Farhat Research Group Seminars

At these meetings, a Farhat Research Group member will give a 20–45 minute talk on their current research. The goal of these seminars is to promote feedback from lab members and improve collaboration.

• The seminars will be held on Wednesdays from 3–4 pm in Durand 026.
  

• Email me at carlberg AT stanford DOT edu if you would like to present!

Schedule

Abstracts

• 10.5.09 | Harsh Menon (hmenon@stanford.edu)
  
  Abstract not available at this time.
  
  

• 9.21.09 | Jon Greatarsson (jontg@stanford.edu)
  
  Abstract not available at this time.
  
  

• 9.9.09 | Sebastien Brogniez (brogniez@stanford.edu)
  
  Stability and accuracy considerations for designing a time integration scheme on moving meshes
  
  Abstract not available at this time.
  
  

• 9.2.09 | Paolo Massimi (pmassimi@stanford.edu)
  
  Abstract not available at this time.
  
  

• 8.26.09 | Julien Cortial (jcortial@stanford.edu)
  
  Time-parallel solution of structural dynamics problems
  
  The time-parallel framework for constructing parallel implicit time-integration schemes (PITA) has two major applications: (a) the exploitation of a number of processors that is far larger than the maximum number of processors that space-parallelism alone can efficiently utilize for reducing the CPU time associated with the solution of partial differential equations, and (b) the parallelization of time-dependent applications for which space-parallelism is unfeasible because of the very few degrees of freedom they involve.
  
  To this effect, the fundamental principles of PITA are derived from classical domain decomposition techniques. The time-domain is divided in time-slices whose boundary define a coarse time-grid. Seed values of the time-dependent solution are introduced on this coarse time-grid in order to initiate embarrassingly parallel time-integrations on the time-slices. A Newton-like procedure is also designed to iteratively reduce the solution jumps on the coarse time-grid. To avoid artificial resonance and numerical instability, a projector is constructed for efficiently propagating a relevant component of the jump correction on the original (fine) time-grid and accelerate convergence.
  
  When the problem is time-reversible, forward and backward-in-time integrations can be carried out simultaneously for an almost doubled parallel potential. In this talk, the application of these ideas to structural dynamics is presented with experimental results and performance assessment.
  
  

• 8.19.09 | Ulrich Hetmaniuk (Guest Speaker)
  
  Special finite element shape functions based on component mode synthesis
  
  Click Here For Abstract

• 8.12.09 | Kevin Carlberg (carlberg@stanford.edu)
  
  Model reduction-based solution methods for real-time nonlinear simulation and repeated analyses
  
  Model reduction is a powerful tool that can be used to dramatically decrease the cost of numerical simulations in one of two contexts: the real-time context and the repeated analyses context. However, satisfying the computational time and accuracy requirements demanded by problems in these categories is not a trivial task. To this end, one model reduction-based solution method for each category is presented that effectively meets the disparate demands of its respective context.
  
  For real-time problems, extremely fast simulation time is required, while moderate errors in the solution can be tolerated. In this talk, a novel Gappy Proper Orthogonal Decomposition (POD) solution method is presented for executing nonlinear simulations in near real-time. The method is based on computing only a few (greedily-chosen) rows of the residual and Jacobian matrix at each Newton iteration.
  
  Repeated analyses problems, such as those arising in design optimization, are characterized by the need to solve consecutive linear systems of equations of the form Aⁱxⁱ=bⁱ, where Aⁱ are nearby matrices. These problems typically have relatively stringent accuracy requirements that typical model reduction techniques can fail to meet. To this end, a novel POD-based iterative method is presented for repeated analyses problems. The algorithm employs a POD basis to accelerate convergence of solving each linear system.
  
  

• 8.5.09 Kevin Wang (icmewang@stanford.edu)
  
  Computation of flow-induced forces on embedded meshes
  
  Non body-conforming CFD grids in which wet surfaces of various obstacles are embedded are gaining popularity in many scientific and engineering applications. Despite many attractive advantages, these methods tend to complicate issues such as enforcing the fluid-structure transmission conditions in general, and the computation of the flow-induced loads on the wet surfaces of obstacles in particular.
  
  In this talk, both of these issues are addressed. More specifically, two load transfer algorithms are presented for fluid and fluid-structure solvers based on embedded methods. In both algorithms, the fluid Dirichlet boundary condition ?—? or for fluid-structure interaction problems, the kinematic transmission condition ?—? is enforced via the analytical solution of an appropriate one-dimensional fluid-structure Riemann problem. As for load transfer, the first algorithm, highlighted by conservation of energy, approximates the structure wet surface by a selected set of control volume faces, while the second algorithm reconstructs the structure wet surface within the fluid mesh in order to obtain better accuracy. The numerical properties of these methods are theoretically analyzed. Their practical significances are also highlighted for several three-dimensional fluid and fluid-structure interaction problems associated with underwater implosion and aeronautical applications.
  
  

• 7.29.09 David Amsallem (amsallem@stanford.edu)
  
  Algorithms for Interpolation on Matrix Manifolds and Application to Model Reduction
  
  A new methodology is presented for interpolating parameterized quantities belonging to matrix manifolds. It relies on identifying the correct manifold for the given application, constructing the appropriate logarithm mapping to move the interpolation data to a tangent space to this manifold where a standard multivariate interpolation algorithm can be applied, and constructing the appropriate exponential mapping to bring back the computed result to the manifold of interest. The proposed methodology is illustrated with its application to real-time computational engineering using databases of reduced-order models.
  
  

• 7.22.09 David Powell (dpowell1@stanford.edu)
  
  Multi-scale modeling and large-scale transient simulation of ballistic fabric undergoing impact
  
  Ballistic fabrics such as Kevlar and Zylon are finding new uses not only as shielding for personnel but also in commercial and military aircraft protecting flight-critical components in the event of a high-speed ballistic impact. Since experimental tests on these materials are often expensive and time consuming, a computational model amenable to large-scale numerical simulations on massively parallel processors would provide an ideal alternative. To this effect, a novel multi-scale approach has been developed for modeling ballistic fabric that extracts information from the micro-scale fibril material properties to build the macro-scale sheet. This approach incorporates the stochastic nature of the material due to the random variations in the yarn caused by the weaving process. Representing these variations is critical to capturing the heterogeneous damage and failure of the material as confirmed by experimental tests. A parallel implementation of the proposed multi-scale method has been developed with particular attention to the contact problem. This talk will focus on all of these issues and report on massively parallel large-scale computations for large sheet applications with both performance and validation results.