Google DeepMind has introduced GenCast, an AI weather forecasting model that marks a notable step in how forecasts may be produced and used. According to the source article, GenCast was more accurate than the current leading forecast system in most of the test cases described, while also showing strengths in wind prediction and some forms of extreme weather forecasting.
The result is not simply another AI benchmark. Weather forecasts guide energy planning, disaster preparation, and daily public decisions. A model that can estimate a range of possible outcomes, rather than deliver only one expected result, could give forecasters and officials a clearer view of uncertainty.
What GenCast does differently
GenCast is an AI-only model from Google DeepMind. The source contrasts it with NeuralGCM, another DeepMind weather model published in July, which combined AI with physics-based approaches similar to those used in existing forecasting systems. NeuralGCM performed similarly to conventional methods while using less computing power.
GenCast takes a different route. It relies on AI methods alone. The source compares the basic idea to ChatGPT, with an important difference: instead of predicting the next likely word, GenCast predicts the next likely weather condition.
During training, the model begins with random parameters, also called weights. It then compares its predictions with real weather data. As training continues, those parameters shift toward patterns that better match actual weather.
The model was trained on 40 years of weather data, from 1979 to 2018. It then generated a forecast for 2019. In that test, GenCast was more accurate than the Ensemble Forecast, ENS, 97% of the time, according to the source article.
Why probabilistic forecasts matter
One of GenCast's most important differences is the type of forecast it produces. The source explains that Huawei's Pangu-Weather model, developed in 2023 after training on 39 years of data, produces deterministic forecasts. A deterministic forecast gives a single expected outcome, such as a specific temperature or rainfall amount.
GenCast instead produces probabilistic forecasts. That means it estimates the likelihood of different weather outcomes. The source gives examples such as a 40% chance of the temperature reaching a low of 30 °F, or a 60% chance of 0.7 inches of rainfall tomorrow.
This matters because weather decisions often depend on risk, not certainty. A single number can be useful, but it can also hide the range of possible outcomes. A probabilistic forecast can help officials understand whether a serious event is unlikely, plausible, or likely enough to require preparation.
That kind of information is especially relevant for extreme weather planning. The source says GenCast was better at predicting wind conditions and extreme weather such as the path of tropical cyclones. Better estimates for extreme weather can support planning for natural disasters.
Energy and disaster planning are key use cases
Wind forecasting is one practical area where improved prediction can matter quickly. The source notes that better wind prediction increases the viability of wind power because it helps operators decide when turbines should be turned on or off.
That point shows why AI weather forecasting is not only a scientific topic. Forecast quality can affect infrastructure decisions. If operators have better information about wind conditions, they can make more informed choices about how to manage turbines.
Extreme weather forecasting has a different but equally direct impact. More accurate estimates can help with disaster planning. GenCast's stronger performance in predicting the path of tropical cyclones suggests that AI systems may become useful tools in high-stakes forecasting workflows.
At the same time, the source is clear that the model still has weaknesses. It may predict where a tropical cyclone goes, but it underpredicts cyclone intensity. The reason given is that there is not enough intensity data in the model's training.
AI still depends on meteorology
The source cautions against treating GenCast as the end of conventional meteorology. The model learns from past weather conditions. Applying those learned patterns too far into the future may lead to inaccurate predictions in a changing and increasingly erratic climate.
GenCast also depends on datasets shaped by physics-based work. Aaron Hill, an assistant professor at the School of Meteorology at the University of Oklahoma, said the model still relies on a dataset like ERA5. ERA5 is described in the source as an hourly estimate of atmospheric variables going back to 1940.
Hill's point is that ERA5 itself has a physics-based model as its backbone. In other words, even an AI-only forecasting model can be built on data that was made possible by conventional meteorology and physics-based systems.
The source also notes that many atmospheric variables are not directly observed. Meteorologists use physics equations to estimate those variables, then combine those estimates with available observations. That information can then feed into a model such as GenCast.
Fresh data remains essential. Ilan Price, a DeepMind researcher and one of GenCast's creators, said a model trained up to 2018 will do worse in 2024 than a model trained up to 2023 will do in 2024. The implication is straightforward: AI forecasting models need updating as the atmosphere changes and new observations become available.
Human forecasters remain central
DeepMind's future plans include testing models directly on data such as wind or humidity readings. The goal is to see how feasible it is to make predictions from observation data alone.
For now, the source describes a more collaborative future. Meteorologists would work with GenCast rather than be replaced by it. Price said experts can review forecasts, make judgment calls, and look at additional data when they do not trust a forecast.
Hill made a similar argument. Human forecasters consider more information than a single AI prediction system can provide, and they can combine those inputs into strong forecasts.
That may be the most realistic lesson from GenCast. AI weather forecasting is advancing quickly, with systems from Google DeepMind, Nvidia, and Huawei all mentioned in the source. But the strongest future for forecasting may not be AI alone. It may be a workflow where AI models produce fast, probabilistic guidance and meteorologists apply expertise, context, and caution.