DD-DA PinT-based model: A Domain Decomposition approach in space and time, based on Parareal, for solving the 4D-Var Data Assimilation model
This work addresses the computational bottleneck of 4D-Var data assimilation for large-scale PDE-constrained optimization, but the contribution is incremental as it combines existing methods.
The paper presents a domain decomposition approach combined with parallel-in-time methods (Parareal) for solving 4D-Var data assimilation, achieving improved accuracy and efficiency through parallel coarse and fine solvers while preserving data locality.
We present the mathematical framework of a Domain Decomposition (DD) aproach based on Parallel-in-Time methods (PinT-based approach) for solving the 4D-Var Data Assimilation (DA) model. The main outcome of the proposed DD PinT-based approach is: 1. DA acts as coarse/predictor for the local PDE-based forecasting model, increasing the accuracy of the local solution. 2. The fine and coarse solvers can be used in parallel, increasing the efficiency of the algorithm.3. Data locality is preserved and data movement is reduced, increasing the software scalability. We provide the mathematical framework including convergence analysis and error propagation.