Fusing experimental and FEM-based knowledge: a transfer learning model for inferring steel corrosion in reinforced concrete structures

• A transfer learning-based framework enables Bayesian inference of SWL from CCW observations under natural corrosion conditions. • GFK and TCA align SWL distributions features across corrosion conditions. • A semi-supervised model introduces FEM-generated pseudo-labels to enhance domain adaptation while mitigating FEM-induced bias. • The proposed model significantly improves CCW prediction under natural corrosion conditions. • Bayesian inference based on the transfer learning model provides robust range estimation of SWL. Corrosion-induced crack width (CCW) can be readily obtained through visual inspection of reinforced concrete structures and serves as a proxy indicator of steel corrosion based on its relationship with steel weight loss (SWL). Although accelerated corrosion tests and FEM simulations provide a cost-effective data source to derive the relationship between CCW and SWL, the calibrated models often fail to generalize under natural corrosion conditions. This study proposes a transfer learning-based model that fuses knowledge from accelerated corrosion tests and FEM simulations to predict CCW from SWL distribution under natural corrosion conditions. This model is then used as the forward model in a Bayesian inference scheme to estimate SWL distributions. Specifically, the proposed approach first uses an unsupervised transfer learning model that combines the geodesic flow kernel (GFK) and transfer component analysis (TCA) to align the marginal distributions of SWL across different corrosion conditions. This unsupervised model is extended to a semi-supervised approach by introducing FEM-generated pseudo labels for CCW under natural corrosion condition, which preserves the statistical dependence between SWL and CCW into the TCA projection in a structure-aware manner. Numerical results demonstrate that the proposed transfer learning models effectively transfer the SWL-CCW mapping learned from a combination of accelerated corrosion tests using the galvanostatic method and FEM simulations to natural corrosion conditions represented by the artificial chloride environment (ACE) method. The unsupervised GFK-TCA method improves CCW prediction accuracy under ACE conditions by 115% compared to the non-transfer learning baseline. Furthermore, the semi-supervised GFK-TCA method achieves an additional 18.0% improvement in accuracy over the unsupervised GFK-TCA method. Based on the transfer learning model, Bayesian inference yields a range estimation of SWL distribution that covers nearly 70.0% of the observations in the RC specimens corroded by ACE method.