Neural Computing & Applications, vol.1, no.19, pp.1-22, 2021 (SCI-Expanded)
In this study, the cement-based parameters affecting CEMI portland cements-polycarboxylate ether-based high-range
water-reducing (HRWR) admixtures compatibility were investigated. For this purpose, eight CEMI cements and three
commercial HRWR admixtures were used. The rheological properties of 112 paste mixtures with different admixture
dosages and water/cement (W/C) ratios were determined in accordance with Herschel–Bulkley model. Then after, using
the experimental data, proper models were established to predict the dynamic yield stress and final viscosity of the pastes.
In addition to cement characteristics (such as fineness, compound composition and equivalent alkali content), waterreducing
admixture content and its solid material content as well as water/cement ratio of the pastes were considered as
input data. Multivariate adaptive regression splines (MARS) and multiple additive regression trees (MART) methods were
used in the models. Besides, artificial neural network (ANN) and conventional regression analysis (CRA) including linear,
power, and exponential functions were applied to determine the accuracy of the heuristic regression methods. Three
statistical indices, root-mean-square error, mean absolute error, and Nash–Sutcliffe, were used to evaluate the performance
of the models. Modeling findings indicated that the model with the lowest error for both of the rheological variables in the
testing set is the MART, followed by ANN, MARS, and CRA-Exponential methods. The most effective cement characteristics
causing incompatibility, hence detraction of paste rheological properties, in decreasing order, were determined
as cement fineness, C3S, C3A and equivalent alkali contents. C4AF and C2S contents of the cement were found to have less
effect on the cement–admixture incompatibility. It will be possible to determine the rheological properties of mixtures
containing different cements without conducting an experimental study by using the model based on MART method.