On the Use of Machine Learning for Mineral Resource Classification

Abstract

Mineral resource classification relies on the expert assessment of a qualified person (QP) to determine which blocks of a 3D mineral resource model are classified as measured, indicated, or inferred. The decision is often based on a combination of quantitative parameters related to the estimation process and qualitative decisions based on previous experience or preconceptions not captured in the numerical model. As such, the procedure is subject to inconsistency, that is, blocks with similar qualities may end up in different categories, mainly due to the subjective nature of the approach. In this paper, we present a methodology to assist the qualified person in this task by clustering blocks with similar parameters and then classifying them into categories in a consistent and automatic manner that only requires the specification of a few hyper-parameters. The result is a consistent classification into resource categories comparable to the result of a classification done by a qualified person but fully consistent and generated in a short time frame. The procedure begins with repeatedly sub-sampling the block model to determine a distance matrix using an unsupervised random forest approach and then clustering the blocks using the associated distance matrix. Then, all the blocks in the model are classified using a supervised random forest approach, which gives a probability of belonging to each class. The initial class can be determined from the class probabilities. Support vector classification with a radial basis function kernel is utilized to smooth the boundaries between classes and define the final classification. An approach to tune the hyper-parameters of smoothing is provided. The methodology is demonstrated with two examples from two gold deposits. Results are comparable to the classification done by the project qualified person using conventional methods.