Emulating Wildfire Plume Injection Using Machine Learning Trained by Large Eddy Simulation (LES)

Abstract

Wildfires have a major influence on the Earth system, with costly impacts on society. Despite decades of research, wildfires are still challenging to represent in air quality and chemistry-climate models. Wildfire plume rise (injection) is one of those poorly resolved processes and is also a major source of uncertainty in evaluating the wildfire impacts on air quality. Studies have shown that current plume rise models are subject to large uncertainties, including the Freitas Scheme, a widely used 1-dimensional, cloud-resolving subgrid model. In this work, a new machine learning-based plume rise emulator is presented, trained using a high-resolution, turbulence-resolving large eddy simulation (LES) model coupled with microphysics. The preliminary results show that this machine learning emulator outperforms the benchmark model, the Freitas scheme, in both accuracy and computational efficiency. Furthermore, a bagging ensemble is built to further increase the robustness and to battle internal variability. Efforts have been made to ensure that the machine learning emulator is robust, transparent, and not overtrained, and the results are interpretable and physically sound. Overall, this Plume Rise Emulating System using Machine Learning (PRESML) is a promising solution for regional and global air quality and chemistry-climate models.