China releases first generative AI language model for meteorological services

China Meteorological Administration (CMA) has unveiled Fenghe, the nation's first generative artificial intelligence language model, dedicated to meteorological services. This innovative system enables users to input queries and receive responses and analytical reports covering diverse weather-related needs -- from wind conditions and clothing recommendations to specialized suggestions for activities like snow viewing.

Fenghe is supported by a high-quality meteorological service corpus comprising 50 million linguistic units, along with 490,000 real-world meteorological Q&A scenarios. It is equipped with data interface tools and personalized applications spanning eight major categories and more than 60 specialized meteorological services.

The system operates through a dual architecture: the Fenghe base model performs intelligent task scheduling, while the meteorological service agent platform provides developers with tool invocation, data access, and agent development support. This integrated approach enables intelligent meteorological services for sectors such as energy and power, transportation, and tourism.

During the 15th National Games of China, Fenghe was integrated into a WeChat mini program of the Guangzhou meteorological department, providing tailored weather services for the major sports event. The system has been deployed and applied by meteorological departments in multiple regions.

Today, artificial intelligence is steadily reshaping people's work and lives. People are no longer satisfied with simply checking weather forecasts; instead, they expect more personalized and customized intelligent services.

In 2024, the CMA Public Meteorological Service Center and Xiong'an Meteorological Artificial Intelligence Innovation Research Institute, in collaboration with the Department of Computer Science and Technology at Tsinghua University, formed a research and development team to independently develop a generative AI language model for meteorological services.

According to the team, in order to achieve a precise, professional, and in-depth understanding of users' needs for meteorological services, they made breakthroughs in four key technologies: corpus construction; knowledge enhancement training and scenario-specific fine-tuning; deep reasoning; and multi-agent collaboration.

Corpus construction technology converts the CMA's vast trove of high-quality meteorological data into dedicated formats and datasets that are "fed" into the Fenghe model for learning, thereby enhancing its professional competence.

Knowledge enhancement training and scenario fine-tuning strengthens the model's fundamental meteorological capabilities and operational decision-making abilities, enabling it to better understand user needs and improve autonomous planning and execution for complex meteorological service tasks.

Deep reasoning enhances the model's reasoning capacity, allowing it to follow logical inference and make decisions much like a meteorological expert. For diverse scenarios such as tourism, transportation, and energy, the team has developed various meteorological service agents. Through multi-agent collaboration, complex tasks are automatically decomposed and planned, with complementary capabilities among agents, enabling them to work in coordination and rapidly generate professional reports.

"Tackling key technologies was never easy. Challenges once stood in the way of our improvement, but we never gave up. We tried over and over, making changes each time, and every step brought something new," said Wang Muhua, senior engineer at the CMA Public Meteorological Service Center.

Today, the application of AI in the meteorological field is becoming increasingly widespread. New applications and scenarios continue to emerge, giving rise to a growing number of AI models that continuously empower the development of the meteorological sector.

Before Fenghe, the CMA had already introduced AI models such as Fenglei, Fengqing, Fengshun, and Fengyu. While distinct from Fenghe, these models are all applied within meteorological forecasting. Fenglei focuses on nowcasting for imminent thunderstorms and heavy rainfall; Fengqing is a global short- to medium-range forecasting system; Fengshun provides global subseasonal-to-seasonal predictions; and Fengyu is a chain-based space weather forecasting model. The integration of AI, combined with established forecasting technologies and forecasters' extensive experience, has made weather forecasts more accurate.

In addition, many AI models have emerged in areas such as extreme precipitation prediction, severe convective weather warnings, meteorological data assimilation and analysis, meteorological observation and data processing, as well as meteorological services and decision support.

These advances have strongly promoted the digitalization, informatization, and intelligent upgrading of meteorological operations, providing important support for people's work and life.