A study published in Science Advances puts GraphCast, Pangu-Weather and Fuxi in difficulty on the rarest events: trained on the past, they underestimate the intensity of records.
Weather models based on machine learning have built their reputations on speed and flattering aggregate scores. A study by Zhang and his co-authors, published in Science Advances and analyzed by Carbon Brief, shifts the focus to what matters most for risk prevention: record-breaking events.
The researchers compared three widely cited systems — Google DeepMind's GraphCast, Huawei's Pangu-Weather and Shanghai-developed Fumi — with traditional physical forecasts for a very large volume of extreme events; In 2020 alone, the analysis lists about 160,000 heat records, 33,000 cold records and 53,000 wind records. In this field, physical models prevail.
The explanation lies in the very nature of learning: these systems depend closely on their training data and tend to underestimate the frequency and intensity of extremes, precisely because these are rare in history. One of the authors speaks of a "warning shot" against the idea of replacing conventional models too quickly, and argues for a combination of the two families of approaches rather than a substitution — especially since warming is pushing the atmosphere towards states with no close equivalent in the archives from which these models have learned.
Read the full article on the original source
Go to the source →