Machine Learning‐Based Prediction of Temperature Evolution in Steel Compartment Fires

Abstract

In performance‐based structural fire design, accurate prediction of compartment gas temperature is essential for evaluating structural response under fire conditions. This study presents a machine learning‐based surrogate model for predicting the time evolution of hot gas layer temperatures at the center of the compartment ceiling in single‐storey steel compartments exposed to fire. An Artificial Neural Network (ANN) is developed using supervised learning with backpropagation and the AdamW optimization algorithm. A dataset comprising 15,000 fire simulations for training and 1,500 simulations for testing is generated using a Python‐based CFAST framework. Input parameters include fire size and location, compartment geometry, vent type and configuration (doors and skylights), opening positions, radiative fraction, and time. The ANN accurately reproduces temperature‐time histories across a wide range of scenarios, showing strong agreement with CFAST predictions. A constrained random sampling strategy is adopted to improve computational efficiency and reduce the number of simulations required for training. Once trained, the surrogate model produces temperature predictions within seconds, several orders of magnitude faster than new CFAST simulations. This enables rapid design iterations, parametric studies, and probabilistic fire risk assessments, demonstrating the potential of data‐driven surrogate models as efficient tools for early‐stage structural fire design.