J8A.6 Advancing Wildfire Research Using Large Eddy Simulation (LES) and Machine Learning

Tuesday, 30 January 2024: 5:45 PM
345/346 (The Baltimore Convention Center)
Siyuan Wang, CIRES, CU Boulder, Boulder, CO; NOAA, Boulder, CO

Wildland fires are a natural phenomenon in many ecosystems, and have become a growing concern in the United States, posing a great risk to human health and properties. Efforts to improve air quality in the majority of US over the past decades show promising trends, except in wildfire-prone regions where air quality has been worsening. Despite decades of research, wildfire modeling still faces challenges, leading to uncertainties in evaluating the wildfire impacts on human health as well as short-term weather forecast skills. One fundamental challenge is that the spatial resolution at which the models are operated is much coarser than the spatial extent of most wildfires. As a result, several key processes in wildfires cannot be explicitly resolved. One of such fundamental yet poorly resolved processes is plume rise. It has been well documented that state-of-the-art plume rise models are subject to large uncertainties, with major impacts on smoke transport and air quality downwind.

Here I present the ongoing development of a machine learning emulator (ensemble-based) for wildfire plume rise. This emulator is trained using a high resolution, turbulence-resolving Large Eddy Simulation (LES) model. This LES model resolves plume rise process considering the impacts of the fire-induced buoyancy, turbulence, entrainment, and moisture processes. The LES model is physics-based and an ideal tool to study wildfire plume dynamics but is too computationally expensive to scale up. The emulator trained using this LES model offers remarkable computational efficiency therefore can be implemented into host models (e.g., air quality forecast systems). Rigorous measures are taken to mitigate overtraining, and the outcomes are physically sound and interpretable. Preliminary results show that this machine learning emulator trained using LES outperforms a widely used physics-based plume rise model in terms of mean smoke injection height and smoke profile shape. The development of this new plume rise emulator is part of a large project leveraging a wide variety of wildfire plume height products from geostationary/polar-orbiting satellites as well as ground radar network; these plume height products will be used to further evaluate the emulator. In summary, this project seeks to improve the wildfire representation in air quality forecast systems as well as chemistry-climate models.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner