- November 17, 2023
- AI Projects
This week, researchers from DeepMind rolled out an incredibly precise AI tool for weather forecasts that they claim signifies a milestone in the field. Featured in a Science publication, staff research scientist from DeepMind, Remi Lam, mentioned that their new algorithm surpasses existing forecasting techniques in both speed and precision, delivering detailed forecasts of atmospheric conditions including pressure, temperature, humidity, and wind direction up to a decade ahead.
Named GraphCast, this model “outdoes the leading operational deterministic models on 90% of 1,380 assessment benchmarks,” as per Lam’s statement. Traditional weather prediction has long depended on NWP (Numerical Weather Prediction) models, elaborate calculations grounded in physics that handle an array of factors. These calculations interpret data from terrestrial weather stations, orbiting satellites, and marine instruments to map out the circulation of warmth, air currents, and moisture on a global scale. Insights from seasoned meteorologists and perpetual enhancement of algorithms add to the preciseness of the forecasts. Nevertheless, even with advanced computers, this method is elaborate, resource-hungry, and expensive.
DeepMind has enhanced this method by teaching their machine-learning models with data spanning over 39 years. Ditching the physics-driven calculations, the neural network behind GraphCast processes historical records at rates immensely surpassing those of traditional forecasting systems. For instance, using 2018’s data, GraphCast was able to churn out forecasts reaching 10 days into the future in under 60 seconds—a task that would have taken hours using the older techniques. The accuracy of GraphCast was significantly improved. “In the lowest layer of the atmosphere, the troposphere, which impacts our daily weather the most, GraphCast outmatched [traditional models] in over 99% of tests,” Lam elaborated.
The AI system also trumped standard tools in 90% of predictions at different atmospheric strata. Computer science expert, Aditya Grover at UCLA, notes that “GraphCast is at the forefront within the field of AI models.” European Centre for Medium-Range Weather Forecasts’ Matthew Chantry conveyed to the Financial Times the surprising advancement in AI-powered meteorological predictions over the last couple of years.
While GraphCast may not excel in hyper-local predictions like rain probabilities in a small area, it shines in broader-scale weather events like tropical storms and notable temperature fluctuations. Lam emphasizes that this technology shouldn’t be seen as a substitute for conventional forecasting techniques but rather as compelling proof that AI-aided predictions can rise to the challenge of real-world forecasting issues, potentially boosting and refining current leading approaches.
For a while, weather prediction has been taken lightly, as echoed by comedian Rodney Dangerfield’s demeanor and characterized by public skepticism about its reliability. George Carlin humorously summarized weather reports as alternating extremes of temperature. Though GraphCast isn’t flawless, and jesters may continue to rib forecasters, the remarkable advancements in AI tech herald a promising future for this domain in weather prediction.
