【CCF-AIR青年基金】High-performance Computing, Numerical Weather Prediction, and Machine Learning

Research Themes

Others

Background

气象预报是国计民生的基础设施。准确快速地预测和跟踪极端天气和气候事件对于国民安全,农业生产等都有重要意义。基于气象预报的风能和太阳能预测对风电厂、太阳能电站的选址、功率预测等能起到最核心的作用,是新能源提升发电量和消纳率,保证其高质量运行的基石。从气候变化科学的角度来看,能够提升温室气体监测评估的能力,为国家实现碳达峰和碳中和目标提供更多的科学支撑。与此同时,气象预测方面的研究能推动基础科学的进步(超算架构、计算科学等),有非常好的科研价值和社会影响力。此外,气象预测在航空、交通、农业、物流、金融等其他行业也有着广泛应用。

 

天气预报包含多个物理模式(每一个对应一组微分方程),通常在超级计算机上进行计算求解。中国的机构和学校多使用气象研究和预报系统(WRF),这套模式相对陈旧,性能也比较局限,所以通常在WRF的预报基础上,还会从国外机构(比如欧洲中期气象预测中心ECMWF和美国全球预报系统GFS)采购其数值模式的结果来改进预测。但是最精确的数据和最先进的模型一般不对外开放,因而难以获取。AI的兴起,给科学计算提供了新的思想方法和数学工具,能帮助我们求解之前难以企及的问题,在数值气象预报模式的计算速度和精度上,也给我们提供了弯道超车的可能性。

 

Weather forecast is an infrastructure of national welfare and people's livelihood. Rapidly and accurately predicting and tracking extreme weather and climate events are of great significance to national security, agricultural production, etc. Wind and solar resource predictions from weather forecasts play an essential role in the site selection and power prediction for wind and solar power plants. It is the cornerstone for new energy to increase power generation and consumption rate, ensuring high-quality operations. Weather forecast capability also improves the ability to monitor and evaluate greenhouse effects from the climate change perspective, providing more scientific support for achieving the carbon peak and neutrality goals. As a research topic, it promotes the progress of fundamental sciences such as supercomputer architecture and scientific computing, with high research values and community impacts. Moreover, weather prediction also benefits many other industries, including aviation, transportation, agriculture, logistics, finance, etc.

 

Weather forecasting has multiple physical modes (each corresponds to a set of differential equations), usually solved on supercomputers. Chinese institutions and universities mostly use the Weather Research and Forecasting (WRF) model system. This model, on the other hand, is rather outdated and has a restricted performance. Therefore, data from other institutions, such as European Center for Medium-Range Weather Forecasts (ECMWF) and Global Forecast System (GFS), are often purchased and deployed to improve the forecasts. Unfortunately, the most accurate data or state-of-the-art models are often under export control and are publicly inaccessible. The rise of AI has provided scientific computing with new thinking perspectives and mathematical tools to help us solve previously unattainable problems. There is a great opportunity for us to overtake in the computational speed and accuracy for numerical weather predictions, with our strengths in AI.

Target

本研究项目的目标是更高效的气象和气候方针,针对多种单一物理模式以及多物理模式的融合,进行快速推演、数据融合、纠偏以及根据时间和地理特点进行准确的模式选择。主要的研究方向包括但不局限于:

  • 针对气象方程的数值算法和底层数值代数的优化;
  • 并行计算的设计和针对HPC的性能优化,尤其是在异构体系上的优化;
  • 利用AI修正物理模型(包括物理参数化过程)及其预测;
  • 针对传统四维变分和卡曼滤波等技术,利用AI加速或者简化数据同化过程;
  • 利用AI加速数值模拟,包括加速物理参数化过程和直接加速气象微分方程的求解;

 

The project aims to provide more efficient weather and climate policy with the integration of many single-physics and multi-physis models, which enables rapid inference, data assimilation, model correction, and accurate model selection based on temporal and geographic characteristics. The research focuses will include but are not limited to the following topics:

  • Specific optimization on the numerical methods and low-level numerical linear algebra routines;
  • Parallelization and performance optimization on High-Performance Computing (HPC) systems, especially on the emerging heterogeneous architectures;
  • AI corrections on the numerical model and its predictions, including physical parameterizations;
  • AI accelerations or simplifications for data assimilation techniques, such as 4D-Var and Kalman filters;
  • AI accelerations for numerical simulations, including physical parameterizations and directly solving the underlying PDEs;

 

Related Research Topics

高精度的气象模拟往往依赖于庞大的空间计算网格,对计算资源的需求极高,因此在高效的气象模拟中如何优化底层数值算法,如何优化计算并行和资源分配,以及如何利用AI改进和加速数值模拟等显得尤为重要。具体的需求方向包括:

  • 针对气象微分方程的底层数值方法,CPU/GPU并行技术;
  • 在算法层面利用AI技术改进物理模型和加速数值模拟;


High-resolution numerical weather prediction often relies on massive computational grids, which poses an extremely high computational burden on the simulation. As a result, a variety of critical factors have to be considered in order to achieve better efficiency, such as how to optimize the low-level linear algebra routines, code parallelization, resources optimization, and how to speed up the solution procedure for differential equations. Specific directions in demand are:

  • Numerical methods for meteorological PDEs, CPU/GPU code parallelization;
  • Use AI to correct the physical models and to accelerate the numerical simulations at the algorithm level.

Scan QR code
关注Ali TechnologyWechat Account