Chapter 01
How mining works
Before any geology or any modelling: who these companies are, where the money comes from, and the one industry statistic that explains why anybody is interested in applying machine learning to this at all.
The value chain
Mining is six businesses wearing one coat, and they have almost nothing in common operationally. Knowing which one you are working for is the difference between a useful model and an irrelevant one.
Everything in this primer sits in the first box, with occasional reference to the second. A resource estimate matters because it is the input to a development decision; a drill hole matters because it either does or does not move that estimate.
Who the players are
| Type | What they are | What they need from you |
|---|---|---|
| Majors | BHP, Rio Tinto, Glencore, Vale, Freeport, Anglo American. Producing mines, real revenue, internal exploration teams. | Scale and rigour. Long procurement cycles, security review, integration with an existing stack. |
| Mid-tiers | One to a few producing mines. Cash-generating but not enormous. | Near-mine exploration to extend mine life — the highest-ROI drilling there is. |
| Juniors | The engine of discovery. No revenue at all. A handful of geologists, a project, and a listing. | Anything that stretches a drill budget. Fast to buy, chronically short of cash. |
| Prospect generators | Juniors with a deliberate model: generate targets, then option them to partners who fund the drilling. Sell ideas, not holes. | Target generation is literally their product — the most natural fit for prospectivity modelling. |
| Royalty & streaming | Franco-Nevada, Wheaton, Royal Gold. Buy a slice of future production instead of operating. | Portfolio-scale technical due diligence — reading a lot of other people's data quickly. |
| Service & consulting | Drilling contractors, assay labs (SGS, ALS, Bureau Veritas), geophysical contractors, consultancies (SRK, AMC). | Labs are where your assay data is physically produced — see Chapter 6. |
Majors mostly do not make grassroots discoveries any more; they buy them. Juniors take the discovery risk on retail and institutional equity, and if they succeed they are acquired. This has a direct modelling consequence: the organisation with the most interesting problem — the junior — has the least money and the least data infrastructure, while the organisation with the best data has the least appetite for frontier risk.
Where the money comes from
A junior explorer has no revenue. It has never sold anything and, statistically, never will. Every dollar it spends was raised by selling a piece of itself. That single fact shapes the industry's behaviour more than geology does.
- Equity raises. Issue new shares for cash, on the TSX / TSX-V (Toronto), ASX (Australia), AIM (London) or JSE (Johannesburg). Toronto and Sydney dominate.
- Private placements. The standard mechanism — a block of shares (often with warrants attached) sold to institutions or high-net-worth investors, typically at a discount to market.
- Flow-through shares. A Canadian tax structure letting a company pass exploration-expense deductions to investors. Enormously influential on where and when Canadian exploration money gets spent — including a year-end spending rush driven by tax deadlines rather than geology.
- Earn-in / farm-in / JV. A partner spends an agreed amount over an agreed period to earn a percentage of the project. How prospect generators monetise.
- Royalties and streams. Sell a percentage of future revenue (NSR — net smelter return) for cash now. Non-dilutive, and permanent.
Every raise issues new shares, so existing holders own less of the same company. A junior that drills for eight years without a discovery has diluted its original shareholders into irrelevance even if the share price held. This is why exploration is time-pressured in a way the geology is not: management needs a result that supports the next raise, and it needs it before the treasury runs out. A "walk away" recommendation is expensive to deliver and unpopular to hear — which is exactly why the decision-theoretic framing in Chapter 10 matters commercially and not just intellectually.
The commodity cycle
Exploration budgets track metal prices with a lag, and the amplitude is brutal — global spending can halve in three years. Because the industry's capacity is people and rigs, and both leave during downturns, the cycle produces a characteristic scar in the data: dense drilling in boom years, nothing in bust years, with method and personnel changing across the gaps. When you later find a step change in your assay database at a date boundary, this is usually why.
Two structural forces currently push the other way. Electrification is pulling copper, nickel, lithium and rare earth demand forward hard, and governments have started treating these as critical minerals — a security category, not just an economic one. That is the tailwind under everything in Chapter 11.
The statistic that justifies this entire primer
Here is the industry's central problem, and it is a productivity problem rather than a geology problem.
Richard Schodde's work at MinEx Consulting is the standard reference here, and the trend it documents is stark: through the 2002–2012 exploration boom, spending rose enormously and discovery rates did not follow. By his count there were 41 new gold discoveries containing 215 Moz in the decade to roughly 2018, against 222 discoveries containing 1.7 Boz over the preceding 18 years. Unit discovery costs doubled, driven by input costs — drilling, labour, land access, administration — all roughly doubling in real terms.
Spending is flat, unit cost is up, and capital has retreated to near-mine drilling — the 45% figure — precisely when the world needs new copper. The industry is not failing to find deposits because it lacks money or rigs. It is failing because the remaining deposits are blind, and the search process is inefficient in a way that money cannot fix. That is an inference problem, and inference problems are tractable. This is the entire commercial thesis for applying decision theory and machine learning here, and it is why the interesting chapters of this primer are the ones about where to look next rather than the ones about prediction accuracy.
Timeframes, and what they do to feedback
Discovery to production runs routinely 10–20 years: several years of exploration, two to four of studies and permitting, two to five of construction. A greenfield copper project can take longer.
For a machine learning practitioner this is the single most uncomfortable fact in the domain. Your ground-truth feedback loop is longer than your tenure. You cannot A/B test target generation. You will not observe whether the ground you deprioritised contained a mine. Every evaluation you run is a proxy, and being honest about which proxy you are using — and what it fails to capture — is most of the professional skill.
Who is on the team
| Role | What they do |
|---|---|
| Exploration geologist | Maps, samples, generates and tests targets. Usually the person who decides where a hole goes. |
| Project / mine geologist | Runs the drill programme, logs core, manages sampling on an active project. |
| Resource geologist | Builds the block model and the resource estimate. The person whose job Chapter 8 describes. |
| Geophysicist | Designs surveys, processes and inverts the data. Your closest natural colleague — already thinking in forward models and inverse problems. |
| Geochemist | Sampling media and methods, QA/QC design, interpretation of element associations. |
| Database manager | Owns the drillhole database. Vastly more important than the title suggests, and usually the only person who knows where the bodies are buried. |
| Qualified Person (QP) | Signs off publicly disclosed technical information and carries professional liability for it. |
You will meet resistance to new methods. It is worth understanding as something other than stubbornness. Decisions are capital-intensive and irreversible; a QP carries personal professional liability for disclosed numbers; the feedback loop is a decade long, so nobody can quickly demonstrate that a new method works; and the industry has been sold a great deal of technology that did not. A method that is auditable, explicable and slightly worse will beat a method that is opaque and slightly better — not because geologists are innumerate, but because they are the ones who have to defend it. Build accordingly.
Sources for this chapter
- S&P Global Market Intelligence — World Exploration Trends and its annual Corporate Exploration Strategies report. The authoritative source for global exploration budgets, commodity splits and the grassroots-vs-mine-site breakdown. Summary coverage of the 2025 numbers: MINING.COM.
- MinEx Consulting publications (Richard Schodde) — the reference dataset on long-run discovery rates, discovery costs and exploration productivity. Nearly every credible statistic about whether exploration is getting harder traces back here. Start with the exploring-under-cover presentation, which is the clearest statement of why the remaining deposits are blind.
- CIM Magazine — "Innovation to improve the exploration odds". Readable summary of the discovery-cost trend and the industry's response.
- PDAC (Toronto, every March) and Mines and Money — the conferences where juniors raise money. Worth attending once early; you will learn more about incentives in two days than from a year of reading.
- Society of Economic Geologists (SEG) and its journal Economic Geology — the professional body for ore deposit science, and the venue where deposit models get argued about.