1052 XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3
overestimate silicate and mica grades in said product and
composite sample. Such errors can be attributed to the large
discrepancies in the flotation behavior observed for the dif-
ferent families (see Figure 7), where the selectivity trends
for sulfide recovery for Family 1 are very different from all
other families. Predictions of product masses are very close
to observed values for all composite samples, indicating the
suitability of the prior adjustment strategy used.
Recovery Interpolation in the TSF
Figure 9 presents the maps of sulfides and gangue mass
pulls to the concentrates and tailings, figures that are used
to quantify the grade and recovery (Figure 10) of these
mineral groups in the different products (Eq. 1). For all
these results, maps in which coloring of the interpolations
(represented by the map) and coloring of the original data
(represented by the dots) merge well characterize reliable
geostatistical models.
The quality of mass pull predictions is good for both
mineral groups in both processing products, indicating the
suitability of the chosen strategy as well as the variogram
models used. On the other hand, predictions of sulfide
grade in the tailings as well as sulfide recovery to the con-
centrate are slightly less accurate, in particular in the north-
ernmost part of the TSF. Additionally, interpolations show
somewhat lower variability than the original data, owing to
the fact that the interpolations are actually smoothing the
data. This discrepancy is, as mentioned, larger in the north
of the TSF, the region of Family 1 (see Figure 2), suggest-
ing that errors may derive from the differences in flotation
behavior among the composite samples (cf. Section Overall
flotation results).
DISCUSSION AND CONCLUSION
In this study, a new methodology to forecast the recov-
ery potential of a tailings storage facility is introduced.
Overall, it involves a series of steps in which uncertainties
accumulate:
1. TSF sampling,
2. Sample clustering,
3. Flotation test work,
4. Automated mineralogy characterization,
5. Particle-based separation models,
6. Geostatistics interpolation.
Errors up to step five can be assessed based on experimen-
tal results collected throughout the work and indicate that
the use of a single PSM for the entire TSF is successful for
all but one of the composite samples (see Figure 8). These
results extend the findings presented by Pereira et al. (2022)
about the generalization performance of PSMs. More spe-
cifically, it indicates that PSMs can be used for predicting
the process outcome of samples that were not used in the
model training phase for flotation, a separation process
where particle-particle interactions are largely expected
(Fuerstenau et al., 2007), as long as flotation mechanisms
are not extremely different (see Figure 7). In this case, bulk
sample properties seem to be controlling the overall behav-
ior of the composite samples in flotation, where process
selectivity is negatively correlated with the content of fine
Figure 7. Selectivity plots between target minerals and silicates. Based on MLA results
overestimate silicate and mica grades in said product and
composite sample. Such errors can be attributed to the large
discrepancies in the flotation behavior observed for the dif-
ferent families (see Figure 7), where the selectivity trends
for sulfide recovery for Family 1 are very different from all
other families. Predictions of product masses are very close
to observed values for all composite samples, indicating the
suitability of the prior adjustment strategy used.
Recovery Interpolation in the TSF
Figure 9 presents the maps of sulfides and gangue mass
pulls to the concentrates and tailings, figures that are used
to quantify the grade and recovery (Figure 10) of these
mineral groups in the different products (Eq. 1). For all
these results, maps in which coloring of the interpolations
(represented by the map) and coloring of the original data
(represented by the dots) merge well characterize reliable
geostatistical models.
The quality of mass pull predictions is good for both
mineral groups in both processing products, indicating the
suitability of the chosen strategy as well as the variogram
models used. On the other hand, predictions of sulfide
grade in the tailings as well as sulfide recovery to the con-
centrate are slightly less accurate, in particular in the north-
ernmost part of the TSF. Additionally, interpolations show
somewhat lower variability than the original data, owing to
the fact that the interpolations are actually smoothing the
data. This discrepancy is, as mentioned, larger in the north
of the TSF, the region of Family 1 (see Figure 2), suggest-
ing that errors may derive from the differences in flotation
behavior among the composite samples (cf. Section Overall
flotation results).
DISCUSSION AND CONCLUSION
In this study, a new methodology to forecast the recov-
ery potential of a tailings storage facility is introduced.
Overall, it involves a series of steps in which uncertainties
accumulate:
1. TSF sampling,
2. Sample clustering,
3. Flotation test work,
4. Automated mineralogy characterization,
5. Particle-based separation models,
6. Geostatistics interpolation.
Errors up to step five can be assessed based on experimen-
tal results collected throughout the work and indicate that
the use of a single PSM for the entire TSF is successful for
all but one of the composite samples (see Figure 8). These
results extend the findings presented by Pereira et al. (2022)
about the generalization performance of PSMs. More spe-
cifically, it indicates that PSMs can be used for predicting
the process outcome of samples that were not used in the
model training phase for flotation, a separation process
where particle-particle interactions are largely expected
(Fuerstenau et al., 2007), as long as flotation mechanisms
are not extremely different (see Figure 7). In this case, bulk
sample properties seem to be controlling the overall behav-
ior of the composite samples in flotation, where process
selectivity is negatively correlated with the content of fine
Figure 7. Selectivity plots between target minerals and silicates. Based on MLA results