A semantic-aware transfer learning framework via natural language processing to predict the cross-material bond strength of composite bars with limited data

This study develops a novel transfer learning (TL) framework integrating natural language processing (NLP) and deep learning (DL) to predict the bond strength of fiber-reinforced polymer (FRP) composite bars in ultra-high-performance seawater sea-sand concrete (UHPSSC) using limited experimental data. The framework incorporates an NLP data processor to handle textual variables, a gated transformer to enhance parameter representations, and DL models for regression. Through leveraging the characteristics of seawater sea-sand concrete (SWSSC) and ultra-high-performance concrete (UHPC), it enables effective TL from large FRP–SWSSC/UHPC datasets to smaller FRP–UHPSSC datasets. Three pre-trained models were developed using FRP–SWSSC, FRP–UHPC, and combined datasets, respectively, and then fine-tuned on FRP–UHPSSC data to create three TL models. Compared with traditional machine learning (ML) models such as XGBoost and random forest (RF) trained directly on FRP–UHPSSC data, the proposed TL models achieved significantly higher predictive accuracy. Specifically, the optimal-performing TL model fine-tuned using combined FRP–SWSSC/UHPC data, achieved an R² of 0.914 and root mean square error (RMSE) of 0.75 on the test set, corresponding to a 13 % increase in R² and a 47.6 % reduction in RMSE relative to the optimal ML model. Furthermore, NLP-based textual representation improved the TL model's test R² from 0.876 to 0.914, highlighting the effectiveness in handling tabular data. The TL model also demonstrates semantic recognition capabilities, addressing variations in parameter expressions and dimensionality, enhancing robustness across scenarios. Overall, the proposed NLP-based TL framework effectively integrates heterogeneous material knowledge and substantially enhances predictive performance and generalization, offering a promising approach for data-limited engineering applications.