Fire resistance rating prediction of timber-to-steel connections and design optimization informed by explainable machine learning

Timber as a construction material is experiencing its renaissance, while fire safety is a critical factor for timber-based building design. Currently, the fire resistance rating of wood-steel-wood (WSW) connections is evaluated using empirical equations derived from experimental results. However, these equations consider a limited set of parameters and lack interpretability. This paper developed an explainable machine learning (ML) model considering comprehensive parameters related to connection’s configuration, based on 140 experimental and experimental-validated numerical data. The performances of various machine learning models are evaluated in terms of predicting the fire resistance rating of connections after hyperparameter tuning. The eXtreme Gradient Boosting (XGBoost) model outperforms other ML models ($R^2=0.93$) and empirical equations. The local sensitivity analysis (LSA), global sensitivity analysis (GSA), and SHapley Additive exPlanations (SHAP) analysis are conducted based on the XGBoost model to investigate the contributions of nine parameters to the fire resistance rating. Both sensitivity analysis and SHAP analysis identify timber thickness and load ratio as the primary factors influencing fire resistance. Finally, the calibrated XGBoost model is incorporated into a non-dominated sorting genetic algorithm (NSGA-II) to optimize the design, aiming to minimize the self-weight of the connection while maximizing the fire resistance rating and load-carrying capacity of the connection subjected to constraints on limited dimensions.