Next Generation Weather and Climate Prediction: Atmospheric Model Dynamical Core

英国气象局(以及英国的天气和气候研究社区)使用的下一代天气和气候模型将在拥有大量并行处理器的超级计算机上运行。在这样一台超级计算机上运行天气模型有一个简单的要求，即将并行处理器的数量增加一倍，就可以将预测所需的时间减半：这被称为“并行缩放”。让计算机模型在大规模的超级计算机上很好地缩放是科学计算的一个巨大挑战；不幸的是，模型中使用的一些算法使预测非常准确，但不能很好地缩放。NGWCP项目的目标是识别新的算法，这些算法可以在超级计算机上很好地扩展，而不会牺牲预报技能(模型正确预报天气的能力)。在这个方案中，我们选择集中在两个方面，水平网格和传输算法。在当前的Met Office模型中，水平方向使用经纬度栅格。这种类型的网格有一个问题，即所有的纬度线都会聚在两极上，所以它们变得越来越紧密。事实证明，这对并行扩展非常不利。在这个方案中，我们提出了三种新的网格类型，它们在全球范围内保持网格线的相同距离，保持了经纬度网格的各种有利性质。模型中的传输算法负责对量(如温度或湿度)随风的移动进行建模。在Met Office Model中，这是通过追溯流体地块在全球各地的路径来实现的。事实证明，这种方法也不利于并行缩放。在本提案中，我们将研究一些最近开发的不需要这些路径计算的替代方案。这些算法必须反映在真实大气中起作用的物理原理，如质量守恒定律。

The next generation of weather and climate models used by the UK Met Office (as well as the community of weather and climate researchers in the UK) will be run on supercomputers with an enormous number of processors operating in parallel. Running a weather model on such a supercomputer has the simple requirement that doubling the number of parallel processors will halve the time required to make a forecast: this is called 'parallel scaling'. It is a great challenge of scientific computing to get a computer model to scale well on massive supercomputers; unfortunately some algorithms used in models make the forecast very accurate, but do not scale well. The aim of the NGWCP project is to identify new algorithms that can scale well on supercomputers without sacrificing forecast skill (the ability of a model to correctly forecast the weather). In this proposal, we choose to concentrate on two aspects, the horizontal grid, and the transport algorithms. In the current Met Office model, a latitude-longitude grid is used in the horizontal. This type of grid has the problem that all the lines of latitude converge on the poles, so they get closer and closer together. It turns out that this is very bad for parallel scaling. In this proposal, we suggest three new types of grids that keep grid lines the same distance apart all over the globe, keeping various favourable properties of the latitude-longitude grid. Transport algorithms in the model are responsible for modelling the movement of quantities (such as temperature or moisture) with the wind. In the Met Office Model, this is done by tracing back the paths of fluid parcels around the globe. It turns out that this method is also bad for parallel scaling. In this proposal, we will investigate some recently developed alternatives that do not require these path calculations. It is essential that these algorithms reflect physical principles at work in the real atmosphere, such as the Law of Conservation of Mass.