The Forecast That Arrives Before Your Coffee Cools

At 5:55 a.m., a grid dispatcher looks at two maps to plan the day’s turbine operations. 

One map shows the traditional physics model that has been used for many years. 

The other map is from GraphCast. Both maps predict a storm coming in by the late afternoon, but they disagree on the timing: one says it will arrive at 4 p.m., while the other predicts 6 p.m.

GraphCast is an AI weather model from Google DeepMind. Instead of calculating physics in real-time, it learns from over 40 years of weather data to predict the next 10 days all at once. In tests, it outperformed Europe’s top model on about 90% of targets and can generate a global forecast in under a minute. This accuracy and speed are why it’s gaining attention.

Those two hours can make a big difference in whether crews need to prepare now or rush to get ready later.

The headline number (and why it matters)

In peer-reviewed tests, GraphCast outperformed ECMWF’s HRES in about 90% of 1,380 verification targets for 10-day global medium-range forecasting. This isn’t just a demonstration; it represents a significant improvement in daily planning.

But what about the 10% where it didn’t perform as well? If that small portion includes the worst days, the overall success rate can be misleading.

Where AI excels: Medium-range forecasts for hazards like tropical cyclone paths, atmospheric rivers, and extreme temperatures often show improved accuracy, providing extra hours to respond.

Where challenges remain: Rapidly intensifying events are tough for everyone. For example, Hurricane Otis reached Category 5 in just about a day, and most models underestimated that rapid increase. The risk of unexpected events still exists.

What actually changed

Traditional models rely on supercomputers to solve physics in real-time.

 In contrast, GraphCast shifts the calculations to a training phase: it learns from decades of ERA5 data and then simulates the atmosphere in six-hour increments. While the training process is intensive, the inference is quick, processing 10-day global forecasts in under a minute on a single TPU v4. This speed allows for more frequent updates, the ability to check multiple scenarios, and still make it to the morning meeting.

Why agencies are hedging

On February 25, 2025, Europe’s weather center (ECMWF) launched its AI Forecasting System (AIFS) alongside the traditional physics-based IFS. By July 1, 2025, they introduced AIFS-ENS, the AI ensemble. Reports indicated improvements of up to 20% in some areas. However, the current AI models operate at a coarser resolution compared to the physics ensemble, meaning that physics still excels in providing precise local details and understanding complex interactions.

What to do on real days

Open both maps.

If they agree, proceed with earlier action and greater confidence.

If they differ, consider the range of predictions as the key information: check local precipitation, monitor for sudden increases in intensity, and prepare for unexpected outcomes. This is where mistakes can happen, and where human judgment remains essential.