Key Takeaways

  • RQD, fracture frequency, point-load strength, clay content, comminution hardness (BWI/DWI), and metal recovery (Cu, Mo, Au, Ag) can all be estimated with machine learning instead of relying purely on sparse physical test coverage.
  • Computer-vision models (e.g. CNN-based “K-Net” pipelines) now read RQD and fracture frequency directly from core photographs, cutting manual logging variance.
  • Published copper-recovery models (XGBoost/Random Forest/ANN) have delivered real production gains — a Cerro Verde case study reported +6.5% Cu production from ML-driven recovery prediction.
  • Most of these submodels aren’t a single commercial product — they’re scikit-learn/XGBoost regressions trained on your own assay + geological + lab-test database, then interpolated into the block model like any other estimated variable.
  • Maturity varies: RQD/recovery models are proven in production; PLT-model prediction is still mostly a custom, in-house build (Emerging).

TL;DR

Train gradient-boosted or neural-network regressors on your existing lab-test results (RQD, PLT, BWI/DWI, clays, recovery) against geology/alteration/mineralogy features, then krige or interpolate the model’s predictions the same way you would a raw assay — giving every block an estimated geotechnical/metallurgical value instead of only the handful with physical tests.

How Do I Predict Geotechnical & Recovery Variables With AI?

The pattern is the same across all seven of these block-model submodels: you have a sparse set of expensive physical measurements (RQD from core logging, PLT/UCS from geotechnical rigs, BWI/DWI from comminution testwork, clay percentages from XRD, and Cu/Mo/Au/Ag recovery from metallurgical testwork), and a much denser set of cheap covariates (lithology, alteration, mineralogy from QEMSCAN/MLA/TIMA, hyperspectral scans, assay grades). The AI approach is to train a supervised regression model — scikit-learn or XGBoost are the standard starting points — that learns the relationship between the cheap covariates and the expensive target, then applies that model everywhere you have covariates but not a physical test.

For RQD and fracture frequency specifically, there’s a more direct route: 2025 research (the “K-Net” approach, published via Springer) trains a convolutional neural network directly on core photographs to segment and count discontinuities, reproducing manual RQD logging to within roughly 1.5% error. If you’re already running TESCAN TIMA or similar automated mineralogy, the mineralogical outputs (clay type, alteration intensity) become strong features for predicting BWI/DWI and clay-content submodels — comminution “hardness proxy” models built this way are now standard geometallurgy practice, not a research curiosity.

The recovery models (Cu, Mo, Au, Ag) are the most validated of the group. Published case studies — an XGBoost model reporting R²=0.83 at the Tizert deposit, and Random Forest/Gradient Boosting work at Cerro Verde reporting a 6.5% production lift — give you a real template: pull historical metallurgical test recovery results, join them against mineralogy/grade/alteration features, and fit a boosted-tree regressor per metal. Point-load-test (PLT) submodels are the least mature of the seven — no dedicated commercial tool exists yet, so this is typically a custom regression you build and validate yourself against your own PLT database before trusting it in the block model.

Whatever you build, treat the ML output the way you’d treat any estimated variable: validate against held-out test results, flag extrapolation risk in domains with no physical test coverage, and don’t let the model silently replace testwork in geotechnically or metallurgically critical zones.

Try It With Geocluster

Wiring together assay databases, automated-mineralogy outputs, and testwork results into a single feature set is most of the actual work here — and it’s exactly the kind of cross-source research and data-harmonization problem Geocluster is built for. If you’re standing up geotechnical or recovery prediction models like these, Geocluster gives you a research harness to pull the pieces together instead of hand-stitching CSVs.