PNNL

Abstract

An ensemble of machine learning based models is used to investigate environmental controls as well as the origins of regional variability of Boreal winter rainfall statistics in the tropics. Precipitation radar observations from field campaigns at Darwin Australia, Manaus Brazil, and Equatorial Indian Ocean along with corresponding environmental variables from ERA5 reanalysis are used to train machine learning models that predict the growth rate of area of a given precipitation intensity. The environmental variables considered are hourly column integrated precipitable water, CAPE, CIN, lower and upper tropospheric wind shear all averaged over 1-degree boxes within the radar domains. The 20-member ensemble of models is comprised of single hidden layer neural network algorithms with comparable accuracies but different hyperparameter choices. The models show an abrupt increase in the growth rate of precipitation area near 60mm PW and at CAPE of about 300J/kg. Growth is also favored by CIN less than 180J/kg. The impact of shear appears to be non-linear in that both weak (1-5m/s) and very strong (>10m/s) being favorable for growth. For the most part PW and CAPE account for the differences between DARWIN and the DYNAMO domains. The strong upper-level shear distinguishes the DYNAMO domain from the other two, while AMAZON on the other hand is comparatively dry and has stronger CIN. Over Amazon upper-level shear is stronger than over DARWIN, while the low-level shear stronger for the latter. the low-level shear is stronger compensating for effect of the former on the precipitation area growth rate.