An end-to-end transformer-based TabPFN framework for predicting mechanical strength degradation of fiber-reinforced polymer composite reinforcements embedded in marine concrete

Abstract

Experimental investigations of fiber-reinforced polymer (FRP) bars in realistic marine concrete environments are costly and time-intensive, resulting in scarce test data that poses a major obstacle to reliable degradation models for FRP’s strength prediction. To address this challenge, this study presents a Transformer-based Tabular Prior-Data Fitted Network (TabPFN) framework for predicting the tensile strength retention (TSR) of FRP bars embedded in marine concrete environments. This approach integrates limited experimental data with the model’s pre-trained prior knowledge, offering a data-efficient and interpretable alternative to conventional methods (e.g., empirical regression and task-specific training approaches). Furthermore, SHapley Additive exPlanations (SHAP) analysis was employed to interpret the contributions of both material and environmental variables on the degradation of FRP bars embedded in concrete. Based on the robust predictions obtained previously, the long-term TSR of FRP bars embedded in concrete was further evaluated, bridging small-sample learning with durability assessment. Finally, the presented model was encapsulated into an open-access, user-friendly prediction platform, facilitating further research and engineering applications.