1. Combustion Science Data Management Needs Jacqueline H. Chen Combustion Research Facility Sandia National Laboratories jhchen@sandia.gov DOE Data Management Workshop SLAC Stanford, CA March 16-18, 2004 Sponsored by the Division of Chemical Sciences Geosciences, and Biosciences, the Office of Basic Energy Sciences, the U. S. Department of Energy

2. Challenges in combustion understanding and modeling Diesel Engine Autoignition, Laser Incandescence Chuck Mueller, Sandia National Laboratories  Stiffness: wide range of length and time scales –turbulence –flames and ignition fronts –high pressure  Chemical complexity –large number of species and reactions  Multi-physics complexity –multiphase (liquid spray, gas phase, soot) –thermal radiation –acoustics...

3. Direct Numerical Simulation (DNS) Approach  High-fidelity computer-based observations of micro-physics of chemistry-turbulence interactions  Resolve all relevant scales  At low error tolerances, high-order methods are more efficient  Laboratory scale configurations: homogeneous turbulence, v-flame turbulent jets, counterflow  Complex chemistry - gas phase/heterogeneous (catalytic) Turbulent methane-air diffusion flame HO2 CH 3 O  CH4 O Oxidizer Fuel

4. . S3D0: F90 MPP 3D. S3D1: GrACE-based. S3D2: CCA-compliant Software design developments. IMEX ARK. IBM. AMR Numerical developments. Thermal radiation. Soot particles. Liquid droplets Model developments CFRFS CCA Post-processors: flamelet, statistical CMCS DM MPP S3D Arnaud Trouvé, U. Maryland Jacqueline Chen, Sandia Chris Rutland, U. Wisconsin Hong Im, U. Michigan R. Reddy and R. Gomez, PSC High-fidelity Simulations of Turbulent Combustion (TSTC) http://scidac.psc.edu

5. 3D DNS Code (S3D) scales to over a thousand processors  Scalability benchmark test for S3D on MPP platforms - 3D laminar hydrogen/air flame/vortex problem (8 reactive scalars)  Ported to IBM-SP3, SP4, Compaq SC, SGI Origin, Cray T3E, Intel Xeon Linux clusters

6. A Computational Facility for Reacting Flow Science (CFRFS) Develop a flexible, maintainable, toolkit for high-fidelity Adaptive Mesh Refinement (AMR) Massively-Parallel low Mach number reacting flow computations Develop an associated CSP data analysis and reduction toolkit for multidimensional reacting flow Use CSP and a PRISM tabulation approach to enable adaptive chemistry reacting flow computations –PRISM = Piecewise Reusable Implementation of Solution Mapping (M. Frenklach) CCA GUI showing connections

7. Motivation: Control of HCCI combustion  Overall fuel-lean, low NOx and soot, high efficiencies  Volumetric autoignition, kinetically driven  Mixture/thermal inhomogeneities used to control ignition timing and burn rate  Spread heat release over time to minimize pressure oscillations

8. Objectives  Gain fundamental insight into turbulent autoignition with compression heating  Develop systematic method for determining ignition front speed and establish criteria to distinguish between combustion modes  Quantify front propagation speed and parametric dependence on turbulence and initial scalar fields  Develop control strategy using temperature inhomogeneities to control timing and rate of heat release in HCCI combustion  deflagration  spontaneous ignition  detonation Chen et al., submitted 2004, Sankaran et al., submitted 2004

9. Initial conditions Baseline symmetric case Cold core gas Hot core gas  Same mean T (1070K)  Different T skewness and variance (15,30K)  Pressure 41 – 55 atm  Lean hydrogen/air

10. Temperature skewness effect on heat release rate Heat release, HighT, positive skewness 2.0 ms 2.4 ms 2.6 ms 2.8 ms SymmHot coreCold core

11. Temperature skewness effect on ignition delay and burn time  Temperature distribution influences ignition and duration of burning.  Hot core gas  Ignited earlier  Burns longer  Cold core gas  Ignited later  Slow end gas combustion DNS CasesBaselineHotCold Burn time [ms]0.8941.0850.953

12. Ignition front tracking method  Y H2 = 8.5x10 -4 isocontour – location of maximum heat release  Laminar reference speed, s L based on freely propagating premixed flame at local enthalpy and pressure conditions at front surface  Density-weighted displacement speed (Echekki and Chen, 1999):

13. Species balance and normalized front speed criteria for propagation mode Black lines – s* d /s L < 1.1 (deflagration) White lines – s* d /s L > 1.1 (spontaneous ignition) A – deflagration B, C – spontaneous ignition AC B Heat release isocontours

14. Fraction of front length and burnt gas area production due to deflagration Solid line front length Dashed line – burnt area production

15. Comparison of experimental and DNS data for ignition/edge flame data H 2 + O = OH + H O 2 + H = O + OH slow OH recombination RP LP DF OH H  st Normalized OH Expt H2H2 Heated air H 2 /N 2 Normalized OH DNS  Flow divergence effect – (Ruetsch et al. 1994) upstream divergence of flow due to increase in normal component of flow resulting from heat release  Curvature – preferential diffusion focusing effect at leading edge

16. Apriori testing of reaction models using DNS of turbulent jet flames Sutherland et al., submitted 2004 CO/H2/air jet flame, scalar dissipation rate

17. Joint experiment/computation of turbulent premixed methane/air V-flame  Stationary statistics required for turbulent premixed flame model development LES/RANS  Flame topology – curvature stretch statistics  Complex chemistry versus simple or tabulated chemistry (heat release, radicals, minor species)  Is preheat zone thickening due to small scales or higher curvatures in thin reaction zone regime? V-flame, expt. Renou 2003 and DNS, Vervisch 2003

18. Lean premixed combustion at Sandia: swirl burner, LES, and DNS LES J. Oefelein Experiment: OH PLIF, PIV R.W Schefer DNS E. Hawkes and J.H. Chen

19. Data management challenges for combustion science 2D complex chemistry simulations today: 200 restart files (x,y,Z 1,…Z 50 ) skeletal n-heptane 41 species, 2000x2000 grid, 1.6 Gbytes/time x200 files = 0.32 Tbyte, 5 runs in parametric study 1.6 Tbytes raw data Processed data: 2 Tbyte data 3D complex chemistry simulations in 5 years: 200 restart files (x,y,Z 1,…Z 50 ) skeletal n-heptane 41 species, 2000x2000x2000 grid, 3.2 Tbytes/time x 200 files = 640 Tbytes per run, 5 runs = 3.2 Petabytes raw data Processed data: 3 Petabytes Combustion regions of interest are spatially sparse Feature-borne analysis and redundant subsetting of data for storage Provenance of subsetted data Temporal analysis must be done on-the-fly Remote access to transport subsets of data for local analysis and viz.

20. Features Feature is an overloaded word A feature in this context is a subset of the data grid that is interesting for some reason. Might call it a “Region of Interest” (ROI) Also might call it a “structure”

21. Why Feature Tracking? Reduce size of data –How do you find small ROI’s in a large 3D domain? –Retrieve and analyze only what you need Provide quantification –Can exactly define ROI chosen & do specific statistics Enhance visualization –Can visualize features individually –Can color code features Facilitate event searching –Events are feature interactions

22. Feature Detection Detection = Identify features in each time step FDTOOLS tests each cell & groups connected ones There are many possible algorithms including pattern recognition

23. Feature Tracking Tracking = Identify relationships between features in different time steps Again, there are many different algorithms, and knowing about how your features interact helps

24. Events Merge (Birth) (Death) Split Other domain specific events like hard-body collision, vorticity tube reconnect, etc. …

25. Design Goal: Flexible & Reusable Callable from running programs Independent of visualization package Modular –Detector plug-ins –Tracker plug-ins –Other plug-ins … CCA compatible Output interface for further analysis

26. DataSet Types fdRegular 2 & 3D of all fdRefined structured (AMR)

27. FDTOOLS Design (Wendy Koegler SNL) FDTOOLS Component DirectorFeature Manager Tracker RepresenterDetector Data Interface Output Interface Analyzer Visualizer

28. Detection and tracking of autoignition features FDTools (Koegler, 2002): evolution of ignition features Hydroperoxy mass fraction

29. Feature graph tracks evolution of ignition features time

30. Feature-borne analysis

31. Ignition feature classification

32. Terascale virtual combustion analysis facility

33. Data management framework for combustion science – I Distributed data mining tools: feature ID and tracking Distributed analysis tools operating on regions of interest –Reaction source term and Jacobian evaluation –Conditional statistics –Isolevel surface of multiply-connected 3D surfaces Interpolate, integrate, differentiate in principle directions to surface –Computational singular perturbation analysis –Reaction flux analysis –Principal component analysis –Spectral analysis

34. Data management framework for combustion science – II Data objects, which interface to metadata and data –Enabling writing and reading data with various flexible formats –Standard data formats –Automatic conversion utilities Flexible, user-configurable, user-friendly GUI’s to enable user to specify desired operations on data General structured and unstructured adaptive mesh data Real-time feature-borne detection, tracking and analysis for computational steering (e.g. adaptive IO, temporal statistics)

35. Data management framework for combustion science – III Distributed visualization tools scalar and non-scalar data Non-scalar data, i.e. vector or tensor Heterogeneous data – combined experimental and computational data Iso-surface rendering and interpolating data onto user-specified slices Streamlines, information overlays Uncertainty Viz reduced-order representations of flow and combustion features