On July 26, at the 2025 World Artificial Intelligence Conference's Meteorological Forum, the National Satellite Meteorological Center (also known as the National Space Weather Monitoring and Warning Center), in collaboration with Nanchang University and Huawei Technologies Co., Ltd., officially launched "Fengyu", the world's first AI-driven chain model for space weather forecasting.

According to Wang Jinsong, Director General of the National Satellite Meteorological Center and head of the Fengyu development team, the release marks a breakthrough in China's space weather monitoring and early warning capabilities, contributing an innovative solution to global space weather forecasting.

A Full-Chain AI Breakthrough for Space Weather

Developed to address the impact of solar storms on satellite operations, radio communication, navigation, and other critical infrastructure, Fengyu overcomes limitations of traditional numerical models in computational efficiency and real-time responsiveness. It is the first AI-powered forecasting system to cover the entire chain of space weather: from the solar wind to the magnetosphere and ionosphere.

While China had previously developed independent AI models for specific regions—such as "Xufeng" (solar wind), "Tianmag" (magnetosphere), and "Dianqiong" (ionosphere)—they lacked a unified mechanism to capture the causal chain from the Sun to Earth.

"To address this, we developed a coupling optimizer that integrates these large models into a modular, upgradable framework," Wang explained. "This allows upstream and downstream models to learn from each other and improve overall forecast accuracy."

Fengyu introduces a pioneering chained training architecture. By building independent models for the solar wind, magnetosphere, and ionosphere, and linking them through an intelligent coupling system, it enhances model coordination and accuracy. This structure enables a more accurate representation of solar-terrestrial interactions and improves the reliability of extreme space weather event predictions.

Built on Domestic AI Infrastructure

Fengyu is built entirely on Chinese-developed infrastructure. Zhang Dixuan, President of Huawei Ascend Computing Business, explained that the model leverages Huawei's Ascend AI infrastructure and the MindSpore deep learning framework to support complex modeling tasks.

On the software side, multiple regional models were constructed using the MindSpore Science toolkit. Parallel processing strategies—such as tensor and pipeline parallelism—were developed specifically for 3D spatiotemporal data, along with scientific computing interfaces and compiler optimizations like auto graph optimization and operator fusion, significantly boosting training and inference performance.

On the hardware side, Huawei's Ascend AI clusters provide industry-leading computing power and reliability. This ensures efficient large-scale training using historical observations and high-resolution gridded data.

According to Chen Zhou, Associate Dean of the School of Artificial Intelligence at Nanchang University, the scale of space weather spans ten orders of magnitude and involves complex physical systems like atmospheric dynamics, electrodynamics, and magnetohydrodynamics—posing challenges to traditional models. Fengyu's deep integration of Earth system science enables major progress in short-term forecasting and extreme event response.

Demonstrated Performance and Future Applications

Fengyu has demonstrated excellent 24-hour forecasting capability across solar wind, magnetosphere, and ionosphere regions in a year-long performance test. During recent geomagnetic storm events, its ionospheric forecasts were particularly accurate.

With 11 national patent applications already filed, Fengyu holds broad application potential—from ensuring satellite communication and navigation stability to improving spacecraft mission safety, protecting power grids, and enhancing aviation security.

"By combining numerical models with AI and expert forecasting experience, Fengyu's accuracy will continue to improve," Wang noted. "We aim to maximize its application value, promote high-quality development of space weather services, and contribute meteorological expertise to space exploration."