Machine learning prediction of corrosion rate of steel in carbonated cementitious mortars

Corrosion rate (i.e., corrosion current density), a crucial kinetic parameter for predicting and modeling service-life performance of reinforced concrete structures, can be estimated using empirical and physics-based models. However, existing corrosion rate predictive models have limited applicability and lowered accuracy to complex scenarios when the material composition and corrosion conditions are substantially varied. In this work, machine learning approach is explored to predict the corrosion rate of steel based on a comprehensive experimental dataset. To consider the wide variation of binder composition and corrosion environment in marine concrete, the experiment involves a broad range of mixture design parameters and relative humidity levels. The result shows that electrical resistivity is the most relevant factor to corrosion rate, and its relationship with corrosion rate is highly dependent on the chloride-to-hydroxide concentration ratio ([Cl−]/[OH−]). Of the various machine learning algorithms tested, support vector regression demonstrates the highest predictability for corrosion rate. Using the feature selection method, electrical resistivity, pore solution composition ([Cl−]/[OH−]), cement proportion, and corrosion potential are identified as the major related features for corrosion rate prediction of steel in carbonated cementitious mortars. The results demonstrate that machine learning is a promising tool for predicting the corrosion rate of steel embedded in cementitious mortars.