Using an explainable neural network to identify tropical drivers of the Northern Hemisphere wave-5 trend pattern

The recent trend in the Northern Hemisphere summer atmospheric circulation resembles a Rossby wave with wave number 5. These quasi-stationary circumglobal Rossby waves are associated with extreme events, such as heatwaves, droughts, and floods, that can have catastrophic societal impacts. Therefore, understanding the drivers of these Rossby waves and evaluating their representation in climate models is a key scientific challenge. However, identifying the drivers of such patterns can be difficult because traditional approaches such as simple correlation analysis may not capture the complex, nonlinear interactions inherent in atmospheric teleconnections. To address this, explainable artificial intelligence (XAI) offers a promising alternative. 

In this study, we test the hypothesis that the observed trend is partially driven by changes in the tropical oceans, which can influence midlatitude weather patterns through tropical-extratropical teleconnections. Using an explainable neural network approach, we aim to identify key tropical regions that drive the midlatitude wave-5 pattern on subseasonal timescales. The methodology is composed of two steps. First, the neural network is trained to predict the wave-5 pattern using tropical outgoing longwave radiation (OLR) fields as input. Next, we apply layer-wise relevance propagation, an explainability technique, to identify which input features are most important for accurate predictions. This process generates heat maps highlighting tropical regions that are important for the generation of a wave-5 pattern. Subsequently, changes in sea surface temperatures (SSTs) and OLR in the identified regions can be assessed as well as their correlation to the trend in the Northern Hemisphere circulation. We will present some preliminary outputs of this analysis.