Scholarlink

<p>Coupled thermo-hydro-mechanical-chemical (THMC) interactions represent the governing multiphysics processes that dictate the fate of a broad spectrum of intricate porous geological media. As physics-based characterization of the system response typically presents significant computational challenges especially for high-fidelity replications, data-driven methodologies have hence garnered growing attention nowadays for their transcendent predictive capabilities. Nevertheless, current surrogate modeling endeavors have focused mostly on conventional geomechanics applications involving only hydraulic and/or mechanical processes, whereas the prospects of deep learning in replicating multiple interacting fields pertinent to complex earth systems have still remained unexplored. We therefore develop a hybrid deep neural network (DNN) in this study which first transforms long-term multiphysics input into high-dimensional hidden states via LSTM, and local temporal dependencies within the feed sequence are subsequently extracted using CNN. Utilizing LSTM’s last hidden state as the query and CNN’s outputs as the key, temporal pattern attention (TPA) is then deployed to dynamically adjust the attention score allocation across feature maps, adaptively focusing on pivotal time-invariant patterns of the interacting multivariate sequence. The resulting TPA-CNN-LSTM architecture’s efficacy is demonstrated by taking hydrating porous cement-based mine waste as a typical example. Our results reveal that the hybrid DNN maintains stable performance across all THMC features, outperforming other baseline models including CNN-LSTM-Self-Attention. We also conducted a comprehensive analysis to elucidate the algorithmic mechanisms underlying TPA-CNN-LSTM’s superior performance, and discussed potential directions for future refinement. The research findings highlight DNN’s promising potential in replicating multiphysics evolutions of generic geological media, and hence have profound implications for parametric design optimization and practical early warning of complex earth 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

Multiphysics Modeling of Porous Media Using Temporal Attention Based Deep Learning Framework

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Yumeng Huang,
  • Jun Tang,
  • Liangdong Huang,
  • Hao Chen,
  • Liangfu Xie,
  • Gongda Lu

摘要

Coupled thermo-hydro-mechanical-chemical (THMC) interactions represent the governing multiphysics processes that dictate the fate of a broad spectrum of intricate porous geological media. As physics-based characterization of the system response typically presents significant computational challenges especially for high-fidelity replications, data-driven methodologies have hence garnered growing attention nowadays for their transcendent predictive capabilities. Nevertheless, current surrogate modeling endeavors have focused mostly on conventional geomechanics applications involving only hydraulic and/or mechanical processes, whereas the prospects of deep learning in replicating multiple interacting fields pertinent to complex earth systems have still remained unexplored. We therefore develop a hybrid deep neural network (DNN) in this study which first transforms long-term multiphysics input into high-dimensional hidden states via LSTM, and local temporal dependencies within the feed sequence are subsequently extracted using CNN. Utilizing LSTM’s last hidden state as the query and CNN’s outputs as the key, temporal pattern attention (TPA) is then deployed to dynamically adjust the attention score allocation across feature maps, adaptively focusing on pivotal time-invariant patterns of the interacting multivariate sequence. The resulting TPA-CNN-LSTM architecture’s efficacy is demonstrated by taking hydrating porous cement-based mine waste as a typical example. Our results reveal that the hybrid DNN maintains stable performance across all THMC features, outperforming other baseline models including CNN-LSTM-Self-Attention. We also conducted a comprehensive analysis to elucidate the algorithmic mechanisms underlying TPA-CNN-LSTM’s superior performance, and discussed potential directions for future refinement. The research findings highlight DNN’s promising potential in replicating multiphysics evolutions of generic geological media, and hence have profound implications for parametric design optimization and practical early warning of complex earth systems.