A convolutional network learns about the North Atlantic storm track to predict heavy rainfall in Western Norway

Neural networks are powerful and widely used tools in weather and climate sciences, but their reliability under climate change remains uncertain as future conditions may be different from their training distribution. One way to build trust in these models is to assess whether they learn physically meaningful relationships rather than spurious correlations. Here, we present a case study investigating whether a simple convolutional neural network (CNN) predicts the occurrence of heavy rainfall in Western Norway for physically interpretable reasons. Since such rainfall is primarily associated with North Atlantic cyclones, we use explainable AI to assess whether the CNN identifies and uses the “correct” cyclones for its predictions.

Using ERA5 reanalysis data, we train a CNN to predict the occurrence of daily heavy rainfall events up to six days ahead from gridded wind and pressure fields. We apply layer-wise relevance propagation (LRP) to identify which regions of the atmospheric input fields contribute most to the model’s predictions. We find that model relevance is spatially aggregated into a small number of coherent patches, with one to three positive relevance patches dominating the prediction in more than 90% of the cases. Physical consistency is assessed by comparing the relevance patterns to objectively tracked cyclones. Interpreting cyclones as being “used” by the network when they spatially overlap with a patch, we show that cyclones contribute positively to the network’s predictions in about 95% of heavy rainfall events. In addition, we show that cyclones highlighted by the network are physically plausible; their trajectories follow the North Atlantic storm track, shifting from the western and central North Atlantic towards the eastern Atlantic and the Norwegian coast as the prediction lead time decreases. These results demonstrate that the CNN learns physically interpretable large-scale dynamics associated with North Atlantic cyclones, providing evidence that explainable AI methods can be used to assess and build trust in machine learning models for weather and climate applications.