Scholarlink

<p>Conglomerate reservoirs are controlled by the interaction of multi-scale sedimentary architectures and diagenetic processes, resulting in highly heterogeneous and spatially discontinuous reservoir properties and geomechanical characteristics. This complexity significantly increases the difficulty of identifying dominant controlling factors and classifying reservoir quality. To address this issue, a novel integrated model framework for dominant factor screening and reservoir quality evaluation is proposed. In this study, a sample similarity graph is constructed to characterize the geological associations among reservoir samples. A Point Graph Attention Network (Point Graph-GAT) model is employed to quantitatively evaluate the importance of feature parameters. Using production well performance as a reference, a hierarchical Transformer-based multi-modal fusion (HT-MMF) model is developed for multi-scale reservoir quality classification. Through a cross-attention mechanism and a hierarchical Transformer architecture, the model captures the intrinsic relationships between dominant controlling factors and production performance. Furthermore, a grade-mapping module is designed to achieve refined reservoir quality classification. The results indicate that the importance ranking of characteristic parameters in conglomerate reservoirs is as follows: Poisson’s ratio &gt; porosity &gt; permeability &gt; pore pressure &gt; brittleness index &gt; density &gt; gamma-ray (GR) &gt; shale content &gt; tensile strength. The proposed model demonstrates stable convergence with a final loss of 0.0006, outperforming both the optimal ablation model (0.001) and conventional benchmark models. In addition, the classification predictions for new wells are highly consistent with actual classifications, with discrepancies in the proportion of each class controlled within 2%. The proposed method effectively captures the heterogeneity of complex conglomerate reservoirs and provides a reliable approach for sweet spot identification and fine-scale reservoir development.</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

An integrated model for conglomerate reservoir quality classification based on graph attention and multiscale multimodal features

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月19日

查看全文
  • Bingjin Zhao,
  • Shanyong Liu,
  • Yishan Lou,
  • Biao Yin

摘要

Conglomerate reservoirs are controlled by the interaction of multi-scale sedimentary architectures and diagenetic processes, resulting in highly heterogeneous and spatially discontinuous reservoir properties and geomechanical characteristics. This complexity significantly increases the difficulty of identifying dominant controlling factors and classifying reservoir quality. To address this issue, a novel integrated model framework for dominant factor screening and reservoir quality evaluation is proposed. In this study, a sample similarity graph is constructed to characterize the geological associations among reservoir samples. A Point Graph Attention Network (Point Graph-GAT) model is employed to quantitatively evaluate the importance of feature parameters. Using production well performance as a reference, a hierarchical Transformer-based multi-modal fusion (HT-MMF) model is developed for multi-scale reservoir quality classification. Through a cross-attention mechanism and a hierarchical Transformer architecture, the model captures the intrinsic relationships between dominant controlling factors and production performance. Furthermore, a grade-mapping module is designed to achieve refined reservoir quality classification. The results indicate that the importance ranking of characteristic parameters in conglomerate reservoirs is as follows: Poisson’s ratio > porosity > permeability > pore pressure > brittleness index > density > gamma-ray (GR) > shale content > tensile strength. The proposed model demonstrates stable convergence with a final loss of 0.0006, outperforming both the optimal ablation model (0.001) and conventional benchmark models. In addition, the classification predictions for new wells are highly consistent with actual classifications, with discrepancies in the proportion of each class controlled within 2%. The proposed method effectively captures the heterogeneity of complex conglomerate reservoirs and provides a reliable approach for sweet spot identification and fine-scale reservoir development.