Chad Burdyshaw

Achieving Automatic Concurrency Between Computational Field Solvers and Adjoint Sensitivity Codes

A Dissertation Presented for the Doctor of Philosophy in Computational Engineering, The University of Tennessee at Chattanooga

Chad Burdyshaw, May 2006

Abstract:
A method has been developed to address the difficulties involved in the implementation, maintenance and extension of an adjoint type sensitivity solver while remaining concurrent with an evolving computational field solver. Through utilization of the Complex Taylors Series Expansion method (CTSE), in conjunction with an object-oriented, polymorphic framework, the proposed method achieves a significant advantage over existing methods by minimizing implementation and maintenance efforts. The application of CTSE, within this framework, facilitates automatic differentiation of the field solver through code transformation based on the data types of input parameters. Differentiation may also be accomplished less efficiently using a code transformation script. This study includes the application of the proposed method to a large scale, unstructured-mesh, multi-regime fluid flow solver. Issues of implementation, maintenance, extensibility, accuracy, efficiency, and parallel scalability are discussed.