AMOR

 

The AD2CompEng Project

Automatic Differentiation and Adjoints applied to Computational Engineering

Sponsored by the EPRSC under Grant  GR/R85358/01

 

Investigators:

Introduction:

The incorporation of engineering simulation software into engineering design is presently hindered by slow simulation turn-around times, poor gradient accuracy and long adjoint code development times. In a number of cases we believe that present AD tools would address such issues and provide valuable input to AD researchers in developing the next generation of AD tools and associated theory. Crucial to the above is having the educated manpower to use the AD tools appropriately and intervening directly into the differentiated code where necessary to improve efficiency.

 

Project Aims:

  1. To provide efficient gradient code using AD for a complicated structural engineering design problem and hence allow for a novel coupling of evolutionary and gradient based optimisation. As well as underpinning new engineering design strategies this will be published as a case study in obtaining AD gradient code for a complex engineering application.
  2. To carefully compare the quality of hand-coded vs. AD-produced adjoint code for aerodynamic adjoints. To date all discrete adjoint code has been developed either by hand-coding or AD. Here we will uniquely compare a previously produced state-of-the-art hand-coded adjoint solver with an AD produced version in terms of both ease of development and run-time requirements.
  3. To apply the lessons learned from the steady flow solver of (ii) to an adjoint associated with an unsteady-like, space-marched flow solver. Here we will further investigate issues of check-pointing i.e., how much of the evolved flow field should be stored and how much recomputed for the adjoint associated with an implicit flow solver.
  4. To clearly show how to utilise AD to provide Jacobian-vector products and preconditioning matrices within a Newton solver for Turbulent Navier-Stokes flows. Hence we will demonstrate how the improved accuracy of forward mode AD (compared to finite-differencing) overcomes the associated doubled expense to result in overall improved efficiency and robustness. This will extend present promising results (Hovland & McInnes Parallel Computing 27,  2001) to a solver of greater complexity and poorer numerical conditioning (stiff system, stretched mesh)
  5. To thoroughly educate a researcher as a U.K. AD application specialist through the above tasks.
  6.  Feedback lessons learned back into the AD community to improve AD tools and theory applied to such problems as (1-4).

Project Meetings:

Journal Papers:

Books:

  

Conference Papers:

Conference & Workshop Presentations:

 

AMOR home