Botswana 路 Investor portal
Methodology
How the targets are made and checked: what the model had to work with, the eight-step workflow, the published score weights, and validation you can re-run.
Share of the country a competent person can set aside for a commodity.
Share of known deposits inside the top-ranked area. 80 percent in 12 percent is capture efficiency, not a discovery claim.
Every score carries its data completeness. Sparse labels widen bounds, never raise scores.
- 1
Mineral-system definition
Competent persons define source, transport, trap and preservation criteria per commodity before any modelling.
- 2
Feature engineering
Each criterion becomes a mappable layer: structural proximity, alteration indices, geophysical responses, geochemical pathfinders, co-registered on a declared grid.
- 3
Label assembly
Known occurrences and deposits with confidence filtering. Where labels are scarce, semi-supervised anomaly detection is used and declared.
- 4
Physics-informed model
Physical constraints are embedded so predictions cannot contradict established geology, for example the redox boundary for sediment-hosted copper.
- 5
Spatial block cross-validation
Training uses geographic blocks, never random splits, so the model must generalise to ground it has not seen.
- 6
Blind-site testing
A set of known deposits is withheld. The model is not trusted on unknown ground until it finds them.
- 7
Ranking
A transparent weighted-composite score, published weights, competent-person adjudication, ROPO-aligned classification.
- 8
Explainability
Feature importance, confidence surface, limitations and a geological hypothesis report on every target.