Deep Learning Modelling and Generalisation of Carbonation Depth in Fly Ash Blended Concrete

Abstract

An accurate concrete carbonation predictive model is needed to ensure effective structural health monitoring of concrete, as a result of nonlinear behaviour of concrete carbonation and uncertainties of environmental factors affecting the progress of carbonation. In reality, the performance of a developed predictive model lies on its ability to accurately generalise (simulate) the input distribution within a given new dataset. The previous machine learning models in the studies for concrete carbonation depth rely on the performance of the validation subset used for updating network weights during training and offer no generalisation performance of the model. This study developed a recurrent neural networks (RNN) model to predict the carbonation of fly ash blended concrete. Eighteen input datasets that focused on the chemical parameters of cement and fly ash were used in addition to established environmental factors. A total of 534 sets of experimental data were extracted from 13 published works of the literature. A leave-one-out cross-validation (training and generalisation subsets) was used. The developed RNN model was compared with the artificial neural networks (ANNs), support vector machine (SVM), radial basis function neural networks (RBFNNs), multi-linear regression (MLR) and bagged and boosted decision trees models that have been previously used to model carbonation depth in concrete. The results shows that the proposed RNN has a high prediction and generalisation capability in the evaluation of carbonation depth, and the highest correlation coefficient of 0.94 than the comparing models. Further, the generalisation subsets show that the results of natural exposed conditions are more reliable than the accelerated test cases.