The Use of an Artificial Neural Network in the Prediction the Compressive Strength of Concrete

Resumo

This study proposes the use of an artificial neural network (ANN) in a mechanical characteristic
prediction of a material widely explored in the design of structures, the concrete. The conventional way to obtain
the mechanical characteristics of such a material is by means of expensive and costly laboratory tests and, as an
alternative, the use of an intelligent simulating system can be proposed. ANNs are based on a bio-inspired model
of the biological neuron, which processes data from simple units. In this study, a well-known (and established in
the academic literature) database was used. The artificial neural network tested used the supervised learning
method and the networks were trained based upon the following algorithms: classical backpropagation,
backpropagation with momentum, backpropagation with learning rate, backpropagation with momentum and
learning rate, and Levenberg-Marquardt. This work includes the use of a preprocessing strategy on the input data
and different backpropagation training algorithms. The main objective of this work was to obtain reliable results
to estimate the compressive strength of concrete by using machine learning. Among the training algorithms tested,
the one that presented the best performance was the Levenberg-Marquardt, which proved to be effective in
predicting the compressive strength of concrete at twenty-eight days, obtaining, as performance metrics, the RMSE
of 4.39 MPa and the coefficient of determination of 0.93. From these results it is possible to verify that this method
proved to be reliable for the calculation of the compressive strength of concrete by reducing possible errors and
amplifying the reliability of the application of computational technology in engineering projects.