Using machine learning to estimate a key missing geochemical variable in mining exploration: Application of the Random Forest algorithm to multisensor core logging data

Mining exploration increasingly relies on large, multivariate databases storing data ranging from drill core geochemical analysis to geophysical data or geological descriptions. Utilizing these large datasets to their full potential implies the use of multivariate statistical analysis such as machine learning. The Random Forest algorithm has proved its efficiency in mining applications. In this study we use it to estimate a key geochemical element, sodium, using a multivariate chemo-physical dataset measured on drill cores in the Matagami mining district of Quebec, Canada. Sodium is important to characterize hydrothermal alteration in volcanogenic massive sulfide settings, since Na depletion can be used to vector towards ore, but this element is not readily measured by portable X-ray fluorescence (pXRF). We first test the algorithm on a database of over 8000 traditional laboratory geochemistry analyses and find a correlation of 0.95 between estimated and measured Na. We then test the algorithm on the multi-sensor core logging data, including density, magnetic susceptibility, and 15 geochemical elements by pXRF, but borrowing Na from traditional geochemistry (n = 260). This yields correlations of 0.66 to 0.75 depending on the training and testing sets. Finally the algorithm is applied to the whole multiparameter database (n = 9675) to estimate Na downcore. There is a good general correspondence with the downcore Na patterns seen through traditional geochemistry, and the estimated Na which has much greater spatial resolution. Random Forest appears to be a very good estimation tool when using large amounts of data and variables, as it uses all variables and automatically prioritizes the most useful. This method also allows visualization of the weight of each variable in the estimation. Future studies should compare RF with other methods.