Scholarlink

<p>Mineral prospectivity mapping plays a critical role in guiding mineral exploration. However, existing data-driven approaches often overlook the spatial dependencies between sampling points and suffer from limited interpretability. To this end, this paper proposes a novel mineral prospectivity mapping method based on an interpretable graph transformer neural network, termed graph transformer (GTF). Compared to graph attention networks, the GTF method effectively captures the informational relationships between sample points across both global and local regions, leading to improved predictive accuracy. Moreover, a gradient descent algorithm is introduced to search for globally and locally optimal subgraphs, enabling the computation of unique contributions from global and local samples. Taking the Changba mineralization area located in Gansu Province, China, as an example, this study evaluated the validity and interpretability of the GTF approach in mineral prospectivity prediction by utilizing 16 mineralization-controlling factors derived from geochemical datasets. The results demonstrate that prospectivity maps generated by the GTF model not only leverage spatial relationships between samples effectively but also analyze the contributions of different features and spatial relationships among samples, considerably improving the precision of delineating potential Pb–Zn mineralization zones.</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

GTF: A New Interpretable Graph Neural Network for Geochemical Anomaly Detection in Mineral Prospectivity Mapping

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Zhengbo Yu,
  • Binbin Li,
  • Fan Zhang,
  • Jiangwei You,
  • Baizhi Li,
  • Zhe Chen,
  • Shuai Zhang,
  • Keyan Xiao

摘要

Mineral prospectivity mapping plays a critical role in guiding mineral exploration. However, existing data-driven approaches often overlook the spatial dependencies between sampling points and suffer from limited interpretability. To this end, this paper proposes a novel mineral prospectivity mapping method based on an interpretable graph transformer neural network, termed graph transformer (GTF). Compared to graph attention networks, the GTF method effectively captures the informational relationships between sample points across both global and local regions, leading to improved predictive accuracy. Moreover, a gradient descent algorithm is introduced to search for globally and locally optimal subgraphs, enabling the computation of unique contributions from global and local samples. Taking the Changba mineralization area located in Gansu Province, China, as an example, this study evaluated the validity and interpretability of the GTF approach in mineral prospectivity prediction by utilizing 16 mineralization-controlling factors derived from geochemical datasets. The results demonstrate that prospectivity maps generated by the GTF model not only leverage spatial relationships between samples effectively but also analyze the contributions of different features and spatial relationships among samples, considerably improving the precision of delineating potential Pb–Zn mineralization zones.