Key Takeaways
- Geometallurgical proxy models are a classic supervised-regression problem: predict an expensive, slow-to-measure variable from cheap, fast-to-measure ones.
- Cancha, built specifically for geometallurgy, and general-purpose Python (scikit-learn, XGBoost) are the two realistic paths: commercial turnkey or custom-built.
- This is a genuinely mature “Yes.” It is standard applied ML with published mining case studies.
- The hard part is not the modelling. It is getting a big enough paired dataset of cheap proxy measurements and expensive ground-truth assays to train on.
TL;DR
To build geometallurgical proxy models with AI, collect paired samples where you measured both a cheap proxy and the expensive test it should predict. Train a regression model on those pairs, using Cancha or scikit-learn and XGBoost. Then apply it wherever you only have the cheap measurement.
What Are Geometallurgical Proxy Models?
Some metallurgical properties are expensive or slow to measure directly. Examples include flotation recovery from a locked-cycle test, or comminution hardness from a full SMC test. You cannot run those tests on every metre of core.
A proxy is something cheap and fast that correlates with the expensive property. Common proxies include XRF assays, point-load index, hyperspectral clay content and blasthole MWD logs. Geometallurgical proxy models learn the relationship between the two. As a result, you can estimate the expensive variable everywhere you have the cheap one.
How Do I Build Geometallurgical Proxy Models With AI?
The process has three steps: build a paired dataset, pick a modelling path, then validate.
Build the paired dataset
Proxy modelling is one of the more approachable ML use cases because it is the textbook supervised-regression setup. You need samples where you measured both the cheap proxy and the expensive ground truth. Each pair becomes one training example. For instance, point-load results can act as a proxy for rock strength, as our guide to predicting UCS from point load testing explains.
Choose a path for geometallurgical proxy models
There are two realistic ways to build the model. Cancha, developed by Transmin specifically for geometallurgy, packages sample selection, prediction modelling and reporting into one integrated tool. Geologists and metallurgists without a data-science background can get a working proxy model without writing code.
Alternatively, build it yourself with standard Python tooling. Use scikit-learn for simpler regression, or gradient-boosted approaches like XGBoost or LightGBM for more complex, non-linear relationships. That gives you more control, at the cost of needing in-house ML capability.
Validate against held-out samples
Whichever path you take, validate against a held-out set before trusting the model operationally. Also check performance within each ore domain, not only the overall score. A model that looks strong on average can still fail badly on one ore type.
Where Proxy Models Go Wrong
The real constraint is rarely the modelling technique. Instead, it is dataset size and quality. You need enough paired examples spanning the range of ore types you will meet. If historical testwork was sparse or biased toward certain domains, the proxy model will quietly fail outside that range. Therefore, plan new testwork to fill the gaps the model reveals.
Proxy models also sit at the centre of the geomet chain. Recovery and hardness prediction is a specific proxy-modelling task. Meanwhile, clay mineral prediction produces one of the most useful proxy inputs.
FAQs
What is an example of geometallurgical proxy models in practice?
A common example predicts ore hardness from point-load tests, assays and logged rock type instead of running a full SMC test on every sample. Another predicts flotation recovery from mineralogy and assays. In both cases, a cheap measurement stands in for an expensive one.
How many samples do geometallurgical proxy models need?
There is no fixed number. What matters is that the paired samples cover every ore type you plan to predict. A small dataset that spans the deposit often beats a large one clustered in a single domain.
Should I use Cancha or build proxy models in Python?
Cancha suits teams that want an integrated, no-code tool built for geometallurgy. Python with scikit-learn or XGBoost suits teams with in-house data-science skills who want full control. Both approaches rest on the same regression principles.
How do I know if a proxy model is reliable?
Test it on held-out samples it did not train on, and check the error in each ore domain. If performance drops sharply in one domain, collect more paired data there. Revalidate whenever the mine moves into new ore.
Try It With Geocluster
Choosing between a turnkey tool like Cancha and a custom pipeline is a research question in its own right. So is working out whether your existing testwork dataset is big enough to train on. Geocluster can help you work through both with sourced evidence.