Identifying Relations Between Deep Convection and the Large-Scale Atmosphere Using Explainable Artificial Intelligence

Deep moist convection is responsible for a large fraction of rainfall in the tropics, but the interaction between deep convection and the large-scale atmosphere remains poorly understood. Here, we apply machine learning techniques to examine relationships between the large-scale state of the atmosphere and two measures of its convective state derived from radar observations in northern Australia: the total area occupied by deep convection and the degree of deep convective organization. Specifically, we use a neural net to predict convective area and convective organization as a function the large-scale state, defined as the thermodynamic and dynamic properties of the atmosphere averaged over the radar domain. Building on research into explainable artificial intelligence, we apply so-called “attribution methods” to quantify the most important large-scale quantities determining these predictions. We find that the large-scale vertical velocity is the most important contributor to the prediction of both convective measures, but for convective area, its absolute and relative influence are increased. Thermodynamic quantities like atmospheric moisture also contribute to the prediction of convective area, but they are found to be unimportant for convective organization. Instead, the horizontal wind field appears to be more relevant for the prediction of convective organization. The results highlight unique aspects of the large-scale state that are associated with organized convection.