Typhoon Krathon's erratic movement and prolonged stay over Taiwan have sparked discussions about which AI models can most accurately predict its trajectory. Research from various companies indicates that the key factors in improving weather prediction accuracy remain data, algorithms, and vast computing resources. A significant challenge lies in integrating new technologies into existing systems without substantially increasing computational demands.
The Facebook page "Weather Observation and Forecast" recently highlighted seven AI models commonly used for weather forecasting. These include the European Centre for Medium-Range Weather Forecasts' (ECMWF) AIFS, Nvidia's SFNO, Microsoft's Aurora, Google's GraphCast, Huawei's Pangu, Fudan University's Fuxi, and the traditional numerical forecasting model IFS used in Europe.
Typhoon Krathon made landfall near Xiaogang, Kaohsiung, around noon on October 3, contrary to earlier forecasts suggesting landfall in Taitung or Hengchun. Over the past few days, different forecasting models have continuously updated their predictions, illustrating the challenges of accurately tracking a typhoon's movement.
Observing the efforts of various software developers in the climate AI field, it's evident that data and algorithms are central to improving prediction accuracy. The focus includes integrating AI technologies into traditional systems, employing 3D simulations to fill data gaps, and leveraging massive computational resources to advance model development.
In November 2023, Google DeepMind introduced its GraphCast model, claiming that its medium-range weather predictions achieved unprecedented accuracy, surpassing the long-standing gold standard—ECMWF's High-Resolution Forecast (HRES). GraphCast, now open-source, is even used by ECMWF. Despite its computational intensity, GraphCast significantly boosts forecasting efficiency, able to generate 10-day predictions in under one minute with a single Google TPU v4 virtual machine. Traditional models like HRES, on the other hand, require supercomputers to run for several hours.
Microsoft launched its Aurora model in mid-2024, which reportedly outperforms Google GraphCast in 94% of key metrics. Compared to the traditional IFS model, Aurora boasts up to a 5,000-fold increase in computational speed.
Microsoft's approach appears to prioritize 3D simulations, as Aurora's training involved approximately one million hours of meteorological simulation data, offering more comprehensive tracking of atmospheric changes. Aurora's spatial resolution reaches 0.1° x 0.1° (about 10km x 10km), compared to GraphCast's 0.25° x 0.25° (28km x 28km). With a model size of 1.3 billion parameters, Aurora also has applications in air pollution monitoring.
It's challenging to definitively determine which AI model performs best for any single meteorological event, as multiple factors like data inputs and usage conditions need to be considered. However, it's clear that competition in this field is intense.
Following Microsoft's release of Aurora, Google unveiled its Neural GCM model in July 2024. Google researchers emphasized that the goal isn't to replace traditional physical models with AI but to find the best ways to integrate them while reducing computational costs. Whether this new model improves spatial resolution and accuracy remains to be seen.
