In four-dimensional variational data assimilation (4DVAR) the majority of the wall clock time is spent calculating the tangent-linear and adjoint. Recent work, but others, indicates that it may be possible to construct at 'time-parallel' version of 4DVAR, in which the full assimilation window is broken down into a nubmer of shorter assimilation windows - each run in parallel. This presentation summarizes the challenges encountered with our ongoing development and use of a software system designed to facilitate exploration of computational optimizations and strategies for Data Assimilation. The software system is designed and constructed from scratch using modern software development methods and tools, though it incorporates components of pre-existing systems where appropriate. We present results of experiments that employ this system to test approaches for assimilation of observations using a four-dimensional variational (4DVAR) scheme. We propose a modular DA system software architecture and demonstrate its utility using a set of models of varying realism and complexity. The software system design and implementation was initially tested and validated using a simple chaotic atmospheric model. A Quasi-Geostrophic (QG) atmospheric model was used to conduct DA experiments of increased difficulty and to validate the software design at larger scales of model complexity. Our QG DA study focused on 2016 winter weather data where a Nature run was used to represent the “true” state of the atmosphere and observations, whereas observation error covariance and observation operator were adapted from pre-existing DA systems. Results will show differences in wall clock time and forecast skill from 3DVAR, 4DVAR, and parallel-in-time 4DVAR assimilation.
This publication was presented at the following:
Authors who have authored or contributed to this publication.