Machine learning identification of dry Intrusion outflows in present and future climates

Dry intrusions (DIs) are the key descending airstreams within extratropical cyclones. They can exacerbate the impacts of mid-latitude weather systems through their interactions with the boundary layer, enhancing atmosphere-surface interactions, and affecting frontal precipitation. DIs have been identified in the past using Lagrangian trajectory analysis, which has enabled studies into the climatology, variability, and characteristics of these airstreams. However, the potential futures of DIs, and the impact of climate change on them, has been unexplored due to the computational and data demands of this approach.

In this work, a convolutional neural network – DI-Net – is trained to identify DI outfow objects from a Lagrangian-identified dataset across the Northern Hemisphere, using information on relative and specific humidity, and topography from ERA5. The model performs well at capturing the main features of the DI climatology. DI-Net is then applied to historical and future climate model data from MRI-ESM2.0 to evaluate the climate model and investigate future changes. We present some of the challenges associated with developing a machine learning model for use with climate data.

The climate model represents the frequency of DIs well. In the most extreme warming scenario (SSP5-8.5), the frequency of DI outflows decreases in general, with increases across western Europe, consistent with the projections of the extratropical stormtracks seen in CMIP6 models. This study demonstrates the utility of the machine learning model to allow us to investigate the future of DIs, and eventually to understand more about how their impacts may change.