Key Takeaways

  • Recovery and hardness prediction from routine measurements is one of the best-documented ML applications in geometallurgy right now, with 2024 to 2026 case studies.
  • Gradient-boosted tree models dominate, not deep learning. The usual tools are XGBoost, LightGBM and Random Forest via scikit-learn.
  • Cancha is the closest thing to a turnkey commercial product built for this use case.
  • This is a solid “Yes”: published, reproducible and increasingly standard practice, not a research curiosity.

TL;DR

For recovery and hardness prediction with AI, train a gradient-boosted regression model on your historical geometallurgical testwork. Use mineralogy, geochemistry and alteration as inputs and measured recovery or hardness as the target. You can build it in Python or use a purpose-built platform like Cancha.

What Is Recovery and Hardness Prediction?

Recovery is the share of a metal, such as copper, that the plant captures into concentrate. Hardness describes how much energy it takes to grind the ore, often measured as the Bond ball mill work index (BWI). Both vary from block to block across a deposit.

Measuring them directly is slow and expensive. Flotation tests and comminution tests need large samples and weeks of lab time. Recovery and hardness prediction uses cheaper, routine measurements to estimate those values everywhere in the block model. As a result, planners can see metallurgical risk before the ore reaches the mill.

How Do I Predict Recovery and Hardness With AI?

The work follows a standard supervised-learning pattern: assemble data, train a model, then validate it.

Assemble the training data

First, gather your historical geometallurgical testwork. Useful features include mineral assemblage from automated mineralogy (TIMA, MLA or QEMSCAN), elemental assays, alteration type and hardness indices. The target is measured copper (or molybdenum, gold, silver) recovery, or a hardness parameter such as BWI. Each row pairs the inputs with a measured result.

Train a model for recovery and hardness prediction

Next, choose the model. The literature favours gradient-boosted trees over deep learning. XGBoost, LightGBM and Random Forest from scikit-learn perform well on tabular, feature-engineered geological data. They are also far easier to interpret than a neural network. That matters when you have to explain to a geologist why the model predicts low recovery in a given domain.

If you would rather not build it in-house, Cancha packages this workflow. Its analysis module uses machine learning to propose prediction algorithms for recovery, product quality and throughput from your geometallurgical sample data.

Validate before trusting the model

Finally, test on held-out samples the model has never seen. Published case studies commonly report R² values of about 0.85 to 0.90 or higher for recovery prediction from mineralogical and geochemical inputs. However, check your own model by ore domain, not just overall. A strong average can hide a domain where it fails badly.

Why This Is More Trustworthy Than Most

Recovery and hardness prediction is arguably the most mature ML application in geometallurgy. It has a real published track record rather than vendor claims to take on faith. Therefore, several operations now treat model output as a genuine input to mine planning, not just a research exercise.

The usual caveat still applies. Model quality depends entirely on how representative your training testwork is of the ore you will mine. The same method also underpins broader geometallurgical proxy models.

FAQs

Which machine learning models work best for recovery and hardness prediction?

Gradient-boosted trees such as XGBoost and LightGBM, and Random Forest, usually perform best. They handle tabular geological data well and need less data than deep learning. They are also easier to explain to geologists and metallurgists.

What data do I need to predict copper recovery?

You need historical flotation test results paired with inputs for the same samples. Typical inputs are mineralogy, elemental assays and alteration type. The more ore types the dataset covers, the more reliable the model.

How accurate is ML recovery prediction?

Published case studies commonly report R² values of about 0.85 to 0.90 or higher. However, accuracy depends on your data. A model can perform well on average and still fail on an ore type missing from training.

Can recovery and hardness prediction feed the block model?

Yes. Once validated, the model estimates recovery and hardness for every block with the required inputs. Several operations already use that output in mine planning and scheduling.

Try It With Geocluster

Pulling together published benchmarks for recovery and hardness prediction accuracy takes time. So does matching them against your own ore body’s characteristics. That literature-grounded research is exactly what Geocluster is built to run.