Before the typhoon 'Doksuri' made landfall in northern the Philippines in July 2023, Microsoft’s artificial intelligence (AI) model 'Aurora World' had already predicted its path. Experts and weather authorities identified northern Taiwan, but Aurora's judgment ultimately proved correct. This was four days ahead of the official forecast from the National Hurricane Center (NHC).
AI is now redefining not just a simple weather assistant tool but the accuracy and speed of weather forecasting itself. Global big tech corporations such as Microsoft, Google, and NVIDIA are increasingly entering areas previously led by government agencies, and countries around the world are demanding a fundamental transformation of forecasting systems.
Last month, Microsoft disclosed the performance of the AI model Aurora developed by its research team on its official blog. The model, trained on over 1 million hours of data collected from satellites, radar, and observation stations, is designed to predict various weather and environmental phenomena simultaneously, including typhoons, cyclones, air pollution, ocean waves, and sandstorms.
The existing numerical forecasting models are based on physical laws and go through multiple layers of calculations, whereas Aurora is a statistical inference-based deep learning model. It utilizes high-performance graphics processing units (GPUs) to achieve computation speeds thousands of times faster than supercomputers, and some of its features are currently implemented in Microsoft's MSN Weather app, providing real-time hourly forecast services.
One of Aurora's strengths is the speed of field application. Wes Bruinsma, a researcher at Microsoft Research, noted, 'Traditional numerical forecasting models take years to develop, while Aurora allows a small number of researchers to fine-tune new features within 4 to 8 weeks, enabling a localized, customized forecasting system tailored to specific regions, industries, and disaster response needs to be rapidly established.'
In addition to typhoon prediction, Microsoft presented actual cases such as the sandstorm that hit Iraq and the typhoon Nanmadol, which made landfall in Japan, stating that Aurora had outpaced the Korea Meteorological Administration’s forecasts. In the case of Nanmadol, Aurora reportedly demonstrated higher precision than existing marine forecast systems by predicting the wave height and direction.
Meghan Stanley, a senior researcher at Microsoft Research, explained, 'Thanks to integrating data from various sources, Aurora not only has high overall accuracy but also excels in predicting extreme weather phenomena. This generalization performance is proven by the fact that it can even predict air quality without any specialized knowledge in atmospheric chemistry or pollutant reactions.' She added, 'We didn’t enforce any rules on the relationships among variables, allowing the deep learning model to learn meaningful relationships on its own,' emphasizing 'this is the true power of deep learning in the simulation field.'
As such, the presence of AI in the field of weather forecasting is growing increasingly significant. In late last year, Google DeepMind announced its weather forecasting AI model 'GenCast,' proving an accuracy of 97.2% higher than the European Centre for Medium-Range Weather Forecasts (ECMWF) model. GenCast predicts 15 days of weather on an hourly basis based on 40 years of data, with a prediction time of just 8 minutes. Currently, the ECMWF is also using part of the GenCast algorithm.
NVIDIA is developing a weather forecasting AI software called 'StormCast,' while the National Oceanic and Atmospheric Administration (NOAA) is developing its own AI forecasting system with federal government support. The China Meteorological Administration is also actively considering ways to integrate its AI model 'DeepSeek' into its forecasting system and has even established a dedicated AI research institute.
Countries are rapidly establishing AI-based forecasting systems aimed at all-round innovation not only in forecast accuracy but also in speed, expense structure, and field applicability.
There are claims that Korea Meteorological Administration also needs to pay attention to this technological trend. Currently, the Korea Meteorological Administration is in the process of introducing the 6th supercomputer, aiming for operational use by 2027, with a total project cost of 115.6 billion won, and interest expenses alone reaching 10.3 billion won. However, during last year's parliamentary audit, it was pointed out that 'even after investing billions of won, the improvement in forecast accuracy has been minimal.'
Kang Jaewoo, a professor in the Department of Computer Engineering at Korea University, stated, 'AI forecasting models can quickly and accurately predict similar situations because they learn patterns based on past observational data,' but noted, 'However, predictions can still falter in extreme unusual weather situations that lie beyond the learning scope, such as those occurring once every 100 years.' He added, 'At this point, a hybrid forecasting system that combines AI with physics-based numerical models is the most realistic alternative, and ultimately, a full transition will only be possible when AI that can make predictions beyond the learning range emerges.'