Transfer learning enables prediction of steel corrosion in concrete under natural environments

Existing machine learning (ML) models for corrosion rate prediction of steel in cementitious materials are typically established based on laboratory datasets obtained under controlled material and environmental conditions, which questions their applicability to more realistic and complex scenarios. Transfer learning (TL), as a branch of ML, can extract knowledge from a source domain, which can be utilized to improve prediction accuracy on a target domain. In this work, a TL paradigm, grounded on an advanced ML model built for steel corrosion in mortars, is proposed to elevate the efficacy of existing ML models in forecasting corrosion rate of steel in concrete under natural environments. The results underscore the prominence of certain features, specifically electrical resistivity, chloride-to-hydroxide concentration ratio ([Cl−]/[OH−]), cement proportion, corrosion potential, porosity, and water content. In addition, the interplay of diverse quantities of features and feature amalgamations exercises a substantial influence on the performance of ML models. It is found that TL strategy enhances the ML model's predictability for corrosion rate in concrete under natural environments. The knowledge pertaining to steel corrosion under controlled laboratory conditions can be transferred to enhance the model's ability to predict steel corrosion in concrete under natural conditions. These results underscore TL's potential in enabling reliable corrosion rate predictions in existing in-service concrete structures, especially with limited data and deficient information for steel corrosion in concrete structures.