Investigations on the influence of variations in hidden neurons and training data percentage on the efficiency of concrete carbonation depth prediction with ann

Abstract

Concrete is undoubtedly one of the most popular construction materials. Carbonation is a well-known concrete durability problem that may negatively affect the performance of reinforced concrete buildings. In this study, efficient prediction of concrete carbonation depth was targetted by employing artificial neural networks with by One-step Secant method of optimization, as an alternative to conventionally used Levenberg-Marquardt algorithm. The effect of varying hidden neuron values and varying train:test data distribution on the evolution of network performance was aimed to be investigated in a sensitive and systematical way, as the scope of this study. Network training was carried out with combinations of 10 different hidden neurons and 11 data distribution ratios. For the task of predicting concrete carbonation depth as the output, the highest coefficient of correlation (R) obtained was 0.99. Results have shown that the variations of training dataset percentage within the range of 30–55% yielded a more significant improvement in the R-value than it is observed within the range of 60–80%. It was also observed that the variation of hidden neurons between the values 5–25 yielded relatively less significant changes on the prediction of accuracy, both in terms of R and MSE, for the range of training data percentages between 60–80%.