| Abstract: |
This paper proposes a framework for the process of implementing and testing variable precision in scientific software using a case study in geophysical modeling. It addresses the question of how to design this framework when you are purposefully introducing reduced-precision numerical representations that could cause computational rounding and propagate other unknown errors. Scientific geophysical models and data assimilation software inherently contain errors from assumptions and simplifications made to solve equations of state, data structures, numerical rounding, interpolation, and limited observational data. Standardizing a framework to enhance the use, effectiveness, and reproducibility in evaluating software where purposeful error is introduced in an error-filled system will help assist developers and scientific domain experts with analysis and validation. |