Abstract: Geochemical exploration is an important method for guiding mineral prospecting by identifying geochemical anomalies. However, due to the high-dimensional and nonlinear nature of geochemical data, traditional data processing methods are inadequate for extracting meaningful weak and subtle anomalies in complex geological environments. Machine learning techniques have been widely applied in the field of mineral exploration, achieving breakthroughs in extracting geochemical anomaly information. Taking the Ashale-Halegai area in Western Tianshan as the study area, this paper systematically compares the performance of three deep learning models—Variational Autoencoder (VAE), Sparse Autoencoder (SAE), and Deep Autoencoder (DAE)—in identifying copper-related geochemical anomalies. First, R-mode factor analysis is used to determine the element assemblage closely associated with copper mineralization. Then, VAE, SAE, and DAE networks are constructed, and reconstruction errors are used as anomaly indicators for anomaly identification. Finally, by comparing the performance of the three models, the applicability of different autoencoder models in the study area is discussed. The results show that the VAE model performs best in extracting weak and subtle anomalies, characterizing spatial patterns of anomalies, and achieving consistency with known mineral deposits, providing effective methodological support for deep and peripheral prospecting of copper polymetallic deposits in the Western Tianshan region. Based on the prediction results of the VAE model and considering the metallogenic geological conditions, eight anomalous areas with good prospecting potential are delineated and evaluated. The results of this study provide a reference for using autoencoder networks to extract geochemical anomalies and offer insights for copper exploration in the study area.