Enhancement of Mechanical and Durability Properties of Recycled Aggregate Concrete Through Partial Cement Replacement With Marble Dust and Finely Ground Recycled Glass Dust: Experimental Investigation and Machine LearningBased Predictive Modelling

The escalating generation of construction and demolition (C&D) waste and the environmental burden imposed by Portland cement production have collectively intensified global interest in recycled aggregate concrete (RAC) as a sustainable structural material. However, RAC inherently exhibits inferior mechanical performance relative to conventional concrete, primarily because of the elevated porosity and compromised interfacial transition zone (ITZ) introduced by residual adhered mortar on recycled aggregate surfaces. This study addresses these limitations through a combined experimental and machine-learning investigation of RAC in which ordinary Portland cement (OPC 53-grade) was partially replaced with marble dust (MD) and finely ground recycled glass dust (GD) — both sized at 70 µm — at total replacement levels of 0% (control M1), 10% (M2), 15% (M3), and 20% (M4), with concomitant introduction of recycled coarse aggregate (RCA) at 0%, 20%, 40%, and 60%. Four mix variants conforming to M30 design strength were cast per IS 10262:2019; compressive strength (3, 7, 14, 28 days), split tensile strength (7, 28 days), flexural strength (7, 28 days), and water absorption were systematically evaluated. The 20% MD+GD blend (M4) achieved a 28-day compressive strength of 35.40 MPa — a 24.5% improvement over the control (28.43 MPa) — while split tensile and flexural strength increased by 55.7% and 50.6%, respectively. Water absorption showed a nuanced response consistent with the competing effects of RCA porosity and pozzolanic pore refinement. The mechanistic basis for these improvements is attributed to dual co-operative phenomena: pozzolanic reactivity of glass dust generating additional C-S-H gel, and the physical micro-filling action of marble dust that densifies the concrete matrix and strengthens the ITZ. To extend predictive capability across an expanded compositional space, an artificial neural network (ANN) with a 7-14-10-14-5 multi-output architecture was developed, achieving R² values of 0.984 and 0.980 for compressive and flexural strength prediction on the held-out test set, outperforming Random Forest, XGBoost, and SVR baselines. SHAP-based feature importance analysis confirmed that RCA replacement percentage and glass dust content are the dominant predictors of all five output responses.