【摘要】 海洋数值模式是定量描述海洋物理现象及其变化的数值模型,也是海洋与气候研究、预测的核心工具。随着海洋观测的不断投入与积累、对海洋认识的不断深入,特别是在高性能计算技术的支撑下,海洋数值模式已有了长足进步,正朝着高分辨率和多物理过程的方向发展。随着分辨率的提高、物理过程的细化,海洋数值模式的发展面临着多个方面的挑战。当前,海洋数据数量和种类不断增多,同时超级计算机、高性能计算和深度学习等技术的快速发展,为海洋数值模式的突破提供了机遇与挑战。本研究回顾了海洋数值模式的发展现状,梳理和分析了其发展中遇到的大规模高效并行计算和参数优化这两个关键问题,探讨和展望了当前海量数据驱动下海洋数值模式的发展趋势。提出计算负载均衡、计算与I/O重叠的并行流水线设计以及降低全局交换的算法改进是当前突破高分辨率海洋模式大规模高效并行效率的关键。从海洋科学、高性能计算以及深度学习深度交叉融合的角度,提出了实现海洋科学与深度学习相结合的6个途径,在此基础上,探讨了基于深度学习的参数化优化可能实现的途径。更多还原

【Abstract】 The ocean model is the key tool for ocean and climate research and prediction, which is a numerical model for the quantitative description of marine phenomena based on the physical laws. With the continuous marine investment, accumulation of ocean observation data, deeply understanding of the ocean process, and high-performance computing technology development, ocean numerical models have made great progress. Now, the main stream of ocean model development focuses on higher resolution and more accurate parameterization of unresolved physical processes. With the finer resolution and the more physical process, the development of ocean numerical models faces several challenges. The increasing ocean data, rapid development of high-performance computing and neural network depth learning technologies provide an opportunity for the breakthrough of ocean numerical models. This paper mainly reviewed the history and research status of the ocean numerical models, clarified and analyzed the two bottlenecks, performance of large-scale parallel computing and physical parameterization scheme, of ocean numerical models’ development and applications. Then, from the perspective of marine science, high-performance computing and deep learning integration, we proposed the six approaches of deep learning and ocean science convegence, the future development and trend of ocean numerical models driven by the mass data.更多还原