A predictive model of concrete corrosion due to sulphuric acid using artificial neural networks

Abstract

This dissertation investigates the level of acid‐resistance of concrete degradation. Concrete specimens obtained from four mixtures (M1, M2, M3 and M4) were prepared with calcareous, siliceous and a blend of calcareous and silica sand; and then, tested in low (30 g/l) and highly (200 g/l) concentrated sulphuric acid solutions. To this end, an architecture of artificial neural networks (ANNs) was implemented to predict the performance of concrete specimens due to sulphuric acid solutions. Neural networks were composed with one hidden layer for one input and output layer. Nine input parameters were: cement composition, proportions of coarse and fine aggregates, water content, and compressive strength, weight loss of concrete, time impacting corrosion, acid concentration and sulphur concentration. Thickness expansion and concrete conductivity are used as output targets to evaluate the degree of deterioration. In this study, the learning through ANNs from training data sets have been proved to be better than measured data. Excellent results were found with a coefficient of determination (R2 ) of 0.9989, 0.9999, 0.9989 and 0.9998, respectively for the four mixtures M1, M2, M3 and M4 using siliceous aggregate. Also, the results show that two ANN models performed with both the thickness (expansion) and the electrical conductivity can successfully learn the prediction of concrete corrosion. In both low and highly concentrated sulphuric acid condition, the model thickness was more accurate in predicting concrete corrosion compared to the model conductivity. The lowest error in neural networks was provided by the mixture (M2) for the concrete using siliceous aggregate. For this purpose, the root mean squared error (RMSE) and the average absolute error (AAE) were of 0.0049 and 0.0048 % respectively.