Identification of Geochemical Anomalies Using a Memory-Augmented Autoencoder Model with Geological Constraint

Abstract

The identification and mapping of geochemical anomaly patterns have emerged as a more precise and efficient approach for mineral exploration, with deep learning algorithms being extensively employed in this realm. However, existing methodologies require further investigation regarding model interpretability and correlation with established mineral control factors. This paper proposes a regional geochemical anomaly identification method based on the memory-augmented autoencoder (MemAE), incorporating geological controlling factors. Firstly, the MemAE model is introduced to address the excessive generalization capability of the traditional autoencoder (AE) model. Secondly, utilizing multifractal singularity theory, a nonlinear functional relationship between faults and mineral deposits is established. This relationship reveals the controlling effect of faults on mineralization and it is incorporated as a constraint term in the MemAE's loss function. Finally, the constructed geochemical anomaly identification model is employed to delineate prospective mineralization areas, with comparative studies conducted on AE, MemAE, and geologically constrained MemAE models. The results demonstrate that the geologically constrained MemAE exhibits superior performance, achieving an AUC of 0.802. The eight delineated mineralization prospective areas show strong concordance with actual distributions. The proposed method, which considers geological controlling factors, effectively enhances model interpretability and demonstrates excellent geochemical anomaly identification capabilities. Consequently, this approach can be considered a viable methodology for mineral exploration.