Scholarlink

<p>Artificial intelligence (AI) has emerged as an increasingly prevalent approach for the constitutive modeling of geomaterials. However, it encounters challenges in generalization when data are scarce, which severely restricts its applications, especially in complex engineering. This research develops a novel data-driven constitutive modeling framework that effectively leverages multi-fidelity data and incorporates two core technologies: (1) encoding the generalized plasticity theory in neural network architectures to develop models with strong generalization capabilities using large amounts of low-fidelity data; (2) implementing the infusion of scarce, high-fidelity experimental data based on transfer learning to enhance prediction accuracy. A comprehensive sensitivity analysis demonstrates the impact of each component in this data-driven model on capturing material strength and deformation features, which guides the integration strategy of high-fidelity physical experimental data. Validation across various stress paths verifies the model’s accuracy and robustness. The engineering feasibility of the framework has been demonstrated through finite element method (FEM) simulations of a concrete face rockfill dam (CFRD), highlighting its ability to address complex geotechnical engineering problems and accurately reproduce stress–strain distributions. This study offers a feasible approach to developing high-fidelity, data-driven constitutive models, even in the context of sparse data.</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

A multi-fidelity data-driven constitutive modeling framework via physics encoding and transfer learning: application to geotechnical engineering problems

  • 期刊论文

  • 发表日期：

    2026年3月13日

查看全文
  • Jingzhou Wang,
  • Gang Ma,
  • Tongming Qu,
  • Daren Zhang,
  • Zhihan He,
  • Wei Zhou,
  • Xiaolin Chang

摘要

Artificial intelligence (AI) has emerged as an increasingly prevalent approach for the constitutive modeling of geomaterials. However, it encounters challenges in generalization when data are scarce, which severely restricts its applications, especially in complex engineering. This research develops a novel data-driven constitutive modeling framework that effectively leverages multi-fidelity data and incorporates two core technologies: (1) encoding the generalized plasticity theory in neural network architectures to develop models with strong generalization capabilities using large amounts of low-fidelity data; (2) implementing the infusion of scarce, high-fidelity experimental data based on transfer learning to enhance prediction accuracy. A comprehensive sensitivity analysis demonstrates the impact of each component in this data-driven model on capturing material strength and deformation features, which guides the integration strategy of high-fidelity physical experimental data. Validation across various stress paths verifies the model’s accuracy and robustness. The engineering feasibility of the framework has been demonstrated through finite element method (FEM) simulations of a concrete face rockfill dam (CFRD), highlighting its ability to address complex geotechnical engineering problems and accurately reproduce stress–strain distributions. This study offers a feasible approach to developing high-fidelity, data-driven constitutive models, even in the context of sparse data.