Data-driven intelligent modeling of unconfined compressive strength of heavy metal-contaminated soil

Abstract

This study focuses on empirical modeling of the strength characteristics of urban soils contaminated with heavy metals using machine learning tools and their subsequent stabilization with ordinary Portland cement (OPC). For dataset collection, an extensive experimental program was designed to estimate the unconfined compressive strength (Qu) of heavy metal-contaminated soils collected from a wide range of land use pattern, i.e. residential, industrial and roadside soils. Accordingly, a robust comparison of predictive performances of four data-driven models including extreme learning machines (ELMs), gene expression programming (GEP), random forests (RFs), and multiple linear regression (MLR) has been presented. For completeness, a comprehensive experimental database has been established and partitioned into 80% for training and 20% for testing the developed models. Inputs included varying levels of heavy metals like Cd, Cu, Cr, Pb and Zn, along with OPC. The results revealed that the GEP model outperformed its counterparts: explaining approximately 96% of the variability in both training (R2 = 0.964) and testing phases (R2 = 0.961), and thus achieving the lowest RMSE and MAE values. ELM performed commendably but was slightly less accurate than GEP whereas MLR had the lowest performance metrics. GEP also provided the benefit of traceable mathematical equation, enhancing its applicability not just as a predictive but also as an explanatory tool. Despite its insights, the study is limited by its focus on a specific set of heavy metals and urban soil samples of a particular region, which may affect the generalizability of the findings to different contamination profiles or environmental conditions. The study recommends GEP for predicting Qu in heavy metal-contaminated soils, and suggests further research to adapt these models to different environmental conditions.