A study conducted by SRK Peru evaluated the application of several machine learning models for mineral prospectivity mapping of copper porphyry–skarn deposits in the Andahuaylas–Yauri metallogenic belt of southern Peru. Public geoscientific datasets, such as those from the Instituto Geológico Minero y Metalúrgico (INGEMMET), were integrated to construct evidential maps representing lithogeochemical, lithological, and structural factors relevant to the formation of the target deposit types. Four models — Random Forest (RF), Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Graph Convolutional Network (GCN) — were compared to benchmark their performance in identifying areas with higher probabilities of mineral occurrence.
All models were assessed using spatial cross-validation based on geographic clustering to mitigate bias, given the limited and imbalanced number of positive mineral occurrence data points. The primary metric was the area under the ROC curve (AUC), with average results of 0.89 (RF), 0.912 (MLP), 0.901 (CNN), and 0.907 (GCN). Although the Multilayer Perceptron achieved the highest Area Under the Curve, the Graph Convolutional Network produced results with greater spatial and geological coherence, assigning high probabilities to geologically favorable areas and delineating new zones of potential interest consistent with known mineralizing processes.
Furthermore, the prospectivity maps generated by the GCN showed a spatial logic more consistent with mineral systems models through their ability to incorporate spatial relationships within the data using graph structures. Unlike alternative approaches, which tended to overfit or failed to reflect significant geological variations, graph-based models offer distinct advantages for integrating complex spatial information in mineral exploration scenarios.
In conclusion, geological data quality and processing remain critical to the success of MPM, but artificial intelligence algorithms such as GCNs represent a significant advance in delineating and prioritizing exploration areas. These methods can effectively integrate both public and private data sources for mineral targeting.
