Scale-Aware Uncertainty-Aware Machine Learning for Gold Prospectivity Mapping: Evidence-Range Audits for Exploration Targeting

Jeffery Opoku , David Banahene

Mineral exploration is expensive because most land has not been tested directly. A missing deposit record is therefore not the same as a true absence record. This paper develops a scale-aware positive-unlabeled learning framework for gold prospectivity mapping using public geological and mineral-occurrence data. The main method is the Non-Gold Evidence Transfer Field (NGETF). NGETF removes every gold- or Au-related record from the critical-mineral context layer, builds multi-scale spatial kernels from the remaining non-gold critical-mineral occurrences, and uses those kernels to rank grid cells by gold prospectivity. On a Nevada pilot grid with 4,392 cells, the conservative NGETF model obtains a quadrant-holdout AUC of 0.774. Cross-state transfer tests train on three western states and forecast a fourth held-out state. We then add public geoscience layers: USGS magnetic and gravity grids, mapped SGMC lithology polygons, and mapped fault traces. With these layers, held-out Arizona reaches AUC 0.890 and held-out Utah reaches AUC 0.926. An independent British Columbia test using the BC Geological Survey MINFILE catalogue gives base AUC 0.958. We then introduce a target-conditioned evidence-range audit: after removing non-gold context records near gold labels, AUC falls to 0.739 with a 25 km buffer and 0.616 with a 50 km buffer, with 50% of above-chance discrimination retained to about 27 km. This audit shows that the method is most useful as a local mineral-system prioritization tool, not as a stand-alone remote gold detector. It can help choose where to review, sample, or drill next, but it does not prove that gold is present.