"Much More Accurate Than Ours"... Google AI Forecasts 10-Day Weather in 1 Minute

Google's AI research subsidiary, DeepMind, is gaining attention for producing results that overwhelmingly surpass the existing weather forecasting models of meteorological agencies using AI.

On the 14th (local time), Google DeepMind published a report on its self-developed weather forecasting model, 'GraphCastAI,' in the U.S. scientific journal Science Journal. According to the report, DeepMind's AI was found to perform significantly better than the European Centre for Medium-Range Weather Forecasts (ECMWF) weather prediction model.

GraphCastAI showed accuracy exceeding existing prediction models in as much as 90% of 1,380 measured items, including temperature, pressure, wind speed and direction, and humidity. The British financial media Financial Times (FT) even described this as a "turning point in weather forecasting."

Google DeepMind's 'GraphCast' in operation [Image source=DeepMind]

원본보기 아이콘

ECMWF also ran weather forecasts using DeepMind's AI model alongside its own integrated prediction system, and as a result, acknowledged GraphCastAI's superiority in terms of accuracy.

Matthew Chantry, ECMWF's machine learning coordinator, told FT, "Google DeepMind's model has developed far more accurately and impressively than we expected two years ago," adding, "It outperformed Huawei's PanguWeather, NVIDIA's ForecastNet, and even our own prediction system."

Data center operating DeepMind's weather forecasting artificial intelligence (AI). [Image source=Google Cloud]

원본보기 아이콘

GraphCastAI was trained on 40 years of accumulated data provided by ECMWF. When the meteorological observation center supplies atmospheric conditions from about six hours ago and the current atmospheric state, the AI derives weather forecast results for the next 10 days in just one minute.

What differentiates GraphCastAI from existing weather prediction models is its "black-box" approach. In other words, data is input into the neural network AI, which then independently computes the results. Existing models have predicted weather by applying equation values from meteorological agencies worldwide.

Chantry explained DeepMind's forecasting method, saying, "Once the data training is complete, the operating cost of GraphCast drastically decreases," and added, "Traditional equation-based calculations are quite energy-intensive, but GraphCast is about 1,000 times more energy-efficient."

GraphCast is already being used in real-world weather observations. In September this year, it predicted the landfall of Hurricane Lee in Canada nine days in advance. Existing meteorological prediction models reportedly forecasted the landfall only six days prior. Thanks to Google DeepMind, there was an additional three days gained for hurricane preparedness.

© The Asia Business Daily(www.asiae.co.kr). All rights reserved.