Multi-fidelity modelling for uncertainty quantification of timber beam-column connections exposed to standard fire

Fire safety design of timber structures requires a comprehensive uncertainty quantification to identify factors that potentially influence the structural fire performance. Prevalent finite element (FE) models, however, have high computational cost to be employed in the uncertainty quantification. This paper presents a multi-fidelity modelling framework for uncertainty quantification of timber beam–column connections exposed to standard fire test, aiming to predict the structural response with limited high-fidelity data points. First, the high- and low-fidelity FE models for sequential thermal-mechanical analysis are introduced. The fire resistance times of the connections with random input variables are evaluated by the high- and low-fidelity models separately. Subsequently, multi-fidelity neural networks (MFNNs) models are trained to correlate both high- and low-fidelity data. The numbers of high- and low-fidelity data used for training the MFNN are determined based on the model’s performance on the validation set. With limited high-fidelity data, the developed MFNN is demonstrated to be considerably accurate in predicting the fire resistance time and displacement evolution of the connection. Then the MFNN is used for the uncertainty quantification including sensitivity analysis, SHapley Additive exPlanations (SHAP) analysis and reliability analysis. The impacts of input variables on the connection’s fire resistance time are quantified. The failure probability of the connection under different load ratios are assessed based on Monte Carlo simulation (MCS).