Testing an Optimized 4D-Var Technique in NWP Models

4-D Var is an effective data assimilation methodology that is used at select operational NWP centers around the world. 4-D Var has been shown to provide an accurate analysis which consequently improves medium range forecasts. Challenges associated with this approach are development and maintenance of the tangent linear and adjoint model, a high computational cost, and the deterministic nature of the analysis. Recent extensions of 4D-Var support stochastic prediction, including Ensembles of DA (EDA) and the Ensemble-Variational Integrated Localized (EVIL) method.

In order to address the computational cost, we evaluate a new formulation of the 4D-Var approach, the Randomized Incremental Optimal Technique (RIOT) in a global atmospheric model. The key aspect of this algorithm is that it takes a widely used operational data assimilation algorithm (incremental 4D-Var) and implements a new dimension of parallelization by replacing a key sequential step (conjugate gradient minimization) with a highly scalable, randomized, and parallel algorithm, e.g., Randomized Singular Vector Decomposition (RSVD). In addition to providing an analysis increment in a shortened wall-time, RIOT greatly improves the scalability of the posterior covariance estimation in 4D-Var. RIOT’s potential for operational scalability in high-dimensional problems has previously been investigated within the chaotic Lorenz-96 model and for chemical emission constraint in WRFDA.

We will test RIOT within the Joint Effort for Data Assimilation Integration (JEDI) framework developed by the Joint Center for Satellite Data Assimilation (JCSDA). The new method will be first evaluated by using a simplified Quasi-Geostrophic (QG) model. The next level of testing will use a model for Numerical Weather Prediction (NWP) whose adjoint and tangent versions are available in the JEDI framework. Within these two models we will evaluate the computational performance of the traditional and alternative approaches. Also we will compare the two methods in terms of the accuracy of analysis fields and corresponding forecasts they produce. Future work will address the approximation quality of the posterior covariance and the performance of RIOT in cycling data assimilation.