Can AI Weather Models Predict the 2024 Dubai Rainfall, a Regional Grey Swan?

AI weather models have recently shown remarkable skills in 1 to 10-day global forecasting, including for some extreme events. However, critical questions about their skills for rare, extreme events that are physically possible but are absent from the short training set (the so-called grey swans), and their learning process remains unclear. The 2024 Dubai rainfall is such an event, having never been observed in that region since the records started in 1949 and well beyond the range of the rainfall in that region in the ERA5 training data used for AI weather models (e.g., Pangu, GraphCast, Fuxi). Here, we first show that, surprisingly, GraphCast predicts this event up to 8 days ahead of time and captures the correct time and location of the rainfall (see the figure). Fuxi also predicts the event, though it underestimates the magnitude of rainfall.

Next, we show that dynamically similar events, with comparable or even larger rainfall, have occurred in the training set but in other regions (e.g., the Mediterranean Sea). We hypothesize that AI models can leverage knowledge from dynamically similar events in different regions despite being trained with location-specific data. This hypothesis is consistent with our parallel study of tropical cyclones, in which we unambiguously demonstrate that the AI weather models can translate learning of dynamically similar events from one region to another.

Further examination of Fuxi’s performance shows that this model’s underprediction of the rainfall is not due to the rarity of the amplitude (related to data imbalance, a generalization error in AI models), but that it is due to Fuxi's stronger spectral bias—an optimization error—that misses small-scale features such as precipitation. This finding underscores the importance of identifying the source of the errors in AI weather models so that their shortcomings can be addressed properly.