Scholarlink

<p>Geochemical anomaly mapping in complex porphyry Cu terrains remains challenging due to the high-dimensional, non-linear, and noisy nature of multi-element stream sediment data. Traditional methods (e.g., univariate thresholding, PCA) and even some unsupervised manifold learning techniques often fail to capture subtle, overlapping alteration-related patterns. To address these limitations, we propose a novel chained deep-learning framework, UAGB, which integrates Uniform Manifold Approximation and Projection (UMAP) for topology-preserving dimensionality reduction, an Autoencoder (AE) for deep regularized compression, a Generative Adversarial Network (GAN) for adversarial feature enhancement, and Bisecting K-means for hierarchical hard clustering. The framework was applied to a 15-element geochemical dataset from the Pariz area, a region hosting several world-class porphyry Cu deposits. The UAGB performance was compared against a baseline UMAP-GMM (UG) model and three alternative methods (PCA+K-means, t-SNE+HDBSCAN, ISOMAP+BIRCH). Quantitative evaluation using the prediction–area plot and Normalized Density (<i>N</i><sub><i>d</i></sub>) index shows that UAGB captures 86% of known Cu deposits within only 14% of the study area (<i>N</i><sub><i>d</i></sub> = 6.14), substantially outperforming UG (<i>N</i><sub><i>d</i></sub> = 3.16) and the best alternative method (PCA+K-means, <i>N</i><sub><i>d</i></sub> = 4.55). Ablation studies demonstrate that removing the AE or GAN reduces <i>N</i><sub><i>d</i></sub> by 21–31%, confirming their irreplaceability and synergistic effect. Spatial cross-validation and hold-out deposit tests (standard deviation = 0.12, withheld deposit capture = 86%) rule out overfitting and confirm generalization. The enhanced feature space becomes highly separable (Silhouette = 0.91, intra-cluster distance ratio = 0.21), justifying the use of hard clustering. The UAGB framework provides a powerful, scalable, and interpretable tool for geochemical anomaly mapping, reducing exploration search space and generating testable targets for porphyry Cu systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

UAGB: a novel chained deep learning architecture for geochemical anomaly mapping in porphyry Cu exploration

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Yousef Bahrami,
  • Abbas Maghsoudi

摘要

Geochemical anomaly mapping in complex porphyry Cu terrains remains challenging due to the high-dimensional, non-linear, and noisy nature of multi-element stream sediment data. Traditional methods (e.g., univariate thresholding, PCA) and even some unsupervised manifold learning techniques often fail to capture subtle, overlapping alteration-related patterns. To address these limitations, we propose a novel chained deep-learning framework, UAGB, which integrates Uniform Manifold Approximation and Projection (UMAP) for topology-preserving dimensionality reduction, an Autoencoder (AE) for deep regularized compression, a Generative Adversarial Network (GAN) for adversarial feature enhancement, and Bisecting K-means for hierarchical hard clustering. The framework was applied to a 15-element geochemical dataset from the Pariz area, a region hosting several world-class porphyry Cu deposits. The UAGB performance was compared against a baseline UMAP-GMM (UG) model and three alternative methods (PCA+K-means, t-SNE+HDBSCAN, ISOMAP+BIRCH). Quantitative evaluation using the prediction–area plot and Normalized Density (Nd) index shows that UAGB captures 86% of known Cu deposits within only 14% of the study area (Nd = 6.14), substantially outperforming UG (Nd = 3.16) and the best alternative method (PCA+K-means, Nd = 4.55). Ablation studies demonstrate that removing the AE or GAN reduces Nd by 21–31%, confirming their irreplaceability and synergistic effect. Spatial cross-validation and hold-out deposit tests (standard deviation = 0.12, withheld deposit capture = 86%) rule out overfitting and confirm generalization. The enhanced feature space becomes highly separable (Silhouette = 0.91, intra-cluster distance ratio = 0.21), justifying the use of hard clustering. The UAGB framework provides a powerful, scalable, and interpretable tool for geochemical anomaly mapping, reducing exploration search space and generating testable targets for porphyry Cu systems.