Scholarlink

<p><?tk 4?>Accurate prediction of Unconfined Compressive Strength (UCS) of clay soils is essential for ensuring safe and economical design in construction and civil infrastructure. Traditional empirical correlations often lack robustness across varying soil conditions and do not capture underlying physical behavior. This study presents a Physics-Informed Machine Learning (PIML) framework that blends domain knowledge with data-driven models to enhance UCS prediction accuracy and reliability. A real-world dataset of 120 samples was expanded using over 1000 synthetic samples generated through Gaussian-based sampling, maintaining statistical fidelity to observed soil properties. Several baseline machine learning models namely Artificial Neural Network (ANN), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and XGBoost were evaluated alongside physics-informed models, including a Physics-Informed Neural Network (PINN) and a hybrid PINN-GPR. The PINN integrates physical relationships (e.g., UCS ∝ γ<sub>d</sub> and UCS ∝ 1/PI) directly into the model loss function. Results show the PINN achieves superior performance (R<sup>2</sup> = 0.941, a20 = 88.6%), demonstrating improved generalization and consistency. SHAP-based feature attribution identifies Plasticity Index and Dry Density as the most influential variables. A real-world deployment example further illustrates how the model enables rapid UCS estimation. This research underscores the value of embedding physics into machine learning for improved accuracy, trust, and usability.</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 machine learning for predicting the unconfined compressive strength of clay soils using real and synthetic data

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Krishna Kumar Ahirwar,
  • Rajeev Jain

摘要

Accurate prediction of Unconfined Compressive Strength (UCS) of clay soils is essential for ensuring safe and economical design in construction and civil infrastructure. Traditional empirical correlations often lack robustness across varying soil conditions and do not capture underlying physical behavior. This study presents a Physics-Informed Machine Learning (PIML) framework that blends domain knowledge with data-driven models to enhance UCS prediction accuracy and reliability. A real-world dataset of 120 samples was expanded using over 1000 synthetic samples generated through Gaussian-based sampling, maintaining statistical fidelity to observed soil properties. Several baseline machine learning models namely Artificial Neural Network (ANN), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and XGBoost were evaluated alongside physics-informed models, including a Physics-Informed Neural Network (PINN) and a hybrid PINN-GPR. The PINN integrates physical relationships (e.g., UCS ∝ γd and UCS ∝ 1/PI) directly into the model loss function. Results show the PINN achieves superior performance (R2 = 0.941, a20 = 88.6%), demonstrating improved generalization and consistency. SHAP-based feature attribution identifies Plasticity Index and Dry Density as the most influential variables. A real-world deployment example further illustrates how the model enables rapid UCS estimation. This research underscores the value of embedding physics into machine learning for improved accuracy, trust, and usability.