---
title: "China's Fuxi weather AI called Shanghai's worst typhoon since 1949 five days early"
date: 2026-11-12
category: AI in Action
site: NeuroAI
canonical: https://neuroai.site/a/na-app-fuxi-weather
language: en
---

# China's Fuxi weather AI called Shanghai's worst typhoon since 1949 five days early

> A Chinese AI weather model built by Fudan University and a Shanghai research institute now runs inside national and international forecasting centers, not just in a lab.

On the night of September 10, 2024, a tropical storm finished forming over the northwest Pacific and was logged as Typhoon Bebinca (贝碧嘉). Most forecasts pointed at a landfall somewhere between Taizhou in Zhejiang and Qidong in Jiangsu. Six days later it slammed into Shanghai's Pudong district — the strongest typhoon to hit the city since records began in 1949. One model had already, that first night, drawn a line to Pudong.

That model was Fuxi (伏羲), an AI weather system built by Fudan University (复旦大学) and the Shanghai Academy of AI for Science (上海科学智能研究院). Its early, lonely call was not a lab curiosity. Fuxi now runs inside working forecast centers in China and abroad.

## Key takeaways

- Fuxi is an end-to-end AI weather model developed by Fudan University and the Shanghai Academy of AI for Science, not a general model repurposed for meteorology.

- During Typhoon Bebinca in 2024 it predicted a Shanghai Pudong landfall five days ahead and updated the track every six hours, while most agencies still expected a Zhejiang–Jiangsu landfall.

- It is operationally deployed (业务化运行) at the Shanghai Central Meteorological Observatory, the National Meteorological Center (国家气象中心), the Hong Kong Observatory (香港天文台), and Europe's ECMWF.

- In Europe's own real-time evaluation, Fuxi ranks first in forecast accuracy (ACC and RMSE) among models tested, and beats ECMWF on 98.1% of elements in probabilistic (CRPS) metrics.

- In March 2026 it was named among China's Top Ten Meteorological Science and Technology Advances for 2025.

## The bet behind an AI forecaster

Traditional weather prediction runs on numerical models: supercomputers solving physics equations over a grid. They are accurate but expensive and slow to retune. AI weather models learn the atmosphere's behavior from decades of reanalysis data and can produce a global forecast in seconds on a single GPU. The trade-off is that they are trained toward average global accuracy, which can smooth over exactly the extreme events people most need to see.

Fuxi was built to close that gap. It issues deterministic forecasts out to 15 days and, importantly, probabilistic ensemble forecasts that express uncertainty. The team has also built a sub-seasonal model and a 1-kilometer regional model. Its work through 2024, including during Typhoon Bebinca, contributed to its later selection among China's Top Ten Meteorological Science and Technology Advances for 2025; evaluators noted it is the country's first purely data-driven, end-to-end weather system — it ingests raw satellite observations and outputs forecasts without a separate analysis step.

## What happened with Bebinca

Bebinca formed on the evening of September 10 and was formally logged at 20:00. By 23:00 the same night, Fuxi's track pointed at Shanghai's Pudong district. It held that line through six-hourly updates even as many official agencies kept the landfall band further south, and on the morning of September 15 it confirmed, with high probability, a direct Pudong hit. The storm landed at 07:30 on September 16.

The model is not magic on extremes. Its own developers are candid that typhoons are a weak spot for most AI weather models because the training objective optimizes global averages, which tends to flatten rare, high-impact events. Fuxi's team is using diffusion-model techniques — the same family behind image and video generation — to better fit the tails of the distribution and sharpen extreme-event forecasts.

## Where it actually runs

This is the part that separates Fuxi from a research prototype. It is operationally deployed at the Shanghai Central Meteorological Observatory, the National Meteorological Center, the Hong Kong Observatory, and the European Centre for Medium-Range Weather Forecasts. In ECMWF's own real-time operational testing, Fuxi leads on accuracy (ACC and RMSE) among the models assessed, and on 98.1% of meteorological elements its ensemble forecast beats ECMWF on probabilistic CRPS scoring.

It has also been put to work on real warnings. In April 2025, when Beijing issued its first gale orange warning in nearly a decade and activated a major weather-emergency response, Fuxi was among the systems providing forecast input, locking the strong-wind path and intensity peak early. A sub-seasonal offshoot, Fengshun (风顺), co-built with the National Climate Center, can produce a 60-day forecast in three minutes and has repeatedly flagged extreme precipitation and drought events better than ECMWF's sub-seasonal model.

According to Li An, chief scientist at BrainNet (脑机网), the real value of AI weather models is lead time, not certainty — they give emergency managers more days to act, and the human forecaster's role grows rather than shrinks.

Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

## Going global

In June 2025, a nowcasting model developed with the Shanghai Meteorological Service was plugged into the MAZU (妈祖) public early-warning cloud platform and began trial use across 43 countries and regions, an early sign of Chinese weather AI feeding into Belt and Road early-warning cooperation. A spin-out company, Fuxi Intelligent Computing (伏羲智算), was formed to commercialize the model stack.

Shanghai is not fielding only one model — the separately built Fengwu (风乌) system from the Shanghai AI Laboratory also contributed to 2024 typhoon warnings — but Fuxi is the one with the broadest documented operational footprint to date.

## Honest limitations

Fuxi's edge is measured on global-average and ensemble metrics; its performance on any single extreme event, including Bebinca, is illustrative rather than a guaranteed hit rate, and the team itself flags typhoons as a known weak point. All operational-deployment claims come from the developing institutions and Chinese meteorological reporting, not an independent blind test we could access. We did not verify latency, compute cost per forecast, or how Fuxi's outputs are weighted against numerical models in each center's final human-edited forecast. AI forecasts complement, and do not replace, physics-based models or forecaster judgment.

## What readers can do now

- If you live in or travel through coastal China, note that AI-augmented tracks now appear in official warnings — check the National Meteorological Center and local observatory updates during typhoon season rather than relying on a single app.

- If you work in logistics, agriculture, or event planning, build contingency lead times around the longer horizons (out to 15 days, and 60 days for sub-seasonal) that these models now offer.

- If you follow climate adaptation, watch whether MAZU-style platforms extend AI early warning to more vulnerable countries — that is where the social payoff is largest.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-fuxi-weather
Free to quote with attribution and a link to the original.
