Evaluating fire resistance of timber columns using explainable machine learning models

The global attention to using timber products as sustainable construction material urges careful assessment of their performance against different hazards, particularly fire. However, the current methods prescribed by design codes for evaluating the fire resistance of timber components tend to underpredict the outcomes of standard fire resistance tests, and lack interpretability due to the use of semi-empirical equations. This study develops explainable data-driven models to predict fire resistance of timber columns using geometry- and material-related properties based on a comprehensive experimental database. Nine different single and ensemble-based machine learning algorithms were trained, and their performance was optimized through rigorous hyperparameter tuning and feature selection. The best models were then interpreted using partial dependence and Shapley plots to infer the underlying relationship between fire resistance and column properties. Lastly, the models' predictive capabilities were compared to available prescriptive equations. The results show that a random forest-based model provides the best performance with an average ratio of predicted to observed fire resistance of 1.03 on the test set. The random forest prediction is mainly governed by column capacity at ambient temperature, and to a lesser degree, columns' cross-section dimension. In addition, the partial dependence plots indicate that the effect of density, modulus of elasticity, length, and compressive strength on fire resistance was not notable. Finally, while the considered prescriptive equations consistently underpredict fire resistance, the random forest model provides a consistently accurate and balanced prediction.