AI in Mineral Exploration: A Framework for Honest, Auditable Prospectivity Mapping

Why we wrote it

A non-trivial number of exploration teams have already sat through a vendor demo for an AI prospectivity tool. The vocabulary was familiar (deep learning, neural networks, end-to-end automation), the heat maps were colourful, and the recommendations were unfalsifiable. The opacity of those systems is not a quirk of the technology. It is a design choice that does not survive contact with a technical committee or a regulator.

This whitepaper is our attempt to set out, in public, what AI prospectivity mapping needs to look like in order to be defensible. The intended audience is people who have to defend a drill program in front of a board, not people shopping for a vendor. We wrote it because the gap between the noise around AI in mining and the work that actually holds up under audit is, at the moment, very wide. We would rather close that gap with a document than with a sales call.

What's inside, chapter by chapter

1. The targeting problem. Exploration runs on incomplete information, and the shortage is not data. Decades of geophysics, geochemistry, and drill logs already sit in client archives in formats and coordinate systems that rarely line up. The shortage is the human bandwidth required to bring those layers into coherent relationship at the scale of a district. This chapter frames the financial weight of that gap and the case for a model that can hold more in working relationship than a person can.

2. Why first-generation AI failed in mining. The failures clustered around four issues: opacity, calibration, deposit-type templates applied to settings they did not fit, and a lack of provenance. None of these are failures of machine learning as a category. They are failures of how machine learning was packaged for a regulated, evidence-driven industry that does not tolerate hand-waving.

3. A framework for honest prospectivity mapping. The technical core of the paper. Positive-unlabelled learning instead of binary classification, gradient-boosted ensembles instead of deep networks, spatial cross-validation with geographic blocks instead of random holdouts, Platt scaling for calibrated probabilities, and bootstrap resampling for uncertainty. Each choice is justified against the alternative and against the four failure modes in chapter two.

4. The role of uncertainty in drill planning. A probability map without an uncertainty map is a single number where two are needed. A high-probability, low-uncertainty pixel is a drill target. A high-probability, high-uncertainty pixel is a request for more data. The chapter sets out how to read the two layers together and why epistemic uncertainty deserves first-class status in a deliverable, not a footnote.

5. What it looks like in practice. How DepoDart applies the framework inside a hands-on engagement that typically runs two to four weeks. Data sharing, ingestion and normalisation handled by our team, the model run, a geoscientist review before anything ships, and the four artifacts the client walks away with (prospectivity surface, ranked targets, uncertainty map, provenance report). At this stage every engagement is collaborative by design; the platform is the tool we use to do the work alongside the client's team, not a self-serve product they log into.

6. Limitations and an honest agenda. Where the model does not help. Greenfields settings with no analogues in the training data, deeply buried deposits below the resolution of available geophysics, and any region where validation data is too sparse to be meaningful. The chapter is deliberately in the body of the paper rather than in an appendix.

What you'll learn

• How to read a prospectivity map produced by a serious AI system, including which questions to ask the vendor before the first pilot.
• Why calibration and uncertainty are not optional features but the difference between a probability map and a relative ranking dressed up as one.
• What a defensible engagement deliverable looks like, and what a geologist should be able to do with the four artifacts it produces.

Download

Download the whitepaper (PDF, 19 pages)

The file is free and does not require a form. If you would like to discuss any section of it with the people who wrote it, the contact form on depodart.com is the fastest route.

The whitepaper is published as version 0.1, a draft for technical review. The case-study sections will be expanded with anonymised pilot data once results are releasable, and a versioned update will be posted here at that point.