XploraGeo

How it works

A blind test for mineral prospectivity mapping.

Every round is a made-up district with gold hidden in it. This page explains how a district is built, what you get, how entries are scored and what keeps the test fair.

How a district is built

A district covers about 20 by 20 kilometres, and it is built in the order the ground was.

  1. Geology

    A 3D history of a greenstone belt: volcanic and sedimentary rocks laid down, folded, cut by a major shear zone and its branches, intruded by granite and dolerite, then weathered and buried under transported cover. The structural geology follows the rules a mapping geologist would expect.

  2. Gold

    Orogenic gold deposits go where the published rules for that deposit style put them: near the branches of a major shear, at bends and jogs, in rock that fractures rather than flows. Sizes vary the way real deposits do, with a few large ones and many small. Some places look right and hold nothing.

  3. Rock properties

    Each rock type gets a density and a magnetic susceptibility from measured ranges, not textbook guesses. Alteration around the gold destroys magnetite, which leaves a subtle low in the magnetics.

  4. The surveys

    The data are simulated from the rocks, with the noise, gaps and levelling errors real surveys have.

  5. Review

    Geologists try to break every district before it goes out. If it would not fool them, it does not go out.

What every entrant gets

The same data pack for everyone: GeoTIFF grids, CSV tables, GeoJSON outlines and a README with the coordinate system and units.

LayerWhat it holdsBackground
Airborne magneticsTotal magnetic intensity along flight lines and as a gridMagnetic surveys
GravityGravity stations and a gridded anomalyGravity surveys
Soil geochemistryGold and pathfinder elements on a sample grid, smeared downslope and muted under coverAssay QA/QC
Outcrop mapGeology where it crops out; the rest is coverGeological mapping
Old drillingCollars, depths and assays from shallow historical holes. A few hit, most miss, some were logged badlyDiamond and RC drilling

How entries are scored

Send either a score for every cell of the district or up to 50 ranked targets, each with a confidence between 0 and 1.

The headline score is the share of the gold, weighted by deposit size, that falls inside your top 5% of ground. Put every deposit in your top 5% and you score 100%. Ranking the ground at random scores about 5%.

Alongside it the board shows the capture curve (how the share grows as you take in more ground), the hit rate of your top targets, and calibration: whether targets you rate at 70% hit about 70% of the time.

Baselines sit on the board so the first bar is clear: random ground, ground nearest the old holes that hit gold, and a simple knowledge-driven model built from the published deposit rules.

Keeping it fair

The rules

What came before: the OZ Minerals Explorer Challenge and other blind tests

Testing exploration methods fairly is an old problem, and machine learning in mineral exploration has made it more pressing.

In 2019 OZ Minerals opened its Mount Woods data in South Australia to the world in the Explorer Challenge, run with Unearthed. More than 1,000 people from 62 countries registered for a prize pool of A$1 million. The targets most entrants agreed on were then tested with five holes totalling 2,503.1 m, and OZ Minerals reported that no significant mineralisation was intersected. What made the challenge worth remembering is that the targets were named before anyone drilled, and the results were published even though they were a miss.

Others have tried the same idea elsewhere. In 2006 the Centre for Computational Geostatistics at the University of Alberta proposed standard test cases so geostatistical software could be checked against known results. The FORCE 2020 lithology competition scored predictions from Norwegian well logs on blind wells, and scores fell on the blind wells, which is the gap a blind test exists to show.

Real ground has one weakness as a test for mineral prospectivity mapping: the later discoveries are public, so a model may already have learned them. A made-up district has no history to learn from. Every round gives the field a fresh, blind test, a result in weeks rather than years, and nothing to hide behind.

Take part

Round news goes out through XploraHub's newsletters. Companies, consultants, research groups and students can register a team now.