The models learn from known deposits — then we prove they generalise by hiding those deposits and checking whether the model still finds them.
Explore the live demos to see it on real ground.
A prospectivity model is only as good as what it learns from, and how honestly it is tested. MineDSS learns from the public geoscience record of the countries it covers — the real deposits that have been found, and the layers of evidence that surround them — and is scored the hard way, on ground it never saw in training.
known occurrences, deposits & mines (AU · USA · Canada)
soil & rock samples in the training foundation
commodities validated across Australia, the USA & Canada
The models are trained on known mineral occurrences, deposits and mine sites — hundreds of thousands of them across Australia, the USA and Canada — as the record of where mineralisation is real. Each is paired with the full stack of evidence at that location: satellite alteration and terrain, airborne and ground geophysics, geochemistry, and mapped geology and rock age. From enough of these real examples, the model learns the multi-layer signature that distinguishes prospective ground from barren ground.
The foundation is public data from the national and state geological surveys — Geoscience Australia, the United States Geological Survey, the Geological Survey of Canada, and the state and provincial surveys — cleaned, harmonised and brought onto common ground across three countries. No proprietary lock-in, and kept up to date as the surveys release new data.
Numbers measured on the same data a model learned from always look impressive and mean almost nothing. So MineDSS hides the known deposits, rebuilds the model without them, and checks whether it finds them anyway — with test areas kept spatially separate from training areas so the model cannot cheat by memorising nearby points. That held-out skill is the score we publish, per commodity and per country. It is model-level skill, never a specific site's measured accuracy, and never a discovery or JORC / NI 43-101 resource claim.
When you bring your own assays or drillholes to sharpen a model for your ground, that data is used for your job only. It is never added to anyone else's model and never leaves your workspace's control — the national training foundation and your private data are kept strictly apart.
The published held-out scores, per commodity and per country, are on the commodities pages — and you can see the same honest test summarised on the home page.
Known mineral occurrences, deposits and mine sites across Australia, the USA and Canada — the public record of where mineralisation is real — each paired with the satellite, geophysical, geochemical and geological evidence at that location. The model learns the multi-layer signature that separates prospective from barren ground.
Public, authoritative geoscience: Geoscience Australia, the United States Geological Survey, the Geological Survey of Canada, and the relevant state and provincial geological surveys — harmonised across the three countries and kept current as new data is released. No proprietary lock-in.
Held-out spatial cross-validation. We hide known deposits, rebuild the model without them, and test whether it still finds them, keeping test ground spatially separate from training ground. The resulting held-out skill is what we publish per commodity and per country — model-level skill, not a specific site's accuracy, and not a discovery claim.
No. Anything you upload is used only for your job, kept apart from the shared national training foundation, and never added to anyone else's model. Your data stays yours.
Draw an area, pick a commodity, and explore the ranked targets and the evidence behind each.
Explore the live demos to see it on real ground.