Machine learning in mineral exploration

Exploiting the advantages of both supervised and unsupervised learning

The potential benefits of AI in mineral exploration are staggeringly large, yet its application is far from simple. In mineral deposit targeting, explorers are trying to identify the location of ore deposits at the core of a very complex system. Evidence of the deposit footprint must be assembled from interpretation of subtle alteration effects, for example, that may extend kilometres from the target. Meeting this challenge requires as much focus on how the problem is set up for generating AI predictive models as on the AI methods themselves. This is where deep domain knowledge and a mining industry-specific, supporting computational framework is required.

We have applied machine learning as part of custom solutions to complex exploration and geotechnical problems since 2015. Our advanced geological modelling and geophysical inversion tools give us the capability to create a fully integrated interpretation, capturing all the relevant features in a single earth model, as input to machine learning algorithms. This makes for sound feature engineering, a key success factor for the application of AI to mineralized system targeting.

Case study: Radiometric alteration footprinting in IOCG environment, Mount Dore project area, AU.

In sparse data or grassroot environments, it is useful to be able to quickly identify alteration footprints of IOCG mineralization. In the case of the Mount Dore region, we tested whether it would be possible to identify the alteration footprint of IOCG deposits using airborne magnetic and radiometric data as a starting point. Using four features and a hierarchical clustering (unsupervised learning) approach, we tested the ability to generate proxies for the mass balance of potassium and uranium linked to alteration effects. The clustering method enabled the generation of groups of data points with similar magnetic, potassium, thorium, and uranium signatures, which exhibit remarkable correspondence to the geological map. Once the groups were established, a 3D regression was used to estimate the expected potassium and uranium signatures for each group of points. Then, a residual value was calculated and used as a proxy for mass balance of both elements. For example, areas with excess potassium were established by comparing the expected potassium for the rock type (data group) to the measured values; they are then considered as being enriched in potassium by alteration processes. Both the potassium and uranium residuals are then combined to produce an alteration map for IOCG deposit footprints. Areas of both potassium and uranium enrichment are considered prospective. The approach performed quite nicely at finding already known deposits in the Mount Dore area.