Applicability Assessment of XGBoost-Based Prediction of Smoke Layer Arrival Time Using Fire Experiment Data under Different Spatial Configurations

Abstract

This study investigated the applicability of a machine learning-based approach for predicting smoke layer arrival time using fire experiment data. Experimental datasets obtained from an infinite ceiling space and an underground passageway were used to develop an XGBoost-based prediction model with space type, ceiling height, heat release rate, and dimensionless radius (r/H) selected as input variables. The model demonstrated high predictive accuracy, achieving an R2 of 0.990, mean absolute error of 0.946 s, and root-mean-square error of 1.293 s. Feature importance analysis revealed that the r/H and space type were the most influential variables affecting the smoke layer arrival time. Compared with the existing prediction correlation, the proposed model exhibited superior predictive performance under each condition and successfully reproduced the experimental results, particularly for the underground passageway configuration. These findings confirm the initial applicability of machine learning-based prediction techniques for estimating smoke layer arrival time.