Abstract: The current design and analysis of fire detection system are mainly based on the simplified unconfined ceiling jet model, which does not consider the effect of building structures on the smoke flow behavior. This work proposed a deep learning model based on conventional neural network (CNN) with UNet architecture which aims to provide quick and accurate prediction of the fire-induced temperature field in rooms with complex layouts. A numerical database with 136 fire scenarios was first established by considering different room layouts, fire locations and room heights. The result shows that the model can provide temperature field for a given building in seconds with an accuracy of up to 88%. This work can contribute to the safety design for buildings with complex architectural plans.

Key words: temperature field, fire detection, fire safety design, complex floorplan, deep learning