Large-Scale Weather Prediction Model Launched by Google’s DeepMind Laboratory

DeepMind’s large model is listed in Science: 10 days of weather data prediction in 1 minute, 90% of indicators surpass the strongest human model.

Continue after the short break

ad: GET $10,000+ Free Elon Musk Money on X (Twitter) Click Here
ad: Monetize your Facebook Account for Monthly $500 Click Here
ad: How to Earn Free $100 on Binance & Withdraw Click Here

Check Also:

The large-scale weather prediction model launched by Google’s DeepMind laboratory has been published in Science magazine.

In less than 1 minute, it can directly predict the weather for the next 10 days.

In terms of accuracy, it surpasses the most advanced human systems in 90% of indicators, a first for AI weather models!

DeepMind’s weather model is called GraphCast and is now open source.

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

Such a resolution is equivalent to dividing the earth’s surface into more than 1 million grids, and each grid can generate hundreds of prediction data, with the total number reaching hundreds of millions.

Different from traditional prediction methods, GraphCast prediction mainly relies on the patterns in the data for prediction, rather than using physical equations established by humans.

Compared with the most accurate human HRES forecast, 90% of GraphCast’s forecast results are more accurate among 1380 test indicators.

If the prediction scope is limited to the troposphere, the index ratio of GraphCast beating HRES is as high as 99.7%.

Some netizens on YC said that “impressive” is no longer enough to describe this achievement.

So, what is the specific prediction performance of GraphCast?

90% of indicators surpass human best methods

On more than 1 million grids divided, each grid divided by GraphCast can produce 227 prediction data.

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

On the Earth’s surface, GraphCast can also predict five variables including temperature, wind speed and direction, and mean sea level pressure.

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

In order to compare the performance of GraphCast and HRES, the researchers selected historical data from 2018 (GraphCast training data ends in 2017) from the ERA5 reanalysis data of the European Center for Medium-Range Weather Forecasts (ECMWF).

The researchers asked HRES and GraphCast to “predict” under the current situation, and then compared their “predictions” with ERA5.

In the 500hPa height field, GraphCast’s RMSE (root mean square error, the lower the value, the better the performance) and ACC (anomalous correlation coefficient) indicators are significantly better than HRES.

Among the 1,380 data points from 50-1000hPa selected by the researchers, 90.3% of GraphCast is better than HRES, and 89.9% have a significant advantage (in group d of the figure below, blue indicates that GraphCast is better than HRES, and the deeper the color, the more obvious the advantage).

In addition to these data, GraphCast also has obvious advantages in predicting extreme weather.

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

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

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

In September this year, 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 is not only highly accurate but also very fast in prediction.

Using GraphCast on a Google TPU v4 machine for 10-day predictions takes less than a minute to complete.

In comparison, using traditional methods such as HRES takes several hours even on a supercomputer.

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

No physical equations are used, predictions rely entirely on data analysis

In the workflow, enter the weather data from 6 hours ago to the current time, and GraphCast can predict the weather in the next 6 hours.

The predicted data can be used as the new “current” state to continue iterative predictions, and the weather conditions can be predicted up to 10 days later.

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

It only achieves prediction by learning existing meteorological data and does not rely on physical equations established by humans.

GraphCast encodes and maps the meteorological data of the 0.25-degree grid to the neural network, and the calculated results are then restored to meteorological data by the decoder.

When training, GraphCast uses weather reanalysis data from 1979-2017 in the ERA5 data set over the past forty years, including satellite images, radar and weather station measurements.

ERA5 is the world’s best reconstructed data generated based on the 4DVar method and assimilated observations, covering the time from the 1940s to the present and the space covering the whole world.

And if more recent data is used, the accuracy of GraphCast’s prediction results can continue to improve.

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

Paper address:

Reference links:

  • [1]

  • [2]

This article comes from the WeChat public account: Qubit (ID: QbitAI) , author: Cressy


Advertising statement: The external jump links (including but not limited to hyperlinks, QR codes, passwords, etc.) contained in the article are used to convey more information and save selection time. The results are for reference only. All Naijatechnews articles include this statement.

Article Editor in Chief: Coker | I want to correct article mistakes.