Table 4

gangue intergrowths, usually high gangue floatability and a low degree of


1

liberation.
*

13
Y. Ghorbani et al. Minerals Engineering 160 (2021) 106667

Furthermore, these relationships provide indirect knowledge of the independent laboratories with a significant data generation component,
structure and/or composition of the feed sources, such that the resulting such as dry laboratories, to ensure that within the competitive private
knowledge is already appropriate to guide the optimisation of the industry setting, important knowledge that can provide universal ben­
operation. For example, one may design optimum blending of ores based efits is not lost due to the tragedy of the commons effect.
on their leaching amenability, promoting synergistic or similarly-
behaved blends and diagnosing the nature of the discrepancy between Declaration of Competing Interest
the thickeners. In essence, we have effectively demonstrated that
existing operational data combined with data analytics produces The authors declare that they have no known competing financial
insightful information and actionable knowledge to drive continuous interests or personal relationships that could have appeared to influence
improvement in the plant’s performance. Therefore, operational data is the work reported in this paper.
an essential mining asset that will only prove to be more valuable, as the
industry transitions towards a data-rich and increasingly analytics- Acknowledgements
driven type of operational feedback. As data becomes more abundant
and complex through the adoption of technology, and its generation We thank Julie E. Bourdeau for help with pre-submission critical
becomes increasingly integrated in plant operation, improving data comments and editing of this paper.
quality through management and governance should increase the vari­
ety, quality and dependability of insights extracted through data ana­ Appendix A. Supplementary material
lytics. Sufficient and high-quality data paves the way for predictive
models and their adoption at the operation level, such as relationships Supplementary data to this article can be found online at https://doi.
between feed source characteristics and recovery efficiency. In an ideal org/10.1016/j.mineng.2020.106667.
future scenario, data analytics could become fully automated and is of
minimal latency to guide plant operations including process refinement
References
and on-line performance monitoring.
The flotation case study highlights several important findings. Agar, G.E., Kipkie W.B., 1978. Predicting locked cycle flotation test results from batch
Firstly, the broad and deep knowledge that can be extracted from an data. CIM Bulletin, 71, No. 799, p. 119.
Agar, G.E., Roy, J., Stratton-Crawley, R., 1982. Bench-scale simulation of flotation plant
abundance of aggregated data and its implications on extraction and
performance. CIM Bulletin, Vol. 75, No. 848.
processing pipeline design. Such knowledge is generally impossible to Çilek, E.C., Umucu, Y., 2001. A statistical model for gangue entrainment into froths in
produce through data gathered at a single plant or simulation. Thus, this flotation of sulphide ores. Miner. Eng. 14 (9), 1055–1066.
first finding also illustrates the importance to aggregate data from Ghorbani, Y., Nwaila, G.T., Chirisa, M., 2020. Systematic framework toward a highly
reliable approach in metal accounting. Miner. Process. Extract. Metallur. Rev.
multiple sources in order to produce generalisable information that https://doi.org/10.1080/08827508.2020.1784164.
transcends the boundaries of a particular ore, plant or business. Aggre­ Hay, M.P., Rule, C.M., 2003. SUPASIM™: a flotation plant design and analysis
gating data in the form of a data bank is an ideal approach. Secondly, methodology. Miner. Eng. 16 (11), 1103–1109.
Hay, M.P., Martin, C.J., 2004. SUPASIM™: a Methodology to Predict Plant Performance
there is a need for rigorous laboratory work with a clear focus and goal from Laboratory Data. In: Proceedings 36th Annual Meeting of the Canadian Mineral
to mend knowledge gaps that are otherwise unfeasible or undesirable to Processors, January 2004, pp 281-299.
address at the plant level. Laboratories can rapidly replicate plant pro­ Hay, M.P., 2005. Using the SUPASIM® Flotation Model to Diagnose and Understand
Flotation Behaviour from Laboratory through to Plant. In: Proceedings 37th Annual
cesses to augment existing data, explore nuances such as ore reaction Meeting of the Canadian Mineral Processors, January 2005, pp 463479.
kinetics that are sometimes impossible to study at the plant level, and Janisch, P.R., 1986. Gold in South Africa. J. South. Afr. Inst. Min. Metall. 86 (8),
provide a neutral, well-managed and data generation source of high 273–316.
Kelsall, D.F., 1961. Application of probability in the assessment of flotation systems. Bull.
quality, free of silos and intended for knowledge production through Inform. Min. Metall. 650, 191–204.
data analytics. Thirdly, the amount of non-public domain data that is Northey, S., Mohr, S., Mudd, G.M., Weng, Z., Giurco, D., 2014. Modelling future copper
essentially inaccessible outside of the industry sector is vast. This issue ore grade decline based on a detailed assessment of copper resources and mining.
Resour., Conserv. Recycl., 83, February 2014, 190–201.
reflects the nature of siloed operations typically conducted in the private
Nwaila, G.T., Ghorbani, Y., Becker, M., Frimmel, H.E., Petersen, J., Zhang, S., 2019.
industry. With changes in the industry, including closures and personnel Geometallurgical approach for implications of ore blending on cyanide leaching and
replacements, the data is gradually lost, especially on exhausted ore adsorption behaviour of Witwatersrand gold ores, South Africa. Nat. Resou. Res.
deposits. The irreplaceable knowledge contained within the data is then 1–24.
Rule, C.M., Knopjes, L., Jones, R.A., 2008. The development of mainstream inert grinding
unfortunately forgotten or inadvertently destroyed. This last finding or MIG IsaMill™ technology at Anglo Platinum. In: Comminution 2008, Falmouth
reinforces the notion of sound data governance and the need for Cornwall, UK.

14

You might also like