North American Monsoon Prediction Using Causality Informed Machine Learning

Deep convection during the North American Monsoon (NAM) frequently initiates along or near higher terrain features and may subsequently descend from the terrain, creating new cells and convective outflows that sometimes have devastating impacts. In some cases, the convection becomes organized as mesoscale convective systems (MCSs) that propagate across urban population centers and create widespread flooding and damage. Regional convection-allowing models, such as the University of Arizona real-time Weather Research and Forecasting (WRF) model, often struggle with predicting the location, timing, and intensity of deep convection in this region. Past work has verified that the probability of detection decreases dramatically to the south and west of the Mogollon Rim in northern Arizona. We hypothesize that the low skill results from: the model’s inability to correctly handle the influence that the complex topography has on convective initiation in the region, and insufficient initialization of critical moisture and land surface temperature data. To address these deficiencies and create a more skillful prediction system, we developed a machine learning model that uses time series of high-resolution dynamical model forecasts with daily satellite observations of soil moisture and land surface temperature. Causation entropy is used to identify the most important spatial relationships amongst the model and satellite predictors. Recurrent neural networks (RNNs) are then activated for each grid point in the domain to produce probabilistic forecasts of rain rate and accumulation. Preliminary validation of model forecasts will be presented. The forecast system will be tested during the 2024 summer NAM and put into operation to support a proposed field campaign in summer 2025.