In numerical fire simulations, the calculation of thermal feedback from the flame to the solid and liquid fuel surface plays a critical role as it connects the fundamental gas-phase flame burning and condensed-phase fuel gasification. However, it is a computationally intensive task in CFD fire modelling methods because of the requirement of a high-resolution grid for calculating the interface heat transfer. This paper proposed a real-time prediction of the flame-to-fuel heat transfer by using simulated flame images and a computer-vision deep learning method. Different methanol pool fires were selected to produce the image database for training the model. As the pool diameters increase from 20 to 40 cm, the dominant flame-to-fuel heat transfer shifts from convection to radiation. Results show that the proposed AI algorithm trained by flame images can predict both the convective and radiative heat flux distributions on the condensed fuel surface with a relative error below 20%, based on the input of real-time flame morphology that can be captured by a larger grid size. Regardless of growing or decaying fires or puffing flames induced by buoyancy, this method can further predict the non-uniform distribution of heat transfer coefficient on the interface rather than using empirical correlations. This work demonstrates the use of AI and computer vision in accelerating numerical fire simulation, which helps simulate complex fire behaviours with simpler models and smaller computational costs. Highlights A total framework between AI model and fire simulation software is designed to further enhance the reliability of AI-based fire simulations. A standard pool fire simulation database is built using numerical model recommend by the IAFSS Computation Group. A deep learning model is developed to predict both the convective and radiative heat flux distributions on the condensed fuel surface using numerical images database. The demonstration showcases the application of AI and computer vision to accelerate numerical fire simulation.
Read full abstract