Can AI weather models track tropical cyclones better?

How the comparison was constructed

The researchers used IBTrACS for observed storm tracks and intensities, ERA5 and ECMWF high-resolution initial-condition analyses as gridded references, and archived IFS-ENS forecasts as the numerical baseline. They initiated forecasts 12 hours before the first reported wind observation, used TempestExtremes to detect storm centres, and scored the probability density that each ensemble placed around the observed position. This approach measures the distribution of possible tracks rather than collapsing an ensemble to one average line.

There were two different evaluation settings because the models and their training data differ. GenCast and FourCastNet 3 checkpoints trained on ERA5 were evaluated on 2021–2023 storms. Operational GenCast and FGN models fine-tuned on higher-resolution ECMWF initial conditions were evaluated on 2025 storms. The paper is clear that because the storms and model versions change, the design does not isolate fine-tuning as the cause of any improvement.

Track scoring also required at least ten ensemble members with an identified position at five or more common lead times. That excluded storms where one or more systems did not maintain enough matched tracks. Detection rate was reported separately, but the benchmark did not include non-cyclone cases. It therefore cannot establish how well a model calibrates the probability that a cyclone forms at all, a critical question for real forecasting operations.[1]

What forecasters and emergency planners should take from it

The result is encouraging for one component of forecasting: where a known storm may travel and how its surrounding pressure and wind fields evolve. Better distributions of plausible tracks could eventually help forecasters communicate the geographic range of risk and help emergency planners avoid treating a single centre line as certainty. The study also finds that learned models reproduce a recognisable relationship between pressure and wind, even when their absolute intensities are biased.

But emergency decisions depend on more than track. The models produced storms that were systematically too weak in some comparisons, and their ensemble spread was not consistently calibrated. A forecast can put the centre near the right place while understating damaging wind, rainfall, surge or the range of plausible outcomes. The authors accordingly warn that the benchmark should not be read as operational guidance. National meteorological agencies remain responsible for warnings and for combining models, observations and specialist judgment.

The commercial context also matters. Several authors are affiliated with insurer AXA, an organisation with a direct interest in catastrophe risk, while others hold academic affiliations. That does not invalidate the methods, but it should be disclosed alongside the fact that the manuscript has not yet passed peer review. The team reports using more than 6,000 H100 GPU-hours for the forecasts it generated and says generative AI assisted code, exploratory plots and text editing, but not data creation, study design or interpretation.[1]

What evidence would change the assessment

A stronger case would test more seasons prospectively, include forecast cycles where no cyclone forms, and compare warnings at the decision thresholds used by operational centres. Evaluation should cover landfall timing, rapid intensification, wind radii, rainfall and surge—not only location, pressure and gridded wind. Regional breakdowns are also important because performance can differ by basin and storm category.

Independent replication would help establish whether the gains survive different tracking algorithms, reference datasets and forecast-availability constraints. The most consequential evidence would be a real-time shadow trial in which AI ensembles run beside operational systems without controlling public warnings, followed by preregistered scoring of accuracy, calibration, timeliness and forecaster workload. Until then, the paper supports careful evaluation of AI ensembles as additional tools, not replacement of physics-based forecasting or human warning services.[1]