A High-resolution Large-eddy Simulation Framework for Wildfire Predictions using TensorFlow

As the impact of wildfires has become increasingly more severe over the last decades, there is continued pressure for improvements in our ability to predict wildland fire behavior over a wide range of conditions. One approach towards this goal is through coupled fire/atmosphere modeling tools. While significant progress has been made on advancing their physical fidelity, existing modeling tools have not taken full advantage of emerging programming paradigms and computing architectures to enable high-resolution wildfire simulations. By addressing this gap, this work presents a new wildfire simulation framework that enables landscape-scale wildfire simulations with physical representation of the combustion at affordable computational cost. This is achieved by developing a coupled fire/atmosphere model in the TensorFlow programming paradigm, which enables highly efficient and scalable computations on Tensor Processing Unit (TPU) hardware architecture. To validate this simulation framework and demonstrate its efficiency, simulations of the prescribed fire experiment FireFlux II (Clements et al., 2019) are performed. By considering a parametric study on the mesh resolution, we show that the global quantities such as volumetric heat release and fire-spread rate are insensitive to the horizontal mesh resolution within a range between 0.5 m and 2 m, which is sufficient for predicting fire intermittency and dynamic fire properties associated with fine-scale turbulent structures in the atmospheric boundary layer.