WeatherNext called Hurricane Melissa's Category 5 landfall five days out

Google DeepMind says its WeatherNext model predicted Hurricane Melissa's Category 5 landfall in Jamaica five days in advance, when the storm was still a Category 1. The standard forecasting trade-off, as DeepMind frames it, is that large global models track storm paths well but blur the small-scale convection that drives intensity, while local high-resolution models do the opposite. WeatherNext, trained on decades of global weather alongside specialized tropical cyclone datasets, is being presented as the first model to do both for an event like Melissa.

The numbers are specific. WeatherNext gave 80% confidence in a Category 5 landfall five days out, rising to near-certainty three days before the storm hit. It can run 50 parallel ensemble scenarios for what-if analysis, and the National Hurricane Center's 2025 verification report ranked it as the top individual model for both track and intensity that season. DeepMind says the NHC used WeatherNext outputs to support its public warnings, not just internal evaluation.

What this fits into is a broader pattern of weather AI moving from research demos to operational tools that forecasters actually consult. WeatherNext is a clearer case than most because the comparison is operational, against the model suite forecasters were already using.

Why it matters

Rapid intensification has been the hardest part of hurricane forecasting and the main reason evacuation orders come late. A model that can call a Category 5 landfall five days out, on an operational schedule, changes what emergency managers can plausibly do. The caveat is one season of results, and forecasters will want several more storms before they treat WeatherNext as the default rather than a useful second opinion.