Lalor geochemical drillhole classification

The study of the Lalor volcanogenic massive sulphide deposit in Snow Lake, Manitoba, illustrates the power of using ML for classifying geological units with drillhole geochemical data.

Project objectives

While the classification of rock types by geochemical variables is widely used in geological mapping and exploration, such traditional methods are restricted in several ways: by class segmentation of scatter diagrams, plotting two to three elements and derived element ratios.

Machine learning (ML) algorithms such as neural networks and support vector machines (SVM) allow for a multivariate approach that includes all available geochemical elements and variables.

For a carefully selected subset of drill core samples, ML algorithms compute the relationships between geological units (as defined by visual observations of the expert) and multi-element geochemical signatures.

This “training” subset is used to compute classification rules that are subsequently applied to all the drillhole geochemical samples, resulting in a predicted 3D distribution of the predefined geological units.

Expected results and key findings

This study demonstrates that automated classification of drillhole geochemical data can predict subsurface distribution of geological units, producing less subjective 3D models than traditional interpretation methods.

If the subset of drillhole data is based on a representative subset of high-quality geological observations, the ML classifiers can be applied at multiple stages of exploration. The classification results may also highlight anomalous geochemical signatures associated with economic mineralization that might otherwise be overlooked.

Other key findings include:

• Using controlled training sets, high success rates — up to 92% — can be obtained with ML for classifying different types of geological units from the same drillhole geochemical dataset (such as lithological and hydrothermal alteration units)
• Although support vector machines provide the highest classification accuracy, other ML algorithms — such as K-nearest neighbour, random forests and Bayesian methods — generated similar results with slightly lower classification accuracies.

If the ML classification of drillhole data results in spatially coherent units, the ML algorithms by themselves do not consider the location and spatial context of the drillhole samples. Future research aimed at incorporating spatial context in the class discrimination functions could improve the accuracy of the ML classification methods.

Digital and artificial intelligence techniques

• Machine learning algorithms for multivariate geochemical classification
• Support vector machines (SVM)
• Ensemble method

Data requirements

• Geochemical analyses of drill core samples georeferenced in 3D space
• Drill log geological descriptions of geological units to compile a representative training set