AI Improves Climate Downscaling for Future Projections

This study addresses the challenge of temporal out-of-distribution shifts in deep-learning-based climate downscaling, where models trained on historical climate data struggle with future projections. Researchers propose a temporal domain-adaptive downscaling framework that integrates supervised high-resolution reconstruction on historical data with domain alignment between historical and future climate distributions. Experiments focusing on daily temperature downscaling over the Continental United States showed that this adaptive model consistently outperformed statistical and deep-learning bias-correction methods. The most significant improvements were observed during periods of strong temporal distribution shift and in topographically complex, high-elevation regions, also reducing upper-tail temperature bias.