An analytical study on the prediction of carbonation velocity coefficient using deep learning algorithm

Abstract

In present paper, we have studied the prediction method to determine the carbonation velocity coefficient of the concrete using deep neural network (DNN). The accelerated carbonation test data for 291 mixtures were used as training data with different experimental variable such as water to binder ratio, admixture (blast furnace slag and fly ash), fine aggregate and coarse aggregate as input data. Therefore, the carbonation velocity coefficient was calculated for 5% CO2, 60% relative humidity, and 20°C temperature. The model-based learning set was 5 hidden layers, 15% validation data ratio and 64 batch data size. Under this study condition, all model-based learnings were trained to the point where the learning was not overfitted. The performance of the DNN model exhibit 9.91% mean absolute percentage error. We compared the DNN model with linear prediction equations proposed by Kishitani, Hamada, and Shirayama equation. The carbonation velocity coefficient (mm/√year) which were calculated using accelerated carbonation experiment and compared with DNN model and linear prediction equations. The mean absolute percentage error of DNN model was 12.00%, which was smaller than that of the linear prediction equations of Kishitani, Shirayama and Hamada. © International Journal of Sustainable Building Technology and Urban Development.