Overcoming the isotropic bias in mineral prospectivity mapping: a direction-aware neural network approach for sandstone-hosted uranium

Mineral prospectivity mapping for sandstone-hosted uranium deposits commonly relies on isotropic distance decay, despite the fact that ore-forming fluids are typically guided by faults and paleochannel systems with strong directional organization. As a result, structurally controlled geological signals may be diffused into quasi-circular anomalies, weakening the geological realism of prospectivity predictions. To overcome this limitation, we propose a Direction-Aware Geographically Weighted Neural Network (DA-GWNN), which embeds expert-defined structural orientation priors into neighborhood weighting to preferentially capture spatial influence along plausible ore-forming directions. In addition, a criteria-based consensus soft-labeling scheme is developed at the grid-cell scale to represent different levels of mineralization evidence under sparse supervision. The method was tested in the Kailu Basin against six benchmark models under nested spatial blocked cross-validation. DA-GWNN achieved the best predictive performance, with a mean AUC of 0.912 ± 0.019 and a mean AUPRC of 0.794 ± 0.026. SHAP analysis indicates that the oxidation–reduction transition zone is the principal control on prospectivity, whereas the role of fault proximity is conditional and varies with the local redox buffering capacity reflected by the gray-sand ratio. Area-constrained analysis shows that the 36.6% P–A envelope captures 78.2% of all positive-labeled cells and represents a basin-scale mineralization-system footprint. Under the stricter economic-positive definition (y = 1.0), DA-GWNN captures 45.0%, 67.5%, and 80.0% of economic positive cells within the Top 5%, Top 10%, and Top 15% cumulative area budgets. These results indicate that DA-GWNN supports a hierarchical exploration workflow from regional system recognition to focused target verification.