"The climate crisis can strike anywhere in the world. But there is still a big gap in weather forecasting capacity between advanced and developing countries. The World Meteorological Organization (WMO) will standardize an artificial intelligence (AI)-based nowcasting system and distribute a six-hour forecasting framework worldwide."
Lee Hae-suk, Director of the Artificial Intelligence Weather Research Division at the National Institute of Meteorological Sciences, said this at the 2nd workshop of the "AI Nowcasting Pilot Project (AINPP)" held at the National Weather Museum in Seogwipo, Jeju, on 24th, noting, "When disasters occur due to the climate crisis, at least six hours are needed to evacuate people or close roads," and, "The goal is for the world to secure this time through AI models."
The WMO, a United Nations (UN) specialized agency established in 1950, oversees the global weather cooperation system. Korea joined the WMO in 1956 and has served on the executive council since 2007. This workshop is the second AINPP meeting organized by the WMO, bringing together experts from around the world to standardize AI weather forecasting technology.
At this meeting, the WMO aims to validate each country's AI weather forecasting models and derive the optimal model applicable to developing countries by 2027. About 70 people attended, including experts from 24 countries such as Korea, China, Japan, and Saudi Arabia, and people from big tech corporations such as Google, Microsoft (MS), and NVIDIA, as well as academia.
◇ Korea develops AI nowcasting precipitation model
As climate change increases forecast uncertainty worldwide, AI weather forecasting is drawing attention. AI is advancing faster than expected, and unlike supercomputers based on central processing units (CPUs), AI is based on graphics processing units (GPUs), allowing it to learn more diverse information more quickly.
Since May, Korea has been developing and using in real-world forecasting an AI nowcasting precipitation model called "NowAlpha." Based on reflectivity data from weather radar, generative AI extracts precipitation patterns, and a transformer (MeBT) learns them to predict precipitation up to six hours ahead at 10-minute intervals with 1-kilometer resolution.
Generative AI such as ChatGPT, developed by OpenAI in the United States, learns massive amounts of text (sentences) data to understand language and generate content users want. This is called a large language model (LLM). The transformer is the foundational principle of LLMs, an AI technology that learns context and meaning by tracking relationships among sentences and words.
In particular, Professor Choi Jae-sik's team at the KAIST Graduate School of AI applied, for the first time, an "explainable AI" forecasting model to NowAlpha. Explainable AI refers to AI that presents the reasons for its judgments in a way humans can understand. It contrasts with "black-box AI," in which even the algorithm's designers cannot explain the reasons for the AI's decisions. NowAlpha presents the basis for its forecast results—such as typhoons, the monsoon front, and terrain effects—to support forecasters' decision-making.
The Korea meteorological Administration and KAIST researchers are working to advance NowAlpha. Yoo Jae-hoon, a KAIST researcher, said, "By using NVIDIA's AI platform 'Cosmos' to add variables such as temperature, humidity, and wind to NowAlpha, we improved predictive accuracy," and added, "We will verify whether it can be applied in operations by 2027." Director Lee said, "In some summer precipitation forecasts, AI was more accurate than existing physics-based models," and explained, "We are comparing and validating big tech models from Google and NVIDIA with the Korea meteorological Administration's physics-based models."
◇ Google, MS, and NVIDIA also join AI weather forecasting
AINPP, launched by the WMO in 2023, is led by the U.S. National Center for Atmospheric Research (NCAR) and the China Meteorological Administration (CMA). The Korea meteorological Administration, the National Institute of Meteorological Sciences, Japan's RIKEN, the meteorological administrations of Vietnam, Thailand, and Hong Kong, and global corporations such as Google, MS, and NVIDIA are also participating. They conduct mutual validation and share data through monthly online meetings and annual in-person meetings.
Weather forecasting requires vast amounts of data because it must integrate complex variables extending from the surface to the upper atmosphere. To this end, Google has developed precipitation and flood prediction systems such as "MetNet3" and "DGMR."
Shreya Agrawal, a Google software engineer, said, "A recent 1-kilometer precipitation prediction project in the United States required more than 8 petabytes (PB; 1 PB is 1 million gigabytes) of data," and added, "Unlike training general language models, training AI-based weather forecasting models requires massive expense and resources."
A single movie is about 6 gigabytes (GB; 1 GB is 1 billion bytes) of data. One PB is the equivalent of about 174,000 movies. She added, "We are working in parallel on lightweighting and optimization to spread AI weather forecasting technology worldwide and, in particular, to make it usable in developing countries."
NVIDIA, an AI semiconductor corporation, introduced a super-resolution (SR) forecasting model that combines GPU computing technology with deep learning. Deep learning is a machine learning method that mimics the process of visual information processing in the human brain.
Jeff Adie, a senior engineer at NVIDIA, said, "NVIDIA has made improving weather forecasting capability and closing global gaps its global mission," and added, "We have developed AI technology that converts low-resolution data into high resolution to precisely analyze localized weather phenomena such as typhoon tracks."
◇ "We must narrow the gap between advanced and developing countries in weather forecasting"
Experts participating in this workshop said that because climate change affects the entire world, narrowing the gap in weather forecasting between advanced and developing countries is urgent. They said international cooperation is essential above all for this.
David John Gagne, leader of the Machine Integrated Learning for the Earth System group at NCAR in the United States, said, "Developing a model applicable to both advanced and developing countries is key." He said, "Because Earth is spherical, the process of transforming it to a plane for AI to learn is very difficult," and added, "Accurately reflecting the geographical characteristics of weather data in AI models is key not only for long-term climate prediction but also for expansion into other fields."
Yuki Honda, Director of the WMO's Integrated Processing and Prediction Systems Division, said, "The WMO's role is global forecasting cooperation and coordination, establishing AI technology guidelines, and supporting member states," and added, "What matters is cooperation that includes not just technology transfer but also training personnel." He assessed, "Korea has outstanding capabilities in AI and weather research and is an important partner in international cooperation."
Director Yoo said, "Public–private–academia cooperation is not easy, but spreading technology for the global community is important," and added, "Each country must ensure timely evacuation through weather forecasts and extreme weather warnings. This is a grave duty to protect people's lives and a task we must all tackle together."