Steel Moment Frames with Reduced Web Section Connections: Bilin Parameters and Application of Machine Learning

This paper investigates the seismic performance of reduced web section (RWS) connections in steel moment frames. Such an outcome is reached by employing advanced machine learning and deep learning techniques to predict the modified Ibarra–Medina–Krawinkler (IMK) Bilin parameters for RWS connections across various configurations. A database of 154 specimens from experimental and finite element studies was analysed, focusing on non-composite RWS connections. The study develops and compares multiple predictive models, including Random Forest, Neural Networks, Support Vector Regression, Gradient Boosting, XGBoost, and deep learning, to predict IMK Bilin parameters with high accuracy. It was found that XGBoost consistently demonstrated superior performance across all assessed metrics, having error metrics (MSE, MAE, Pearson and Spearman correlations, MAPE, and ROC area under the curve) at least an order of magnitude below its counterparts or significantly outperforming. Feature importance analysis revealed complex interactions between geometric properties and connection behaviour, with web slenderness and span-to-depth ratio playing critical roles. Additionally, pushover analyses were conducted on 2-, 4-, 8-, 12-, and 20-story benchmark frames to associate the global seismic behaviour of frames with RWS and RBS connections. Results indicate that RWS connections often provide higher strength and comparable or superior ductility to their counterparts, especially in low to mid-rise structures. All RWS configurations achieved interstorey drifts exceeding 4%, meeting ANSI/AISC and EC8 performance targets. Hence, the work showcased contributes to the adoption of RWS connections for retrofitting purposes by providing a robust tool for structural engineers to incorporate RWS connections into performance-based seismic design.