Aethermine Primer · v2

Machine learning for mineral exploration

You already know how to fit models. This primer is about the part nobody teaches: what happens when your training set has eleven positive labels, your negatives are unverified, each new observation costs a quarter of a million dollars, and the thing you are predicting is buried a kilometre underground.

Mineral exploration looks, from the outside, like a classification problem — a map, some features, and a binary deposit / no deposit target. That framing survives about a week of contact with the industry. What it really is: sequential Bayesian experimental design under a prior that is probably wrong, executed against an irreversible capital budget, on data assembled from eighty years of inconsistent record-keeping.

That makes it one of the most interesting applied ML problems available. Almost every hard idea in modern machine learning shows up here in its most consequential form: label scarcity, covariate shift, spatial leakage, calibration, active learning, uncertainty quantification, and the difference between a model that is accurate and a model that changes a decision.

Fundamentals first

No prior geology is assumed. The first four chapters are the industry itself — how mining companies are structured and financed, how ore physically forms, how an exploration programme is actually run, and the vocabulary and data schema everything else sits on. Only then does the primer turn to modelling.

That ordering is deliberate rather than polite. Nearly every way a machine learning project fails in this domain is a domain failure wearing a statistical costume: treating unexplored ground as barren, averaging grades without length weighting, validating a spatial model with a random split, modelling a laboratory artifact as geology. You cannot catch any of those without knowing how the rock, the rig and the lab behave.

Those four numbers are the whole business case. Spending is flat, discovery is getting more expensive, and capital has retreated to safe near-mine drilling exactly when the world needs new copper, nickel and lithium. The industry is not short of money or rigs — the remaining deposits are blind, buried under cover with no surface expression, and the search process is inefficient in a way money cannot fix. That is an inference problem. Chapter 1 makes the case properly.

How to read this

Thirteen chapters in four parts. They build, so read in order the first time. Every chapter carries a purple box stating the ML framing explicitly, a copper box giving the geological reality that framing has to survive, and its own annotated source list at the end so you can go deeper on whichever part grabs you.

Part I · The industry

Part II · The data

Part III · Machine learning

Part IV · Practice

Three habits to bring with you

1. Distrust every label

In this domain a "negative" almost always means nobody looked, and a positive often means someone looked hard because it was easy to get to. The label distribution is a map of historical human attention, not of geology. Chapter 9 makes this formal; it is the single most common way ML projects in exploration quietly fail.

2. Ask what decision the number changes

A prospectivity map with an AUC of 0.94 is worth nothing if it ranks the same ground the geologists already had at the top. The unit of value is a changed drill hole. Chapter 10 gives you the formal machinery — value of information — for asking that question properly.

3. Respect the provenance chain

Every grade in a database traces to a physical piece of rock, crushed, split, dissolved and read by an instrument, then transcribed — often more than once. Each step has a failure mode, and the failures are not random. Chapter 6 is the unglamorous one, and it separates people who ship models from people who ship models that get trusted.