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

TL;DR

The study investigates whether AI weather models can forecast unprecedented regional extremes (gray swans) by disentangling extrapolation from translocation. Using Dubai's 2024 rainfall as a stringent test, it analyzes GraphCast, AIFS, and FuXi forecasts with ERF diagnostics, Lagrangian tracking, and a targeted fine-tuning experiment. The results indicate that GraphCast and AIFS achieve accurate timing, location, and substantial amplitude through translocation from dynamically similar events elsewhere, while FuXi tends to underestimate peak rainfall; there is little evidence for extrapolation from weaker local events. The work highlights spectral bias and data imbalance as key limitations for tail forecasts and outlines pathways—diffusion models, optimized loss strategies, oversampling—to enhance extrapolation and tail accuracy, with practical implications for regional warning systems and AI-enabled climate emulation.

Abstract

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.

Figures (14)

• Figure 1: Forecasting Dubai's 2024 extreme rainfall event using AI weather models.(A) Time series of 12-hour accumulated precipitation averaged around Dubai, comparing ERA5 reanalysis (black line) with forecasts from GraphCast (blue line, 5 days ahead), AIFS (light blue line, 5 days ahead), and FuXi (magenta line, 5 days ahead; dashed line, 1 day ahead). The horizontal dashed line indicates the maximum 12-hour accumulated precipitation (33 mm) in this region ever observed from 1979 to 2021, covering the training period for GraphCast, AIFS, and FuXi. (B) Circulation pattern and rainfall over the Arabian Peninsula on April 11th, 2024, at 12:00 UTC (5 days before the event), showing 500 hPa geopotential height (contours), 850 hPa wind vectors (arrows), and 12-hour accumulated precipitation (shading), all from ERA5 reanalysis. (C) Same as (B), but for 5 days later, April 16th, 2024, at 12:00 UTC, when the instantaneous rainfall value is close to its peak. (D) Same as (C), but for a 5-day forecast by GraphCast initialized on April 11th (panel B). (E) Same as (D), but for a 5-day forecast by FuXi. See Figure S2 for circulation forecasts from AIFS, and Figure S3 for circulation forecasts from two other public AI models, Pangu and FourCastNetV2, which do not provide precipitation.
• Figure 2: The Dubai rainfall event as a regional gray swan. (A) Histograms of 12-hour accumulated precipitation within a $100$ km $\times 100$ km box around Dubai in the training dataset for GraphCast, AIFS, and FuXi models (ERA5, 1979-2020). (B) Same as (A), but focused on the 20°N-50°N latitude band, considering the full training set and the same dataset but with tropical cyclones excluded or only dynamically similar events (extratropical cyclones) included; see Methods for details. The vertical lines indicate the 12-hour accumulated precipitation during the 2024 Dubai extreme event. (C) Moisture transport around Dubai in the training set is shown as a 2D histogram of 850 hPa meridional wind ($v850$) and specific humidity ($q850$). The red '+' symbol marks the values of the 2024 Dubai event (April 15-16). (D) Same as (C), but for all dynamically similar events in the 20°N-50°N latitude band.
• Figure 3: Dynamically similar events in other regions within the training dataset.(A) Conceptual model of an extratropical cyclone-driven heavy precipitation event, similar to the 2024 Dubai event. (B) Circulation and precipitation of the 2024 Dubai event in ERA5 (same as Figure \ref{['fig: Dubai-prediction']}C). (C) Composite of dynamically similar events that have precipitation comparable to or even stronger than the Dubai event, identified from the training set in ERA5 over 20°N $-$50°N latitude band of the Northern Hemisphere using a Lagrangian tracking algorithm (see Methods for more details). The patterns are centered on the city of Dubai for the composite.
• Figure 4: Effective receptive field (ERF) of GraphCast and AIFS for forecasting precipitation (6 hours later) in the Dubai area. Four selected variables are shown: geopotential at 500 hPa, zonal velocity at 850 hPa, specific humidity at 850 hPa, and meridional wind at 850 hPa. The forecast is initialized at 12:00 UTC on April 15, 2024, with ERF calculated after one time-step (6 hours). The star symbol in each panel shows the location of Dubai. GraphCast uses two previous time steps for predictions; we show results for the later time step, as the earlier step exhibits a similar but weaker pattern. Dot stippling in (B) indicates regions where the field is zero (absolutely no information is used for the prediction). The maps are scatter plots where point size and color intensity both scale with the standardized ERF amplitude, with darker blue regions indicating stronger influence on precipitation predictions over Dubai. Different from the actual gradient calculation, only relative spatial variations of the ERFs are meaningful, so we omit a colorbar and physical units. The multi-icosahedral mesh structure is distinctly visible in GraphCast, while AIFS results show a latitude band, consistent with the model's attention window design lang2024aifs. See Methods for details of ERF calculations.
• Figure 5: Fine-tuning GraphCast with samples that have the precipitation of dynamically similar events outside Dubai, capped degrades the forecast skill for the 2024 Dubai event.(A) Forecast of the 2024 Dubai rainfall event by the original GraphCast model versus the version fine-tuned on data where precipitation for dynamically similar events outside the Dubai region was capped at 30 mm/12-hour (approximating Dubai's historical maximum). (B) Circulation pattern of the Dubai event predicted by the fine-tuned model (analogous to Figure \ref{['fig: Dubai-prediction']}D), showing that the circulation forecast skill remains intact despite the decline for rainfall forecast skill. (C) Global Anomaly Correlation Coefficient (ACC), and (D) Global Root Mean Square Error (RMSE) for the original and fine-tuned models. Global metrics are averaged over 72 test cases sampled every 5 days throughout 2024, demonstrating that the fine-tuning process did not degrade general global forecast skill.