J. Ray
|
|
J. Ray |
|
|
A Computational Facility for Reacting Flow Science This project is focused on advancing capabilities for scientific studies of low Mach number chemically reacting flows on massively parallel computational hardware. We are developing and using a flexible component-based toolkit to enable simulations of reacting flows with detailed chemistry and transport. Key challenges involve developing: (1) a scalable high-order adaptive mesh refinement (AMR) strategy to address the issue of spatial scales; (2) a scheme to reduce the computational loads associated with large stiff chemical mechanisms; and (3) a scalable software approach that accommodates contributions by researchers of diverse software expertise as well as incorporation of legacy code to a very complex codebase. We use a parallel C++ AMR framework (GrACE) to address the issue of the large range of length scales in reacting flow. Multiple levels of refinement, required to achieve an acceptable resolution of flame structure, impose a serious load-balancing (and consequently parallel scalability) problem. This is being addressed by using high-order (greater than second order) discretization schemes which deliver the same accuracy as conventional second-order AMR schemes but at significantly coarser resolution. This enables resolution of laboratory-sized flames with few (3-4) levels of refinements in the mesh, a configuration that can be properly load-balanced by current domain decomposition technology. Preliminary studies with model problems have shown the recovery of ideal (high-order) convergence on AMR meshes. The numerical construction uses the AMR structure for solving the species and energy conservation equations, coupled with a fixed-mesh low Mach number momentum solver. The problem of chemical complexity and stiffness (which imposes large computational loads) is being addressed using computational singular perturbation (CSP) theory. We use CSP to identify slow and fast chemical processes at a given spatiotemporal location. This information on the underlying chemical manifold is then utilized to project out the fast modes, arriving at a reduced non-stiff chemical source term. Results show that such a simplification allows us to stably evolve each point explicitly at relevant slow timescales. Given the significant cost of the CSP analysis we use the Piecewise Reusable Implementation of Solution Mapping (PRISM) tabulation procedure to construct and reuse local response surface models of the CSP basis vectors in discretized hypercubes in chemical phase space. The software structure follows the Common Component Architecture paradigm whereby specific functionalities are embodied in separate components which are assembled into simulation codes. The bulk of the components involve wrappers around legacy codes and external publicly available libraries. Contributors to this research project include H. Najm, C. Kennedy, S. Lefantzi, J. Lee (Sandia National Laboratories), M. Valorani (University of Rome, “La Sapienza”), D. Goussis (Greece), and M. Frenklach (Lawrence Berkeley National Laboratory).
Sandia National Laboratories, Livermore, CA |