Integrating machine learning and experimental characterization for predicting the compressive strength of dune-sand-based high-performance concrete

Abstract

Dune sand (DS) is readily accessible in southeastern Algeria and can be serves as a viable substitute for poorly available of crushed sand (CS) widely used in Saharan construction. This study aims to the valorization of dune sand collected from Mih-Ouensa of El Oued region which is weakly used in concrete field, for development of high-performance concrete based this raw material (HPC-DS). In order to identify the main composition of used DS, physico-chemical analyses were conducted by X-ray diffraction (XRD), X-ray fluorescence (XRF) and SEM/EDX observations. Based on the characterized materials, several HPC-DS mixtures were formulated with different contents of DS, CS, Portland cement, and silica fume (SF). The experimental data were obtained by the application of compressive strength test on the 100 mm cubic specimens. In addition to the experimental analysis, nine supervised machine learning algorithms were applied to forecast the compressive strength of HPC-DS based on the mixture composition. The obtained results demonstrated that the addition of dune sand up to 40% as a partial substitute for crushed sand improved the mechanical properties of HPC-DS, while greater dune sand content resulted in a significant decrease in strength. Furthermore, of all assessed models, the Extra Trees Regressor achieved the highest predictive accuracy (R² = 0.9243), exhibiting robust alignment with the experimental trends. Feature-importance analysis indicated that hydration age is the primary factor influencing compressive strength, followed by superplasticizer content and dune sand addition. Moreover, a positive correlation was statically deduced between the hydration age and dune sand content.