Interpretable machine learning model for corrosion kinetics of steel in mortars

Existing methods for predicting steel corrosion lack a balanced approach that offers both high predictive accuracy and physical interpretability, along with explicit mathematical expressions suitable for practical engineering applications. This work utilizes a comprehensive corrosion dataset in conjunction with machine learning (ML) and the corrosion mechanisms of steel in cementitious media, to develop a corrosion state classification expression and a corrosion rate prediction expression. The results demonstrate that the established ML benchmark model achieved good performance in predicting both the corrosion state and rate of steel in mortars, achieving a predictive test accuracy of 0.96 and a test coefficient of determination (R 2 ) value of 0.92, respectively. The corrosion state classification expression, based on an interpretable ML method - symbolic learning (SL), combines the chloride-to-hydroxide concentration ratio ([Cl − ]/[OH − ]) and corrosion potential (CP) to assess the likelihood of corrosion occurrence, with a test accuracy of 0.83. The corrosion rate prediction expression, derived from SL and steel corrosion mechanisms, accounts for the corrosion rate of steel in pore solution (a function of [Cl − ]/[OH − ] and CP), as well as the effects of pore structure (function of porosity) and ion diffusion (function of electrical resistivity), yielding a prediction performance with a test R 2 of 0.95. The proposed corrosion rate expression captures key trends consistent with steel corrosion mechanisms, including the chloride threshold effect, a negative correlation with CP and electrical resistivity, and a positive correlation with porosity. The proposed ML models with explicit expressions provide a quantitative and physically interpretable description of corrosion behavior, offering a simple tool for practical corrosion assessment.