Geoscience Reference
In-Depth Information
Table 14.21
Data available
Data type
Description
Notes
Head assays
This data contains the % content of various important elements,
including: Co, As, Mo, Ni, Pb, Zn, Zr, Sr, Bi, Cd, Cs, Ga, In, Sb,
Se, Te, Th, Tl
This data is compositional in nature
Mineralogy
A total of 10 identified minerals and minestals set make up the
bulk of the deposit. These include: brannerite, coffinite, urani-
nite, pyrite, chalcopyrite, bornite, chalcocite, other sulphides,
acid soluble gangue and acid insoluble gangue
This data is also compositional in nature
Association data
A number of thin sections are available. These have been analyzed
and the complete matrix of associations between minerals is
available. This describes the contact area between two adjacent
minerals within a single grain of crushed material
This data is also compositional in nature
Table 14.22
Description of predictive models generated
Model
Input variables
Output
Comments
Full model
Head assays (i.e. %cu, %U …)
10 Mineralogy
10 × 11 matrix of associations
Specific gravity
Cu, U, Au, Ag recoveries
Acid consumption
Net recovery (U)
BMWi and DWi
This model represents the maximum data
available
Typical model
Head assays (i.e. %cu, %U …)
10 Mineralogy
Specific gravity
Cu, U, Au, Ag recoveries
Acid consumption
Net recovery (U)
BMWi and DWi
This is the base case model. Field data will
most likely contain all these variables
Limited model
Limited head assays
7 Mineralogy variables
Specific gravity
Cu, U, Au, Ag recoveries
Acid consumption
Net recovery (U)
BMWi and DWi
Only head assays that have many samples
in the available database are considered
data in Table
14.21
as input to a regression model, these six
plant performance variables can be predicted at all locations
in the deposit.
3. Merge the variables (level 2). This step reduces the 23
merged variables to 4.
4. Regression on the four variables.
5. Back transform the estimated variables (DWi, BMWi, Cu
recovery, U
3
O
8
recovery, acid consumption and net re-
covery).
6. Determine uncertainty in the model.
14.7.2
Methodology
A linear regression model is used to predict the plant per-
formance variables. One drawback with a linear regression
model is that all input variables are required for prediction.
Thus, if a single input variable is missing from a sample, the
regression model cannot be applied. For this reason, three
regression models are generated (Table
14.22
). Each model
represents a decreasing number of input parameters. For ex-
ample, for locations in the deposit where association data is
not known the “full model” cannot be applied and the “typi-
cal model” would be appropriate.
The regression models are based on a large set of input
variables. The variables are merged into super secondary
variables based on the correlations between variables. This
is done because there are too few sample data available to ac-
curately determine regression coefficients for the 204 input
variables available. The final model is a linear regression on
four super secondary variables. The methodology consists
of six steps:
1. Normal score the input variables.
2. Merge the variables (level 1). This step reduces the 112
input variables to 23 merged variables.
Step 1: Normal Score Data
First, the number of variables
must be reduced. Variables are removed from the analysis
because (1) they have a low correlation to the six output
variables or (2) they are highly redundant with one of the
other input variables. A variable was considered to have a
low correlation if the maximum correlation to any of the out-
put variables was less than 0.13. A variable was considered
redundant with another input variable if it had a correlation
greater than 0.94. This reduces the number of input variables
to 112.
There are a total of 841 samples available for modeling;
however, not all samples contain all 112 variables used in the
calibration of this model. Due to the nature of a regression
model, it is necessary that all 112 variables be present for a
sample to be used for calibration. Of the 841 samples, 328
samples were retained for modeling. More data are available
if the mineral associations are ignored (for example, if using
the “typical” model).
All 118 variables (112 input + 6 output) are indepen-
dently normal score transformed. A visual assessment of the
