Mineral prediction research using the Geology-based Cyclic Adversarial Network for Anomaly Detection

Abstract

Deep learning algorithms represented by generative adversarial networks (GANs) have been extensively applied in mineral prospectivity mapping research. However, the inherent instability and “black-box” nature of these models have constrained further advancement in such studies. This research proposes an anomaly detection model integrating geochemical data with geological knowledge (Geology-based Cyclic Adversarial Network for Anomaly Detection, G-CANomaly). First, taking the Mila Mountain area of Tibet as the study area, we quantified the relationship between deposit density and ore-controlling factors (fault structures) using multifractal singularity theory. Second, we restructured the network architecture of traditional GANomaly models and embedded geological knowledge to establish the G-CANomaly anomaly detection model, achieving a transition from purely data-driven approaches to hybrid “data + knowledge”-driven methodologies. Finally, we conducted comparative analyses of model performance among GANomaly, CANomaly, and G-CANomaly. The results demonstrate that: (1) The anomaly zones delineated by G-CANomaly exhibit the highest spatial correlation with known mineral occurrences in the study area. (2) The top 3 % anomaly areas can encompass 80 % of known mineral deposits. (3) The main assessment metrics of G-CANomaly, AUC, Accuracy, and Recall, were 0.97, 0.89, and 0.93, showing 3–6 % improvements over baseline models. These findings indicate that the G-CANomaly model demonstrates strong feasibility in geochemical anomaly identification research, with notably enhanced training stability and convergence rate. This study provides valuable references for constructing hybrid geochemical-geological knowledge-driven deep learning models and enriches the methodological framework of mineral prospectivity mapping.