Explainable Deep Learning Reveals Future Changes in East Asian Extreme Winter Predictability

Abstract

Deep learning-based approaches have emerged as powerful tools for seasonal prediction, yet their skill remains limited for extreme anomalies, and warming climates may alter predictability sources. Here we present a convolutional neural network (CNN)-based framework to assess how seasonal predictability changes in a future climate, focusing on East Asian extreme winters. Our CNN, trained on CESM2-LENS simulations for historical and SSP3-7.0 scenarios and optimized with Amplitude Focal Loss, reliably captures seasonal mean and extreme temperatures. Overall prediction skill increases in the future climate, with warm winter extremes becoming more predictable while cold extremes show little change. Using Integrated Gradients, an explainable AI method, we show this future asymmetry is linked to enhanced sensitivity of East Asian winter temperatures to tropical Pacific SST under background warming, where warm-phase SST more effectively drives extratropical circulation changes. Our framework provides a useful tool for exploring how seasonal predictability mechanisms evolve under climate change.