Key Takeaways

  • A proxy model is 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 vs. custom-built.
  • This is a genuinely mature “Yes” — nothing speculative about it, this is standard applied ML with published mining case studies.
  • The hard part isn’t the modelling, it’s getting a big enough paired dataset of cheap proxy measurements and expensive ground-truth assays to train on.

TL;DR

Train a regression model (Cancha for a turnkey path, or scikit-learn/XGBoost if you want to build it yourself) on paired examples of a cheap/fast measurement and the expensive assay it’s meant to predict, then use the model to estimate the expensive variable everywhere you only have the cheap one.

How Do I Build Proxy Models With AI?

Proxy modelling is one of the more approachable ML use cases in this entire poster because the problem is exactly the textbook supervised-regression setup: you have a variable that’s expensive or slow to measure directly (say, flotation recovery from a locked-cycle test, or comminution hardness from a full SMC test) and you want to predict it from something cheap and fast (XRF assays, point-load index, hyperspectral clay content, blasthole MWD logs). If you have enough paired examples where you measured both the cheap proxy and the expensive ground truth, that’s a training set.

There are two realistic ways to build this. Cancha, built specifically for geometallurgy by Transmin, packages sample selection, prediction modelling, and reporting into one integrated tool aimed at exactly this problem — geologists and metallurgists without a data-science background can get a working proxy model without writing code. The alternative is building it yourself with standard Python ML tooling — scikit-learn for simpler regression, or gradient-boosted approaches like XGBoost or LightGBM for more complex, nonlinear relationships between your proxy variables and target — which gives you more control at the cost of needing in-house ML capability.

Whichever path you take, the actual constraint is rarely the modelling technique — it’s dataset size and quality. You need enough paired (proxy, ground-truth) examples spanning the range of ore types you’ll encounter, and if your historical testwork was sparse or biased toward certain ore domains, your proxy model will quietly fail outside that range. Validate against a held-out set before trusting it operationally.

Try It With Geocluster

Deciding between a turnkey tool like Cancha and a custom-built pipeline — and figuring out whether your existing testwork dataset is actually big enough to train on — is exactly the kind of research question Geocluster can help you work through.