Explainable multi-output AI modeling for Fly Ash-Ground Granulated Blast Furnace Slag geopolymers: Predicting strength, workability, and embodied carbon

As attention shifts from operational to embodied emissions, fly ash–slag (FA–GGBFS) geopolymers offer a credible low-carbon binder option. Their formulation requires simultaneous consideration of strength, workability, and embodied carbon, yet most AI models are opaque and address these properties separately. This study develops a multi-output Random Forest with a global to local explainability pipeline to jointly predict and interpret these three properties for 885 FA–GGBFS mixes using 17 input features. The model achieved test set R² values of 0.84 for compressive strength (CS), 0.85 for slump flow (SL), and 0.91 for embodied carbon (EC), comparable to leading single target studies. Global SHapley Additive exPlanations (SHAP) analysis identified dominant drivers for each property, and clustering of SHAP values revealed formulation regimes per target, with governing parameters shifting between performance contexts. A thermodynamics-inspired local explainer (TERP) applied to regime medoids reduced decision logic to 2–6 features with local fidelity (|ρ| = 0.50–0.78). Partial dependence plots defined regime-specific thresholds: in high-strength mixes, CS plateaus near 44 MPa at GGBFS = 200 kg/m³ and curing age = 50 days; in high-flow mixes, SL peaks at 215 mm for water-to-binder = 0.25 with SiO₂/Al₂O₃ = 1.8; in high-carbon mixes, EC rises above 60 °C; and in low-carbon mixes, with Na₂SiO₃ > 110 kg/m³ . Integrating multi-target prediction with layered interpretation moves beyond single-target black-box models, exposing formulation-specific patterns and providing transparent, data-driven guidance for FA–GGBFS mix design. XAI-driven insights into the formulation and performance (CS, SL, EC) of FA-GGBFS geopolymers. • Multi-output RF predicts CS, SL, EC (R² = 0.84–0.91, test set). • Global SHAP ranks target-specific drivers across outputs. • SHAP clustering identifies distinct formulation regimes. • TERP on medoids extracts 2–6 local drivers (fidelity ≤ |ρ| 0.78). • PDPs define numeric thresholds for CS, SL, and EC.