Predicting carbonation coefficient using Artificial neural networks and genetic programming

Concrete carbonation is considered an important problem in both the Civil Engineering and Materials Science fields. Over time, the properties of concrete change because of the interaction between the material and the environment and, consequently, its durability is affected. Conventionally, concrete carbonation depth at a given time under steady-state conditions can reasonably be estimated using Fick's second law of diffusion. This study addresses the statistical modelling of the concrete carbonation phenomenon, using a large number of results (827 specimens or samples, i.e. 827 is the number of data concerning the measurement of the carbonation coefficient in concrete test specimens), collected in the literature. Artificial Neural Networks (ANNs) and Genetic Programming (GP) were the Soft Computing techniques used to predict the carbonation coefficient, as a function of a set of conditioning factors. These models allow the estimation of the carbonation coefficient and, accordingly, carbonation as a function of the variables considered statistically significant in explaining this phenomenon. The results obtained through Artificial Neural Networks and Genetic Programming were compared with those obtained through Multiple Linear Regression (MLR) (which has been previously used to model the carbonation coefficient of concrete). The results reveal that ANNs and GP models present a better performance when compared with MLR, being able to deal with the nonlinear influence of relative humidity on concrete carbonation, which was the main limitation of MLR in modelling the carbonation coefficient in previous study. ANNs are commonly seen as a black box; in this study, an attempt is made to address this issue through Knowledge Extraction (KE) from trained weights and biases. KE helps to understand the influence of each input on the output and the influences identified by the KE technique are in accordance with general knowledge.