Transfer learning-based method for predicting axial and eccentric compressive capacity of concrete-filled steel tube columns with iron tailings sand

Concrete-filled steel tube (CFST) columns incorporating iron tailings sand (ITS) provide dual environmental benefits by reducing natural sand consumption and promoting the utilization of iron tailings waste. However, the limited availability of experimental data hinders the accurate prediction of their compressive capacity. To address this challenge, a transfer learning-based prediction framework was developed using a source dataset of 333 conventional CFST specimens and a target dataset of 41 CFST specimens with ITS. Based on conventional machine learning models, transfer learning models were developed by incorporating a Sigmoid scheduling function and a boosting strategy to enhance knowledge transfer and prediction accuracy. Model hyperparameters were optimized using five-fold cross-validation, and the overall performance was evaluated through TOPSIS. The results indicate that the CFST columns achieved the optimal compressive performance at an ITS replacement ratio of approximately 60%. Among all models, the transfer learning-based model Tr-HGBT exhibits the best comprehensive performance, achieving an R 2 of 0.978, RMSE of 117.658 kN, MAE of 63.072 kN, and MAPE of 4.2%. Compared with the baseline model HGBT, Tr-HGBT improves R 2 by 4.9% and reduces RMSE by 31.4%. Feature importance analysis reveals the following overall ranking: D>e>L>t>f y >f cu >r>cross section type , where the SHAP values of D are 517.30 for r = 0 and 345.52 for r ≠ 0. The novelty of this study lies in the development of a Dual-Learner Boosting-based Transfer Learning framework for CFST columns with ITS under small-sample conditions and the establishment of an interpretable GUI prediction tool, providing an efficient approach for the design and engineering application of sustainable CFST structures.