A deep learning-based land-atmosphere coupled model for heatwave prediction

Extreme heatwaves are intensifying under climate change, yet their prediction remains limited by inadequate representation of land–atmosphere (L–A) interactions. Most deep learning–based weather models rely solely on atmospheric variables, overlooking the influence of land surface conditions on heat extremes. Here, we present an L–A coupled prediction framework for Northern Hemisphere summer that incorporates multi-layer soil moisture (SM) and temperature into atmospheric forecasting. To better capture delayed land surface feedbacks, the model is trained with a multi-step loss. This approach improved the representation of L–A interactions across 1–7 day lead times. Using multi-step loss, the L–A coupled model achieved a 5.9–11.2% improvement in heatwave forecast accuracy relative to the atmosphere-only model, as measured by root mean squared error, whereas single-step loss achieved only 0.4–2.4% improvement. Skill gain was strongest at short leads (~ 3 day) when both SM and circulation predictability were high, and sustained through 7 days by L–A coupling driven by SM predictability. Case studies of recent heatwaves further demonstrated its ability to capture land surface drying and associated temperature extremes. These findings underscore the importance of incorporating L–A coupling with multi-step optimization for advancing data-driven heatwave prediction.