Artificial neural network-based prediction of compressive strength in sustainable one-part GGBS/POFA-based geopolymer mortar with synthesized sodium silicates

This study examined how multicollinearity affects the predictive modelling of compressive strength in sustainable ground granulated blast slag (GGBS)/palm oil fuel ash (POFA)-based one-part geopolymer mortars using homebrewed sodium silicate (SS) activators. SS was developed from silica-based materials: eco-processed pozzolan, illitic clay, and rice-husk ash using the thermochemical method. One-part geopolymer mortars were produced from binary precursors of POFA and GGBS and homebrewed SS (SSRHA, SSEPP, SSIC), generating 308 data points on mix compositions, testing ages, and corresponding compressive strength. Multicollinearity existed among POFA, GGBS, and Sodium hydroxide (SH). Artificial neural network (ANN)-I (all inputs) and ANN-II (excluding multicollinear inputs), along with a stepwise linear regression (SLR) model, were used. ANN-II performed the best (R2 = 0.9869), followed by ANN-I (0.9825) and SLR (0.6906). Taylor and boxplot diagrams validated the models. Sensitivity analysis ranked age as the most important factor and SSEPP as the least. Removing multicollinear inputs improved the ANN performance in predicting compressive strength.