Unraveling Winter Precipitation Predictability over CONUS via Deep Learning and Explainable Artificial Intelligence

Abstract Seasonal precipitation variability is among the most consequential aspects of weather and climate, affecting society and regional economies over contiguous United States (CONUS) and around the globe. Better understanding of precipitation predictability and its sources remains a pressing challenge, despite recent advances in physics-based modeling and forecasting. The use of deep learning models to boost seasonal forecasts has been explored; however, implementing explainable artificial intelligence (AI) tools to gain physical insights remains underexplored. In this study, for the first time, we use a diverse set of deep learning models with varying levels of complexity [linear models, linearized convolutional neural networks (CNNs), CNNs, vision transformers] and explainable AI methods to enhance understanding and answer three key questions: 1) Which CONUS regions exhibit higher precipitation predictability, and how much additional predictability can deep learning models yield compared to linear counterparts? 2) What are the main sources of predictability that deep learning models rely on? 3) How can we use explainable AI (XAI) ensembles (XAI tools applied