GenCast predicts weather and extreme risk with cutting-edge accuracy

New AI models advance predictions of weather uncertainty and risk, providing faster and more accurate forecasts up to 15 days in advance.

Weather affects us all and shapes our decisions, safety, and way of life. As climate change increases extreme weather events, accurate and reliable forecasts are more important than ever. However, weather cannot be predicted perfectly, and forecasts are uncertain, especially over a few days.

Because perfect weather forecasts are impossible, scientists and weather agencies use probabilistic ensemble forecasts, where models predict a variety of likely weather scenarios. Such ensemble forecasts are more useful than relying on a single forecast because they provide decision makers with a complete picture of likely weather conditions and the likelihood of each scenario in the coming days and weeks.

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

Evolution of AI weather models

GenCast represents a significant advance in AI-based weather forecasting, building on previous weather models that are deterministic and provide the single best forecast of future weather. In contrast, GenCast forecasts consist of an ensemble of 50 or more forecasts, each representing an expected weather trajectory.

GenCast is a diffusion model, a type of generative AI model that powers recent rapid advances in image, video, and music generation. However, GenCast differs from these in that it is adapted to the spherical shape of the Earth and, given the latest weather conditions as input, learns how to accurately generate complex probability distributions of future weather scenarios.

To train GenCast, we provided 40 years of historical weather data from ECMWF’s ERA5 archive. This data includes variables such as temperature, wind speed, and air pressure at different altitudes. The model learned global weather patterns directly from this processed weather data at 0.25 degree resolution.

Setting a new standard in weather forecasting

To rigorously 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 skill than ECMWF’s ENS, the top operational ensemble forecasting system that many national and local decisions rely on every day. .

We comprehensively tested both systems and looked at the predictions of different variables at different lead times (1,320 combinations in total). GenCast was more accurate than ENS on 97.2% of these targets and 99.8% for lead times greater than 36 hours.

Ensemble forecasting expresses uncertainty by making multiple forecasts representing different possible scenarios. If most forecasts indicate that the cyclone will reach the same area, the uncertainty is low. However, predicting different locations increases uncertainty. GenCast strikes the right balance between overestimating and underestimating confidence in its predictions.

It takes a single Google Cloud TPU v5 just 8 minutes to generate one 15-day forecast with the GenCast ensemble. All predictions in an ensemble can be generated in parallel at the same time. Traditional physics-based ensemble predictions, such as those produced by ENS, occur at 0.2° or 0.1° resolution and take hours on supercomputers with tens of thousands of processors.

Advanced prediction of extreme weather

By more accurately predicting the risk of extreme weather events, authorities can save more lives, avoid damage, and save money. When we tested GenCast’s ability to predict extreme heat, cold, and high wind speeds, GenCast consistently outperformed ENS.

Next, consider tropical cyclones, also known as hurricanes and typhoons. Having better and more advanced warning of where you will land is invaluable. GenCast provides excellent predictions about the path of these deadly storms.

Better forecasting could also play an important role in other aspects of society, such as renewable energy planning. For example, improved wind power forecasting could directly increase the credibility of wind power as a sustainable energy source and accelerate its adoption. In a proof-of-principle experiment that analyzed predictions of the total amount of wind power produced by a group of wind farms around the world, GenCast was more accurate than ENS.

Next generation forecasting 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 starting to power the user experience in Google Search and Maps, improving predictions of precipitation, wildfires, flooding, and heatwaves.

We deeply value our partnership with weather agencies and will continue to work with them to develop AI-based methods to enhance forecasts. On the other hand, traditional models are still essential for this task. One is to provide the necessary training data and initial weather conditions for models such as GenCast. This collaboration between AI and traditional meteorology highlights the power of a combined approach to improve predictions and better serve society.

To encourage broader collaboration and help accelerate research and development in the weather and climate community, we are making GenCast an open model and making its code and weights publicly available, similar to deterministic medium-range global weather prediction models. did.

We will be releasing real-time and historical forecasts from GenCast and previous models soon. This allows anyone to integrate this weather information into their own models and research workflows.

We want to actively engage with the broader climate community, including academic researchers, meteorologists, data scientists, renewable energy companies, and organizations focused on food security and disaster response. Such partnerships provide deep insight and constructive feedback, as well as valuable opportunities for commercial and non-commercial impact. All of these are important to our mission of applying models to benefit humanity.