New AI model improves prediction of weather uncertainty and risk, delivering faster and more accurate forecasts up to 15 days ahead

Weather affects us all – it influences our decisions, our safety and our way of life. As climate change leads to more extreme weather events, accurate and trustworthy forecasts are more important than ever. However, the weather cannot be predicted perfectly and forecasts are uncertain, particularly beyond a few days.

Because perfect weather forecasting is not possible, scientists and weather agencies use probabilistic ensemble forecasts, in which the model predicts a range of likely weather scenarios. Such ensemble forecasts are more useful than relying on a single forecast because they provide decision makers with a more comprehensive picture of possible weather conditions in the coming days and weeks and the likelihood of each scenario.

Today, in a paper published in Nature, we introduce GenCast, our new high-resolution (0.25°) AI ensemble model. GenCast provides better forecasts of both daily weather and extreme events up to 15 days in advance than the leading operating system, the European Center for Medium-Range Weather Forecasts (ECMWF) ENS. We will publish our model’s code, weights, and predictions to support the broader weather forecasting community.

The development of AI weather models

GenCast represents a significant advance in AI-based weather forecasting, building on our previous weather model, which was deterministic and provided a single, best estimate of future weather. In contrast, a GenCast forecast includes an ensemble of 50 or more forecasts, each representing a possible weather pattern.

GenCast is a diffusion model, the type of generative AI model that underpins recent rapid advances in image, video and music generation. However, GenCast differs from these in that it is adapted to Earth’s spherical geometry and learns to accurately generate the complex probability distribution of future weather scenarios when given the most current weather condition as input.

To train GenCast, we provided it with four decades of historical weather data from ECMWF’s ERA5 archive. This data includes variables such as temperature, wind speed and pressure at different altitudes. The model learned global weather patterns at a resolution of 0.25° directly from this processed weather data.

We’re setting a new standard for weather forecasting

To accurately evaluate GenCast’s performance, we trained it on historical weather data up to 2018 and tested it on data from 2019. GenCast demonstrated better forecasting capabilities than ECMWF’s ENS, the leading operational ensemble forecasting system on which many national and local decisions depend daily.

We tested both systems extensively, examining forecasts of various variables at different lead times – a total of 1320 combinations. GenCast was more accurate than ENS on 97.2% of these targets and 99.8% on lead times greater than 36 hours.

An ensemble forecast expresses uncertainty by making multiple predictions that represent different possible scenarios. If most forecasts indicate that a cyclone will hit the same area, the uncertainty is low. However, if they predict different locations, the uncertainty is higher. GenCast strikes the right balance and avoids overestimating or underestimating its confidence in its forecasts.

A single Google Cloud TPU v5 takes only 8 minutes to generate a 15-day forecast in the GenCast ensemble, and each forecast in the ensemble can be generated simultaneously and in parallel. Traditional physics-based ensemble forecasts, such as those produced by ENS at 0.2° or 0.1° resolution, take hours on a supercomputer with tens of thousands of processors.

Advanced forecasts for extreme weather events

More accurately predicting the risks of extreme weather can help authorities protect more lives, prevent damage and save money. When we tested GenCast’s ability to predict extreme heat and cold as well as high wind speeds, GenCast consistently outperformed ENS.

Now let’s look at tropical cyclones, also called hurricanes and typhoons. Getting better and more advanced warnings about where they will make landfall is invaluable. GenCast provides excellent forecasts of the tracks of these deadly storms.

Better forecasts could also play a key role in other areas of society, for example in renewable energy planning. For example, improvements in wind power forecasting directly increase the reliability of wind power as a source of sustainable energy and will potentially accelerate its adoption. In a proof-of-principle experiment analyzing predictions of all wind power generated by groups of wind farms around the world, GenCast was more accurate than ENS.

Next generation forecasts and climate understanding at Google

GenCast is part of Google’s growing suite of next-generation AI-based weather models, including Google DeepMind’s AI-based deterministic medium-range forecasts and Google Research’s NeuralGCM, SEEDS and Flood models. These models are beginning to improve the user experience in Google Search and Maps and improve prediction of precipitation, wildfires, floods, and extreme heat.

We greatly value our partnerships with weather agencies and will continue to work with them to develop AI-based methods that improve their forecasts. In the meantime, traditional models remain essential to this work. On the one hand, they provide the training data and initial weather conditions that models like GenCast need. This collaboration between AI and traditional meteorology highlights the power of a combined approach to improve forecasts and better serve society.

To encourage broader collaboration and accelerate research and development in the weather and climate community, we made GenCast an open model and published its code and weights, as we did for our medium-range deterministic global weather forecast model.

We will soon release real-time and historical forecasts from GenCast and previous models, allowing anyone to integrate these weather inputs into their own models and research workflows.

We are committed to collaborating with the broader weather community, including academic researchers, meteorologists, data scientists, renewable energy companies, and organizations focused on food security and disaster relief. Such partnerships provide deep insights and constructive feedback, as well as invaluable opportunities for commercial and non-commercial impact, all critical to our mission to apply our models for the benefit of humanity.