Title: Adaptive and Parallel Computational Simulations in Various Engineering Applications

Speaker: Xiaojuan (Sarah) Luo, Scientific Computation Research Center (SCOREC), Rensselaer Polytechnic Institute

Date/Time: Monday, July 27, 2009 at 9:00am – 10:00am

Location: CSRI Building, Room 90 (Sandia NM)

Brief Abstract: Computational science and engineering has matured as a design tool that is synergistically used along with theory and experiment. Its ability to tackle problems with complicated three-dimensional effects at multiple interacting scales of length and time has attracted tremendous attention of scientists and engineers. Realizing such a potential requires development and implementation of advanced numerical techniques capable of solving large-scale engineering problems. The first part of this talk presents several adaptive meshing technologies for various engineering applications such as accelerator design, fusion modeling and viscous flow analysis. More specifically, unstructured mesh adaptation, mesh curving and high-order boundary layer meshing will be discussed that have been developed and used to perform accurate finite element analysis. Examples include usage of valid curvilinear meshes in designing next generation accelerators (that resulted in substantial savings such as 30% improvement in simulation time due to a better-conditioned system) and in high-order viscous flow analysis (that achieved the optimal rate of convergence in shear force computation).

The second part of this talk is focused on parallel multi-scale analysis and its applications for bioengineered tissues. A collagen multi-scale structural model has been developed to analyze the mechanics of bioengineered tissues that involve compositions and structures at several scales such as fibers (m) and tissues (m) and cannot be modeled accurately using a single-scale method. The model employs fiber elements to statistically represent the microscopic collagen fiber network in representative volume elements (RVEs).  Further, the Cauchy stress tensor is calculated using volume averaging theory for each RVE and linked to solve the macro-scale stress balance. The simulation results are of significant importance to the outcome of clinical applications and the laboratory design and industrial production. The computational cost for the simulations is substantial and a domain decomposition parallel computing technique is employed. The approach applies dynamic load balance partition tool (Zoltan) to decompose the macroscopic finite element mesh into multiple partitions, each of which as well as the underlying microscale RVEs are associated with a unique processor. Such a technique allows computing both the RVEs and the partitioned mesh independently in each processor. The communication is accomplished at the macroscopic scale through sending and receiving degree of freedom information between processors. Results demonstrated that the parallel approach can achieve good scalability on IBM Blue Gene/P machine up to 8k processors.

CSRI POC: Pat Knupp, (505) 284-4565