Institutions & Research Projects developing and using OpenAD

OpenAD is a tool for automatic differentiation (AD) of numerical computer programs. The initial development was part of the ACTS (Adjoint Compiler Technology & Standards) project funded  by the NSF Information Technology Research (ITR) program. Current research and  development is supported by

• ECCO2 under the NASA Modeling, Analysis, and Prediction (MAP) program
• CSCAPES under the DOE (Office of Science) Scientific Discovery through Advanced Computing  (SciDAC) program
• Uncertainty Quantification in Geophysical State Estimation ( Part I /  Part II ) under the NSF's ITR program.
  

The main goals for OpenAD initially defined for the  ACTS project are: Develop a flexible, modular, open source tool that can generate adjoint codes of numerical simulation programs,  Establish a platform for easy implementation and testing of source transformation algorithms via a language-independent abstract intermediate representation  Support for source code written in C and Fortan, Generate  an adjoint for the MIT general circulation model.  The example on  the right (animated gif)  shows a sensitivity map of the heat transport in the north atlantic to temperature in a depth of 1590 meters over a period of 10 years going backwards in time. There are also results of a 100 year simulation at three depth levels [mpg] [mov].

The intention of this project summary is to give potential users of adjoint compiler technology an overview of the current capabilities of OpenAD together with pointers to more detailed information in the form of online documentation and published papers. It should substantiate the following important points:

• OpenAD has been used successfully to implement a source transformation tool for the automatic generation of adjoint Fortran code.
• The automatic generation of adjoint code for large-scale numerical simulations is a highly complex and work intensive task in terms of algorithmic and software-technological problems.
• The ACTS project made an important first step in the right direction. Further funding is required to continue the work on open software for the automatic generation of robust and efficient adjoint code.

A number of questions are likely to occur while browsing this website. The authors, Uwe Naumann and Jean Utke, are looking forward to answering them. The links under "Summary" in the frame to your left  are a brief guided tour. More details can be found following the links under 'Technical Details'.