Predicting and Interpreting Atmospheric Blocking Using Convolutional Neural Networks and Explainable-AI Techniques

Atmospheric blocking is closely associated with various surface extreme weather events. Identifying potential precursor patterns of this phenomenon can significantly improve our forecasting capabilities. Explainable machine learning methods have proven to be promising tools for predicting future blocking occurrences in addition to providing physically-informed interpretations. By leveraging these methods, valuable insights can be gained to better understand the underlying mechanisms of atmospheric blockings and similar extreme weather events. In this study, we employ a Convolutional Neural Network (CNN) to predict atmospheric blocking events using streamfunction and geopotential height from the Large Ensemble Community Project (LENS) and ERA5 datasets. Combining both the big data from climate model outputs and real-world data from ERA5 allows us to strike a balance in terms of accuracy and relevance, which would optimize the predictive power of neural networks. To leverage the structure of the model to decompose its prediction, we applied Layer-Wise Relevance Propagation (LRP), an Explainable AI technique, to identify those spatial patterns and geographic regions guiding the CNN’s predictions, providing insights into how the model interprets the data and identifies blocking events. The CNN predicts atmospheric blocking with accuracies of 85%, 78%, 75%, and 68%, at 3, 5, 7, and 10 lead days, respectively. The results indicate that the machine learning approach for predicting and explaining atmospheric blocking enhances the accuracy of 7+ day weather forecasts and demonstrates consistency with physical interpretations.