Neural network-based prediction of mechanical properties in high-strength fly ash-based geopolymer mortars: a comparative analysis of model architectures and optimizers

This study investigates various machine learning models, namely multi-layer perceptron (MLP) and generalized regression neural network (GRNN), for predicting the mechanical properties of high compressive strength geopolymer mortars. Both classification (MLPC and GRNNC) and regression (MLPR and GRNNC) based models, with MLP architectures comprising 1 and 2 hidden layers, are developed. Furthermore, three optimization algorithms, namely Levenberg–Marquardt (LM), momentum (M), and resilient backpropagation (R), are utilized. The models’ inputs are alkali concentrations, heat-curing temperatures, and curing periods. The results showed that the classification-based MLP with one hidden layer and resilient optimizer (MLPC-1-R) outperformed the other models by recording lower prediction deviations and high prediction accuracy. On the other hand, the regression-based models showed promising results and less sensitivity to the optimization type, unlike the classification-based ones. Finally, the resilient backpropagation (R) optimizer tends to provide consistent and high performance for both classification and regression-based models.