Data-driven machine learning approach for stress history evaluation in cohesive soils using cone penetration test data

Daeun Gwak, Taeseo Ku,
Data-driven machine learning approach for stress history evaluation in cohesive soils using cone penetration test data,
Engineering Geology,
Volume 355,
2025,
108246,
ISSN 0013-7952,
https://doi.org/10.1016/j.enggeo.2025.108246.
(https://www.sciencedirect.com/science/article/pii/S0013795225003424)
Abstract: Accurately assessing the geostatic stress history is crucial for predicting the deformation characteristics and engineering properties of soils, as it is influenced by various geotechnical and geological factors such as varying loads, groundwater fluctuations, and environmental conditions. Although a traditional laboratory method using consolidation tests still provides a direct reference of stress history, it often has clear limitations, particularly with silts and sands, and is also time-consuming. As a result, alternative indirect approaches, such as analyzing field test data from the Cone Penetration Test (CPT), have also been developed and widely adopted due to their advantages such as fast and easy practical application and continuous profiling. However, despite these advantages, concerns remain regarding the reliability of CPT-based stress history estimation. This study proposes a robust data-driven approach to enhance stress history prediction using CPT data, addressing the existing reliability concerns. An extensive investigation was conducted by applying advanced machine learning techniques, including Deep Neural Network (DNN), Support Vector Regression (SVR), Random Forest (RF), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM). The approach utilizes a large-scale and high-quality global database compiled from CPT testing sites worldwide, focusing on key measurable parameters such as cone tip resistance, porewater pressure, and depth. This study also addresses essential methodological steps, such as data preprocessing, hyperparameter tuning, and 5-fold cross-validation. The results demonstrate that the machine learning-based models achieve remarkably improved accuracy in predicting the overconsolidation ratio (OCR) and preconsolidation stress (σp’) compared to conventional methods.
Keywords: Cone penetration test; Machine learning; Overconsolidation stress; Preconsolidation stress; Stress history