Dr. Jens-Dominik Mueller 'Shape optimisation with discrete adjoint solvers'
Date: Wed 15 Feb 2012, 15:00 - 16:00
Location: Francis Bancroft 3.21
Dr. Jens-Dominik Mueller
Senior Lecturer in Bio-Fluids (took this from the webpages, Jens please correct if it is wrong)
TITLE: Shape optimisation with discrete adjoint solvers
Multi-parameter numerical optimisation of shapes and topologies using CFD is currently done e.g. with parametric studies at huge computational cost. As an alternative, gradient-based optimisation algorithms can converge with very few design steps, but the cost of computing the gradient with simple methods such as finite-differences is linear in the number of design variables.
Adjoint methods are an essential ingredient for industrial CFD optimisation as they allow to compute gradients at constant cost, irrespective of the number of design variables. Recent focus in the field has been on discrete adjoint methods where an exact derivative of the discretised equations is used. The process can be viewed as a mechanical application of the chain rule, hence it can be automated using Automatic Differentiation (AD) tools.
However, these tools are limited in scope and maturity, and examples in the literature typically require substantial alteration of the CFD code before AD can be successfully applied and typically are limited to simple languages such as Fortran77.
The focus of the work in our group has been to develop adjoint codes using AD tools. In earlier work we applied this to compressible unstructured codes and have recently applied AD to an existing incompressible CFD solver (CFD-ACE+), as well as a bench-code based on the same pressure-correction discretisation. These codes use advanced language features in F90 such as array intrinsics, derived data types and modules. While our in-house codes were written to be suitable for AD, our aim with ACE+ was to derive an adjoint version with minimal changes to the original code.
An introduction to the adjoint approach and its potential in industrial CFD optimisation will be given. The theory and application of AD tools will be briefly covered, including the difficulties of applying AD to CFD codes will be presented. Adjoint-based optimisation results will be presented.