Knowledge-driven machine learning for predicting corrosion rate of steel in mortar

Existing machine learning (ML) models for predicting the corrosion rate of steel in concrete are typically trained on data collected under steady state conditions, since the micro-environments around the steel under real world conditions are dynamic and difficult to monitor in real time. Capturing these dynamics for robust ML prediction remains a major challenge. In this work, corrosion experiments under cyclic drying-wetting conditions were conducted to replicate dynamic chloride ingress induced corrosion processes in reinforced concrete. By integrating mass transport modeling, sensor measurements, and experimental analysis with ML methods, a knowledge-driven ML model was developed. The results indicate that the use of the convection-diffusion equation effectively simulated chloride transport, providing time-varying chloride profile data around the steel. The chloride-to-hydroxide ratio and corrosion potential exhibit a strong correlation with the corrosion rate, highlighting their significance in corrosion prediction. Among the tested ML algorithms, the random forest model achieved the highest accuracy, further improving when time data was included as an input feature. Furthermore, model performance declined when training and evaluation were conducted using time-series based data partitioning rather than random partitioning, underscoring the strong temporal dependency of corrosion rate prediction. These findings demonstrate that incorporating time-dependency and physical insights alongside data-driven approaches can significantly enhance prediction accuracy and robustness, providing a promising pathway for reliable corrosion rate prediction in dynamic environments.