Interpretable Deep Learning for the Identification of Sudden Stratospheric Warming Events

An advanced understanding of stratospheric variability and its coupling to the troposphere is critical to improving the prediction of near-surface fields at subseasonal-to-seasonal timescale. In the most extreme case, a stratospheric sudden warming (SSW) event occurs and substantially perturbs the stratospheric circulation and, subsequently, exerts profound surface impacts. Interpretable deep learning could be a powerful tool in recognizing SSW spatial details and better categorizing the type of disrupted vortices. Here we apply a deep learning approach to identify SSW events from nonSSW ones using a global climate model with large ensembles. We start with a 1-dimensional case by using the stratospheric zonal wind of SSW events along the 60°N latitude to train neural networks with different complexity: logistic regression network, shallow neural network, and deep neural network. All neural networks can identify SSW events with a fairly high accuracy. To address the interpretability of how these neural networks learn to distinguish SSW from nonSSW events, we mask out the zonal wind fields with varying longitudinal windows to test if the spatial structure of disrupted vortices is decisive for the network performance. Neither shallow nor deep neural networks show apparent spatial dependence when the masking window is short, while logistic regression network gives strong spatial dependence centering around 160°W, where small variation and negative mean value of zonal wind appear. The dependence of shallow and deep networks emerges as the window length increases. To further explore the 2-dimensional spatial dependence, we further train a convolutional neural network exploiting the two-dimensional zonal wind fields in the Northern Hemisphere. Similar tests are performed by systematically masking out the zonal wind fields by a rectangular region with varying size. The spatial dependence of 2-dimensional neural network is largely consistent with 1-dimensional networks, but the spatial extents expand wider to the north of south of 60°N. The results highlight the capability of interpretable deep learning tools in learning the SSW spatial information and revealing the spatial dependence, which may carry out important implications for the prediction of SSW genesis.

Key words: interpretable deep learning, stratospheric sudden warming