GenCast Forecasts Weather and Extreme Condition Risks with Cutting-Edge Accuracy

A new AI model improves the prediction of weather uncertainties and risks, providing faster and more accurate forecasts up to 15 days in advance.

Weather affects everyone, influencing our decisions, safety, and lifestyle. With climate change leading to more extreme weather, precise forecasts are crucial. However, predicting weather perfectly isn’t possible, especially beyond a few days.

Since perfect weather forecasts are unattainable, scientists and weather agencies use probabilistic ensemble forecasts. These models predict a range of possible weather scenarios, offering a broader view than single forecasts, which helps decision-makers understand potential weather conditions and their likelihood.

Today, we introduce GenCast in a paper published in Nature. GenCast is a high-resolution (0.25°) AI ensemble model that provides better forecasts for both everyday weather and extreme events compared to the leading system, the European Centre for Medium-Range Weather Forecasts’ (ECMWF) ENS, with predictions up to 15 days ahead. We are sharing the model’s code, weights, and forecasts to support the broader weather forecasting community.

The evolution of AI weather models

GenCast represents a significant advance in AI-based weather prediction. Unlike our earlier model, which was deterministic and offered a single best estimate, GenCast produces an ensemble of 50 or more predictions, each representing a possible weather path.

GenCast is a diffusion model, a type of generative AI model that has recently advanced image, video, and music generation. However, GenCast is tailored for the Earth’s spherical shape, learning to generate the complex probability distribution of future weather scenarios using the latest weather data.

We trained GenCast using four decades of historical weather data from ECMWF’s ERA5 archive, which includes variables like temperature, wind speed, and pressure at different altitudes. This enabled the model to learn global weather patterns at 0.25° resolution directly from the processed data.

Setting a new standard for weather forecasting

To thoroughly assess GenCast’s performance, we trained it on historical data up to 2018 and tested it on data from 2019. GenCast outperformed ECMWF’s ENS, the leading operational ensemble forecasting system relied upon for numerous national and local decisions daily.

We extensively tested both systems across 1,320 combinations of forecasts for different variables and lead times. GenCast was more accurate than ENS in 97.2% of these scenarios and in 99.8% at lead times exceeding 36 hours.

An ensemble forecast expresses uncertainty by making multiple predictions for different possible scenarios. If most predictions show a cyclone hitting the same area, uncertainty is low. But if they suggest different locations, uncertainty is higher. GenCast strikes the right balance, neither overstating nor understating its forecast confidence.

On a single Google Cloud TPU v5, GenCast takes just 8 minutes to generate one 15-day forecast in its ensemble, with all forecasts created simultaneously in parallel. Traditional physics-based ensemble forecasts, like ENS, at 0.2° or 0.1° resolution, require hours on a supercomputer with thousands of processors.

Advanced forecasts for extreme weather events

More accurate forecasts of extreme weather risks can help protect lives, prevent damage, and save money. When testing GenCast’s predictions for extreme heat, cold, and high winds, it consistently outperformed ENS.

Consider tropical cyclones, also known as hurricanes and typhoons. Enhanced and timely warnings of their landfall are invaluable. GenCast delivers superior predictions of these dangerous storms’ paths.

Improved forecasts can also significantly impact other societal areas, like renewable energy planning. Better wind-power forecasts enhance wind power’s reliability as a sustainable energy source, potentially accelerating its adoption. In a proof-of-concept experiment analyzing global wind power predictions, GenCast was more accurate than ENS.

Next generation forecasting and climate understanding at Google

GenCast is part of Google’s expanding suite of next-gen AI-based weather models, including Google DeepMind’s deterministic medium-range forecasts and Google Research’s NeuralGCM, SEEDS, and flood models. These models are starting to enhance user experiences on Google Search and Maps by improving precipitation, wildfire, flooding, and extreme heat forecasts.

We highly value partnerships with weather agencies and will continue collaborating with them to develop AI-based methods enhancing their forecasts. Meanwhile, traditional models remain crucial, supplying training data and initial conditions for models like GenCast. This collaboration between AI and traditional meteorology demonstrates the power of a combined approach to improve forecasts and better serve society.

To promote broader collaboration and accelerate research and development in weather and climate, we’ve made GenCast an open model, releasing its code and weights, just as we did with our deterministic medium-range global weather forecasting model.

We’ll soon release real-time and historical forecasts from GenCast and previous models, enabling anyone to integrate these inputs into their models and research workflows.

We are eager to engage with the broader weather community, including academic researchers, meteorologists, data scientists, renewable energy companies, and organizations focused on food security and disaster response. These partnerships offer valuable insights and feedback, alongside opportunities for commercial and non-commercial impact, supporting our mission to apply our models for humanity’s benefit.