Announcement_9

Check out the paper led by former research scientist Dr. Qiang Sun (now a faculty member at Nanjing University) titled “Predicting regional gray swans via translocation: AI weather models and Dubai’s unprecedented 2024 rainfall” published in Science Advances. AI weather models learn from past weather, so unprecedented extremes (“gray swans”) are especially hard for them. Yet GraphCast and AIFS accurately forecast Dubai’s devastating April 2024 storm up to 8 days ahead, even though it brought nearly twice the heaviest rainfall the region had seen in their training data. The paper shows why: Dubai is a dry region, but similar weather patterns with even heavier rain are frequent elsewhere in the world, and the AI models can learn from these analogues on their own (i.e., they “translocate,” as our earlier PNAS paper hinted at for tropical cyclones). When GraphCast was fine-tuned on data with the rainfall of those distant storms capped near Dubai’s previous record, it still forecast the storm, but not its record rainfall. This is encouraging for regions with short weather records and for a climate change signal mostly in the form of shifting weather patterns (rather than intensifying to new global records). The paper is highlighted by the UChicago Climate Institute.