Predicting performance of self-compacting concrete mixtures using artificial neural networks

Self-compacting concrete (SCC) is highly workable concrete that can flow th rough congested structural elements under its own weight and adequately fil l voids without segregation and excessive bleeding. Because of its complex mixture proportions, research on SCC has been highly empirical, and no mode ls with reliable predictive capabilities for its behavior have been develop ed. Thus, its rheological and mechanical properties are often described usi ng traditional regression analysis and statistical methods. The absence of a theoretical relationship between mixture proportioning and measured engin eering properties is overcome by subjectively assuming certain empirical re lationships based on limited experimental data, which are not applicable fo r conditions located outside the experimental domain, or when different mat erials are used. This paper demonstrates that artificial neural networks (ANN) can be used t o predict the performance of SCC mixtures effectively. Inspired by the inte rnal operation of the human brain, the ANN method has learning, self-organi zing and auto-improving capabilities. Thus, it can capture complex interact ions among input/output variables in a system without any prior knowledge o f the nature of these interactions, and without having to explicitly assume a model form. Indeed, such a model form is generated by the data points th emselves. This paper describes the database assembled, the architecture of the network selected, and the training process of the ANN model used. Initi al tests show that the ANN method can accurately predict the slump flow, fi lling capacity, segregation, and compressive strength test results of SCC m ixtures. A model for the acceptance/rejection of SCC mixtures based on know ledge of their mixture proportions is proposed and may be used after suffic ient development of a more comprehensive database on an industrial scale fo r the proportioning of SCC with tailor-made properties.