Data-Driven Selection of Parametric and Travelling Fire Models for Performance-based Fire Design

Abstract Selecting appropriate design fire models remains a key unresolved challenge in performance-based structural fire engineering, where the applicability of the Eurocode EN 1991-1-2 parametric fire curve and travelling fire (TF) models is not clearly defined. This study proposes a quantitative threshold for selecting between the EN and TF models based on full-scale experimental results. A total of 222 available full-scale natural fire experiments are summarised and used as benchmark data. For each experiment, average compartment fire temperature–time curves are predicted using the EN and TF models, respectively, and their predictive accuracy is quantitatively assessed to determine the best-performing model for each fire test. A logistic regression (LR) model is then derived from the available dataset to predict the best-performing model for a given set of fire inputs. As the required TF model inputs, namely the peak fire temperature, heat release rate per unit area, and fire spread rate, were not reported in the experimental data, a sensitivity study is conducted to evaluate their influence on the derived LR model. The results show that the LR model is most sensitive to the assumed fire spread rate. Therefore, three applicability segments are introduced into the LR model to cover a wider range of fire scenarios. The proposed LR model achieves an average training accuracy of 0.736 and an ROC–AUC of 0.808. A probabilistic risk assessment case study using Monte Carlo simulation demonstrates that the proposed LR model reduces both over- and under-design compared with EN-only or TF-only methods, providing a more balanced and reliable framework for current performance-based structural fire design.