These remarkable discoveries were recently featured in the latest issue of the peer-reviewed journal Science, as detailed in a study conducted by a team of 18 experts from Google DeepMind and Google Research.
This breakthrough signifies a new phase in the competition between China and the US in the realm of artificial intelligence, specifically in the field of machine learning-based weather prediction (MLWP), where mutual recognition for achievements has been prevalent among businesses.
Notable progress in this domain encompasses Microsoft’s ClimaX, NVIDIA’s FourCastNet, and DeepMind’s GraphCataS. China has also made significant contributions with Huawei’s Pangu-Climate and the Shanghai Artificial Intelligence Laboratory’s FengWu.
DeepMind, known for tackling complex challenges, shifted its focus to atmospheric conditions following victories in strategic games like Go and StarCraft.
In September 2021, the organization introduced a model for short-term rainfall prediction, demonstrating an impressive 89% accuracy rate compared to alternative methods, as reported in a Nature publication.
However, the intricacies of extended weather forecasting, spanning four to ten days, posed a more significant challenge.
These forecasts are crucial for sectors such as agriculture, infrastructure, and commerce, necessitating vast amounts of data.
Weather institutions like ECMWF rely on numerical weather prediction (NWP), which leverages sophisticated algorithms based on intricate physical principles, utilizing both present and historical data from satellites and meteorological stations.
Despite its reliability, NWP is known for its high costs and computational demands.
Huawei introduced its Pangu-Sunshine model in November 2022, competing directly with DeepMind’s GraphCast. Image: Epa
The unveiling of the Pangu-Climate model by Huawei in November 2022 marked a milestone in technological advancements. For the first time, this model surpassed traditional NWP methods in accuracy by employing 3D neural networks.
Subsequently, DeepMind enhanced its GraphCast architecture, closely trailing Huawei’s progress. DeepMind has reclaimed its lead in accuracy over time through continuous innovations.
DeepMind acknowledged the significant benchmark set by Huawei while highlighting its own achievements.
According to the research, “GraphCast outperformed its primary competitor, the Pangu-Weather model, in 99.2% of the 252 benchmarks presented in July,” and “it surpassed conventional HRES systems in 90% of the 1,380 test cases.”
Given the similarities between GraphCast and Pangu-Weather, it is uncommon to witness a Chinese company serving as the benchmark in Google’s AI competition.
Both models were fine-tuned using temperature data provided by ECMWF from 1979 to 2017. They are now accessible as experimental models on the ECMWF platform, enabling researchers to forecast various factors such as precipitation and wind speed.
According to Huawei Cloud personnel, AI models generate faster and more precise weather forecasts.
The Earth’s surface is segmented into over a million grids, each operating at latitude and longitude resolutions of 0.25 degrees. While Pangu-Climate provides weekly updates, GraphCast updates its projections every six days.
In terms of model complexity, GraphCast, with 36.7 million parameters, is dwarfed by the billion-level parameters of Pangu-Climate.
Both models exhibit minimal computational requirements during operation. GridCast, trained on 32 of Google’s custom-designed Cloud TPU v4 units over three weeks, can produce 10-day forecasts in less than a minute. On a V100 graphics card, Pangu-Weather can forecast global weather for 24 hours in just 1.4 seconds, over 10,000 times faster than traditional methods.
Despite the accolades in scientific publications, GraphCast and Pangu-Weather are not yet positioned to replace conventional forecasting techniques.
According to Huawei scientist Tian Qi, the design of Pangu-Weather was never intended to supplant industries, businesses, or individuals but rather to serve as an auxiliary tool, enhancing productivity.
This sentiment was echoed in a paper by DeepMind research engineer Remi Lam, emphasizing that the institute’s approach should not be seen as a substitute for established techniques that have undergone extensive real-world testing and development.
