Exploring semi-supervised generation of mineral tables for automated mineralogy systems

The quality of results obtained by a scanning electron microscope-based automated mineralogy systems strongly depend on the project-specific library of spectra of the minerals considered to form the samples analysed, the mineral table. Mineral tables can be selected from existing universal libraries according to the expected or suspected minerals, and/or can be constructed based on spectra collected in the samples of a specific project. One can prefer extense mineral tables to enhance the chances of capturing variations in chemical composition of specific minerals; or else smaller, compact mineral tables to reduce misclassification. No general, perfect balance within generality-specificity or extension-compaction exists, so that users end up interactively and iteratively building the spectral library for each project in tedious steps of adding and removing mineral spectral candidates. This process is also project- and ore-specific. While some automated mineralogy devices provide operating modes for automatically constructing mineral lists throughout a measurement, these commonly offer only limited settings and are not clear about the data processing steps.

The goal of this contribution is to compare the performance of several components of a strategy to automatically construct automated mineralogy mineral lists - making use of several machine learning algorithms, for the specific case of dataset collected with the Mineral Liberation Analyser (MLA). The strategy has five steps: (1) preliminary data transformation, (2) dimension reduction, (3) endmember detection, (4) phase detection, and (5) unmixing.

For each of these steps, several options were tested. These included for data transformation peak extraction and Box-Cox transformations, which at the same time embraces logarithm/log-ratio transformations, square root transformations and the identity transformation. Regarding dimension reduction, principal component analysis and minimum-maximum autocorrelation factors were tested. In step three, we considered QHull convex hull detection, and N-FINDER, a conventional linear endmember detection method. In step four, the goal is to find the groups of spectra that can be identified with the members of the mineral table, not all of them being necessarily endmembers. Algorithms tested here correspond to model-free unsupervised classification algorithms, such as k-means, hyerarchical clustering methods and topological spectral clustering among other. Finally, in step five we tried several sparse and unconstrained linear unmixing algorithms. This unmixing was done within the sample of spectra forming the training data only in order to determine the number of necessary groups (or clusters) to extract from step four, as the actual final phase attribution will be done by the MLA software for the whole project after delivering the mineral table. The strategy presented here offers not only improvements to the workflow of scanning electron microscope-based automated mineralogy systems but also is a step stone for compiling mineral lists in analytical devices such as µX-Ray Fluorescence automated mineralogy, where spectra mixing is a bigger issue.