Machine learning is creating value in all facets of the mining industry, from exploration to production.
The authors provide an accessible, high-level introduction to artificial intelligence (AI), machine
learning (ML), and deep learning (DL); the latter being widely recognised as one of the most powerful
forms of ML. In particular, the authors will introduce deep learning models known as convolutional
neural networks (CNNs), how they are applied, and the economic considerations necessary for
determining when DL may be the right solution to de-risking complex block modelling problems.
The authors present preliminary results from the mineral resource modelling study of the Jundee
orogenic gold deposit, Yandal belt/gold province, Western Australia. The primary goal was to identify
the direction and location of the narrow gold veins more accurately and demonstrate how non-linearly
correlated elements are used as direct inputs to the resource model to assist with the target
element’s grade prediction. This demonstrates that: (1) existing techniques for finding correlations
between assayed elements do not adequately reflect the complex geology of the asset, (2) nonlinear
correlations that are difficult to model as simple mathematical functions are representative of
geological patterns in a deposit, and (3) non-linearly correlated assayed data, fed as inputs, increase
the performance of the resource model as reconciled through blind tests.
To conclude, the authors hypothesise that the patterns represented by the DL block models may be
revealing the results of overprinting geological processes that generated mineral deposits. For
example, the primary hydrothermal processes that deposited metals created depositional patterns
that become particularly complex as a result of being overprinted, in part or in whole, by secondary
physicochemical processes. This may explain why non-linear geochemical relationships have the
capacity to generate more accurate block models of mineral deposits.