Prediction of stress concentration factors for concrete-filled steel tubular (CFST) T-joints and K-joints using artificial neural network

• Artificial neural network (ANN) models were developed to predict the stress concentration factors (SCFs) of concrete-filled steel tubular (CFST) T-joints and K-joints. • Geometrical effects on SCF were analyzed using Shapley additive explanations (SHAP) interpretability methods. • The impact of ANN hyperparameters on model performance was systematically evaluated. • ANN predictions were validated against finite element analysis (FEA) and regression based parametric equations. • The minimum dataset required for the ANN model to achieve acceptable accuracy was investigated. Concrete-filled steel tubular (CFST) T-joints and K-joints are widely used in CFST trussed arch bridges. Stress Concentration Factors (SCFs) are critical for evaluating joint fatigue life, as they capture local stress amplification that may initiate cracks under cyclic loading. This study develops an Artificial Neural Network (ANN) model to predict SCFs based on data from Finite Element Analysis (FEA) for 140 and 272 samples of CFST T-joints and K-joints, respectively. Input parameters include the diameter ratio ( β ), chord diameter-to-thickness ratio (2 γ ), thickness ratio ( τ ), and the brace-to-chord angle ( θ ). The data were split into 80 % training and 20 % testing subsets for each joint type. The ANN, implemented in Python using TensorFlow and the Keras API, showed high predictive accuracy, aligning closely with both FEA results and regression-based parametric equations, particularly when trained on a complete dataset. Shapley additive explanations (SHAP) analysis revealed that the influence of geometric parameters on SCF varies by joint type and loading condition, with the thickness ratio (τ) consistently exerting the most substantial effect. Even when trained on the minimal dataset comprising the maximum, middle, and minimum values, the ANN models maintained high accuracy, indicating their capability to capture nonlinear parameter interactions effectively. Additionally, multiple regression analysis confirmed that dataset diversity significantly influences prediction accuracy, underscoring the importance of selecting well-distributed dataset. The results suggest that a well-trained ANN model offers an efficient and reliable alternative to complex equations for predicting SCFs in CFST T-joints and K-joints.