Predicting Beyond Training Data via Extrapolation versus Translocation: AI Weather Models and Dubai's Unprecedented 2024 Rainfall

Artificial intelligence (AI) models have transformed weather forecasting, but their skill for gray swan extremes is unclear. Here, we analyze GraphCast, AIFS, and FuXi forecasts of the unprecedented 2024 Dubai storm, which had twice the training set’s highest rainfall in that region. Remarkably, GraphCast and AIFS accurately forecast this event up to 8 days ahead. FuXi forecasts the event, but underestimates the rainfall. Fine-tuning and receptive field analyses suggest that these models’ success stems from “translocation”: learning from comparable/stronger dynamically similar events in other regions during training. Evidence of “extrapolation” (learning from weaker events) is not found. Even events within the global distribution’s tail are poorly forecasted, which is not just due to data imbalance (generalization error) but also spectral bias (optimization error). These findings demonstrate the potential of AI models to forecast regional gray swans and the opportunity to improve them through understanding the mechanisms behind their successes/limitations.

💡 Research Summary

**
The paper investigates how state‑of‑the‑art AI weather models—GraphCast, AIFS, and FuXi—handled the unprecedented April 2024 rainfall over Dubai, an event whose 12‑hour accumulated precipitation (≈60 mm) was roughly twice the maximum observed in the models’ training period (33 mm). Despite this being a “regional gray‑swan” (an extreme that is physically possible but absent from the training data), GraphCast accurately forecasted the timing, location, and magnitude of the storm up to eight days in advance, while FuXi captured the timing and circulation but severely underestimated the peak rainfall.

The authors frame two possible mechanisms for such out‑of‑distribution success:

1. Extrapolation – the model learns from weaker, more frequent local events and generalizes to stronger, unseen cases.
2. Translocation – the model learns from comparably strong or stronger events that are dynamically similar but occur in other regions of the globe.

Through a series of quantitative analyses they find strong evidence for translocation and none for extrapolation. Using a Lagrangian tracking algorithm they identify “dynamically similar” extratropical cyclone‑type events across the 20°–50° N band that exhibit precipitation equal to or greater than the Dubai storm. Composite circulation fields of these events closely match the Dubai case, and joint histograms of 850 hPa meridional wind (v850) and specific humidity (q850) show that the key moisture‑transport dynamics of the Dubai event are well represented in the global training set, even though they are out‑of‑distribution locally. Consequently, GraphCast appears to have leveraged a large effective receptive field that enables it to “translocate” knowledge from distant, strong events to the Dubai region.

The study also examines why FuXi underperforms. Both models are trained on the same ERA5 reanalysis, yet their architectures differ: GraphCast uses a graph‑neural‑network with global attention, whereas FuXi relies on a transformer‑based design. The authors argue that FuXi’s internal representation may be less capable of capturing the long‑range dynamical relationships needed for translocation, leading to a bias toward the training‑set maximum.

Beyond data imbalance, the authors highlight spectral bias—the tendency of deep networks to learn low‑frequency (large‑scale) patterns faster than high‑frequency (small‑scale) ones. Extreme precipitation involves high‑frequency structures; the models retain larger errors in the high‑frequency spectral band, indicating an optimization error rather than a pure generalization issue. This explains why even events that lie within the global distribution’s tail are poorly forecasted.

Practical implications are discussed. Translocation suggests that AI models can provide early warnings for regional extremes even in data‑sparse areas (e.g., deserts, polar regions) if similar dynamics exist elsewhere in the training corpus. To mitigate spectral bias and improve extrapolation capability, the authors recommend: (i) pre‑training on higher‑resolution observational datasets, (ii) incorporating multi‑scale loss functions that explicitly penalize high‑frequency errors, and (iii) hybridizing data‑driven models with physical constraints. They also stress the need for controlled experiments that deliberately remove strong, dynamically similar events from training to isolate the contribution of translocation versus extrapolation.

In summary, the Dubai 2024 storm demonstrates that AI weather models can successfully forecast regional gray‑swans when they have learned from globally distributed, dynamically analogous events—a process the authors term translocation. However, the persistent spectral bias limits the ability to capture the finest‑scale features of extreme precipitation, pointing to clear avenues for future model development and for building more reliable AI‑driven early‑warning systems.