Intelligent prediction of concrete carbonation depth using neural networks

Abstract

The carbonation problem in concrete has the potential to cause severe degradations in structures and therefore, its accurate prediction is critical in the field of civil engineering. This study involves the results obtained from the preliminary investigations on the use of arti_cial neural networks (ANN) as a non-destructive method for the prediction of carbonation depth in concrete. A total of 225 experimental cases obtained from the related literature have been used as the training and testing data set, with 18 different input parameters identified to be influencing the output, which is the carbonation depth measured in concrete. Two learning schemes were suggested with varying training: testing data distributions and three di_erent values for hidden neurons were tested in combination. Results show that the use of ANN for the prediction of carbonation depth has the potential to provide predictions with satisfactory accuracy. Variations in the coe_cient of determination (R), the mean squared error (MSE), and the number of iterations required for learning, within the proposed changing combinations of training: testing data distribution and the number of hidden neurons has been discussed. The combination of the highest coe_cient of determination (R), and the lowest mean squared error (MSE) were determined as 0.975 and 0.0018 respectively, which was observed when the CGP method is used with 50%:50% training: testing data distribution) and with 10 hidden neurons.