Google DeepMind's AI has the upper hand; predicts faster and more accurate weather and severity

Search

  • Home
  • TECH
  • Google DeepMind’s AI has the upper hand; predicts faster and more accurate weather and severity

Google DeepMind’s AI has the upper hand; predicts faster and more accurate weather and severity

According to studies published, Google DeepMind's GraphCast model outperformed the existing gold standard for weather forecasting by both accuracy and speed up to 10 days in the future.

|
November 15, 2023 12:01 IST
Google DeepMind’s AI has the upper hand; predicts faster and more accurate weather and severity
Google DeepMind’s AI has the upper hand; predicts faster and more accurate weather and severity

New Delhi: Scientists and experts have struggled for millennia to develop reliable techniques for predicting weather patterns. DeepMind, Google’s artificial intelligence division, however, asserts that its new AI model significantly improves upon previous weather-predicting methods.

Google has introduced GraphCast, an AI model that can provide medium-range weather predictions with “unprecedented accuracy,” in the research journal Science. If they could be predicted more quickly and accurately, it would allow us to better prepare for natural catastrophes and maybe save lives.

According to studies published, Google DeepMind’s GraphCast model outperformed the existing gold standard for weather forecasting by both accuracy and speed up to 10 days in the future.

When compared to the model from the European Centre for Medium-Range Weather Forecasts (ECMWF), GraphCast performed better in more than 90% of over 1,300 test locations. More than 99 per cent of meteorological variables, including precipitation and air temperature, were correctly predicted by GraphCast, whereas the ECMWF’s model was off by a wide margin.

Importantly, GraphCast can provide meteorologists with early warnings of severe circumstances like temperature and cyclone tracks, much before the time when normal models would predict them. According to Remi Lam, a staff research scientist at Google DeepMind, GraphCast properly forecasted Hurricane Lee in Nova Scotia nine days beforehand in September. Only six days before the storm hit Nova Scotia was the location determined by conventional weather predicting models.

Traditional weather forecasting relies heavily on enormous computer models. Because the simulations include many various meteorological factors, including temperature, precipitation, pressure, wind speed and direction, humidity, and cloud cover, running them consumes a lot of resources and takes a long time.

The machine learning in GraphCast allows for these computations to be completed in under a minute. It makes forecasts based on forty years of weather data instead of equations based on physics. Using graph neural networks, GraphCast divides the planet up into a grid of more than a million points. The model projects conditions like temperature, wind speed and direction, mean sea level pressure, and humidity at each grid point.

Aditya Grover, an associate professor of computer science at UCLA and the creator of ClimaX, a foundation model that facilitates a wide range of weather and climate modelling activities, calls GraphCast a “reckoning moment” for weather prediction.

GraphCast has several limitations, however. According to ECMWF’s director of Earth system modelling Peter Dueben, the organisation is still behind in several areas compared to traditional weather forecasting models, particularly in terms of precipitation. In order to improve their forecasts, meteorologists will still need to employ both traditional and machine-learning algorithms.

  • Share:
Latest