Our vision
Building the foundation models for minerals.
A foundation model learns the ground once, from everything measured on it, and can then be asked many questions: where a metal is likely, what a drill core holds, what a satellite scene shows. That is what we are building toward, one tested step at a time.
Data published by 80 organisations ·Names and logos belong to their owners; no endorsement or partnership is implied. All 80 sources
Why the data comes first
A model can only learn what has been measured.
Most of the world's geoscience is already published, but it is scattered across agencies, formats and decades. So the first job is the record itself: gathered from the authorities that publish it, reconciled onto one map, and counted.
- 18.2M
- Geoscience features
- Geology, structure and deposits across 40 countries
- 8.7M
- Tenure records
- From the registers of 48 countries
- 489.3M
- Drill-core spectra
- Raw spectra from 6,546 scanned holes
- 28
- Country atlases
- Each built the same way, so any question can be asked of any of them
How we are moving toward it
From the record to one model.
- 01
The record
Built, and growingGeology, structure, deposits, tenure, geochemistry and drill-core scans, gathered from the authorities that publish them and reconciled onto one map.
The data - 02
Models for one question
In serviceMachine-learning prospectivity for 51 minerals, trained on 13.8M samples: national models in Australia, the United States and Canada, and a global model for 44 minerals everywhere else. Each is tested on ground it never saw before it is used. Next, they are retrained on the full record and new sources, such as hyperspectral imagery.
Prospectivity - 03
Models for one project
At project scaleField data from the ground itself, integrated with the national evidence into one model for the project, and into three dimensions once holes are drilled. Where the data cannot support a reliable model, we say so.
Prospectivity - 04
Reading the drill core
Next489.3M raw spectra from 6,546 scanned holes, 2,746 of them matched to laboratory assays. Next, we will train a model to read mineralogy and grade from the spectrum.
- 05
One model across the record
The goalA model that learns from the whole record at once (geology, geochemistry, geophysics, satellite imagery and drill-core spectra), so each new question starts from what it already knows, not from nothing. It will be tested on ground held back from training, then against history: trained only on data from before a cut-off date, and checked against the deposits found since.
On the way
What will not change.
Tested before it is used
Every model is scored on ground it never saw. A model that fails is not used, for any mineral, in any country.
Evidence you can check
Every result carries the evidence behind it, and says where that evidence is thin.
Clear about what exists
We say what is in service, what is next and what is still a goal, as this page does.
A partner's data stays theirs
What a partner shares with us is used for that work only, never to train models for anyone else.
-
Partners
Exploration and mining companies interested in working with us.
-
Investors
The record, the models and the ground they point to.
-
Publishers and researchers
Geological surveys, universities and programmes whose work the record is built on.