Reliable and Interpretable AI for CFST Column Safety Assessment

This chapter explores the application of a unified AI-based framework for assessing the reliability of concrete-filled steel tube (CFST) columns, a critical component in modern construction. The framework combines Monte Carlo Simulation (MCS) for reliability assessment, CatBoost for predictive modeling, and SHAP for interpretability. The study utilizes a dataset of 663 experimental CFST column specimens, modeling input variables as independent random variables with normal distributions to reflect material and geometric variability. The CatBoost algorithm, a gradient-boosting technique, is employed for regression, with hyperparameter tuning achieved through five-fold cross-validation. The results demonstrate a strong correlation between the predicted and true failure probabilities, validating the model's accuracy. The framework classifies CFST columns into three safety levels based on the reliability index, achieving 100% classification accuracy. SHAP analysis identifies the outer diameter as the most critical determinant of failure likelihood, followed by wall thickness and column length. The study concludes by proposing future research directions, such as extending the framework to eccentric and slender columns and incorporating time-dependent degradation effects for long-term safety evaluation. This comprehensive approach offers a practical and efficient method for data-driven, interpretable, and reliable assessments of structural design, making it an invaluable resource for professionals in the field.