Sustainable Stabilization of Expansive Soils Using Waste Marble Powder and Expanded Polystyrene Beads: Experimental Evaluation and Predictive Modelling

Abstract

Expansive soils exhibit considerable volume changes with moisture fluctuations leading to serious challenges for civil infrastructure, causing structural instability, pavement distortion, and foundation damage. While lime and cement remain widely used stabilizers, recent research has increasingly focused on waste-derived materials such as marble powder (MP) and expanded polystyrene beads (EPSb) as promising alternatives. These materials provide a practical approach to soil stabilization while contributing to the reuse of industrial by-products. In this study, the engineering behavior of high-plasticity clay was improved through the inclusion of MP and EPSb as additive materials. MP was added at 0%, 5%, 10%, 15%, and 20%, and EPSb at 0%, 0.3%, and 0.9% by dry weight of the high plasticity clay. Both additives were used alone and in combination. Laboratory tests, including Standard Proctor, free swell (FS), and unconfined compressive strength (UCS), were conducted. The results confirmed that the additives effectively reduced the liquid limit (LL) by 20.1% and the plasticity index (PI) by up to 22.4%. Results showed that EPSb effectively reduced FS and UCS, while MP decreased FS and increased UCS up to an optimal content. The most effective mixes achieved a maximum reduction of 54.7% in free swell (FS) (at 20% MP and 0.9% EPSb content) and a maximum increase of 13.1% in unconfined compressive strength (UCS) (at 5% MP content) compared to the untreated soil. The compaction tests further revealed a general decrease in optimum moisture content (OMC) and a slight increase in maximum dry density (MDD) with increasing MP content. Accordingly, the free swell (FS) and unconfined compressive strength (UCS) of the treated soils were predicted using multiple linear regression (MLR) and artificial neural network (ANN) models, developed from both the current experimental dataset and previously published studies. Input variables included untreated FS and UCS values, additive percentages, and one index property. The ANN model demonstrated superior predictive capability, achieving R2 values of 0.955 and 0.874 for FS and UCS, respectively, compared to 0.411 and 0.618 obtained with MLR. These results highlight the robustness of ANN in capturing nonlinear soil behavior and underscore its reliability and accuracy, particularly under limited data conditions.