Machine Learning-Based Evaluation of Bearing Capacity in Geosynthetic-Reinforced Working Platforms over Soft Subgrade

Abstract

Accurate prediction of bearing capacity in geosynthetic-reinforced working platforms over soft subgrade remains a challenge in geotechnical engineering practice. Existing methods provide practical solutions to estimate the bearing capacity in two-layered systems with soft subgrades but produce inconsistent results under varying soil conditions. This study introduces machine learning (ML) methods as an alternative framework to evaluate the bearing capacity of such a system and investigate the relative importance of governing parameters. A database of 50 cases, compiled from laboratory, centrifuge, and field studies, was utilized to train and test multiple ML models, including Random Forest, Gradient Boosting, Support Vector Regression, and Artificial Neural Networks. Input features included granular layer thickness and properties, undrained shear strength of the lower layer, geosynthetic tensile strength, and other geosynthetic properties. The ML models demonstrated strong capability in capturing nonlinear interactions and identifying the dominant influence of critical parameters across diverse test conditions. Comparative evaluation of the ML models using repeated cross-validation demonstrated that ensemble tree-based methods provide robust and accurate predictions for this dataset, while permutation-based feature importance offers physically interpretable insight into the governing parameters. These findings support the integration of ML into geotechnical design practice and provide new insights for the refinement of existing bearing capacity calculation methods and design guides.