Emulating Wildfire\nPlume Injection Using Machine Learning\nTrained by Large Eddy Simulation (LES)

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