Artificial Intelligence for Drill Core data Analysis
Performance Testing of the integrated Neural Network
CoreSmart is an Artificial Intelligence Predictor to enrich drill core geochemical data based on hyperspectral scans. The system is using corescan database Version 2.0 including 1000km scanned core from Australia, Europe, Asia and Africa and one million corresponding geochemical records.
In a performance test the team analyzed the database with standard classification methods, deep learning approaches and the unique for CoreSmart developed neural network. Tested was the prediction accuracy for four classes of Gold, Copper and Iron contents in a 60% learning and 40% testing dataset of 110.000 records. The following table shows the achieved prediction accuracies achieved with the different tested methods for Iron and a rating combined over all commodities tested.
Class |
Comparison Methods (Accuracy in Percent) |
CoreSmart Neural Network |
||||||
SVM4 |
RBF |
NB |
MLG |
SVM2 |
MLP |
KNN |
||
Fe0 – below 1ppm |
62.3 |
62.7 |
80.9 |
68.0 |
73.0 |
82.8 |
76.0 |
|
Fe1 – below 1% |
31.3 |
0.0 |
15.7 |
9.6 |
1.7 |
65.4 |
87.0 |
|
Fe2 – below 10% |
28.4 |
77.6 |
67.4 |
93.6 |
60.8 |
84.4 |
85.9 |
84.0 |
Fe3 – higher concentr. |
45.6 |
48.6 |
57.5 |
64.9 |
72.4 |
68.9 |
77.5 |
82.0 |
Accuracy overall Iron |
42.7 |
64.9 |
68.3 |
72.6 |
63.6 |
75.1 |
80.7 |
82.3 |
Average Performance Ranking over all Commodities |
11.00 |
9.67 |
9.33 |
8.00 |
7.52 |
6.67 |
4.67 |
1.33 |
Best Fit Rank all commodities |
8 |
7 |
6 |
5 |
4 |
3 |
2 |
1 |
Abbreviations: SVM4 – Supported Vector Machine (4 classes problem) / RBF – Radial Base Classifier / NB – Naïve Bayes / MLG – Multinomial Logistic Regression / SVM2 – Supported Vector machine (2 classes problem) / MLP – Multi layer Perceptron / KNN – K-Nearest-Neighbor
The results showing a significant better performance of the CoreSmart Neural network (average over all commodities above 80%) against classical prediction methods (average below 70%). Therefore, the CoreSmart Neural Network allows even for a 4-class analysis good prediction of geochemical properties over the complete core and can be used for more targeted geochemical probing, deposit modelling and as base for artificial intelligence areal exploration models.