GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy

The new artificial intelligence model develops prediction of uncertainty in weather and risks, providing faster and more accurate predictions up to 15 days before

The weather affects us all – the formation of our decisions, our safety and our way of life. Since climate change pushes the most extreme weather events, accurate and confidential predictions are more important than ever. However, weather cannot be perfectly predicted, and predictions are not particularly confirmed after a few days.

Since the ideal weather forecasts are not possible, scientists and weather agencies use the probability of the potential band, as the model predicts a set of possible weather scenarios. The expectations of this group are more beneficial than relying on one prediction, because they provide decision makers with a more complete picture of the possible weather conditions in the coming days and weeks and the extent of the possibility of each scenario.

Today, in a paper published in Nature, we offer Gencast, the new AI group model (0.25 degrees). GENCAST provides better expectations for both daily events and extremist events from the upper operational system, the European Medium Range weather forecast (ECMWF), 15 days ago. We will issue a symbol, our weights and our style expectations to support the broader weather community.

The development of artificial intelligence weather models

GENCAST represents a decisive progress in predicting a weather -based weather that relies on the previous weather model, which was inevitable, and presented the best appreciation for future weather. On the contrary, GENCAST expects a range of 50 predictions or more, each of which represents a possible weather path.

GENCAT is a spread model, and it is the type of obstetric artificial intelligence model that supports fast rapid developments in generating images, videos and music. However, GENCAST differs from this, as it is adapted to the spherical engineering of the Earth, and learns to be accurately born distributing the complex possibilities for future weather scenarios when giving the latest weather condition as an entrance.

To train GENCAST, we have provided it four decades of historical weather data from ECMWf. These data include variables such as temperature, wind speed and pressure on different altitudes. Learning the model of global weather patterns, with a resolution of 0.25 degrees, directly from these processing weather data.

A new standard for weather forecast

To accurately evaluate GENCAST, we trained it on historical weather data until 2018, and we tested it on data from 2019. GENCAST showed a better prediction skill than ENS ENS, a higher practical prediction system on which many national and local decisions depend every day.

We have tested both systems comprehensively, given the expectations of different variables at different times – 1320 groups in total. GENCAST was more accurate than ENS on 97.2 % of these goals, 99.8 % in bullets larger than 36 hours.

Expectations that the group express express uncertainty by making multiple predictions that represent possible different scenarios. If most predictions show a hurricane hitting the same area, then uncertainty is low. But if they expect different sites, then uncertainty is higher. GENCAST draws the right balance, avoiding exaggeration or reducing its confidence in its expectations.

Google Cloud Tpu V5 requires only 8 minutes to produce 15 -day forecasts in GenCast, and all expectations of the group can be created at the same time, in parallel. Traditional physics -based expectations such as those produced by ENS, at 0.2 degrees or 0.1 degrees, expect a super computer with tens of thousands of processors.

Advanced expectations for harsh weather events

The most accurate expectations for harsh weather risks can help officials protect more lives, avoid damage, and save money. When we tested Gencast’s ability to predict intense heat, cold, and severe wind speeds, GENCAST continuously outperformed ENS.

Now think about tropical hurricanes, also known as hurricanes and hurricanes. Warnings are better and more advanced than where they will strike the ground is invaluable. GENCAST offers superior predictions of these deadly storms paths.

The best predictions can also play a major role in other aspects of society, such as renewable energy planning. For example, the improvement in wind marriage increases directly from the reliability of wind power as a source of sustainable energy, and it is possible to accelerate its adoption. In the experience of proving the principle that has analyzed the predictions of the total wind energy resulting from groups of wind farms all over the world, GENCAST was more accurate than ENS.

Predicting the next generation and understanding the climate in Google

GENCAST is part of the growing Google group of the next generation weather models, including medium -range expectations based on Google DeepMind, Neuralgcm, SEEDS and Floods for Google Research. These models began to run the user’s experiences on Google Search and their maps, and improving the prediction of rain, forest fires, floods and intense heat.

We deeply appreciate our partnerships with weather agencies, and we will continue to work with them to develop methods based on artificial intelligence that enhances their prediction. At the same time, traditional models remain necessary for this work. For one reason, they provide training data and initial weather conditions required by models such as GenCast. This cooperation highlights between artificial intelligence and traditional custody, the strength of the common approach to improving predictions and the best community service.

To enhance the broader cooperation and help accelerate research and development in the weather and climate community, Gencast made a model open and issued its symbol and weights, as we did to the model of the medium -term global global prediction.

We will soon issue actual and historical forecasts from Gencast, and previous models, which will enable anyone to integrate these weather inputs into their workflow models.

We are keen to communicate with the broader weather community, including academic researchers, meteorologists, data scientists, renewable energy companies, and institutions that focus on food security and disaster response. These partnerships provide deep visions and constructive reactions, in addition to invaluable opportunities for commercial and non -commercial influence, all of which are our mission to apply our models in favor of humanity.