Strength of a cement-improved clay and a machine learning evaluation of interactions between observed mechanical behaviours

• Ensemble tree-based models, especially Random Forest, outperformed others. • PCA effectively reduced dimensionality and multicollinearity. • Density and cement content were key features influencing strength in treated clays. • Neural networks revealed local feature sensitivity. Cement is a widely used chemical additive in soil improvement. Studies focusing on the interaction between sample properties, test conditions, and observed mechanical behaviours are valuable for gaining a thorough understanding of the improvement process. This study uses a machine learning-based approach to systematically investigate the improvement in the mechanical properties of a clay soil, evaluating the interaction between multiple factors, including cement content, change in specimen densities during curing, saturation levels during triaxial testing, confining pressures, and stress conditions, to uncover combined effects and key influencing features. Using data-driven techniques of correlation analysis, principal component analysis, and tree-based regression models, the study reveals that specimen density critically governs the strength properties of cemented soils, in addition to cement content. The results further demonstrate the superior modelling capability of ensemble tree-based machine learning algorithms over linear models. These findings enrich the understanding of cement-improved soils and indicate essential recommendations for developing machine learning models by pointing out critical features, such as density, and best performing algorithms.