The vertical distribution of biomass burning aerosols (BBAs) is important in regulating their impacts on weather and climate. The plume-rise process affects the injection height of wildfire aerosols and interacts with air parcel lifting and cloud processes, but these processes are not physically represented in most global climate models. In this study, we designed two approaches to address this issue: a physical representation based on process understanding and a machine learning (ML) technique.

In the former approach we replaced the fixed vertical profiles of monthly BBA emissions in the DOE’s Energy Exascale System Model (E3SM) with an interactive fire plume-rise model, which depends on the ambient thermodynamic conditions and fire properties from observations. Daily BBA emission, superimposed with a fire diurnal cycle retrieved from satellite observation, was included in model simulations. The model showed improved agreement with satellite retrievals and in-situ observations. We also find that the new model produces a larger carbonaceous aerosol burden, leading to 0.13 W m^-2 warming at the top of atmosphere compared to the default E3SM.

Although the physically-based plume-rise model can explicitly simulate plume injection heights, it is computationally expensive. Therefore, in the second approach, we employed artificial intelligence and machine learning (AI/ML) techniques to develop an emulator for the plume-rise process. Specifically, we trained a multilayer perceptron (MLP) artificial neural network using fire properties and atmospheric thermodynamic fields as predictors, with the plume injection heights calculated by the first approach as the target variable. The training and evaluation results demonstrate that the MLP can reasonably predict plume injection heights. Finally, we compared the performance of experiments using E3SM with: 1) fixed; 2) plume-rise model simulated; and 3) ML emulated BBA emission vertical profiles in simulating smoke plumes and the radiative effects of BBAs.