Dakotathon

The CSDMS Dakota Interface

Dakota is a software toolkit developed at Sandia National Laboratories that provides an interface between models and a library of analysis methods, including support for sensitivity analysis, uncertainty quantification, optimization, and calibration techniques. Dakota is a powerful tool, but its learning curve is steep: the user not only must understand the structure and syntax of the Dakota input file, but also must develop intermediate code that allows Dakota to set up and run a model, read outputs from the model, and calculate a response statistic from the outputs.

The CSDMS Dakota Interface, or Dakotathon, is a Python package that wraps and extends Dakota’s file-based user interface. It simplifies the process of configuring and running a Dakota experiment, and it allows a Dakota experiment to be scripted. Dakotathon creates the Dakota input file and provides a generic analysis driver. Any model componentized in the CSDMS modeling framework automatically works with Dakotathon. Dakotathon has a plugin architecture, so models not wrapped into the CSDMS modeling framework can be accessed by Dakotathon by programmatically extending a template; an example is provided in the Dakotathon distribution.

Currently, six Dakota analysis methods have been implemented from the much larger Dakota library:

• vector parameter study,
• centered parameter study,
• multidim parameter study,
• sampling,
• polynomial chaos, and
• stochastic collocation.

For instructions on installing and using Dakotathon, please see Dakotathon's CSDMS Model page.

Links

• The CSDMS Model page for Dakotathon
• https://github.com/csdms/dakota
• http://csdms-dakota.readthedocs.io
• https://dakota.sandia.gov/
  • https://dakota.sandia.gov/download.html
  • https://dakota.sandia.gov/content/install-dakota
• https://dakota.sandia.gov//sites/default/files/docs/6.5/html-ref/index.html
• Slides and repository for AGU talk https://github.com/mdpiper/AGU-2016

Mpiper (talk) 14:09, 27 December 2016 (MST)