Scholarlink

<p>Three-dimensional mineral prospectivity modeling is essential for locating concealed orebodies at depth, yet it frequently faces the challenge of scarce exploration data condition. Surface geochemical exploration data are easily accessible, have extensive spatial coverage, and contain crucial signals indicative of buried hydrothermal mineralization. Due to dimensional heterogeneity, effectively integrating surface geochemical data into 3D deep prediction remains a challenging task. Here, we propose a multiscale mixture-of-experts graph convolutional network (multiscale MoE-GCN) to synergistically integrate heterogeneous surface geochemical data and 3D geological models, in which a specifically designed heterogeneous graph unifies multi-dimensional predictive variables and spatial voxels within a single topological system. Specifically, the model features a dual-scale graph convolution module to separately extract representations from fine-scale 3D structures and ore-fluid simulations, as well as coarse-scale geochemical data. Further, a cross-modal mixture-of-expert (MoE) fusion mechanism is employed to decipher the intrinsic correlations between surface geochemical indicators and deep structural–hydrothermal mineralization processes. In this context, the edges within the heterogeneous graph express the genetic and spatial associations between surface geochemical data and deep-seated alteration–mineralization. Adaptive weighing of these heterogeneous features is employed to mitigate the inherent uncertainty associated with 3D deep mineral prospectivity. Results from the Xiadian gold deposit, China, demonstrate the model’s efficacy. By incorporating surface geochemistry as a constraint, the multiscale MoE-GCN significantly outperforms baseline methods reliant solely on sparse 3D structural models, achieving a superior area under the curve of 0.97. Crucially, the incorporation and fusion of surface geochemical data enabled the model to identify mineralization targets that were overlooked by single-source 3D structural models, demonstrating high consistency with known ore-controlling geological structures. Thus, the multiscale MoE-GCN can incorporate the surface geochemical data as a vital complement into the deep 3D mineral prospectivity modeling, thereby alleviating the persistent challenge of data scarcity in deep-seated exploration.</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

Multiscale MoE-GCN for 3D Mineral Prospectivity Modeling Integrating Surface Geochemical Exploration Data: Application in the Xiadian Gold Deposit, China

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Jiaxuan Song,
  • Xiancheng Mao,
  • Zhankun Liu,
  • Xiaoqi Yang,
  • Jia Sun,
  • Hao Deng,
  • Jin Chen,
  • Shuyan Yu,
  • Juexuan Huang

摘要

Three-dimensional mineral prospectivity modeling is essential for locating concealed orebodies at depth, yet it frequently faces the challenge of scarce exploration data condition. Surface geochemical exploration data are easily accessible, have extensive spatial coverage, and contain crucial signals indicative of buried hydrothermal mineralization. Due to dimensional heterogeneity, effectively integrating surface geochemical data into 3D deep prediction remains a challenging task. Here, we propose a multiscale mixture-of-experts graph convolutional network (multiscale MoE-GCN) to synergistically integrate heterogeneous surface geochemical data and 3D geological models, in which a specifically designed heterogeneous graph unifies multi-dimensional predictive variables and spatial voxels within a single topological system. Specifically, the model features a dual-scale graph convolution module to separately extract representations from fine-scale 3D structures and ore-fluid simulations, as well as coarse-scale geochemical data. Further, a cross-modal mixture-of-expert (MoE) fusion mechanism is employed to decipher the intrinsic correlations between surface geochemical indicators and deep structural–hydrothermal mineralization processes. In this context, the edges within the heterogeneous graph express the genetic and spatial associations between surface geochemical data and deep-seated alteration–mineralization. Adaptive weighing of these heterogeneous features is employed to mitigate the inherent uncertainty associated with 3D deep mineral prospectivity. Results from the Xiadian gold deposit, China, demonstrate the model’s efficacy. By incorporating surface geochemistry as a constraint, the multiscale MoE-GCN significantly outperforms baseline methods reliant solely on sparse 3D structural models, achieving a superior area under the curve of 0.97. Crucially, the incorporation and fusion of surface geochemical data enabled the model to identify mineralization targets that were overlooked by single-source 3D structural models, demonstrating high consistency with known ore-controlling geological structures. Thus, the multiscale MoE-GCN can incorporate the surface geochemical data as a vital complement into the deep 3D mineral prospectivity modeling, thereby alleviating the persistent challenge of data scarcity in deep-seated exploration.