Comparative evaluation of random forest and graph neural networks for gold prospectivity in the Hatu–Sartohay Belt, China

• Built a multi-source geoscience database for gold targeting. • Systematically compared RF, GCN and Deep GAT models. • Graph models improved spatial coherence and accuracy. • Identified three structurally controlled exploration targets. The Hatu gold district in the Western Junggar region, China, located in the southern Central Asian Orogenic Belt (CAOB), represents a structurally complex and metallogenically significant segment of one of the world’s largest accretionary orogens. Predicting concealed orebodies and deep mineralization remains challenging due to the multiple scales of lithological, structural, and magmatic-hydrothermal factors. To address this issue, we constructed a high-resolution multi-source geoscience database comprising 20,796 geochemical samples, geological maps, structura1 anomalies, and 128 known gold deposits and occurrences across an area of 2872 km2. Three models random forest (RF), graph convolutional network (GCN), and deep graph attention network (DeepGAT) – were compared for mineral prospectivity modeling. RF, trained on 46 feature-selected variables, provided a solid baseline (area under the receiver operating characteristic curve, AUC = 0.9443; macro-averaged F1-score = 0.83) but limited recall. GCN improved performance (AUC = 0.9804, macro-F1 = 0.9024) by capturing spatial dependencies and generating fault-controlled belts. DeepGAT achieved the highest accuracy (AUC = 0.9882, macro-F1 = 0.9236), delineating high-confidence anomalies consistent with lithological and structural features. The predictive maps identified three metallogenic belts: (Ⅰ) Bieluagaxi trend, along the Bieluagaxi pluton; (Ⅱ) Hatu–Gezigou zone, along the Anqi fault; and (Ⅲ) Sartohay–Gemengou belt, in the hanging wall of the Darbut fault. The results demonstrate that graph-based learning in a geological context not only enhances predictive accuracy but also improves geological plausibility, offering a robust framework for exploration targeting in structurally complex orogenic gold systems