Artificial intelligence for climate teleconnections: a graph neural network approach with interpretable attention mechanisms

Climate teleconnections are long-range relationships among climate phenomena that underpin the foundation for explaining the climate dynamics of the planet and forecasting extreme events. Although there has been remarkable progress in climate science, many teleconnections remain inadequately characterized or even unidentified, partly due to the limitations of conventional statistical methods that cannot capture complex spatio-temporal dependencies and causality. A machine learning model can address these limitations. One approach to representing spatio-temporal relationships in climate data is through graphs, where nodes are points in space and time projected onto a spherical approximation of Earth, and edges connect each node at a specific time step to its neighbors in the subsequent time step. In this way, a model is directed to learn causal directions of influence. Graph neural networks (GNNs), especially graph attention networks (GATs), can extract causal patterns in the graphs through message-passing mechanisms. However, we need a reliable approach to assess the significance of those patterns. A straightforward solution is to assume that attention weights computed by the model yield estimates of causal strength. However, teleconnection analysis lacks ground truth data, making it difficult to determine whether model choices are random or reflect genuine causal relationships. Therefore, this represents an unsupervised task that requires modeling the probability of inference correctness. To guide the neural network in learning significant graph relationships, we employ a GAT as the encoder in a variational graph autoencoder (VGAE) architecture, with the training objective of learning meaningful data representations while minimizing the entropy of attention weight distributions. If the model chooses a same long-range connection at different scales of input data, then we can assume this is a real teleconnection pattern. Results show that the model can potentially detect known teleconnection patterns on ERA5 reanalysis dataset. However, current limitations - including low spatial resolution, neglect of inter-variable relationships, and absence of appropriate causality tests - indicate clear directions for future work in discovering new teleconnections.