Shulou(Shulou.com)11/24 Report--

The big weather prediction model launched by Google's DeepMind Lab has been published in Science magazine.

It can directly predict the weather for the next 10 days in less than a minute.

In terms of accuracy, it surpasses the most advanced human system in 90% of the index, and is the first time in the AI weather model!

DeepMind's weather model, called GraphCast, is currently open source.

Its resolution is 0.25 degrees longitude / latitude (about 28 × 28 kilometers at the equator), while the current highest resolution is 1 degrees.

This resolution is equivalent to dividing the earth's surface into more than 1 million grids, each of which can produce hundreds of pieces of predictive data, totaling hundreds of millions.

Different from the traditional prediction methods, GraphCast prediction mainly depends on the laws in the data, rather than using the physical equations established by human beings.

Compared with the most accurate HRES forecast of human beings, 90% of the 1380 test indicators of GraphCast are more accurate.

If the forecast is limited to the troposphere, the percentage of GraphCast defeating HRES is as high as 99.7%.

Some netizens on YC said that the word "impressive" is no longer enough to describe this achievement.

So what exactly is the performance of GraphCast's forecast?

The best way for people to surpass the 90 per cent index is that each of the more than 1 million grids divided by GraphCast can produce 227 pieces of prediction data.

It includes 6 atmospheric variables (including specific humidity, wind speed and wind direction, temperature, etc.) at 37 different heights.

On the earth's surface, GraphCast can also predict five variables, including temperature, wind speed and wind direction, as well as mean sea level pressure.

The complete variable types and specific heights (in air pressure, in hPa) are shown in the following table:

To compare the performance of GraphCast and HRES, the researchers selected historical data from 2018 (GraphCast training data up to 2017) from ERA5 reanalysis data from the European Centre for medium range Weather Forecast (ECMWF).

The researchers asked HRES and GraphCast to make "predictions" in the current situation, and then compared their "predictions" with ERA5.

In the 500hPa height field, the RMSE (root mean square error, the lower the value, the better) and ACC (anomaly correlation coefficient) of GraphCast are significantly better than HRES.

Among the 1380 data points selected by the researchers in 50-1000hPa, GraphCast has a significant advantage over HRES,89.9% in 90.3% (in group d below, blue indicates that GraphCast is superior to HRES, and the deeper the advantage is, the more obvious).

In addition to these data, GraphCast also has a clear advantage in predicting extreme weather.

For tropical cyclone tracks, the median error of GraphCast is lower than that of HRES, especially in the first 4.75 days, the advantage begins to become obvious (figure a, b below).

When predicting the water vapor flux based on atmospheric rivers (Atmospheic River), the RMSE value of GraphCast is also significantly lower than that of HRES (figure c below).

When predicting heat wave, GraphCast is also more accurate than HRES at 12 hours, 5 days and 10 days in advance (figure d below).

In September, GraphCast successfully predicted Hurricane Lee in the North Atlantic nine days before landfall, compared with a maximum of six days in advance using traditional methods.

GraphCast has not only high accuracy, but also very fast prediction speed.

It takes less than a minute to make a 10-day forecast using GraphCast on a Google TPU v4 machine.

By contrast, using traditional methods such as HRES can take hours even on a supercomputer.

So, how does GraphCast achieve efficient and accurate weather forecasts?

Do not use physical equations, the prediction depends entirely on the data analysis workflow, input from 6 hours ago to the current meteorological data, GraphCast can predict the weather in the next 6 hours.

The predicted data can be used as the new "current" state, continue to iterate to predict the weather conditions after 10 days at most.

At the principle level, GraphCast uses machine learning and graph neural network (GNN) architecture, which includes one layer of encoder and decoder, and 16 layers of middle layer with 36.7 million parameters.

It realizes the prediction only by learning the existing meteorological data and does not rely on the physical equations established by human beings.

GraphCast encodes and maps the meteorological data of 0.25 degree grid to neural network, and the results are restored to meteorological data by the decoder after transmission and calculation.

During the training, GraphCast used 1979-2017 weather reanalysis data from the ERA5 data set for nearly 40 years, including satellite images, radar and weather station results.

ERA5 is a global optimal reconstruction data generated based on 4DVar method and assimilation observations, covering time from the 1940s to the present, while space covers the whole world.

If more recent data are used, the accuracy of GraphCast's predictions can continue to improve.

In the next step, DeepMind plans to build an ensemble forecast model to adapt to the actual weather uncertainty and further enhance the accuracy of the forecast.

Paper address:

Https://www.science.org/doi/10.1126/science.adi2336

Reference link:

[1] https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/

[2] https://www.ft.com/content/ca5d655f-d684-4dec-8daa-1c58b0674be1

This article is from the official account of Wechat: qubit (ID:QbitAI). Author: Creasy.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.