Enabling Smart Dynamical Downscaling of Extreme Precipitation Events With Machine Learning

The projection of extreme convective precipitation by global climate models (GCM) exhibits significant uncertainty due to coarse resolutions. Direct dynamical downscaling (DDD) of regional climate at kilometer-scale resolutions provides valuable insight into extreme precipitation changes, but its computational expense is formidable. Here we document the effectiveness of machine learning to enable smart dynamical downscaling (SDD), which selects a small subset of GCM data to conduct downscaling. Trained with data for three subtropical/tropical regions, convolutional neural networks (CNNs) retained 92% to 98% of extreme precipitation events (rain intensity higher than the 99th percentile) while filtering out 88% to 95% of circulation data. When applied to reanalysis data sets differing from training data, the CNNs' skill in retaining extremes decreases modestly in subtropical regions but sharply in the deep tropics. Nonetheless, one of the CNNs can still retain 62% of all extreme events in the deep tropical region in the worst case.