Recycled aggregate concrete (RAC) has emerged as a promising sustainable material in construction, replacing natural aggregates with recycled ones to reduce environmental impact. However, the incorporation of recycled aggregates (RA) introduces significant heterogeneities and uncertainties in concrete performance due to its influence on spatial distribution of the microstructures. A key determinant of RAC mechanical integrity is the interfacial transition zone (ITZ), which frequently governs stress concentrations and damage initiation. These motivate a multiscale predictive framework that resolves ITZ heterogeneity while quantifying how uncertain inputs propagate to macroscopic stiffness and strength. A semi-analytical multiscale micromechanical framework is employed in which the ITZ is represented as spatially heterogeneous rather than as a uniform layer. The ITZ is parameterized through non-uniform porosity profiles and thickness, informed by experimental data from nanoindentation and SEM–EDX. RAC-specific uncertainties, including variability in RA stiffness, hydration kinetics, and internal curing effects, are modeled probabilistically and propagated across scales to quantify their influence on macroscopic stiffness and strength. Variance-based global sensitivity analysis is used to identify dominant parameters and key interactions, enabling attribution of strength variability to microstructural drivers, with particular emphasis on ITZ morphology and properties. To overcome the significant computational expense typically associated with numerous analyses using detailed micromechanical models, a key strategic component of this investigation involves developing an efficient surrogate model which will help to drastically reduce computational effort without significantly compromising accuracy. The framework enables uncertainty-aware identification of the dominant microstructural and hydration-related drivers of RAC stiffness and strength, with particular emphasis on ITZ heterogeneity. Efficient surrogate modeling makes large-scale probabilistic analyses tractable, allowing systematic exploration of mix designs under realistic RA variability. The resulting sensitivity-informed insights support more reliable performance prediction and guide targeted interventions for sustainable RAC design.