Google uses AI to improve cyclone forecasts

Google is using artificial intelligence to predict cyclone paths and intensity more accurately, and to run those forecasts on a single TPU, rather than always relying on building-sized supercomputers.

The goal is not just to improve a weather map. A more accurate forecast can give authorities and communities more time to evacuate, protect infrastructure, and prepare emergency teams.

From supercomputers to a single TPU

For decades, meteorologists have used models based on the laws of physics. These systems simulate how the atmosphere moves using data collected by satellites, sensors, and weather stations.

The problem is that this data can be incomplete or noisy. Calculating how the atmosphere will evolve also requires enormous computing power. Traditional models have needed machines the size of shipping containers or even several-story buildings.

According to Ferran Alet, a Google DeepMind researcher, the meteorological community historically gained around one additional day of forecast accuracy per decade through advances in models, satellites, and computing capacity.

Google says its WeatherNext models have concentrated similar progress into a single generation of models by combining physics and artificial intelligence.

What artificial intelligence adds

Google trained WeatherNext on 50 years of historical weather data. The AI learns patterns from the past and combines them with current conditions to estimate how a storm will develop.

Physics-based models remain important because they describe the state of the atmosphere and the rules that govern it. AI adds another capability: producing forecasts more quickly and helping calculate future scenarios without having to repeat every physical simulation from scratch.

For cyclones, the improvement aims to answer two key questions:

• Where will the storm make landfall?
• How intense will it be when it arrives?

Speed matters too. A model that can run on a single TPU makes it easier for researchers and weather agencies to generate more forecasts and test different possibilities with fewer resources.

One more day can change the response

Weather agencies are considering extending their public forecasts from five to seven days. That extra day does not guarantee that every prediction will be accurate, but it can be decisive when organizing a response to a dangerous storm.

Google cites Hurricane Melissa as an example. Before hurricane season, the company worked with the National Hurricane Center to support its forecasts. Although Melissa initially appeared weak, AI predictions indicated with high confidence that it could intensify rapidly, from a Category 1 storm to a Category 5 storm.

According to Google, the models detected that development several days before the hurricane made landfall. The warning gave Jamaican authorities more time to prepare emergency measures.

What it means for you

AI does not replace meteorologists or eliminate uncertainty. A cyclone can change course or intensify because of factors no model can anticipate perfectly. These systems are meant to improve the information available and provide scenarios early enough to support decisions.

For the public, the most important change may be receiving more accurate alerts with more time to prepare. For scientific teams, it also means that tools that once required large computing centers could become available to more institutions.

Google has openly released WeatherNext 2 and its cyclone models, as well as created the interactive Weather Lab site. The next step will be to see how they perform outside research settings, with different data and needs in each region. The technology's real value will be measured by the decisions it helps people make before the storm arrives.