Machine learning optimized dry lean geopolymer concrete with reclaimed asphalt pavement for sustainable pavement applications

This study develops a process-based optimization strategy for dry lean geopolymer concrete (DLGC) using reclaimed asphalt pavement (RAP) material via response surface methodology (RSM) and machine learning (ML) approaches. A Box–Behnken design with 156 experimental runs considered six parameters with the following levels: sodium hydroxide concentration (8–16 M), Na2SiO3/NaOH molar ratio (1.5–3.0), RAP content (0–40%), alkaline liquid-to-binder (AL/B) ratio (0.35–0.55), binder content (280–380 kg/m3), and total aggregate content (1600–1900 kg/m3). The ternary binder system was composed of fly ash (FA, 60–70%), ground granulated blast furnace slag (GBFS, 20–30%), and metakaolin (MK, 10–15%). Optimum RAP content of 25–30% achieved 35–42 MPa (mean ± SD: 38.5 ± 2.3 MPa) compressive strengths that were 85–95% that of the control mix, while conferring high sustainability gains. ML models exhibited superior performance compared to statistical models with the Random Forest model providing R2 = 0.961 and RMSE = 2.14 MPa for compressive strength (CS), flexural strength (FS), and split tensile strength (STS) prediction. The most important influential parameters were binder content (28.3%) and NaOH concentration (22.1%). The developed second-order polynomial models showed excellent correlation for all three responses (R2 > 0.94). Pareto-optimal solutions for the DLGC mixes were identified for both economical and high strength application, with optimum mixes attaining 36.4 MPa strength at 0.85 cost index and 41.8 MPa for high strength use. DLGC mixes displayed high durability performance with 1200–1400 coulombs of chloride penetration, 3.2–4.1% water absorption, and 95% freeze–thaw durability factor. The life cycle assessment of DLGC mixes revealed 45% lower CO2 emissions than conventional concrete and 25–30% lower maintenance costs over a 25-year service life.