On-the-fly postprocessing and feature extraction of flame and flow properties obtained by Direct Numerical Simulations
Software Engineering and Programming Languages
Fluid Mechanics
Final Report Abstract
High-precision simulations of turbulent combustion on modern supercomputers can produce hundreds of terabytes of raw data. In order to obtain insight from this data, it is traditionally stored on a hard disk and then analyzed after the fact. Due to the sheer size of the data, storing it in its entirety has become practically impossible. This limits the applicability of traditional analysis as a post-process and makes it very challenging to extract meaningful insight from the simulation. This project aimed at developing methods for analyzing and visualizing the simulation data on-the-fly. This means that simulation results are processed while they are being produced, and only the interesting results, which are typically much smaller in size, have to be stored on disk. During the course of the project, we developed several approaches that tackle the problem of on-the-fly processing of simulation results in different ways. A Sparse Representation of the Flame Structure: We developed a way to extract a “skeleton”, which captures the structure of the entire flame, but at greatly reduced storage size. Using this representation, we can visualize important features of the flame and enable researchers to gain an understanding of how the behavior of the flame is related to its shape and deformation over time. If necessary, we can reconstruct the full raw data at a reasonable accuracy, which makes it possible to employ the wide range of existing post-processing tools for analyzing the data. On-The-Fly Tracking of the Flame Surface: The surface of the flame and its behavior over time is one of the most interesting features in a turbulent combustion simulation. We developed an algorithm for tracking this surface during the simulation and capturing its detailed temporal behavior. This is challenging because of the highly parallel nature of simulations on supercomputers, and because we need to be able to follow single points attached to the surface for long periods of time. Using our tracking algorithm, it becomes possible for the first time to perform detailed studies of the temporal behavior of the flame surface in large-scale 3D combustion simulations. The resulting data can be used to develop models of complex combustion mechanisms such as local flame extinction and reignition. Detecting Twist in Mechanical Parts: An important part of developing methods for on-the-fly processing of simulation data is the identification of features that are well-suited for computation in an on-the-fly environment. We identified a feature in structural mechanics simulations that indicates which parts of a complex mechanical part under stress experiences twist. Without our feature extractor, finding such locations is not intuitive and can be very time-consuming. Our algorithm for computing this feature is extremely well suited for parallelization, which is an important prerequisite for its application as an on-the-fly tool.
Publications
-
(2018) Core Lines in 3D Second-Order Tensor Fields. Computer Graphics Forum 37 (3) 327–337
T. Oster, C. Rössl, and H. Theisel
-
Towards direct numerical simulations of low-mach number turbulent reacting and two-phase flows using immersed boundaries. Computers & Fluids, 131:123–141, 2016
A. Abdelsamie, G. Fru, T. Oster, F. Dietzsch, G. Janiga, and D. Thévenin
-
Sparse representation and visualization for direct numerical simulation of premixed combustion. Computer Graphics Forum (Proc. EuroVis), 33(3):321–330, 2014
T. Oster, D. J. Lehmann, G. Fru, H. Theisel, and D. Thévenin
-
Probability of hotspot ignition and ignition spot tracking in turbulent hydrogen-air mixtures using direct numerical simulations. In 8th European Combustion Meeting, pages 925–930, 2017
C. Chi, A. Abdelsamie, T. Oster, and D. Thévenin
-
On-the-fly tracking of flame surfaces for the visual analysis of combustion processes. Computer Graphics Forum, 37(6):358–369
Oster T., Abdelsamie A., Motejat M., Gerrits T., Rössl C., Thévenin D., and Theisel H.