A metric-learning enhanced variational autoencoder for unsupervised detection of geochemical anomalies in the Hatu gold belt, Xinjiang, China

• Metric-enhanced variational autoencoder detects multivariate geochemical anomalies. • The model improves anomaly separation and identifies most known gold mineralization. • Detected anomalies align with Au-As-Sb patterns and major faults in the Hatu gold belt. Geochemical anomaly detection is a key technique for identifying concealed ore bodies and assessing mineral potential through the analysis of spatial patterns in elemental distributions. However, traditional methods are often constrained by the high-dimensional and nonlinear nature of geochemical data, as well as by their reliance on labeled samples. To address these challenges, this study develops a novel metric-learning enhanced variational autoencoder (MeVAE) model that integrates generative modeling with discriminative metric learning for unsupervised multivariate geochemical anomaly detection under complex geological backgrounds. The MeVAE model employs a variational autoencoder (VAE) to decouple nonlinear correlations in high-dimensional data while dynamically optimizing the Mahalanobis distance matrix to enhance the discriminative capacity of the latent space for capturing geochemical anomalies. Application of the MeVAE model on 1:50,000-scale geochemical data from the Hatu gold belt (Xinjiang, China) demonstrates its effectiveness in detecting anomalies, as evidenced by an area under the curve (AUC) of 0.88 and a lift index of 5.4. The delineated anomalous zones cover 17.4 % of the study area and include 88.5 % of known gold deposits. These anomalies closely align with regional fault structures and key metallogenic element assemblages (Au-As-Sb). Compared to baseline models, including standard VAE and the denoising autoencoder (DAE), MeVAE exhibits improved performance in both detection accuracy and spatial delineation. This study provides an effective unsupervised solution for geochemical anomaly detection and automatic anomaly extraction in geologically complex regions like the Hatu gold belt, with promising applications in mineral prospectivity mapping.