Scalable Algorithms for Multiscale Modeling and Analysis of Turbulent Combustion

Pascucci 0904631湍流燃烧的精确模拟是一个主要的开放性问题，需要千万亿次计算来解决在很宽的长度和时间尺度上的物理过程的高度非线性耦合。PI的方法，开发新的建模和算法的方法，这个问题，有效地解决高性能计算（HPC）的燃烧模拟在Petascale。PI的方法结合了三种技术：自动算法并行化，用于模型简化的多维数据分析，以及多尺度建模与拓扑分析，以连接不同尺度的模型。算法并行化基于算法分析，该算法分析检测计算阶段之间的依赖关系，使用图论来检测和利用并行性比当前算法更有效。这种方法独立于MPI分布式并行并与之互补，并且允许实现利用每个计算节点上可用的多核资源所需的更细粒度并行。PI还计划采用一种强大的新方法来模拟多物理场流动，例如利用直接数值模拟（DNS）和一维湍流（ODT）来提供代理“真值集”的湍流燃烧。高维DNS数据集，包含TB的数据，可以被分析以提取已知存在的低维流形。诸如主成分分析之类的技术可以识别用于表示这种高维数据中的流形的最佳基础。一旦一个基础已经确定，并从ODT产生的数据集中提取，传输方程的变量形成的基础上可以导出和解决的大涡模拟（LES）。然后，LES可以用于生成新的ODT模拟，其可以反馈到LES，从而创建动态建模方法，该方法使用下尺度的、高分辨率的统计信息来构建要在更大尺度（LES）上使用的模型。这种建模方法是在千万亿次系统上进行早期测试的主要候选者。这项研究中的研究人员已经证明了将DNS和LES扩展到万亿次计算系统的能力，而千万亿次计算的可用性将直接支持这些建模方法。算法和建模的进步将应用于天然气的富氧燃烧。富氧燃烧是一种促进碳捕获和封存以减轻来自燃烧化石燃料的发电厂的二氧化碳排放的技术。虽然将应用于天然气系统，但这里开发的技术和算法将直接应用于其他系统，包括煤炭和柴油和汽油等运输燃料。该项目将为学生提供独特的教育经验，包括在国家实验室的暑期实习。在普通班级中讲授这个项目中学到的经验教训将有助于教育未来的劳动力。此外，该研究将加强大学研究人员与参与模拟和模型开发的国家实验室工作人员之间的合作，他们也将参与指导学生。

Pascucci0904631Accurate simulation of turbulent combustion is a major open problem requiring petascale computing to resolve highly nonlinear coupling of physical processes over a wide range of length and time scales. The PIs approach to develop new modeling and algorithmic approaches for this problem to tackle effectively High Performance Computing (HPC) for combustion simulation at the Petascale. The PI's approach combines three techniques: automatic algorithm parallelization, multidimensional data analysis for model reduction, and multi-scale modeling with topological analysis to connect models at different scales. The algorithm parallelization is based on an algorithmic analysis that detects dependencies among computing stages, using graph theory to detect and exploit parallelism more effectively than current algorithms. This approach is independent from and complimentary to MPI distributed parallelism and allows achieving the finer grain parallelism necessary to exploit the multi core resources available on each computing node. The PIs also plan a powerful new approach to model multiphysics flows, such as turbulent combustion that leverages direct numerical simulation (DNS) and one-dimensional turbulence (ODT) to provide surrogate 'truth sets'. High-dimensional DNS data sets, containing terabytes of data, can be analyzed to extract lower-dimensional manifolds known to exist. Techniques such as principal component analysis can identify the optimal basis for representing manifolds in this high-dimensional data. Once a basis has been identified and extracted from the data sets generated by ODT, transport equations for the variables forming the basis may be derived and solved in a large-eddy simulation (LES). The LES can then be used to generate new ODT simulations which can feed back to the LES, thereby creating a dynamic modeling approach that uses down-scale, highly resolved statistical information to construct models to be used on larger scales (LES). This modeling approach is a prime candidate for early testing on petascale systems. The researchers in this study have already demonstrated the ability to scale DNS and LES to terascale computing systems, and availability of petascale computing will directly enable these modeling approaches. Application of the algorithmic and modeling advances will be made to oxyfuel combustion of natural gas. Oxyfuel combustion is one technique to facilitate carbon capture and sequestration to mitigate carbon dioxide emissions from power plants burning fossil fuels. While application will be made to natural gas systems, the techniques and algorithms developed here will apply directly to other systems including coal and transportation fuels such as diesel and gasoline. This project will provide unique educational experiences for students, including summer internships at national laboratories. Incorporating in regular classes the lessons learned in this project will help educate the future work force. Additionally, the research will strengthen collaborations between university researchers and national laboratory staff involved in simulation and model development, who will also participate in mentoring students.