Gencast provides the weather and the risk of extreme conditions with the latest accuracy

The new AI model develops predicting uncertainty and weather risk, providing more accurate forecasts up to 15 days forward

The weather affects all of us – shaping our decisions, our safety and our lifestyle. Because climate change is driven by more extreme weather events, accurate and trustworthy forecasts are more important than ever. However, the weather cannot be perfectly predicted, and the forecasts are particularly uncertain over a few days.

Since the perfect weather forecast is not possible, the scientists and weather agencies use probabilistic team forecasts in which the model provides for a number of likely weather scenarios. Such team forecasts are more useful than relying on one forecast, because they provide decision -makers with a more complete picture of possible weather conditions in the coming days and weeks and the probability that every scenario.

Today, in paper Published in natureWe present Gencast, our new AI model with high resolution (0.25 °). Gencast provides better forecasts of both daily weather and extreme events than the highest operating system, the European Medium -range Weather Forecasts “(ECMWF) Ens, up to 15 days. We will release the code, weight and forecasts of our model to support the wider weather forecasting community.

Evolution of weather models AI

Gencast means critical progress in a weather forecast based on artificial intelligence, which is based on our previous weather model, which was deterministic, and provided one, best respect for future weather. On the other hand, the Gencast forecast includes a team with 50 or more forecasts, each of which represents a possible weather trajectory.

Gencast is a diffusion model, a kind of generative AI model, which underlies the last, quick progress in generating image, video and music. However, Gencast differs from them because it is adapted to the spherical earth geometry and learns to accurately generate the complex distribution of the probability of future weather scenarios when he received the latest weather as a contribution.

To train Gencast, we provided him with four decades of historical weather data from ECMWF Archive era5. These data include variables such as temperature, wind speed and pressure at various heights. The model has learned global weather patterns, with a resolution of 0.25 °, directly from these processed weather data.

Establishing a new weather forecasting standard

In order to strictly assess Gencast results, we trained it on the basis of historical weather data by 2018 and tested it on data from 2019. Genciast showed better forecasting skills than ENE ECMWF, the highest system of forecasting the operational team on which many national and local decisions depend.

We have comprehensively tested both systems, analyzing the forecasts of various variables at different implementation times – a total of 1320 combinations. Gencast was more accurate than ENE on 97.2% of these goals and at 99.8% during the implementation of more than 36 hours.

Better extreme weather forecasts, such as heat waves or strong winds, allow timely and profitable preventive measures. Gencast offers greater value than ENS when making decisions regarding preparation for extreme weather, in a wide range of decision -making scenarios.

The team's forecast expresses uncertainty by making many forecasts that represent various possible scenarios. If most forecasts show a cyclone that hit the same area, uncertainty is low. But if they predict different locations, uncertainty is higher. Gencast hits the right balance, avoiding both overestimation and understanding of his trust in his forecasts.

Extracting one 15-day Genciast forecast in the Genciast team in the Genciast team occupies a single Genciast forecast, and each forecast in the team can be generated simultaneously at the same time. Traditional team forecasts based on physics, such as those produced by ENS, with a resolution of 0.2 ° or 0.1 °, take many hours for a supercomputer with tens of thousands of processors.

Advanced forecasts on extreme weather events

More accurate risk forecasts of extreme weather can help officials secure more lives, avoid damage and save money. When we tested Gencast's ability to predict extreme heat and cold and high wind speeds, Gencast consistently exceeded ENS.

Consider now tropical cyclones, also known as hurricanes and typhans. Getting better and more advanced warnings about where they reach the earth is invaluable. Gencast provides excellent forecasts of traces of these deadly storms.

The forecast of the Gencast band shows a wide range of possible paths for Typhoonon Hagibis seven days in advance, but the spread of the expected paths is tightened for several days to great confidence, an accurate cluster when the destructive cyclone approaches the coast of Japan.

Better forecasts can also play a key role in other aspects of society, such as renewable energy planning. For example, the improvement of wind power forecasting directly increases the reliability of wind power as a source of sustainable energy and potentially accelerates its acceptance. In an evidence, which analyzed the forecasts of total wind energy generated by groups of wind farms around the world, Gencast was more accurate than ENS.

New generation forecasting and understanding of the climate on Google

Gencast is part of the growing Google Package for a new generation of weather models based on artificial intelligence, including Google Deepmind's artificial intelligence Medium range deterministic forecastsand Google Research's NeuralgcmIN SEEDSAND Flood models. These models start impressed for power supply in the Google search engine and maps and improving forecasting precipitateIN firesIN flood AND Extreme heat.

We deeply value our partnerships with weather agencies and we will continue to cooperate with them to develop AI -based methods that increase their forecasting. Meanwhile, traditional models remain necessary for this work. First, they provide training data and initial weather conditions required by models such as Gencast. This cooperation between AI and traditional meteorology emphasizes the power of a combined approach to improving forecasts and better serving to society.

To support wider cooperation and help in accelerating research and development in the weather and climate community, we made Gencast an open model and published him code AND weightsSimilarly to our deterministic global model of medium range forecasting.

Soon we will issue real -time and historical forecasts from Gencast and previous models that will allow everyone to integrate this input data with their own models and study of work flows.

We are happy to engage in a wider weather community, including academic researchers, meteorologists, data scientists, energy companies and organizations focused on food safety and reacting to the disaster. Such partnerships offer deep observations and constructive feedback, as well as invaluable possibilities of commercial and non -commercial influence, all of which are crucial for the use of our models for humanity.

Thanks

We would like to recognize Raia Hadsell for supporting this work. We are grateful to Molly Beck for providing legal support; Ben Gaiarin, Rozeions and Chris Apps to provide license support; Matthew Chantry, Peter Dueben and the dedicated ECMWF team for their help and feedback; and our nature reviewers for their careful and constructive feedback.

This work reflects the contribution of the co-authors of the newspaper: Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam and Matthew Willson.

LEAVE A REPLY

Please enter your comment!
Please enter your name here