Numerical Simulation of Particulate Flows
A. Baggag and A. Sameh
The Computing research institute (CRI) is an interdisciplinary unit whose overall mission is the advancement of the whole computational process. It seeks to advance the role of information technology research in Purdue's science and engineering disciplines. CRI has two foci:
- Creation of information technology that enables broad use of computing and digital communications. This includes numerical and non-numerical parallel algorithms, software systems, architecture of high performance computing platforms, embedded real-time computing, databases and data mining, broadband networks, wireless networking, and sensor/actuator systems; and
- Applications of information technology for effective development of vital computational research areas in science and engineering disciplines.
These include bioinformatics, combustion, electromagnetics, fluid dynamics, materials science and engineering, molecular biology, and nantechnology.
Parallel Linear System Solvers in Particulate Flows
Numerical simulation of particulate flows is of great interest in some industrial applications such as enhancing productivity of oil reservoirs and the manufacturing process of polymers. It is computationally intensive as it involves the numerical solution of coupled Navier Stokes equations and Newton's equations of motion. Finite element discretization of this coupled system on an unstructured grid, using an arbitrary Lagrangian-Eulerian moving mesh, leads to to very large sets of algebraic nonlinear equations, and hence large sparse linear systems, that have to be solved at each time step. These linear systems are nonsymmetric and indefinite. Unless the time steps are extremely small, standard linear system solvers based on Krylov subspace methods fail to converge with classical preconditioners. In this study, we designed a hybrid scheme for solving these indefinite systems that proved to be both robust and ideally suited for parallel computing platforms even with suitably large time steps. [This work is supported by an NSF grant]