Parallel, Adaptive Finite Element Methods for Chemically Reacting Flow Simulations



K.D. Devine, J.N. Shadid, A.G. Salinger, H.K. Moffat, and S.A. Hutchinson, G.L. Hennigan

Sandia National Laboratories
Albuquerque, New Mexico 87185

Abstract

MPSalsa [5, 7] is a computer program developed at Sandia National Laboratories that solves coupled three-dimensional fluid flow and detailed reaction chemistry systems for modeling chemically reacting flow, such as flow in Chemical Vapor Deposition (CVD) reactors. MPSalsa uses an unstructured finite element mesh to model complicated geometries. Thermal diffusion, variable physical properties, and mixture-averaged and true multicomponent diffusion are included. Both gas and surface reactions can be modeled using the Chemkin and Surface Chemkin libraries [1, 4]. An inexact Newton method is used to solve the nonlinear system of fully-coupled equations [2]. The Aztec library [3] of preconditioned Krylov iterative solvers is used to solve the resulting linear systems. MPSalsa and Aztec have been optimized for high performance on massively parallel computers. Peak speeds of 65 gigaFLOPS have been achieved on Sandia's Intel Paragon computer [6].

To capture features such as steep gradients in species concentrations in chemically reacting flow solutions, an automated adaptive mesh refinement strategy is used. Adaptive methods have been used with great success for a variety of mechanics and fluid flow problems. Their use on parallel computers, however, is complicated by the need for distributed data structures, dynamic load balancing, data migration, and maintenance of a distributed element database. In this talk, we describe the data structures and strategies used for adaptive mesh refinement and dynamic load balancing in MPSalsa. Results demonstrating the performance of adaptive refinement and load balancing will be shown.