Combining spatio-temporal neural networks with mechanistic interpretability to investigate teleconnections in S2S forecasts

Subseasonal-to-seasonal (S2S) forecasts are crucial for decision-making and early warning systems in extreme weather. However, the chaotic nature of atmospheric dynamics limits the predictive skill of climate models on S2S timescales. Teleconnections can provide windows of improved predictability, but leveraging these external drivers to enhance S2S forecast skill remains challenging. This study introduces a spatio-temporal neural network (STNN) designed to predict weekly North Atlantic European (NAE) weather regimes at lead times of one to six weeks during boreal winter. The STNN integrates a stacked vision transformer (ViT) encoder and a long short-term memory (LSTM) decoder to capture short- and medium-range variability. By incorporating spatio-temporal data on the stratospheric polar vortex, tropical outgoing longwave radiation, and 1D NAE regime time series, the network can access patterns linked to teleconnections of key drivers of European winter weather. Its modular design enables the application of mechanistic interpretability, providing novel neuron-level insights into the prediction behavior. The improved predictive skill beyond lead week three and enhanced accuracy for specific regimes suggest novel learned patterns of external drivers. Using Activation Maximization (AM), we analyze these learned representations, and by incorporating gradient-based explanations of correct predictions, we infer additional insights into prevalent teleconnections.