Explore the geology
Switch between depth slices. Inspect the fictional domains, routine assays and which intervals already have BBWi.
INTERACTIVE CHALLENGE / FICTIONAL GEOLOGY
Beat the Sampler.
What would you test?
You have an invented iron-ore deposit, plenty of geological observations and a limited testwork budget. Can you pick a more informative collection of core specimens than a sampling algorithm?
Choose up to three Bond Ball Mill Work Index (BBWi) tests, then compare your programme with Random, Representative and Information Gain approaches. The interesting part is why the choices differ, not just who “wins”.
Free-to-explore, fictional public beta hosted on Posit Connect Cloud. No account or actual mine-site data is needed for the challenge.
Previously drilled16 holes
Fictional geological domains and assays
Existing BBWi36 results
Already returned laboratory observations
Untested core12 intervals
BBWi genuinely missing in this game
Fictional budgetAUD 4,500
Up to three additional tests
ONE SMALL EXPERIMENT
Explore. Select. Compare.
Make a testwork basket
Select one to three currently untested specimens. Watch the hypothetical laboratory quote accumulate against the fixed budget.
Compare expected information
Lock your basket and compare your choices with five reproducible sampling baselines under the same model and budget cap.
Locking a basket scores possible tests; it does not purchase material or reveal any missing laboratory result. A sampled location's expected information depends on the fitted model and the targets you want to predict.
THE MATHEMATICAL IDEA
The most unusual core isn't always the best addition to the batch.
A geologically interesting specimen and an informative testwork programme are not necessarily the same thing.
GEOLOGICAL COVERAGE
Where are the gaps?
Different domains, spatial settings and assay compositions may need representation. Selecting more samples from the same geological neighbourhood can also create redundancy.
BAYESIAN INFORMATION GAIN
Where might we learn most?
A Gaussian-process model estimates how a proposed test could reduce uncertainty at specified prediction targets. After adding one candidate to the batch, another may contribute less because their information overlaps.
In the frozen challenge, an extremely SiO₂-rich specimen can rank highly as an individual information-gain candidate yet be omitted from the cost-aware combination. That is not a contradiction: the batch also depends on its costs, alternative specimens and information overlap.
The score is model-expected learning, not realised prediction error. One seeded Random comparison does not establish that IG is generally superior.
FROM AN INTERACTIVE CHALLENGE TO A PLANNING WORKFLOW
What if you also had to verify the physical core?
Beat the Sampler leaves out some important operational decisions. Does the material still exist? Is enough core available for the laboratory method? How should geological coverage be planned before using information-theoretic acquisitions?
The separate Geometallurgical Testwork Optimiser demonstrates physical specimen eligibility, budgeted coverage planning and a fictional sequential laboratory mode where selected results are returned before the model updates.
Research demonstration, not operational advice. All geological data and quoted costs are fictional. The 12 untested BBWi responses are genuinely missing and are never invented or revealed. Expected uncertainty reduction does not prove improved holdout RMSE, calibrated predictive uncertainty, economic EVSI or actual mine-site savings.
Ready to choose your samples?
Start with your geological intuition. See what the Sampler chooses, then explore the mathematics behind the difference.