Formulation and prediction of ready-mix concrete performances using neural networks

Abstract

The objective of the research is to investigate the possibility of using artificial neural networks (ANN) during concrete mix design either to produce the optimal ingredient proportions to meet certain performance criteria or to predict the properties of an already proportioned mix. For this purpose, 81 different concrete mixes were prepared in the laboratory of a ready-mixed concrete plant. Sixty-nine randomly selected mixes (85% of the prepared mixes) were used to train two developed ANN. The first ANN produces optimal proportions of a concrete mix based on specified properties. The second ANN predicts the properties of a concrete mix based on its proportions. The remaining 12 mixes were used to validate the developed ANN and to compare their outcome with those obtained using three existing proportioning methods. It was found that both developed ANN produce results with root-mean-squared error lower than those obtained using the other studied methods. It is therefore recommended that concrete producers develop similar ANN for the set of their local materials and for the properties that they deem important for them. The expected better accuracy of the proposed procedure justifies its implementation and the money spent in the original testing programme used for the network training.