Computational Fluid Dynamics and Artificial Intelligence for Building Fire Safety Design

By Ivan Defaz

As shown in Fig. a, due to its time consuming nature, the fire modelling analysis is often performed at a later design stage when most of the architectural details are determined. The analysis is only conducted to validate the safety requirements of the proposed design, influenced by the conventional idea of pass/no pass (Zeng et al., 2024).

Computational fluid dynamics (CFD), a branch of fluid mechanics that uses numerical analysis and data structures to analyze and solve problems related to fluid flow, can reproduce the transient fire development and thus provides a better predicion of the fire gas characteristics. However, it is seldom used in current engineering practices due to its time-consuming nature.

To facilitate the fire engineering design, a deep learning model with long short-term memory (LSTM) network is used to rapidly predict the fire gas temperature and velocity under the growing scenario. Fig. d shows a T-square growing fire scenario in the typical civil building, which is modelled with CFD to create a database. To apply long short-term memory (LSTM) network, this database is split into 75 % and 25% of data for training and validation, respectively.

Fig. b-c compare the temperature and velocity predictions by the CFD and artificial intelligence (AI) model at different radial positions in one of the scenarios. The CFD model reproduces the gradual increase of the gas characteristics with the growth of the fire heat release rate (HRR), although strong fluctuations can be found during both growth and early stages since the burning is an inherently turbulent process, especially in the near-fire region. On the other hand, the predicted results by the AI model are comparable to the CFD simulations. While each CFD simulation takes from 24 h to 48 h with a 32-core server, the AI model can provide a similar result of ceiling jet evolvement within 20 s, which is even faster than applying empirical correlations.

Although the AI model has good performance in the prediction of the trained scenarios, the performance of the AI model in unseen scenarios should be evaluated. In conclusion, the proposed AI model can provide prompt and accurate predictions of gas temperature and velocity under both steady and growing fire stages, which can enhance the design of the fire detection and sprinkler system.

To learn more, please contact Ivan Defaz, P. Eng.