Explainable AI for subseasonal forecasting of the north atlantic oscillation

Abstract Subseasonal forecasting of the North Atlantic Oscillation (NAO), which spans two to eight weeks, is essential for climate-sensitive sectors such as energy and agriculture. However, it remains challenging due to the chaotic nature of atmospheric processes and the influence of tropical conditions, resulting in limited forecast skill and interpretability beyond approximately two weeks. This study aims to address the limited predictive accuracy in forecasting the NAO beyond 14 days by focusing on decadal representations of the NAO and utilizing "forecasts of opportunity." We propose two machine learning models: an Artificial Neural Network (ANN)-based model and a Gradient Boosting Decision Tree (GBDT)-based model. These models use daily mean sea level pressure (SLP) and 50 m wind (V50) data from the CESM2-LE dataset and are enhanced by explainable artificial intelligence (XAI) techniques, namely SHapley Additive exPlanations (SHAP) and Integrated Gradients (IG) to ensure transparency and interpretability. The models achieve accuracies of 0.769 (ANN) and 0.773 (GBDT) at a 14-day lead time for the top 20% most confident predictions, surpassing the baseline models across lead times. Validation against ERA5 reanalysis data shows accuracies of 0.732 (ANN) and 0.739 (GBDT), confirming robust predictive relationships. By analyzing the NAO over decadal timescales, we uncover patterns in variability and, through the use of XAI, identify key spatial predictors that drive these predictions. These findings not only enhance transparency and predictive skill in subseasonal NAO forecasting but also establish a solid foundation for future research. This future research can build upon our work by integrating additional climate variables known to influence the NAO.