Comparison of regression based functions and ANN models for predicting the compressive strength of geopolymer mortars

Abstract

This study investigated the predictability of the compressive strength (CS) of geopolymeric mortars based on blast furnace slag (BFS) and steel mill slag (SMS). For this purpose, the study consists of two parts. In the first part of the study, BFS and SMS, two different types of slag were used as binders in 11 different proportions. At the end of the curing period, the weight, ultrasonic pulse velocity (UPV) and compressive strength of the mortars were determined. In the second part of the study, the compressive strength was predicted using regression analysis (CRA), multivariate adaptive regression spline (MARS), random forest (RF), multiple additive regression trees (TreeNet) and artificial neural networks (ANN). The model performance of the methods was compared using root mean square error (RMSE), mean absolute error (MAE) and Nash-Sutcliffe efficiency (NSE) performance statistics. When comparing the performance of the developed prediction models, the power function method was found to produce the best predictions among the regression-based methods. For the MARS, TreeNet and RF models, the TreeNet model produced the best prediction, while for the ANN_5 and ANN_10 models, the ANN_5 model produced the best prediction. In general, it can be concluded that the models developed with ANN can predict the compressive strength of mortars with a very high accuracy. Significant economic and time savings can be achieved with the developed models. In addition, the CS values of geopolymeric mortars prepared with different proportions of slag types and activator can be predicted without waiting for 7–28 days of curing.