Big data-driven global modeling of cohesive soil compaction across conceptual and arbitrary energies through machine learning

Abstract

This study aims to propose an intelligent modeling framework for accurately predicting compaction characteristics of cohesive soils across compaction energy (CE) levels. A comprehensive database of 1001 observations falling within the theoretical bounds was created through experimental investigation encompassing sieve analysis, hydrometer analysis, liquid limit (w_L), plastic limit (w_P), specific gravity (G_s), and compaction tests on natural soil samples and literature review, encompassing diverse cohesive soils, CE levels, and compaction characteristics. Multiple machine learning techniques, including Extreme Gradient Boosting (XGBoost), Random Forest (RF), Gene Expression Programming (GEP), Multi Expression Programming (MEP), Artificial Neural Networks (ANN), and Multiple Linear Regression (MLR), were applied to develop predictive models. XGBoost demonstrated superior performance in predicting maximum dry density (γ_dmax) and optimum moisture content (w_opt) as evaluated by statistical indicators and external validation and compared with existing models in the literature. The proposed models effectively captured the influence of key parameters, highlighting the primary role of CE and w_L, the secondary role of plastic limit (w_P), the tertiary role of plasticity index (I_P) and fines activity (A_F), and the quaternary role of soil gradation in predicting and influencing the compaction characteristics of cohesive soils. This approach enables accurate global modeling of cohesive soil compaction across varying CE levels, providing a valuable tool for geotechnical engineers and researchers to determine compaction characteristics for a known CE level using basic soil properties used for soil classification.