Analyses suggest that AI weather models outperform supercomputer counterparts, but gaps remain

The world’s most powerful weather supercomputers and next-generation AI models are at odds over the projected path of Typhoon Champi, though some analyses suggest that AI has already surpassed supercomputers in its typhoon tracking and prediction abilities.

Typhoon Champi was about 620 kilometers (385.3 miles) southeast of Taipei, Taiwan, as of 3 p.m. on Thursday, according to the Korea Meteorological Administration (KMA). The season’s 30th typhoon was moving northwest at 10 kilometers per hour (6.2 miles per hour) after weakening from Category 5, the highest level, to Category 4. Its eventual course could bring direct or indirect effects to Korea, China and Japan.

Conventional supercomputer-based weather models and AI have differed somewhat in their forecasts since the typhoon first formed. AI models have consistently predicted that Champi will turn sharply northeast before reaching Taiwan, then pass south of Jeju Island on its way to Japan.

The European Centre for Medium-Range Weather Forecasts (ECMWF) model and other conventional forecasts, however, have projected that Champi would travel farther west, either crossing northeastern Taiwan or reaching waters well off China’s coast before changing course.

The U.S. Joint Typhoon Warning Center sided with AI. It attributed its support for the AI forecast to the proven accuracy of models, such as Google DeepMind’s, in predicting northeastward turns in past typhoons under atmospheric conditions similar to Champi’s.

The KMA’s forecast also largely aligned with the AI projections.

“AI models have shown strong accuracy in forecasting typhoon paths, and we favored [their prediction of Champi’s] northeastward turn after assessing the storm’s movement and surrounding pressure systems,” a KMA National Typhoon Center official said. “However, the forecast remains highly uncertain, and we are closely monitoring [Champi’s] trajectory.”

Korea’s weather agency began incorporating AI into typhoon forecasts on a large scale last year. The technology has since evolved from a supplementary tool into a central part of typhoon forecasting.

Research suggests that AI may already outperform supercomputers in forecasting the expected paths of typhoons.

AI models produced relatively smaller errors in predicting typhoon tracks compared to conventional numerical models in most cases, according to an analysis of typhoons over the past two years by the Typhoon Research Center at Jeju National University.

The ECMWF’s Integrated Forecasting System, one of the world’s leading numerical weather prediction models, recorded an average typhoon track error of 205.5 kilometers across one- to five-day forecasts, compared to less than 200 kilometers for most AI models.

AI also outperformed conventional numerical models by as much as 150 kilometers in forecasts five days ahead.

“AI demonstrated a decisive edge in accuracy over traditional numerical models for long-range predictions,” said Moon Il-ju, the director of Jeju National University’s Typhoon Research Centre. “Using forecasts from multiple high-performing AI models together could make typhoon predictions more accurate.”

But AI has proved less reliable in forecasting a typhoon’s strength and speed. Limited data can lead the models to underestimate storm intensity, according to some analyses.

Experts say that a hybrid system that harnesses the advantages of AI and supercomputers is essential to improving weather forecasts.

“Numerical models predict typhoon paths by analyzing the forces at work in the atmosphere, while AI can use vast amounts of data to identify patterns those models miss,” said Kang Nam-young, a professor in the Department of Geography at Kyungpook National University. “The two should be used to complement each other.”

BY CHON KWON-PIL [[email protected]]