A new particle-based approach to process modelling and diagnostics

For a long time, the mining and minerals processing communities have developed process models based on the bulk chemical compositions of ores and processing products. Recently, the emergence of the geometallurgy field and the advent of new characterization techniques have shifted the focus to mineralogical composition and certain particle-based properties (e.g. mineral liberation). Automated mineralogy systems based on scanning electron microscope platforms, such as the Mineral Liberation Analyzer or TIMA-X, were essential for this advance. They are capable of producing large data sets with detailed information on the sizes, shapes and compositions of individual ore particles in a sample.
Methodologies that use this particle-based information to model the outcomes of a specific processing unit or an entire operation are summarized as “particle tracking”. This is a useful diagnostic technique for mineral separation processes. However, limitations in the statistical methods applied in existing approaches mean that users are required to summarize considerably the information contained in the particle datasets by pre-selecting explanatory variables (expert input) and binning particles. This means that a distinct recipe, including specific assumptions, is required for each new case. It also makes the resultant models liable to human bias.
This work describes a new method for particle tracking that combines automated mineralogy data with machine learning to automate the variable selection process and eliminate the need for particle binning. In doing so, it maximizes the utilization of detailed particle data, minimizes the effects of human input, and provides the flexibility to assess different cases with minimal prior adjustments. Its utility is demonstrated using data from both a flotation and a magnetic separation units of an operating niobium mine. In both cases, the method is able to accurately predict the mineralogical compositions of the concentrate and tailings streams. It clearly has the potential to be extended as a diagnostic tool for the optimization and operation of processing plants.