A Machine Learning Framework for Estimating Lightning Flash Rate and Optical Energy from Meteorological Variables

Lightning is the largest natural source of nitrogen oxides (NO_x) in the upper troposphere and plays a crucial role in atmospheric composition through its role as an ozone (O3) precursor and sink via OH of methane. Uncertainty regarding the production of lightning NO_x (LNO_x) is still prevalent despite numerous advancements in observations and atmospheric modeling. In this study, a convolutional neural network (CNN) is being trained to predict flash rate and mean optical energy per lightning flash, using various meteorological variables from NASA GEOS-5 simulations as input and data from the Geostationary Lightning Mapper (GLM) as ground truth. The training period will span the summer months of 2019 and 2020 and focus on the southeast United States. The mean optical energy per flash predictions generated by the CNN will be on a 0.5ºx0.5º grid, similar to the GEOS-5 grid. This approach will enhance existing lightning parameterization schemes for GEOS-5 (and other coupled Earth system models) by elucidating complex relationships between GEOS-5 meteorological variables (e.g., convective available potential energy, convective cloud fraction, updraft velocity) and GLM lightning characteristics. Meteorological variables to be used are those that have a known direct and strong relationship to flash rate, length and peak current. Lightning optical energy is directly related to the production efficiency of LNO_x and therefore the CNN will be used as a parameterization scheme that also estimates LNO_x per unit of predicted flash optical energy. This new parameterization scheme aims to improve lightning representation and advance our understanding of tropospheric chemistry through its eventual integration into the NASA GEOS model.