AI has shown the ability to predict weather 10 days in advance with more accuracy than current state-of-the-art simulations. This was achieved by Google DeepMind, although the use of AI for weather prediction has raised concerns among meteorologists.
Existing weather forecasts rely on mathematical models that are based on physics and computer simulations to predict future weather events. These models have improved over time but require powerful supercomputers and high energy demands. Google DeepMind’s approach, however, focuses on training their GraphCast AI model using historical weather data from various sources such as satellites, radar, and ground measurements.
The GraphCast AI model uses real meteorological readings from over a million points on the planet and predicts the weather six hours ahead. By continuously using these predictions as inputs for further forecasting, DeepMind claims that their model outperformed the ECMWF’s high-resolution forecast in more than 90% of the tested data points, with accuracy reaching as high as 99.7% at some altitudes.
Matthew Chantry at the ECMWF expressed excitement over the potential of AI in improving and lowering the energy costs of weather forecasts. However, some meteorologists have raised concerns about relying solely on AI for weather forecasting. Ian Renfrew of the University of East Anglia highlighted the current limitations of AI in handling data assimilation and expressed concerns about completely abandoning deterministic models based on physics and chemistry.
The greater accuracy provided by AI in weather forecasting is promising. However, the potential tradeoffs in terms of public trust and the ability to interrogate and improve forecasts are important factors to consider.
DeepMind’s approach to weather forecasting using AI differs significantly from traditional models by relying on historical data patterns. The potential of AI to improve forecast accuracy and reduce computational time is acknowledged. However, the challenges of data assimilation and the importance of public trust and model interrogatability need to be carefully considered before fully relying on AI for weather forecasting.
