Predicting reinforcement corrosion in concrete based on CT data using a two-stage multilayer perceptron model

Reliable prediction of internal corrosion in reinforced concrete based on observable surface cracks and porosity remains a critical challenge. This study investigates the relationship between cracks, pores, and corrosion in reinforced concrete exposed to an aggressive environment. The patterns of cracks and corrosion are clearly revealed by CT scanning. The effect of cracks on corrosion is found to be stronger than that of pores. To reflect the nonlinear evolution trend of corrosion, a two-stage multilayer perceptron (MLP) model was developed to predict corrosion depth using both crack width and porosity as initial inputs. Subsequently, only crack width was employed with pseudo-labels for further predictions. The model demonstrated relative errors under 10% for most specimens, yielding results that were smoother and closer to the measured values compared with traditional regression methods. Sensitivity analysis revealed that the influence of crack width was more significant than that of porosity. To enhance accuracy, the model was also validated against an external dataset. Overall, this study presents a preliminary approach for linking concrete surface features to internal corrosion, highlighting its potential and the need for further validation. © 2026 Elsevier Ltd.