102nd American Meteorological Society Annual Meeting

A Machine Learning Plume-Resolving Model Implementation over North America for Mega-Wildland Fire Smoke Impacts on Distant Planetary Boundary Layers

, , , , , , , , , , , and

Recent persistent droughts and extreme heatwave events over the Western states of the US are creating highly favorable conditions for mega wildland fires. The IPCC AR6 report suggests that such extreme events will continue occurring with increasing frequency and intensity over forested regions, globally. We have shown that smoke produced from such wildland fires, which contain the burnt biomass carbon-based particulate matter at 2.5 microns, can be dynamically transported 1000s of km to affect the composition of distant atmospheric planetary boundary layers. Under normal wildfire events, the smoke plume would essentially rise, and dissipate but mega wildfires can generate Pyro Cumulonimbus clouds that can extend up to 6 miles into the troposphere, even into the stratosphere, and persist for long time periods. As a result, atmospheric states can transport wildfire smoke from the Western US to cities over the Eastern US, significantly affecting the air quality of distant communities. We propose to initially target the long-distance impact of smoke from the expected annual reoccurrence over N. America mega wildland fires, which are poorly forecast today by current air quality predictions. We employ the NASA unified Weather Regional Forecast (NuWRF) model with the SFIRE coupled surface fire spread and fuel moisture model, which is structured to meet the WRF guidelines for incorporation as a physical process module, along with a fuel data assimilation scheme in addition to assimilation of multiple satellite and ground observing systems. We bypass for now the computationally intensive WRF-SFIRE-CHEM by implementing the NuWRF-GOCART-SFIRE model using a tri-nested grid setup employing a neural net model acceleration of a 3km x 3km outer grid over N. America, a nested sub-kilometer resolution plume resolving inner grid of 350m x 350m over US and a 2D SFIRE 39m x 39m grid over the MODIS/VIIRS FRP identified region of the Western US, to forecast the transport of wildland fire smoke and their impacts on air quality over the Eastern US. We will present the preliminary test results of machine learning emulations and accelerations of this ultra-high spatial resolution of the NuWRF-GOCART-SFIRE model, which directly resolves the spread and intensity of mega wildland fire plumes and improves smoke transport. Current available operational air quality models, of relatively coarse resolutions of 3km or more, rely on simplified sub-grid plume parameterizations to account for the plume dynamics and often result in significant transport deficiencies resulting from poorly resolved plume intensities and heights. We will present NuWRF-GOCART-SFIRE performance results for the numerical experiments coupling fire progression with smoke production and resolved plume dynamics. These computations, using a combination of neural architecture searches and reverse knowledge diffusion, have the potential for accelerating our model implementation performed at an order of magnitude higher spatial resolution than the operational models thereby enabling the simulation studies of potential smoke penetrations of PM2.5 on distant planetary boundary layers over the Eastern US.


American Meteorological Society

UMBC ebiquity