Key Takeaways

  • With block model machine learning, RQD, fracture frequency, point-load strength, clay content, comminution hardness (BWI/DWI) and metal recovery (Cu, Mo, Au, Ag) can be estimated beyond the sparse physical test coverage.
  • Computer-vision models (for example, 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 are scikit-learn or XGBoost regressions trained on your own assay, geological and lab-test database, then interpolated into the block model like any other variable.
  • Maturity varies: RQD and recovery models are proven in production, while PLT prediction is still mostly a custom, in-house build (Emerging).

TL;DR

Block model machine learning means training a regressor on your existing lab tests against cheap, dense features such as lithology, alteration, mineralogy and grade. You then predict every block that lacks a test and krige or interpolate the result like a raw assay. Validate against held-out tests before anyone plans with it.

What Is Block Model Machine Learning?

A block model divides a deposit into thousands of 3D cells. Each block carries estimated values: grade, density, and increasingly geotechnical and metallurgical properties. Planners use those values to decide what to mine, how to blast and how the plant will perform.

The problem is coverage. Geotechnical and metallurgical tests are expensive, so only a small share of samples ever get them. However, cheap measurements such as assays, logging codes and hyperspectral scans exist almost everywhere. Block model machine learning learns the link between the cheap data and the expensive tests. It then fills every block with an estimate.

How Do I Predict Geotechnical & Recovery Variables With AI?

The pattern is the same across all seven submodels. Here it is in three steps.

Training regressors on sparse lab tests

Start with the expensive targets: RQD from core logging, PLT and UCS from geotechnical rigs, BWI and DWI from comminution testwork, clay percentages from XRD, and metal recovery from metallurgical tests. Then gather the cheap covariates: lithology, alteration, mineralogy from QEMSCAN, MLA or TIMA, hyperspectral scans and assay grades. Next, train a supervised regression model. scikit-learn and XGBoost are the standard starting points.

Reading RQD from core photos and mineralogy

For RQD and fracture frequency, there is a more direct route. The 2025 “K-Net” research, published via Springer, trains a convolutional neural network on core photography to segment and count discontinuities. It reproduces manual RQD logging to within roughly 1.5% error. Similarly, if you run TESCAN TIMA or similar automated mineralogy, its outputs become strong features. Comminution hardness models built this way are now standard geometallurgy practice, as our page on predicting BWI and DWI with machine learning explains.

Block model machine learning for metal recovery

Recovery models (Cu, Mo, Au, Ag) are the most validated of the group. Published case studies include an XGBoost model reporting R² = 0.83 at the Tizert deposit, and Random Forest and Gradient Boosting work at Cerro Verde reporting a 6.5% production lift. Follow the same template: pull historical recovery results, join them to mineralogy, grade and alteration features, and fit a boosted-tree regressor per metal. By contrast, PLT submodels are the least mature. No dedicated commercial tool exists yet, so you build and validate a custom regression against your own PLT database.

Limits of Block Model Machine Learning

Treat the ML output like any estimated variable. Validate it against held-out test results. Flag extrapolation risk in domains with no physical test coverage. Above all, don’t let the model silently replace testwork in geotechnically or metallurgically critical zones.

Model quality also depends on the domains you train within. Good geometallurgical domains keep each model inside rock that behaves consistently. As a result, many teams fit one model per domain rather than one for the whole deposit.

FAQs

What variables can block model machine learning predict?

Common targets include RQD, fracture frequency, point-load strength, clay content, BWI and DWI hardness, and Cu, Mo, Au and Ag recovery. Any property measured on a subset of samples can be a target, as long as dense covariates exist. Accuracy varies by variable and deposit.

Which algorithms work best for block model machine learning?

Gradient-boosted trees such as XGBoost are the usual starting point, with Random Forest and neural networks as alternatives. Boosted trees handle mixed data types well and train quickly. The right choice depends on your data volume and validation results.

How do ML predictions get into the block model?

You predict the variable at each sample location that has covariates, then krige or interpolate those predictions into blocks. Alternatively, you predict directly at block level from block-estimated covariates. Either way, the result is treated like any other estimated variable.

Can machine learning replace metallurgical testwork?

No. Models extend testwork coverage; they do not remove the need for it. Critical zones still need physical tests, and new tests keep the models calibrated as mining moves into new ground.

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. That cross-source data-harmonisation problem is what Geocluster is built for. Instead of hand-stitching CSVs, you get a research harness to pull the pieces together. Read how the Geocluster Research Harness works to learn more.