Key Takeaways

  • Both calculated mineralogical models and geometallurgical (geomet) models are, in practice, regression/classification problems trained on automated-mineralogy data — not distinct disciplines requiring separate tooling.
  • QEMSCAN, MLA, and TIMA supply the ground-truth mineral abundance and texture data that feeds the models.
  • scikit-learn and XGBoost are the standard open-source libraries for building the regression/classification layer — recovery, hardness, throughput, and mineral-abundance predictions are commonly built as gradient-boosted regressions.
  • Leapfrog’s Geomet extension folds these predictions back into the same 3D block model as your geology, so geomet variables interpolate alongside grade.
  • This is confirmed, currently-used practice (ai_relevant: Yes) — not speculative.

TL;DR

Train a gradient-boosted regression (XGBoost or scikit-learn) on your QEMSCAN/MLA/TIMA automated-mineralogy results and metallurgical test data, then interpolate the model’s predictions into your block model via Leapfrog’s Geomet extension to get spatially continuous mineralogy and geomet domains.

How Do I Predict Mineralogy and Geometallurgical Domains With AI?

Both of these model types solve the same underlying problem: you have automated-mineralogy results (mineral abundance, texture, liberation) and metallurgical test results (recovery, hardness, throughput) at a limited set of drillhole/sample locations, and you need continuous spatial predictions across your whole deposit. The standard workflow is a two-step pipeline. First, you train a supervised model — typically a gradient-boosted regression using XGBoost or a classifier/regressor from scikit-learn — on your labeled dataset, where the labels come from QEMSCAN, MLA, or TIMA automated mineralogy plus whatever comminution/flotation testwork you’ve run (BWI, DWI, recovery tests). The model learns the relationship between cheaper, more abundant inputs — assay grades, alteration codes, hyperspectral signatures — and the expensive-to-measure mineralogical/metallurgical targets.

Second, you interpolate the model’s predictions spatially. This is where it plugs back into your existing geology software: Leapfrog’s Geomet extension (part of the Leapfrog ecosystem) is built specifically to carry geometallurgical variables through the same block-modelling pipeline as your grade estimates, so a mine planner sees predicted hardness or recovery sitting right next to Cu grade in the same block. The practical payoff is that you stop having to physically test every block — the ML model extends your limited (and expensive) automated-mineralogy and metallurgical testwork across the whole resource.

This is one of the more mature AI applications on this list — it’s not an emerging research direction, it’s a documented, current practice with real published case studies (R² figures around 0.89 for recovery prediction models are reported in industry literature). The main caveat is data quality: these models are only as good as your QEMSCAN/MLA/TIMA sample coverage, so sparse or unrepresentative sampling will produce a model that extrapolates poorly.

Try It With Geocluster

Wiring together mineralogy data, ML regression, and your block model is exactly the multi-step workflow that benefits from a dedicated research harness rather than a pile of disconnected notebooks and scripts. Geocluster is built to help you reason across that whole geomet-modelling pipeline. Check it out on GitHub.