Mechanistic-data-driven prediction of axial compression behavior of circular concrete-filled steel tube at different ages after freeze-thaw cycles

This study investigates the axial compression performance of circular concrete-filled steel tube at different ages after freeze-thaw cycles. Axial compression tests were conducted on 24 specimens with varying concrete ages and numbers of freeze-thaw cycles. Results indicate that the mechanical properties of specimens deteriorated increasingly with more freeze-thaw cycles and earlier ages. Parametric analysis based on the validated finite element model reveals that enhancing the steel tube strength and wall thickness effectively mitigates freeze-thaw damage in specimens, while increasing the concrete strength improves frost resistance. A predictive formula is developed to estimate the ultimate strength of concrete-filled steel tube at different ages after freeze-thaw cycles. Machine learning models based on the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) algorithm are proposed for predicting the ultimate strength. The comparative results indicate that the WGAN-GP enhanced Artificial Neural Network (ANN) model achieves the highest predictive accuracy (Coefficient of Determination (R²) = 0.956, Root Mean Square Error (RMSE)= 11.772, Mean Absolute Error (MAE)= 9.036). Three hyperparameter optimization algorithms: Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Parrot Optimizer (PO) are employed to refine the enhanced ANN model, among which the PSO-optimized model demonstrates superior performance (R²=0.985, RMSE=6.99, MAE=5.4). Meanwhile, a Long Short-Term Memory (LSTM) model is developed to predict the load-displacement curves of concrete-filled steel tube at different ages after freeze-thaw cycles, achieving high predictive accuracy, with load predictions yielding R² of 0.957 and displacement predictions reaching R² of 0.97.