On-the-fly postprocessing and feature extraction of flame and flow properties obtained by Direct Numerical Simulations

直接数值模拟是目前公认的研究湍流流动的最佳方法。然而，高雷诺数的数值模拟需要数十亿个网格点，并且需要进行数千次迭代。在考虑涉及化学反应的更复杂的流动时，为了测试和开发简化模型，许多变量必须在空间和时间上进行分析和关联。这导致了数TB甚至数PB的原始数据，这些数据不能再存储在现有的磁盘空间上，也不能使用任何现有的计算机网络在可接受的时间内传输。因此，有必要开发新的方法。在相应的DNS模拟期间对生成的流动数据进行同时分析和后处理(即时分析)在这方面似乎特别有希望，因为需要以这种方式存储和处理的数据要少得多。这种分析是通过即时特征提取实现的：流和标量场的特征被提取，与由域名系统生成的数据并行，因此不再需要存储数据文件。然而，这与在数据分析、特征提取、并行化和相关的数值技术方面存在着相当大的差距。

Direct Numerical Simulations (DNS) are presently recognized as the best possible solution to investigate numerically turbulent flows. However, DNS at high Reynolds numbers require billions of grid points and are pursued during thousands of time iterations. When considering even more complex flows involving chemical reactions, many variables have to be analyzed and correlated in space and time in order to test and develop reduced models. This leads to Terabytes or even Petabytes of raw data, which cannot be stored any more on existing disk space, nor transferred within an acceptable time using any existing computer network. It is therefore necessary to develop new approaches. A simultaneous analysis and postprocessing of the generated flow data during the corresponding DNS simulations (on-the-fly analysis) appears to be particularly promising in this regard, since significantly less data has to be stored and handled this way. This analysis is realized by on-the-fly feature extraction: features of the flow and scalar fields are extracted in parallel to data generation by DNS, such that the data files do not need to be stored any more. However, this is associated with considerable chal-lenges concerning data analysis, feature extraction, parallelization and relevant numerical techniques.