Scholarlink

<p>This study presents a novel, integrated physics-informed machine learning framework that simultaneously addresses UCS prediction, uncertainty quantification, and mix design optimization for chemically stabilized soils, combining a rigorously curated 990-mixture dataset, a Physics-Informed Neural Network (PINN) embedding geochemical monotonicity constraints, Monte Carlo-based reliability quantification, and a multiobjective Pareto optimization pipeline within a single unified framework. Various machine learning models including linear regression, random forests, support vector regression, gradient boosting, and artificial neural networks (ANN) were tested. Out of the models tested, the ANN produced the best results with an <i>R</i><sup>2</sup> value of 0.919 and RMSE of 1.124&#xa0;MPa. A Physics-Informed Neural Network (PINN) was further developed to enforce physically consistent trends, providing improved interpretability despite marginally reduced accuracy. Reliability-based UCS estimation was possible as predictive uncertainty was quantified using Monte Carlo simulation. Along with defined geo-chemistry principles, SHAP analysis identified the three most important controlling factors to be curing time, GGBS content, and alkaline activators. A multiobjective optimization framework revealed the possible design of economically viable and environmentally friendly stabilization mixes. The developed framework represents a real-world decision-making support system for chemically stabilized soil 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

A Physics-Informed Machine Learning Framework for Strength Prediction, Uncertainty Quantification, and Optimization of Chemically Stabilized Soils

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Ajay Pratap Singh Rathor,
  • Abdollah Tabaroei,
  • Arindam Dey,
  • Jitendra Kumar Sharma

摘要

This study presents a novel, integrated physics-informed machine learning framework that simultaneously addresses UCS prediction, uncertainty quantification, and mix design optimization for chemically stabilized soils, combining a rigorously curated 990-mixture dataset, a Physics-Informed Neural Network (PINN) embedding geochemical monotonicity constraints, Monte Carlo-based reliability quantification, and a multiobjective Pareto optimization pipeline within a single unified framework. Various machine learning models including linear regression, random forests, support vector regression, gradient boosting, and artificial neural networks (ANN) were tested. Out of the models tested, the ANN produced the best results with an R2 value of 0.919 and RMSE of 1.124 MPa. A Physics-Informed Neural Network (PINN) was further developed to enforce physically consistent trends, providing improved interpretability despite marginally reduced accuracy. Reliability-based UCS estimation was possible as predictive uncertainty was quantified using Monte Carlo simulation. Along with defined geo-chemistry principles, SHAP analysis identified the three most important controlling factors to be curing time, GGBS content, and alkaline activators. A multiobjective optimization framework revealed the possible design of economically viable and environmentally friendly stabilization mixes. The developed framework represents a real-world decision-making support system for chemically stabilized soil systems.