Huawei Cloud's Pangu-Weather AI model, faster and more accurate than physics forecasts in tests, published in Nature

Huawei/Pangu-Weather

Nature published Huawei Cloud's Pangu-Weather, an AI model trained on 39 years of global weather data that beat ECMWF's operational physics forecast on every variable tested, for forecasts from 1 hour to 7 days. It runs more than 10,000 times faster than that supercomputer model.

• A 24-hour global forecast takes about 1.4 seconds on a single V100 GPU.
• Built on a 3D Earth-Specific Transformer trained on hourly data from 1979 to 2021.
• China's Meteorological Administration said it foresaw Typhoon Mawar's turn five days ahead.
• The first Nature paper with employees of a Chinese tech company as sole authors.