This study presents the development of composite binders for hardening backfill mixtures using low-activity Portland cement, crushed gravel-sand materials, and local loams, together with a machine learning-based approach for predicting their mechanical and rheological performance. Experimental investigations were carried out to determine water demand, setting time, compressive strength, and mixture mobility for different binder compositions containing mineral and clay fillers. The results confirmed that partial replacement of Portland cement with locally available materials reduces cement consumption while maintaining the required technological and strength characteristics for gravity-flow pipeline transportation and formation of a stable backfill massif. To extend the practical applicability of the obtained experimental data, a data-driven predictive model was introduced using supervised machine learning techniques. The experimental dataset derived from laboratory measurements was used to establish relationships between binder composition parameters and resulting strength properties. Model evaluation demonstrated reliable prediction accuracy and confirmed the feasibility of applying digital methods for optimization of composite binder design in mining backfill technologies. The proposed integration of experimental materials science and machine learning provides a foundation for intelligent selection of binder compositions, reduction of material costs, and support for digital transformation in construction and mining engineering.