Automatic Differentiation applied to Edge in NFFP5 project MADEF

Authors:

  • Olivier Amoignon

Publish date: 2014-04-09

Report number: FOI-R--3689--SE

Pages: 29

Written in: Swedish

Keywords:

  • function derivatives
  • sensitivities
  • adjoint method
  • automatic
  • differentiation
  • reverse mode.

Abstract

In the project MADEF we apply automatic differentiation (AD) as a complement to the adjoint flow solver already present in the CFD code Edge, in the framework of aerodynamic aircraft design. We used the code Tapenade from INRIA for the development of a program called "dforce" that calculates derivatives of a function F depending on the flow solution given by the CFD solver such as aerodynamic forces, moment or criteria like the DC60 for internal flows. The derivatives of F, with respect to the nodal values of the primitive or conservative variables for instance, are partial derivatives used at different stages when calculating the shape derivative of F. Our focus is thus on applications in design carried out by gradientbased optimization where F is either the cost function being minimized or a constraint. Following the development of the program dforce, the adjoint solver in Edge has been adapted in order to enable using right-hand sides calculated by AD.