Ambient-cured GGBS geopolymer concrete modified with alccofine, silica fume, or nano-silica: experiments and machine learning

Ambient-cured geopolymer concrete (GPC) offers a promising route toward lower-carbon infrastructure by utilizing industrial by-products while avoiding the thermal energy required for conventional heat curing. However, achieving mechanical and durability performance comparable to heat-cured GPC under ambient conditions remains a major challenge. This study investigates two multi-scale binder-modification routes for ambient-cured GGBS geopolymer concrete using Alccofine (AF) together with either silica fume (SF) or nano-silica (NS). AF was fixed at 10% of binder mass, while SF was investigated at 10–30% and NS at 1–3%. The experimental program first establishes 100% GGBS benchmark mixtures cured under ambient conditions and at 60 and 90 °C and subsequently evaluates the modified ambient-cured mixtures in terms of compressive, split tensile and flexural strengths, sorptivity, chemical resistance, freeze–thaw resistance and rapid chloride penetrability. The best-performing AF10SF20 mixture, containing 10% Alccofine and 20% silica fume, achieved compressive strengths of 51.8 and 66.0 MPa at 28 and 90 days, respectively, with corresponding 90-day split tensile and flexural strengths of 4.0 and 6.7 MPa. The mixture exhibited initial and secondary sorptivity values of 0.095 and 0.035 mm/min 0.5 , respectively, and showed strength losses of only 9.6%, 7.6%, 5.8% and 3.1% after exposure to H 2 SO 4 , HCl, Na 2 SO 4 and NaCl, respectively. After freeze–thaw exposure, the mass and strength losses were limited to 3.0% and 6.0%, while the rapid chloride penetrability was 1450 C. Notably, AF10SF20 exceeded the 90 °C-cured benchmark in several key performance indicators, including long-term compressive and flexural strength and resistance to transport and environmental deterioration. In parallel, machine-learning models were developed for compressive-strength prediction, with particle-swarm-optimization-assisted XGBoost providing the best reported predictive performance. Explainable artificial intelligence using SHAP, LIME, partial-dependence and individual-conditional-expectation analyses identified the major influencing mixture parameters and revealed nonlinear responses consistent with the experimentally observed best performed dosages. The findings demonstrate the potential of multiscale binder modification to reduce dependence on thermal curing under the investigated laboratory conditions, while further field-scale validation is required to establish its applicability under variable construction environments.