Prediction of Compressive Strength of Activated Binary Cement Mortars Using Response Surface Methodology and Artificial Neural Networks

This study investigates the prediction of the compressive strength of activated binary cement mortars containing ground granulated blast-furnace slag (GGBFS) using two complementary modelling approaches: response surface methodology (RSM) and an artificial neural network (ANN). A database of 96 mix–curing combinations was assembled from the published literature on activated slag cement systems. Four input parameters were considered: slag replacement ratio (0–70 %), slag fineness (Blaine, 320–650 m²/kg), curing temperature (20–80 °C) and curing time (1–90 days). The output variable was the 1- to 90-day compressive strength (9–64 MPa). A second-order RSM polynomial was fitted and analysed by ANOVA, while a back-propagation feed-forward ANN with a 4–9–6–1 architecture and hyperbolic tangent activation was trained using the L-BFGS algorithm. The ANN reached R² = 0.996 and RMSE = 0.92 MPa over the whole dataset, clearly outperforming the RSM model (R² = 0.785, RMSE = 6.83 MPa). Sensitivity analysis showed that curing time was the dominant predictor (62 %), followed by slag fineness (22 %), curing temperature (8 %) and slag replacement ratio (8 %). The results confirm that ANN is a more reliable surrogate for the highly non-linear hydration kinetics of activated slag mortars and can be used to support mix-design decisions and reduce experimental effort.