A comprehensive analysis of autoencoder variants for identification of multivariate geochemical anomalies linked to hydrothermal copper mineralization in Feizabad district, NE Iran

Geochemical anomaly detection plays a pivotal role in mineral exploration at various scales. This process necessitates the integration of a conceptual model of mineral deposit type sought, alongside the utilization of data-driven methodologies to identify subtle anomalies within intricate multivariate geochemical datasets. Autoencoders (AEs), as unsupervised neural networks and reconstruction based anomaly detection algorithms, are suitable for this purpose. Several different AE variants can be used for geochemical anomaly detection, that each can potentially lead to the recognition of different anomalous patterns, complicating the selection of a singular best variant. In the present work, we implement and evaluate four AE variants, i.e., AE, sparse AE (SAE), variational AE (VAE), and convolutional AE (CAE), to compare their effectiveness in detecting geochemical anomalies in the Feizabad region, NE Iran. Our analysis, based on prediction-area (P-A) plots, indicates that the AE outperforms the others with a normalized density index score of 2.85, while SAE, VAE, and CAE scored 2.57. Interestingly, although VAE scored lower than AE, it provided more accurate and meaningful spatial zoning than its peers, even surpassing CAE, which is specifically designed to capture spatial patterns. These findings highlight that an improved model does not necessarily ensure superior perfoemance, highlighting the critical nature of comparative analysis in this field.