NJU-Earth Joins FDP: Building a Family of AI Weather Forecasting Models
NJU-Earth（南京大学）现已参加 FDP。NJU-Earth 是由南京大学大气科学学院联合阿里巴巴达摩院，基于 Baguan 架构共同打造的全球 0.25° 高分辨率 AI 气象预报模型，面向全球中短期天气推演任务。
项目展示：FDP 官方网站
项目信息
| 项目 | 信息 |
|---|---|
| 模型名称 | NJU-Earth |
| 参赛单位 | 南京大学 |
| 联合研发 | 南京大学大气科学学院、阿里巴巴达摩院 |
| 空间分辨率 | 全球 0.25° |
| 主要任务 | 全球中短期天气预报 |
模型介绍
NJU-Earth 融合了多尺度天气特征预编码技术与交叉注意力特征融合技术。模型通过学习不同空间尺度上的天气特征，并在特征融合阶段建立多类气象变量之间的联系，实现对全球天气演变过程的智能推演。
在全球 0.25° 高分辨率预报模型的基础上，团队进一步构建了系列化 AI 预报体系，将研究范围从全球天气预报拓展到集合预报、区域精细化预报、强降水预测、台风研判以及行星大气模拟等任务。
NJU-Earth 模型体系
- BaguanHR：面向全球中高分辨率集合预报。
- RainCast：面向区域强降水预测。
- RegionalCast：面向区域精细化天气预报。
- BaguanCyclone：面向台风路径与强度研判。
- NJU-Mars：面向火星大气环境模拟。
- CBC-Diffusion：用于增强降水预报，进一步提升强降水概率预测能力。
参与 FDP
参加 FDP 为 NJU-Earth 提供了一个集中展示和检验 AI 气象预报能力的平台。通过统一平台上的预报展示与应用检验，可以更直观地观察模型在不同天气过程中的表现，也有助于推动 AI 预报技术面向真实气象业务持续改进。
NJU-Earth 将以系列化模型体系连接全球预报与专业气象任务，为气象科研、天气预报和防灾减灾提供智能化技术支撑。
NJU-Earth (Nanjing University) is now participating in FDP. Jointly developed by the School of Atmospheric Sciences at Nanjing University and Alibaba DAMO Academy, NJU-Earth is a global AI weather forecasting model built on the Baguan architecture. It produces medium-range and short-range forecasts worldwide at 0.25° resolution.
Project showcase: FDP Official Website
Project Information
| Item | Details |
|---|---|
| Model | NJU-Earth |
| Participating institution | Nanjing University |
| Joint developers | School of Atmospheric Sciences, Nanjing University; Alibaba DAMO Academy |
| Spatial resolution | Global 0.25° |
| Primary task | Global medium-range and short-range weather forecasting |
Model Overview
NJU-Earth combines multiscale weather feature pre-encoding with cross-attention feature fusion. It learns weather patterns at different spatial scales and models relationships among meteorological variables during feature fusion, enabling data-driven simulation of global weather evolution.
Building on its global 0.25° forecasting model, the team has developed a family of specialized AI forecasting systems. The research now spans global ensemble forecasting, regional high-resolution prediction, heavy-rainfall forecasting, tropical cyclone analysis, and planetary atmosphere simulation.
The NJU-Earth Model Family
- BaguanHR: global medium- and high-resolution ensemble forecasting.
- RainCast: regional heavy-rainfall prediction.
- RegionalCast: fine-grained regional weather forecasting.
- BaguanCyclone: tropical cyclone track and intensity assessment.
- NJU-Mars: simulation of the Martian atmospheric environment.
- CBC-Diffusion: precipitation enhancement for improved probabilistic forecasting of heavy rainfall.
Participation in FDP
FDP provides NJU-Earth with a platform for presenting and evaluating AI-based weather forecasts. Forecast visualization and application-oriented evaluation on a shared platform make it easier to examine model behavior across weather events and support continued improvement toward operational meteorological use.
Through its family of forecasting models, NJU-Earth connects global prediction with specialized meteorological applications and provides intelligent technical support for atmospheric research, weather forecasting, and disaster risk reduction.