A machine learning-based approach to evaluate the fire resistance of timber columns

Abstract

Timber construction offers substantial advantages in terms of sustainability, ease of modular construction, and aesthetics. In recent years, structural members made from timber have been increasingly used in residential and commercial buildings. When used in building applications, structural timber members must meet required fire resistance ratings. Fire resistance of timber members can be evaluated by standard fire testing, which requires sophisticated and expensive testing facilities, and is labor-intensive and time-consuming. An alternative to fire testing is using advanced numerical models, which is computationally intensive. This study leverages machine learning (ML) methods to overcome the complications of fire testing and complex numerical modeling. Therefore, an ML workflow is developed and applied to a database of 70 fire tests reported in the literature to accurately evaluate the fire resistance of timber columns. The input parameters considered for training the models comprise geometric and material properties and loading conditions during the fire tests. Four different ML algorithms were implemented, namely, multiple linear regression, support vector machines, light gradient boosting, and random forest. The ML models benefited from an automated training procedure comprising hyperparameter tuning and cross-validation. Furthermore, Shapley additive explanations were used to interpret the relationship between timber column geometry, material properties, and fire resistance. The results show that random forest provides a higher accuracy with an R-squared of 0.84 on the test set, where column capacity, width, depth, and load level are the most critical parameters.