Bayesian Deep Learning Workflow Maps 3D Copper Mineralization at Kogodai Prospect from Sparse Drilling and Geophysical Data

Researchers desurveyed historical drillholes and trenches into a unified 3D frame, composited assays to common support, and trained Bayesian neural networks to predict continuous fields of copper grade, chargeability, and apparent resistivity. Monte Carlo dropout quantified epistemic uncertainty, treating the task as spatial field learning rather than pointwise regression. The resulting maps highlight a principal mineralized corridor and secondary targets aligned with induced polarization anomalies.

The approach directly addresses brownfield exploration challenges where dense drilling is costly and geophysical inversions remain ambiguous. By flagging high-uncertainty zones, the model prioritizes infill drilling locations that can most efficiently reduce risk. This integrates AI with established geophysical workflows to accelerate time-to-discovery while lowering environmental footprint from unnecessary drill pads.

Key limitations include incomplete provenance metadata for legacy geophysical products and inconsistent historical sampling protocols. Stronger evidence would require a controlled field validation program comparing model-guided versus conventional targeting success rates over multiple prospects. The study remains a preprint without peer review.

Future work could embed these predictions into real-time drill planning dashboards, testing whether uncertainty thresholds improve hit rates above baseline exploration statistics.