Adjoints of large simulation codes through Automatic Differentiation

Laurent Hascoët Benjamin Dauvergne

Revue Européenne de Mécanique Numérique / European Journal of Computational Mechanics (24 pages)

Abstract: Adjoint methods are the choice approach to obtain gradients of large simulation codes. Automatic Differentiation has already produced adjoint codes for several simulation codes, and research continues to apply it to even larger applications. We compare the approaches chosen by existing Automatic Differentiation tools to build adjoint algorithms. These approaches share similar problems related to data-flow and memory traffic. We present some current state-of-the-art answers to these problems, and show the results on some applications.

Keywords: Automatic Differentiation, Reverse Mode, Checkpointing, Adjoint methods, Gradient

Full text (pdf)


@article{HascoetDauvergne07,
  author = {Hasco{\"e}t, L. and Dauvergne, B.},
  title ={Adjoints of large simulation codes through Automatic Differentiation},
  journal = {REMN Revue Europ{\'e}enne de M{\'e}canique Num{\'e}rique / European Journal of Computational Mechanics},
  volume = {17} ,
  pages = "63-86",
  year = 2008
}