Using a Convolutional Neural Network to Disentangle Environmental Differences between Developing and Non-Developing African Easterly Waves

African Easterly Waves (AEWs) are the precursor atmospheric disturbances to most hurricanes in the North Atlantic basin. Previous studies have shown favorable environments for AEWs to develop into tropical storms include warm sea surface temperatures off the west coast of Africa, moisture in the atmosphere at mid-levels, and weak vertical wind shear. This work will explore the ability of a convolutional neural network (CNN) to distinguish between AEWs that will or will not develop into tropical storms when the disturbances are still over Africa. We will use environmental variables from the ERA5 reanalysis and a database of objectively tracked AEWs in ERA5 (Lawton & Majumdar 2023) to train the CNN. We will input different environmental variables at various levels of the atmosphere to see which variables optimize the success of the CNN in discriminating between developing and non-developing AEWs. Results from two different CNN experiments will be shared: (1) CNN's ability to discriminate at a set time after AEW formation (e.g., between 12 and 60 hours) and (2) CNN's ability to discriminate at a set lead time before the maximum intensity of the AEW. Of particular focus will be which environmental moisture variables provide the most skillful outcomes and whether the CNN performs better with native variables like U and V winds or derived variables like relative vorticity and convergence.