Evaluation of carbonation progress using AIJ model, FEM analysis, and machine learning algorithms

Abstract

Increasing costs due to failure and reconstruction highlight the importance of concrete durability research. Carbonation of concrete, which can accelerate corrosion, is one of the major deterioration mechanisms in reinforced concrete structures. Experimental data has been used to develop carbonation prediction models, however, the service life predicted from various models can differ significantly. A potential solution is the application of an artificial neural network algorithm, which simulates the human nervous system, to evaluate concrete carbonation. In this study, the possibility of applying machine learning to predict concrete carbonation behavior is evaluated. A deep learning model, which has the best learning power among various machine learning models, was applied. This model is structured such that hidden layers of hierarchical artificial neural networks are formed in several layers. Existing carbonation experimental data (water-to-cement ratio 0.55 and 0.65, temperature 20 °C, relative humidity 60%, and CO2 concentrations 5% and 20%) was predicted by using the deep learning model which was also compared with the results of two other models – AIJ model and FEM analysis. Under the test conditions, the differences in carbonation rate coefficient between experimental data and the deep learning results ranged from 0.01 mm/year to 0.10 mm/year for the different water-to-cement ratios and CO2 concentrations. These results were comparable though somewhat better than results from FEM analysis, which showed corresponding differences ranging from 0.08 mm/year to 1.04 mm/year. The results were significantly better than the AIJ model, which showed differences ranging from 0.32 mm/year to 2.34 mm/year. These preliminary results suggest that a deep learning algorithm can be used to accurately predict concrete carbonation results.