The problem
Each dataset gives a partial view. A soil anomaly says metal reached the surface, not where it came from. A chargeability high says something polarisable sits at depth, not that it is ore. Read one at a time, the layers give a list of anomalies; read together, they give a short list of places where independent evidence agrees.
Doing that by eye, with maps laid over each other, favours whichever layer is most colourful. Doing it explicitly makes the reasoning visible and repeatable.
What you send
Any combination of drilling, surface geochemistry, geophysics, elevation, remote sensing and geological mapping, ideally as one database (see data integration). No layer is compulsory. The method is chosen for the data you have, not the other way round.
What we do
- Turn each dataset into evidence layers tied to the deposit model you are exploring for
- Combine them with a method suited to the data: expert-weighted where known mineralisation is scarce, statistical where there is enough to learn from
- Test the result by hiding known data and checking whether the method finds it again
- Report which layer drives each high-ranking area, so the map can be argued with
- Once a zone is drilled, model it under alternative geological interpretations and state the outcome as ranges
What you get
Before drilling: a prospectivity map ranking the areas most likely to host mineralisation, showing which data layer drives each one.
After 18 holes in the northern zone, for example: “contained copper between 35 and 90 kt (low and high cases), under two alternative geological interpretations”, with the holes that would narrow the range most.
What it cannot say
A prospectivity score is a ranking, not a probability of discovery. It says area A has more supporting evidence than area B under the stated model; it does not say there is a deposit in either. Ranges after drilling are conceptual and are not mineral resource estimates under NI 43-101 or JORC.
Questions
Which data layers are required?
None in particular. The analysis uses whichever layers exist and reports how much each contributes. A missing layer widens the uncertainty; it does not stop the work.
Is this machine learning?
Sometimes. Statistical and machine-learning methods need known examples to learn from, and early-stage districts often have too few. In that case an expert-weighted method is more honest. We say which was used and why.
How do you know the ranking is any good?
By testing it. Known occurrences or drill results are withheld, the analysis is rerun, and we check whether the withheld data falls in highly ranked ground. The test result is part of the report.