Rock-chemistry-to-mineral-properties conversion: Machine learning approach

Abstract

A problem of prediction of ore minerals properties by whole-rock chemistry (47 elements) has been stated and solved for a case of the Kovdor baddeleyite-apatite-magnetite deposit (Russia, Murmansk region). Four regression methods have been used: linear multiple regression, artificial neural networks, random forests, and multivariate adaptive regression splines. The latter has turned out to be the best. Content of major and trace elements in the economic minerals, average grain size, and textural properties (intergrowths of the ore minerals with gangues) have been predicted with fairly high accuracy. Subsequently, a problem of finding an optimal number of predictors (i.e., a number of elements in whole-rock chemistry analyses) has been investigated to simplify regression models and, therefore, reduce time and cost of the prediction of ore mineral properties. We have found that 5–9 predictors are enough to predict one parameter of an ore mineral. It allows using the approach both for geometallurgical modelling of a deposit and for real-time control of ore quality and/or mineral processing.