Scholarlink

<p>Accurate prediction of collapse potential (CP%) in unsaturated gypseous soils is fundamental for safe and cost-effective design. Lime and silica fume have proven their effectiveness to reduce the collapsibility of the soil. Modelling the collapsibility of the soil is complex due to the highly nonlinear hydro-mechanical interactions and data limitations. This study proposes a fundamentally novel Physics Informed Nural Network (PINN) architecture that combines physical laws directly into the learning process to predict the collapsibility of chemically stabilized unsaturated gypseous soils. A dataset containing 600 experimental data was used in the analysis. Inputs are soil suction, gypsum content (%), degree of saturation (%), vertical stress, lime (%), silica fume (%) and their effect on the output collapse potential (%). The model achieved excellent predictive performance in testing phase (R² = 0.9896, Root Mean Squared Error RMSE = 1.41%) while sensitivity and partial derivative analyses confirmed the dominance of gypsum content and applied stress on soil collapsibility. Furthermore, a user-friendly graphical interface (GUI) was developed to help geotechnical engineers without Artificial Intelligence (AI) expertise to predict CP% under varying field conditions. The proposed PINN method proposes a trustable and explainable tool for predict collapse behavior in treated gypseous soils and provides a foundation for physics-consistent machine learning applications in geotechnical engineering.</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

Physics-informed neural networks for predicting collapse potential in chemically stabilized unsaturated gypseous soils: interpretable machine learning framework

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Omar H. Jasim,
  • Mohammed A. S. Al-Hitawi,
  • Mohammed Y. Fattah,
  • Nameer A. Kareem

摘要

Accurate prediction of collapse potential (CP%) in unsaturated gypseous soils is fundamental for safe and cost-effective design. Lime and silica fume have proven their effectiveness to reduce the collapsibility of the soil. Modelling the collapsibility of the soil is complex due to the highly nonlinear hydro-mechanical interactions and data limitations. This study proposes a fundamentally novel Physics Informed Nural Network (PINN) architecture that combines physical laws directly into the learning process to predict the collapsibility of chemically stabilized unsaturated gypseous soils. A dataset containing 600 experimental data was used in the analysis. Inputs are soil suction, gypsum content (%), degree of saturation (%), vertical stress, lime (%), silica fume (%) and their effect on the output collapse potential (%). The model achieved excellent predictive performance in testing phase (R² = 0.9896, Root Mean Squared Error RMSE = 1.41%) while sensitivity and partial derivative analyses confirmed the dominance of gypsum content and applied stress on soil collapsibility. Furthermore, a user-friendly graphical interface (GUI) was developed to help geotechnical engineers without Artificial Intelligence (AI) expertise to predict CP% under varying field conditions. The proposed PINN method proposes a trustable and explainable tool for predict collapse behavior in treated gypseous soils and provides a foundation for physics-consistent machine learning applications in geotechnical engineering.