Tim Albring, SciComp

Title:
Semi-automatic Transition from Simulation to Optimization

Abstract:

PDE-constraint optimization often relies on the adjoint-based sensitivity evaluation, where one distinguishes between the discrete and the continuous methods. Both approaches have their advantages and disadvantages, but they are both difficult to apply to complex models and require involved development. However, based on the abstract structure of the primal fixed-point solver often applied for the numerical solution of PDEs, we will demonstrate in this talk that it is possible to construct a discrete adjoint solver which enables the computation of consistent gradients in a robust way. While the development and maintenance of the adjoint solver is automatically performed along with the development of the primal solver it also directly inherits its convergence properties. Since the implementation is heavily based on advanced techniques of Algorithmic Differentation (AD), we will give additionally some introductory notes on this method of evaluating gradients in a computer program. Furthermore, application to the open-source multi-physics framework SU² used for aerodynamic shape optimization will finish the talk.