─ 004Case Study

Case Study: Coe Fairbairn

A blind test of AI mineral discovery

Given only exploration data from before a discovery was made, could our models find the deposit anyway?

Eigenform

The test

We wanted to test our models’ ability to find undiscovered mineral deposits without spending the millions required for a real-world drilling campaign. If this were a trading model we could backtest on real data. If it were a cybersecurity model we could have it search for known exploits. Mining is different. Known deposits are very well covered in the training data, so it is hard to judge whether a model can rediscover them from data alone; and a negative result never means nothing is there, only that nothing has been found.

Our test case was the Cue Victory goldfield in Western Australia. It has been mined for gold since the late 19th century, so there is a wealth of publicly available data. In 2021 significant rare earth deposits were also found there. That combination let us build a clean test: take only the data from before the rare earth discovery, and see whether the models could point to where those deposits would turn out to be.

To stop the models using hindsight from their training data, we replaced well-known place names and mining companies with pseudonyms and swapped tenement IDs for placeholders. “Cue Victory” became “Coe Fairbairn.”

Stage 1: Data preparation

First we used our own suite of proprietary map-processing tools to convert scanned PDFs into machine-readable map files and tables, chunking and indexing everything. This is not simply digitising images. The system has to deal with broken lines, faded shading and multilingual legends, and turn them into polygons where every edge and vertex actually connects.

The same area vectorised into clean, connected lithology polygons with crisp boundaries.
A scanned geological survey map with weathered colouring, broken lines and faded shading.
Digitisation: drag to compare the scanned map with the machine-readable polygons extracted from it.

Stage 2: Inductive world modelling

The model then searched the data, writing code to identify and confirm relationships between features. Most geological AIs stop here. Ours goes further, using a mineral-systems and information-geometry approach to build the most informative single model it can. If all the features in a region emerged from the same geological processes, then learning more about one feature carries some information — however small — about the others, which is what turns scattered datapoints into a coherent mineral-systems thesis.

Workflow diagram: geological maps, drill logs and geochemistry feed a hypothesis-test-evaluate loop that builds a voxel-based probabilistic belief model, a knowledge graph and a training signal.
From scattered geological data to a single probabilistic world model.

Stage 3: Where to dig

The inductive approach to data taken by the models allowed them to compose probability layers — mineral pathfinders, interesting fault lines, even neighbouring mines’ behaviour — into prospectivity heat maps and physics-aligned coherent structural subsurface projections.

Satellite view of the Coe Fairbairn tenements with a gold-and-magenta prospectivity heat map overlaid, beside a panel of twenty-three weighted feature layers.
Composed probability layers become a prospectivity heat map over the tenements.
Interactive 3D structural cross-section of the subsurface with stratigraphic layers colour-coded, and a gravity profile plotted along the survey line beneath it.
A physics-aligned structural projection of what lies beneath the survey line.

Stage 4: Learning to discover more

The reasoning traces from successful experiments can be turned into focused training examples for the agent. The model is rewarded for reaching a correct final answer, but it is also taught the intermediate skills of geological feature engineering. In early rounds we have seen a 15–17% improvement on our internal geological baselines.

Model-lineage graph of successive training generations, with a layer inspector showing a hydrothermal gold prospectivity feature and its scores, and a sliding-window fine-tuning schematic.
Reasoning traces from successful experiments become training examples for the next generation.

Result: a new old discovery at Cue

Given only gold-exploration data from before the 2021 rare earth discovery, the models still pinpointed where in the region rare earths would eventually be found. The system reported it honestly — as a testable hypothesis that would need field validation, not a settled conclusion — which is exactly the register we want.

An AI agent's summary table of rare-earth targets by tenement, each rated primary, secondary or tertiary confidence, with a verdict flagging the prediction as a hypothesis that still needs field validation.
The model's own verdict: a testable hypothesis, not a data-supported conclusion.

Work with us

What might be lurking unnoticed on your ground?

If you hold exploration data that has never been worked through properly, we would like to hear about it. Our technical contribution can form part of an earn-in or joint venture rather than a conventional software engagement.

alex@send.eigenform.ai