Real-time prediction of structural fire responses: A finite element-based machine-learning approach

Finite Element (FE) methods have been widely adopted in structural design to analyze the mechanical performance quantitatively, while most comprehensive FE models are computationally intensive and time-consuming, thus failing to be applied in a real-time context. To facilitate real-time prediction of structural fire response, this study develops an FE-based machine learning (ML) framework to predict structural displacement based on temperature data, which integrates the advantages of accuracy from FE methods as well as efficiency from ML techniques. FE models of a steel frame structure are developed for fire analysis in the present study. A numerical database of structure responses subject to hundreds of fire scenarios is established by extracting the results from FE models. Four ML models are trained based on the numerical database with temperature data as input and displacement data as output, among which Random Forest (RF) and Gradient Boosting (GB) models are found to outperform other models in terms of predictive accuracy. From the numerical results, the R square of the predictive mid-span displacement from RF and GB models could reach up to 0.99 when 1000 fire scenarios are included in the training dataset. All models are robust against the noise in temperature data when the signal-tonoise ratio is greater than 15. The FE-based ML framework developed in this study is of high potential to be applied to real-time structural response prediction during fire emergencies.

Automated summary

This study developed an FE-based machine learning framework for predicting structural displacement based on temperature data, with RF and GB models outperforming others in terms of predictive accuracy. The developed framework shows high potential for real-time structural response prediction during fire emergencies.