Abstracting the environment: automating geoscientific simulation

这个项目将使地球科学家实现计算机模拟的能力发生革命性的提高，特别是对新兴的并行硬件，并使用模拟的结果。地球系统中过程的计算机模拟已经成为科学的关键工具之一。在大气和海洋中，从冰冻的冰盖到地幔熔岩，流体和固体的模拟是无处不在的基本工具。这些关键过程：世界上许多最大的计算机都在模拟它们。用来产生这些模型的数值方法正迅速变得更加复杂。与此同时，大规模并行计算机硬件的出现为实现前所未有的分辨率水平提供了机会。然而，计算和新硬件的复杂性使得研究人员很难编写出正确的、足够高的性能和足够可用的计算机代码。传统的软件开发本质上要求超人开发人员同时是地球科学家、数学家和计算机科学家。从本质上说，这些困难的出现是因为传统计算机将数值和并行实现混为一谈。取而代之的是，可以通过用类似于数学的高级计算机语言指定数值方法来开发模型，并且并行实现可以自动生成。这将使数值建模专家能够指定他们的算法，并以当前开发成本的一小部分得出正确的并行代码。对于致力于较小规模和简化模拟问题的科学家和工程师来说，这种模型的自动生成是当今的现实。然而，地球的曲率、地球物理域的极端平坦性以及涉及的域的规模意味着地球科学家有额外的需求，这需要对模拟代码生成系统进行深刻的改变。我将扩展代码生成技术来应对这些特殊的挑战，从而为地学模型开发提供自动化。这些科学不仅仅依赖于模拟过程：它还依赖于研究系统的敏感度，优化输入和参数，稳定性分析和误差分析。所有这些过程都需要一个伴随模型：本质上是原始模拟的梯度。开发毗邻模型是如此复杂，以至于只有最大的国家中心通常才能负担得起开发它们的费用。通过使用代码生成，我已经演示了对于某些类型的模型，这几乎可以自动完成。我将把这种能力扩展到地学中更常见的其他离散处理，从而将伴随项这个强大的工具交到从事许多尖端地学研究的个别科学家和学生手中。最大的模拟，特别是气候系统的模拟，产生了许多复杂的数据，许多重要的科学都是通过研究存档模拟的输出来实现的。对于来自许多模型的大量数据，即使是计算统计数据的过程也是繁琐和容易出错的。目前也不可能核实公布的数据分析是否计算正确。我将扩展模拟软件的自动生成，以支持自动数据查询语言。这使得这种形式的数据科学的劳动密集度大大降低，将允许数据科学使用适当发布的方法，将减少错误来源，并将使科学家能够有效地处理未来的海量数据集。

This project will deliver a revolutionary increase in the ability ofgeoscientists to implement computer simulations, especially foremerging parallel hardware, and to work with the results of simulations.Computer simulation of processes in the Earth system has become one ofthe key tools if science. In the atmosphere and ocean and from frozenice sheets to the molten rock of the Earth's mantle, simulations offluids and solids are ubiquitous and essential tools. These arecritical processes: many of the world's largest computers are engagedin simulating them.The numerical methods used to produce these models are becomingrapidly more sophisticated. At the same time the emergence ofmassively parallel computer hardware presents the opportunity forunprecedented levels of resolution. However, the complexity of thenumerics and the new hardware is such that it is becoming verydifficult for researchers to write computer code which is correct,sufficiently high performance and sufficiently usable. Conventionalsoftware development essentially requires superhuman developers whoare simultaneously geoscientists, mathematicians and computerscientists.In essence these difficulties occur because conventional computer codemixes the numerics and the parallel implementation. Instead, modelscould be developed by specifying the numerical methods in a high-levelcomputer language similar to the maths, and the parallelimplementation could be generated automatically. This would enableexperts in numerical modelling to specify their algorithm, and arriveat correct, parallel code at a tiny fraction of current developmentcosts. This automatic generation of models is a reality today for scientistsand engineers working on smaller-scale and simplified simulationproblems. However the curvature of the earth, the extreme flatness ofgeophysical domains and the scale of the domains involved mean thatgeoscientists have additional needs which require deep changes insimulation code generation systems. I will extend code generationtechniques to meet these special challenges, and therefore deliverautomation to geoscientific model development.Much science does not just depend on simulating processes: it alsodepends on studying the sensitivity of systems, optimising inputs andparameters, stability analysis and error analysis. All of theseprocesses require an adjoint model: essentially the gradient of theoriginal simulation. Developing adjoint models is so complex that onlythe largest national centres can typically afford to developthem. Using code generation, I have already demonstrated that this canbe made almost automatic for some types of model. I will extend thiscapability to other discretisations which are more common in thegeosciences, and thereby put the powerful tool that adjoints are intothe hands of the individual scientists and students who conduct muchof the cutting edge geoscience.The largest simulations, particularly of the climate system, produceso much complex data that much important science occurs by studyingthe output of archived simulations. For large collections of data frommany models, even the process of calculating statistics is labouriousand error-prone. It is also currently impossible to verify ifpublished data analyses are correctly calculated. I will extend theautomated generation of simulation software to allow for an automateddata query language. This make this form of data science far less labour-intensive, will allow data science with properly published methods, will reduce sources of error and will allow scientists to work effectively with the massive data sets of the future.