Reliable Neural Networks for Proportioning of Concrete Mixes Containing Cement Replacement Materials

Artificial neural network (ANN) is one of the most suitable analytical methods used for modeling complex relationships between concrete properties and concrete mix ingredients. In the present study, common ANN is first trained to predict concrete mix proportions using experimental data sets. Experimental data sets include the content of water, cementitious materials (cement, silica fume, rice husk ash, and natural pozzolanas), coarse and fine aggregates, superplasticizer and the compressive strength, the slump, and the maximum size of the aggregates. Also, statistical analyses are used to correlate the results and the acceptable error. Then, the neural network is retrained using experimental data sets together with analytical data sets that are produced using the Iranian national concrete mix design method, empirical relationships, or both. Finally, the performance of the retrained neural networks is compared with the first one. The results show that the neural networks, which are trained using both analytical and experimental data sets, are more reliable and can be reliably used for modeling of complex relationships.