AI weather models are changing forecasting – one street at a time, and GBA is the laboratory

Analysis

Artificial intelligence is transforming weather forecasting, making it possible to deliver predictions with a level of detail and speed that traditional systems cannot match. Where public weather agencies once relied on massive supercomputers to simulate the atmosphere through physics-based models, AI now enables startups and commercial forecasters to generate rapid, hyper-local results — from predicting floods on individual streets to monitoring wind speeds for turbine efficiency.

As Bloomberg reports, “The weather forecasting industry has made big leaps in accuracy but has struggled with hyper-local predictions.” The increasing frequency of extreme weather, driven by climate change, has made these granular forecasts even more critical. AI is offering a low-cost, high-speed solution to meet this demand.

“The application of a previously trained machine-learning weather forecasting model, in terms of computing, costs nothing,” said Peter Bauer, a scientist at the Max Planck Institute for Meteorology and a former official with the European Centre for Medium-Range Weather Forecasts. Unlike traditional global models that try to simulate everything from temperature and rainfall to soil moisture and ozone, AI allows for “specialist systems that do something specific very, very well,” Bauer explained. That shift, he said, marks “a big, big step” forward and creates “endless opportunities for businesses.”

Hong Kong startup leads the way

One of the clearest examples of this new forecasting potential comes from Hong Kong-based startup Stellerus, developed from research at the Hong Kong University of Science and Technology (HKUST). The company has created a system that processes rainfall forecasts using machine learning to produce street-by-street flood predictions in under three minutes.

“Public weather agencies cannot do very customized predictions for a particular industry or business,” said Hui Su, chair professor at HKUST and co-founder of Stellerus. Su’s team is also using AI techniques to analyze satellite data and monitor emissions from specific factories and ships.

Stellerus has partnered with Taiping Reinsurance, a Hong Kong-based subsidiary of China Taiping Insurance Holdings Co., to use these high-resolution forecasts for client alerts. The goal is to provide early warnings with real precision. “A car owner might receive a text message to move their vehicle out of a specific garage set to be flooded in a rainstorm,” Bloomberg reported.

“This is the first time Taiping Reinsurance is bringing high-resolution flood modeling in-house,” said Sheldon Yu, CEO of Taiping Reinsurance. Until now, the company had relied on less detailed models from U.S. risk analytics firms. The decision to shift came after multiple severe rain events in recent years caused widespread flooding in Hong Kong.

“We are pushing the industry to think outside the box of traditional ideas,” Yu told Bloomberg. “And take a more proactive way to deal with catastrophe events and the losses caused to the community.”

Flood modeling work focusing on GBA

Flooding is particularly difficult to model. Localized weather, combined with unpredictable factors like drainage capacity, makes it hard to forecast impacts accurately. But as rainfall intensifies with climate change, floods are responsible for a growing share of insured losses.

More than one in five people around the world are estimated to face major flood risk. South and East Asia are among the most exposed regions. In southern China’s Greater Bay Area – where Taiping Reinsurance is focusing its flood modeling work – nearly one-third of coastal residential zones are projected to experience significantly higher flood threats in the coming decades.

Yu said the company plans to expand its flood modeling work into Macau and other cities in Guangdong province, where high-resolution data could shape policy discussions around creating a province-wide flood insurance program.

Still, some experts warn that greater use of AI in forecasting also brings new challenges. Many commercial forecasting systems operate as closed-source models, meaning their internal mechanics cannot be independently verified.

“In an era of AI forecasting, it’s hard to trace the connection between observations, like satellite images and radar readings, and forecast accuracy,” said Sarah Dance, professor of data assimilation at the University of Reading. “You don’t really understand how changing one weight in the neural network affects the forecast very well.”

A global trend

According to the report, the trend is global. Nvidia Corp. has developed an AI weather model with the Taiwanese government that uses coarse global data sharpened through machine learning. “The more granular you can understand when and where events are going to occur, we believe the better reaction and planning you can make,” said Dion Harris, Nvidia’s senior director for high-performance computing and AI factory solutions.

Taiwan’s Central Weather Administration is currently evaluating the model, according to researcher Cheng-Chin Liu. Meanwhile, the European Centre for Medium-Range Weather Forecasts in February became the first major forecasting agency to launch its own AI model. The Hong Kong Observatory is also referencing multiple AI-based models — including Huawei Technologies Co.’s Pangu-Weather – as it assesses future forecasting tools.

Not all companies are leaving physics behind. Meteomatics, a Swiss-based commercial forecaster, launched a 1-kilometer resolution weather model for the U.S. in January, following its European launch in 2022. It is based on classical physics, supported by AI-enhanced post-processing and drone-gathered atmospheric data. CEO Martin Fengler said running their models requires over 100,000 computing cores, and going global would take “something in the order of 1 million.”

While AI models are less energy-intensive, Fengler believes they can oversimplify complex dynamics. “They tend to take certain shortcuts,” he said. “AI is a great tool to enhance the classical way of modeling, but not fully replace it.”

In the end, most AI weather models still rely on traditional, physics-based datasets collected and maintained by public agencies. Bauer emphasized that governments should continue managing these foundational archives while private companies innovate on top.

“The distribution of duties and responsibilities will slightly shift,” he said. “But I don’t think it’s a problem. I think it creates more opportunities than it creates risks.”