FOCUS: AI Models Make Inroads into Weather Forecasting

Tokyo, Sept. 15 (Jiji Press)–Artificial intelligence is gaining ground in weather forecasting. Weather forecasts have traditionally relied on numerical weather prediction, in which supercomputers process observational data and simulate atmospheric conditions on the basis of the laws of physics. In recent years, however, AI-based weather models have emerged. These systems analyze vast amounts of historical data to predict future weather patterns. Meteorological agencies and private-sector companies around the world are now developing proprietary AI forecasting models. Improved Accuracy “Meteorological agencies around the world have spent 70 years staking everything on physical models and competing fiercely to improve their accuracy,” said Daisuke Abe, an executive officer at private weather forecasting company Weathernews Inc. “It was shocking to see them overtaken almost instantly by AI, which emerged from a completely different direction.” Research into AI weather modeling has advanced rapidly over the past four years. In addition to private-sector companies such as U.S. semiconductor powerhouse Nvidia Corp. and technology giant Google LLC, public institutions including the European Centre for Medium-Range Weather Forecasts and the U.S. National Weather Service are developing and refining their own models. Some AI models have proved more accurate than conventional numerical weather prediction systems in capturing large-scale atmospheric patterns and have reduced errors in typhoon track forecasts by around 40 pct. Nvidia’s Earth-2 is envisioned as a digital twin of the planet–a virtual replica of the physical world. It currently provides services including medium-range weather forecasts extending up to 15 days and Earth-2 Nowcasting, which delivers highly accurate forecasts of severe weather up to six hours in advance. Stan Posey, program manager for the Earth System Science Domain at Nvidia, indicated that his team was prepared to consider regional customization to account for differences in geography and climate. “You all have different scenario conditions depending on the geography of your region, etc. … We might have to customize our models for that region,” he told Jiji Press. The Japan Meteorological Agency is working to introduce its own AI weather model around 2030. Officials said the agency decided to pursue independent development to establish a framework that would allow it to maintain stable operations and improve forecasting accuracy on its own over the long term. The agency has identified improvements to information for disaster preparedness and response as a key policy priority. “We need to actively adopt promising technologies (for that purpose),” said Yoichi Hirahara, the agency’s director for AI strategy planning and coordination. He expressed hope that “more accurate information will enable appropriate disaster preparedness and response measures.” Strengths of Conventional Systems AI weather models have not yet reached the point where they can fully replace conventional numerical forecasting systems, however. While AI has demonstrated high accuracy in relatively large-scale forecasts, conventional models retain an advantage in predicting rare phenomena for which training data are limited, as well as small-scale events such as localized torrential rainfall. The agency plans to combine AI-based weather models with conventional systems to capitalize on the strengths of both. Hirahara emphasized that humans will remain in control. “What will not change is that knowledgeable forecasters will be responsible for making the final decisions about what information to issue, including forecasts,” he said. The application of AI technology is expected to extend beyond weather forecasting, with growing hopes that it can also be used to predict river flooding and wildfires. Weather forecasting, long centered on physics-based models, may advance further through the integration of AI technology. END [Copyright The Jiji Press, Ltd.]