Compressive strength prediction of ternary blended geopolymer concrete using artificial neural networks and support vector regression

The development of ternary blended geopolymers is one of the recent advancements in geopolymer concrete technology, which utilizes different source materials in various proportions. The researches based on the behavior of ternary blended geopolymers are very limited. A new attempt has been made in this paper to anticipate the behavior of ternary blends with the past performance of the binary blends, i.e., the third binder content of the binary blended data has to be kept as zero and based upon the number of data sets, the ternary blends have to be anticipated by the predictive model. To do so, artificial neural networks and support vector regression were adopted in this study, since a high degree of accuracy is required for such predictive attempts. Thirteen input parameters of mix design factors were considered to predict the compressive strength. In order to validate and check the predicting potential of developed models, a database based on the behavior of ternary blended geopolymer obtained from the own past study was also incorporated for training (303 data sets) and validation (50 data sets). A few significant parameters like heat curing regime and liquid parameters like molarity and modulus were also considered as inputs, where lies the novelty of this research. Several predictive models were developed with different training parameters, and selected few models for further validation processes based on training accuracies and tested on regression parameters and error analysis. It was identified that Gradient descent ANN models showed the predicting potential of 88% upon coefficient of determination with errors of about 3.9 and 4.4, respectively. ANN models proved to be efficient in assessing the behavior of ternary blends from the data of binary blends with high degree of accuracy which can be applied efficiently wherever the data inputs are inadequate but with large databases. Support vector regression models showed inferior results compared to ANN models, as they showed a predicting accuracy of only 78%, with an average error of 6.4.