Meteorological demands ranging from typhoon movements over the Pacific Ocean 5,000 kilometers away, to scattered heavy rains during the flood season in North China, and climate trends of dryness or wetness two months later—spanning from hourly to seasonal and local to global scales—are now integrated into the forecasting scope of AI model systems. The Xiong'an Meteorological AI Innovation Institute, jointly established by the China Meteorological Administration, the Hebei Provincial Government, and the Xiong'an New Area Management Committee, brings together top-tier scientific research forces. It has developed the "Feng" series of AI meteorological models covering hourly to seasonal timescales, linking short-term warnings with climate predictions into a complete, full-time domain chain. From its roots in Xiong'an to serving the entire nation, empowering industries, and reaching a global audience, this AI "national team" in meteorology is accelerating China's smart weather progress at a new pace.
On July 30, at the Xiong'an Meteorological AI Innovation Institute, a staff member introduced the "Feng" series AI meteorological models.
AI Empowerment: Redefining the Technical Path of Weather Forecasting
On July 30, at the Xiong'an New Area Science Park, inside the office of the Xiong'an Meteorological AI Innovation Institute (referred to as the Xiong'an Institute), Deputy Director Wang Yaqiang focused on the flickering code on his screen. Groups of meteorological data flowed through neural networks, colliding repeatedly, with each iteration narrowing the gap to real atmospheric conditions. "Previously, weather forecasting relied on solving equations. Now, we let AI large models learn and deduce," Wang said. This statement captures the Xiong'an Institute's core mission: exploring the use of AI technology to rebuild the technical pathway for weather forecasting. As a new research institution co-established by the China Meteorological Administration, the Hebei Provincial Government, and the Xiong'an New Area Management Committee, the institute's positioning is clear: with self-developed meteorological model series as its core, it drives the integration and application of AI in areas like disaster prevention and mitigation, climate change, ecological environments, energy, and transportation.
However, understanding why this institution chose Xiong'an requires grasping the basics of traditional weather forecasting. Traditional forecasting, also known as numerical weather prediction, essentially involves solving equations. Scientists divide the atmosphere into countless grids, starting from the current initial state and using fluid dynamics and thermodynamics equations to advance calculations second by second—calculating this second, then the next, and so on. Predicting the weather for the next five days requires tens of thousands of rolling calculations. "Numerical models integrate second by second, with errors accumulating like a snowball, increasing uncertainty with each step," Wang explained. In recent years, numerical forecasting has faced challenges in terms of computational costs and fine-scale depiction of extreme weather, opening new technical space for AI methods.
AI meteorological forecasting follows a completely different logic. Instead of calculating equations second by second, it "learns" the intrinsic patterns of atmospheric evolution from vast historical data. After repeated training, given an initial state, the AI model can directly output weather forecasts. Comparing the two, AI's key advantage lies in significant savings in computational power and efficiency gains. "Currently, the world's most advanced numerical weather prediction systems, for a single 10-day global forecast, consume computational resources amounting to millions of CPU core hours," Wang noted. In contrast, a trained AI model can output a global forecast in just a few minutes on a single GPU card. However, this efficiency comes with the extreme complexity of the training process. AI forecasting requires not only massive, high-quality meteorological reanalysis data but also continuous iterative optimization and rigorous operational testing. This is a vast systems engineering task, far beyond the capacity of any single team. Previously, relevant research forces were dispersed across various operational and research units, with room for further coordination in data, computing power, technical routes, and results transformation. The China Meteorological Administration recognized the need to consolidate these forces for targeted, organized research.
Thus, a talent mobilization effort unfolded in Xiong'an. Where did the talent come from? On one hand, the China Meteorological Administration selected operational and research backbone personnel from its direct units, including the National Meteorological Center, National Climate Center, National Meteorological Information Center, Chinese Academy of Meteorological Sciences, and Public Meteorological Service Center. On the other hand, it conducted nationwide talent selection from provincial and municipal meteorological departments. Additionally, the Xiong'an Institute recruited specialists in fields like AI through open social hiring, ultimately forming a research team of nearly 100 people. This team is a true "national team" in the meteorological AI field: including Dou Zesheng, a chief architect recruited from Baidu to lead core technology directions; high-level talents such as National Key R&D Program chiefs, China Meteorological Administration leading talents, and young talents; and backbone members from various meteorological units nationwide, bringing years of operational experience to technical challenges. To attract talent, the Xiong'an New Area also offered substantial policy support: including the institute in Category Two relocation unit management, issuing "Xiong'an Talent Cards" to staff, with eligible experts receiving direct home-buying qualifications; implementing a "special post, special salary" policy, with the New Area providing matching support at 50% of the salary for the chief architect. Today, the Xiong'an Institute has established six major centers: weather forecasting, climate prediction, meteorological science, meteorological service language, general technology, and innovative applications. It is moving towards the goal of becoming a world-class center for meteorological forecasting technology R&D, an international cooperation hub for early warning, and a talent innovation highland, based on a mechanism of shared talent management, shared resources, co-built projects, and shared outcomes.
On July 29, at the Xiong'an Meteorological AI Innovation Institute, Senior Engineer Nie Gaozhen used the "Fengqing" model to analyze the trajectory of Typhoon "Dolphin."
The "Feng" Series Matrix: Covering Full-Time Domain Forecast Scenarios
On the morning of July 29, in an office at the Xiong'an Institute, Senior Engineer Nie Gaozhen called up the latest typhoon forecast imagery. At that moment, deep in the Pacific, the 13th typhoon of the year, "Dolphin," was slowly moving northwest, about 5,000 kilometers from China's coastline. The "Fengqing" model Nie used, led by the China Meteorological Administration and a key focus of the Xiong'an Institute, is an AI global medium-range weather forecast model. It uses an "AI + physics" dual-drive design, taking just 3 minutes per forecast on a single AI computing card to generate global weather forecasts for the next 15 days, updated every 6 hours, with a resolution of 25 kilometers. "In recent years, with the Xiong'an Institute as its core R&D force, the China Meteorological Administration has woven a matrix of 'Feng' series AI meteorological models covering different time scales," Wang Yaqiang explained. These models have distinct roles, from hourly to monthly scales, from data assimilation to forecast services, and from professional analysis to public interaction, interlocking to form a closed loop.
As the first "Feng" series model successfully developed and put into use, "Fengqing" has performed admirably in recent practice. In September 2024, Typhoon "Yagi" formed in the South China Sea, and "Fengqing" accurately locked onto its path 6 days before landfall. Throughout 2025, "Fengqing's" typhoon path forecast accuracy surpassed the European Centre for Medium-Range Weather Forecasts (ECMWF). During the 2025 flood season, when heavy rains in North China were widespread and scattered, "Fengqing" outperformed traditional models in accuracy scores for heavy and torrential rain levels, achieving the highest rating. The longer the forecast, the greater the uncertainty. From 15 to 60 days, this is a "forecast desert" acknowledged by the global meteorological community—traditional numerical models see a sharp decline in accuracy here. The "Fengshun" model was developed to fill this gap. "Fengshun" is the world's first real-time operational model system based on AI methods for sub-seasonal to seasonal global climate anomaly prediction. It updates daily, producing a large-sample ensemble forecast of 100 members for the next 60 days. Since August 2025, "Fengshun" has represented China in the Global AI Weather Forecasting Competition jointly organized by the ECMWF and the World Meteorological Organization. Among over 100 competing models, including from Microsoft and NVIDIA, it performed strongly and recently won the championship in the third season. Now upgraded to version V1.5 and evolving towards 2.0, "Fengshun" aims to further enhance probabilistic forecast levels, providing more reliable support for risk predictions of high-impact events like floods and human health.
If the first two models are for professional forecasters, the "Fenghe" model directly engages the public. "Fenghe" is the world's first open-source meteorological large language model with billions of parameters, officially released and globally opened in July 2026. Trained on 50 million tokens of high-quality meteorological service data, it integrates over 60 types of authoritative meteorological data. It features functions like weather queries, science popularization Q&A, travel planning, and can perform logical reasoning for meteorological services and risk impact analysis. "For example, if you ask 'Is it suitable to climb Mount Tai this weekend?', it won't just say 'it will rain,' but will provide comprehensive advice considering altitude, temperature, and wind," Wang said. Currently, "Fenghe" has been deployed in localized meteorological service scenarios across multiple provinces and cities.
However, all these models share a common prerequisite: they need someone else to "organize the data" before they can work. What does this mean? The first step in weather forecasting is to compile vast amounts of raw data from satellites in space, ground-based radars, and various weather stations into a unified, complete "snapshot" of the atmospheric state. This step, called "assimilation," is quite complex and traditionally handled by numerical models. The second step is for the AI model to use this organized "snapshot" to predict future weather trends. This led to a bold idea: could AI models directly start from raw satellite signals, radar echoes, and surface observation data, completing the entire "organize-judge-forecast" process themselves? If achieved, the entire forecast chain would be significantly shortened, and the model could directly "consume" China's high-density local observation data, potentially making forecasts for China more detailed and accurate. This idea eventually gave birth to the "Fengyuan" model. Released on December 19, 2025, "Fengyuan" is China's first meteorological AI forecast model base with completely independent intellectual property rights. "The name 'Fengyuan' has three layers of meaning: source of innovation, source of observation, and source of openness," Wang explained. It can directly read real-time observation data from satellites, radars, and weather stations, and after self-analysis and deliberation, directly provide global weather forecasts. Thus, "Fengyuan" can be seen as an all-weather, high-precision meteorological AI analyst and forecaster. "The release of the 'Fengyuan' model marks a critical step for China's meteorological department in achieving an end-to-end scientific model base with independent intellectual property rights, securing technical autonomy in key links, and laying an engineering foundation for continuous iteration and rolling evolution," Wang said.
A recent photo of the exterior of the Xiong'an Meteorological AI Innovation Institute.
Open Collaboration: Driving Widespread Implementation of Technological Achievements
On July 17, in Shanghai, at the meteorological session of the 2026 World AI Conference, Wang Yaqiang took the stage to deliver a keynote report on behalf of the China Meteorological Administration, systematically introducing the progress of AI model development in China's meteorological department. The audience was filled with global meteorological experts and AI practitioners. On the same day, the world's first bilateral international cooperation laboratory in the meteorological AI application field—the China-Thailand Joint Laboratory for Intelligent Prediction and Early Warning of Meteorological Disasters—was officially launched. "The laboratory is co-led by the Xiong'an Institute and the Climate Center of the Thailand Meteorological Department, with a dual director system from China and Thailand. The research team consists of over 40 experts, covering fields like atmospheric science, AI, power systems, and agricultural meteorology," Wang said. This collaboration between China and Thailand was not spur-of-the-moment. The two countries' meteorological departments had previously co-hosted the 21st Asian Regional Climate Monitoring, Assessment, and Prediction International Symposium and regularly conducted joint climate consultations during the flood season each year. "Leveraging China's global meteorological datasets and AI models adapted through transfer learning, we can specifically enhance Thailand's forecasting and early warning capabilities for disasters like droughts and heavy rains," said Dr. Chalermpong Unnariya, the Thai director of the lab. Such international cooperation is not an isolated case. In recent years, models like "Fengqing" and "Fengshun" have been successively included in bilateral cooperation framework agreements with countries including the UAE, Pakistan, Thailand, Ethiopia, and South Africa. The "Feng" series is steadily stepping onto the global stage.
Turning back to domestic efforts, the Xiong'an Institute is actively building a cooperation network. In July, the institute's visitor schedule was packed: on July 16, the Yibin Municipal Government, Sichuan Provincial Meteorological Bureau, and Yibin Meteorological Bureau visited; on July 27, the Digital Intelligence Integration Innovation Center of Beijing University of Posts and Telecommunications came for exchanges; on July 28, the Anhui Provincial Meteorological Bureau visited; and on July 29, the Jilin Provincial Meteorological Bureau arrived for research. In preceding months, meteorological departments from Shanghai, Shenzhen, Hunan, Qinghai, Tibet, and elsewhere had already visited. Why are provincial meteorological bureaus flocking here? Wang Yaqiang highlighted the key reason: "A major feature of AI weather forecasting is its relatively low barrier to entry. Developing traditional numerical models is very difficult for provincial meteorological departments, but AI is different. The top-level model frameworks are open-source and easily accessible." A low barrier doesn't mean everyone can do it well; challenges vary by province. The Tibet Autonomous Region Meteorological Bureau hopes to use AI to improve forecasting and warning capabilities in complex terrain areas—with sparse observation stations and large terrain fluctuations, traditional models often fail to "calculate accurately." The Guizhou Provincial Meteorological Bureau, which visited earlier, has similar needs. "The more complex the terrain and the lower the quality of observation data, the more urgent the demand for AI," Wang said. Local areas can base their work on the "Feng" series model frameworks and capabilities, combining local data and application scenarios to tailor models to their specific needs. Cross-industry collaborations are also expanding. In April 2025, the Xiong'an Institute cooperated with China Mobile and PowerChina to successfully establish the Hebei Provincial Key Laboratory of Meteorological AI. This is the province's first AI key research platform focusing on cross-disciplinary areas like "Meteorology + New Energy" and "Meteorology + Low-Altitude Economy." While weaving a dense national cooperation network and expanding its international cooperation map, the Xiong'an Institute remains firmly rooted in Xiong'an itself, promoting the deep implementation of meteorological AI technology in the New Area's construction scenarios. In April this year, the Xiong'an National Climate Observatory was officially inaugurated, with a "one main, eight auxiliary" layout covering the New Area's urban and rural spaces, including four typical ecological zones: urban, forest, wetland, and farmland. Currently, the "Feng" series models are deeply integrating with the observatory's observational data, serving key tasks like meteorological support during the critical "late July to early August" flood season and winter heating. Urban waterlogging is another key focus. Xiong'an is a "city of the future" built from scratch, with design concepts like underground pipe corridors and sponge cities integrated throughout, but extreme rainfall still poses risks. Currently, the Xiong'an Institute is conducting technical research and project feasibility studies on urban waterlogging risk forecasting, exploring ways to extend precipitation forecasts to include flood extent, depth, and impact assessments on transportation infrastructure, aiming to turn weather forecasts into disaster risk forecasts.
From the accelerated iteration of the "Feng" series models to innovations spreading nationwide and globally, in just over two years, the Xiong'an Institute has embarked on a journey "rooted in Xiong'an, serving the nation, and looking to the world." Looking ahead, the institute will accelerate the construction of an Earth system AI forecast model and a crowdsourcing platform, promoting further integration of data and models across multiple spheres like weather, climate, oceans, and land surfaces. At the same time, it will continue to expand AI application scenarios in meteorology, bringing more scientific achievements from the laboratory into production and daily life.
Reporter's Observation: AI Technology Opens New Space for Meteorological Forecast Development
During interviews at the Xiong'an Meteorological AI Innovation Institute, the most profound impression was that a discipline built on long-term theoretical accumulation and engineering practice is now opening new development spaces due to the integration of AI technology. For China's meteorological department, this signifies not only further improvements in forecasting technology and operational capabilities but also a significant opportunity to accelerate innovation, achieve breakthroughs, and form distinctive advantages. For a long time, numerical weather prediction has been the core technology of modern weather forecasting operations. International advanced institutions like the ECMWF, after decades of continuous accumulation, have formed clear advantages in areas like data assimilation, model development, and operational operation. China's numerical prediction capabilities are also steadily improving, but there remains room for further catch-up and breakthroughs in certain core technologies, foundational data, and fine-scale forecasting skills. Numerical prediction is a highly complex systems engineering task. From the dynamic framework and data assimilation to various physical process parameterizations, each improvement requires long-term research accumulation, engineering iteration, and operational verification. Consequently, forecast skill improvements are often incremental and rarely achievable through a single point technology in the short term. The development of AI provides a new technical choice for weather forecasting. It is not a simple replacement for traditional numerical prediction but offers new possibilities for improving forecast efficiency, enhancing certain element forecast results, and expanding application scenarios by learning atmospheric evolution patterns from massive data. The deep integration of AI and numerical prediction in data, mechanisms, and operational processes is becoming a key direction in international meteorological science and technology development.
On this new track, China has a strong development foundation. In recent years, China's AI industry has developed rapidly, accumulating significant capabilities in algorithm development, computing infrastructure, and engineering applications. At the same time, relying on a comprehensive observation system composed of Fengyun meteorological satellites, weather radars, and ground-based meteorological stations, China has continuously accumulated a wide range of rich meteorological observation data, providing important support for AI meteorological model R&D. Of course, AI weather forecasting is still in a rapid development phase. When facing extreme weather, the models' ability to depict intensity, location, and evolution processes still needs improvement; model interpretability, physical consistency, and stability need strengthening; and autonomous high-quality datasets, unified evaluation systems, and operational application workflows also require continuous refinement. For AI models to truly enter operations, they must undergo long-term, rigorous testing, and cannot be judged simply by a single case or a few indicators. Therefore, the opportunities brought by AI do not mean starting from scratch or completely denying traditional technical routes. The future of weather forecasting is likely to be a new form where physical mechanisms and data-driven approaches are integrated, and numerical models and AI models develop synergistically. In this sense, what the Xiong'an Institute is exploring is not just a few models or technologies, but a new way of organizing scientific research and a new technological evolution path. Facing the technological changes brought by AI, as long as we persist in focusing on operational needs, strengthen foundational capabilities, and promote multi-domain collaborative innovation, it is possible to form more achievements with independent characteristics in new technological directions, adding new momentum to China's high-level self-reliance and self-improvement in meteorological science and technology.
Comments