3 Presentations

• 
  03:30 PM - 03:55 PM

  Static Aeroelastic Design Optimization of Lightweight Structures: Computational Challenges and Opportunities

  • Graeme Kennedy, presenter, University of Toronto

  Static aeroelastic design optimization of flexible, lightweight aircraft involves the simultaneous design of both aerodynamics and structures. Static aeroelastic analysis itself is a computationally expensive problem that often requires the use of high-performance computing techniques. Simulation-based design optimization of static aeroelastic systems presents many computational challenges and opportunities for the application of novel optimization techniques that have the potential to reduce computational times.

• 
  03:55 PM - 04:20 PM

  Quasi-Newton Jacobian Estimates for Matrix-Free Structural Optimization

  • Andrew Lambe, presenter, University of Toronto
  • Joaquim Martins, University of Michigan

  In structural optimization problems with failure constraints, the computational cost is dominated by computing the gradients of all the constraints. Using a "matrix-free" optimizer can reduce this cost significantly by requiring only appropriate matrix-vector products with the full constraint Jacobian. To keep the total number of matrix-vector products small, we will advocate estimating the constraint Jacobian within the matrix-free optimizer using a quasi-Newton method. Results from structural test problems demonstrate that the computational cost scales well compared to traditional SQP algorithms.

• 
  04:20 PM - 04:45 PM

  Implementation of a Matrix-Free Augmented Lagrangian Algorithm

  • Sylvain Arreckx, presenter, Polytechnique Montréal
  • Dominique Orban, GERAD - Polytechnique Montréal

  In many applications, problems are so large that we cannot compute/store explicit Jacobians and don't have access to Hessian information. We will outline a matrix-free algorithm for solving nonlinear problems with both
  equalities and inequalities. Our algorithm is based on an augmented Lagrangian approach and relies on matrix-vector products only. We also show some numerical results on the CUTEr and COPS collections.