In a mine, knowledge of rock types is often desired as they are important indicators of grade, mineral processing complications, or geotechnical attributes. It is common to model the rock types with visual graphics tools using geologist-generated rock type information in exploration drillhole databases. Instead of this manual approach, this paper used random forest (RF), a machine learning (ML) algorithm, to model the rock type at Erdenet Copper Mine, Mongolia. Exploration drillhole data was used to develop the RF models and predict the rock type based on the coordinates of locations. Data selection and model evaluation methods were designed to ensure applicability for real life scenarios. In the scenario where rock type is predicted close to locations where information is available (such as in blocks being blasted), RF did very well with an overall success rate (OSR) of 89%. In the scenario where rock type was predicted for two future benches (i.e., 30 m below known locations), the best OSR was 86%. When an exploration program was simulated, performance was poor with a OSR of 59%. The results indicate that EMC can leverage RF models for short-term and long-term planning by predicting rock types within drilling blocks or future blocks quite accurately.

A new machine learning algorithm, BoxRF, was developed specifically for estimating grades from drillhole datasets. The method combines the features of classical estimation methods, such as search boxes, search direction, and estimation based on inverse distance methods, with the robustness of random forest (RF) methods that come from forming numerous random groups of data. The method was applied to a porphyry copper deposit, and results were compared to various ML methods, including XGBoost (XGB), k-nearest neighbors (KNN), neural nets (NN), and RF. Scikit-learn RF (SRF) performed the best (R² = 0.696) among the ML methods but underperformed BoxRF (R² = 0.751). The results were confirmed through a five-fold cross-validation exercise where BoxRF once again outperformed SRF. The box dimensions that performed the best were similar in length to the ranges indicated by variogram modeling, thus demonstrating a link between machine learning and traditional methods. Numerous combinations of hyperparameters performed similarly well, implying the method is robust. The inverse distance method was found to better represent the grade–space relationship in BoxRF than median values. The superiority of BoxRF over SRF in this dataset is encouraging, as it opens the possibility of improving machine learning by incorporating domain knowledge (principles of geology, in this case).

Machine learning (ML) is increasingly being leveraged by the mining industry to understand how rock properties vary at a mine site. In previously published work, the rock type, granodiorite, was predicted with high accuracy by the random forest (RF) ML method at the Erdenet copper mine in Mongolia. As a result of the optimistic results (86% overall success rate), this paper extended the research to determine if ML would be successful in modeling rock domains. Rock domains are groups of rocks that occur together. There were two additional goals. One was to determine if the variograms could predict or help understand how ML methods would perform on the data. The second was to determine if 2D modeling would perform well given the disseminated nature of the deposit. ML methods, multilayer perceptron (MLP), k-nearest neighborhood (KNN) and RF, were applied to model six rock domains, D0–D5, in 2D and 3D. Modeling performance was poor in 2D. Prediction performance accuracy was high in 3D for the domains D1 (92–94%), D2 (94–96%) and D4 (85–98%). Note that the domains D1 and D2 together constituted about 80% of the samples. Conclusions drawn in this paper are based on the results of 3D modeling since 2D modeling performance was poor. Prediction performance appeared to depend on two factors. It was better for a domain when the domain was not a minuscule proportion of the sample. It was also better for domains whose indicator semi-variogram (ISV) range was high. For example, though D4 only contributed 15% of the samples, the range was high. MLP did not perform as well as KNN and RF, with RF performing the best. The hyperparameters of KNN and RF suggested that performance was best when only a small number of samples were used to make a prediction. One overall summary conclusion is that the two most important domains, D1 and D2, could be predicted with high accuracy using ML. The second summary conclusion is that semi-variograms can provide insight into ML performance.