The World's First Full-Chain Space Weather AI Forecast Model 'Fengyu'! Led by the National Satellite Meteorological Center, Co-Developed with Nanchang University and Huawei}

In just a few minutes as a communication satellite flies overhead at the first cosmic velocity, over a million people are connected worldwide through its network. In reality, there are thousands of such satellites orbiting Earth. While we enjoy fast and convenient satellite network services, a system called “Fengyun Space” silently sends early warning signals to these satellites, alerting them of an impending impact from solar activity that will arrive in about 24 hours. After receiving the warning, ground control activates emergency plans, calmly responding to the solar storm and averting a space weather crisis.

This scenario exemplifies China’s advancement toward intelligent space weather forecasting, with one of the core technologies being the “Fengyu” model. Wang Jinsong, director of the National Satellite Meteorological Center (National Space Weather Monitoring and Early Warning Center), states that this is the world’s first full-chain space weather AI forecast model.

1. Invisible 'Cosmic Tsunami' — Why Do We Need a Space 'Meteorologist'?

Currently, the Sun is in a high activity phase, with eruptions and other unpredictable events like invisible 'cosmic tsunamis' threatening satellites, aircraft, and critical ground infrastructure.

Predicting these storms, which traverse 150 million kilometers, is extremely challenging. Traditional models rely on numerical simulations, but space weather involves complex physical interactions across layers like the Sun, interplanetary space, magnetosphere, and ionosphere. This makes models computationally intensive, slow, and often unable to meet real-time demands, with some like COSMOS taking up to 22 million GPU hours for training.

2. 'Fengyu' Debuts — The World’s First Full-Chain Space Weather AI Forecast Model

Faced with these challenges, AI technology has opened new avenues. On July 26, 2025, at the World Artificial Intelligence Conference’s meteorology session, the National Satellite Meteorological Center (National Space Weather Monitoring and Early Warning Center), in collaboration with Nanchang University and Huawei, officially released the “Fengyu” model.

Wang Jinsong believes that the success of the “Fengyu” model has established a triple approach of physical models, numerical forecasting, and AI, greatly enhancing China’s space weather prediction capabilities.

Zhang Dixuan, President of Huawei’s Computing Cloud Business, states that the “Fengyu” space weather model, based on the MindSpore Science suite and Ascend hardware, achieves a full-process application from training to inference, covering solar wind, magnetosphere, and ionosphere, with superior efficiency, accuracy, and system adaptability compared to traditional platforms.

Revolutionary Architecture: From 'Independent Battles' to 'Joint Operations'

Historically, space weather models were often isolated, focusing on specific regions like solar wind or ionosphere. Wang Jinsong points out that this 'battle alone' approach fails to reflect the causal physical relationships across the entire Sun-Earth system, limiting forecast accuracy.

To address this, the “Fengyu” model pioneered a “chain training structure”, integrating forecasts into a collaborative system. This includes three key technological innovations.

First, world’s first end-to-end AI modeling of the entire chain. “Fengyu” is the first system globally to realize end-to-end AI modeling from the Sun to Earth, including models for solar wind “Xufeng”, Earth’s magnetic field “Tianmu”, and ionosphere “Diqu”. These regional models use a chain training mode and pluggable architecture, enabling flexible and efficient updates, with new models like auroras under development.

Second, innovative space weather upstream-downstream intelligent coupling technology. The “Fengyu” model’s unique “intelligent coupling optimization mechanism” (also called the coupling optimizer) is key to the collaboration of the three regional models. Chen Zhou emphasizes that this is a multi-region model coupling optimization method based on deep neural networks, which achieves collaborative optimization and hour-level fast forecasting through perception responses and structural adaptive adjustments across regions.

For example, the output of the “Xufeng” model feeds into the downstream “Tianmu” and “Diqu” models. The coupling optimizer calculates multiple loss functions (Loss1, Loss2, Loss3, LossX1, LossX2) to optimize all models collaboratively.

This allows the “Fengyu” model to more accurately simulate the process of solar wind impacting Earth’s environment and depict complex interactions between magnetic fields and the ionosphere, fundamentally improving understanding and prediction of space weather.

Wang Jinsong believes that the practice of the “Fengyu” model provides a valuable example of how AI can utilize diverse data sources to describe and interpret complex physical phenomena.

Fourth, self-controlled AI operator optimization technology. Zhang Dixuan explains that, on the software side, “Fengyu” is built on the MindSpore Science suite, implementing models for ionosphere, magnetosphere, and other regions, with tensor parallelism, pipeline parallelism, and other strategies. It develops scientific computing interfaces for 3D spatiotemporal data, and uses automatic graph optimization and fusion techniques to enhance training and inference efficiency.

Hardware-wise, “Fengyu” runs on Ascend AI clusters. It leverages system-level high-reliability design and hardware-software co-optimization to boost computational power, supporting large-scale historical and high-resolution data training.

Data-Driven Foundation: “Integrated Heaven and Earth” Observation System

High-quality data is essential for advanced AI models. China’s “Integrated Heaven and Earth” space weather monitoring system provides this. Space-based, the “Fengyun” satellites monitor key elements like the Sun, magnetosphere, and ionosphere. Ground-based, 73 stations operated by the China Meteorological Administration and 31 stations from the “Zhongwu Project” with nearly 300 devices conduct continuous observation. These massive, multi-dimensional data sources fuel the “Fengyu” model.

The model innovatively combines numerical space weather models with observational data, creating a high-quality data foundation that supports full-chain intelligent monitoring, modeling, and early warning.

Chen Zhou highlights that the ionosphere component of “Fengyu” is highly adaptable, capable of integrating data from different sources and resolutions.

3. From Forecasting to Protection – Application Examples and Performance of “Fengyu”

“Fengyu” not only features architectural innovations but also demonstrates breakthrough forecasting capabilities. Over a year of testing shows it can accurately predict solar wind, magnetosphere, and ionosphere conditions within 24 hours.

During recent major geomagnetic storms, “Fengyu” excelled in ionospheric predictions, with errors in total electron density predictions kept around 10%, the best result globally, according to Wang Jinsong.

The model has applied for 11 national patents.

Application Example: Guiding Spacecraft “Towards Benefits and Avoiding Harm”

“Fengyu” can be used beyond forecasting, such as in spacecraft design, management, and operation. For instance, predicting solar activity helps estimate radiation exposure for satellites, guiding protective measures. Accurate forecasts also assist in orbit management and mission safety, like adjusting fuel use and attitude control during space weather events.

4. Next Stop – 'Edge Intelligence' in the Stars and Oceans

The release of “Fengyu” marks a major breakthrough in China’s space weather monitoring. As Wang Jinsong notes, its innovations in architecture, data fusion, and application set a benchmark in AI for science, inspiring further integration of space science, machine learning, and high-performance computing.

Looking ahead, the goal is to deploy AI directly on satellites for autonomous decision-making, moving from cloud-based models to onboard edge computing. This evolution will drive lightweight AI models, edge inference, and highly reliable intelligent systems, lighting a smarter, safer path for human exploration of the stars.