Work to implement the Object-Oriented Prediction System (OOPS) as a new 4D-Var data assimilation framework for ECMWF’s Integrated Forecasting System (IFS) is making good progress, with first OOPS-IFS test results showing good performance. OOPS was conceived at ECMWF by Yannick Trémolet and Mike Fisher as a unified, easy-to-use framework for running different variational data assimilation formulations with a variety of forecast models. It will replace the control layer of data assimilation code that has supported ECMWF for the past 20 years. OOPS is an international effort, involving major input from ECMWF, Météo-France and the HIRLAM-ALADIN community.
The OOPS code is available under an Apache-2 licence. This enables wider collaboration with other centres. The collaboration with the European numerical research centre CERFACS has already been valuable, and there are prospects for working closely with the US Joint Center for Satellite Data Assimilation (JCSDA). Academic partners also benefit from the simplified models in OOPS to test new data assimilation approaches. It is then relatively straightforward to test ideas demonstrated in this way with the full system, in collaboration with ECMWF.
Benefits of OOPS
OOPS is an abstract control layer that can manipulate elements of the data assimilation system without needing to know their model-specific implementation details. OOPS includes simplified models (Quasi-Geostrophic and Lorenz-95) to enable early testing of data assimilation algorithms. This enabled the early demonstration of new concepts such as the saddle-point formulation of weak-constraint 4D-Var. OOPS-IFS will bring some significant benefits:
- it will be easier to develop and test alternative minimisation algorithms
- it will be possible to test approaches such as the saddle-point formulation with a full system
- OOPS will provide a common framework for the development of coupled data assimilation
- OOPS reduces interdependencies in the code that make it hard to change one area without causing unexpected impacts elsewhere
- the multi-incremental assimilation will be run as a single executable, reducing I/O costs.
Towards implementation
Interfacing OOPS with the IFS, including atmospheric 4D-Var and NEMOVAR ocean 3D-Var, has necessitated significant refactoring of the Fortran code. The major goal in 2016 and 2017 has been to enable the IFS 4D-Var system to be run from OOPS with the same performance and capabilities as in operations, so that we can replace the current IFS control layer with OOPS. This is a major undertaking due to the complexity of the IFS: in addition to the core 4D-Var algorithm, the IFS has many essential components which must work consistently under OOPS. Late in 2016, a highly simplified 4D-Var system was working under OOPS. Since then progress has been very encouraging. In 2017 many more elements of the full 4D-Var system have been added. During this work some issues in the IFS itself have been discovered, notably an error in the virtual temperature conversion in the thermodynamic balance operator of the background error covariances. A correction for this has been applied in the IFS with a notable positive impact. In this way the OOPS development is already benefiting operational scores.
However, some elements remain to be implemented, including the variational quality control (VarQC), second level preconditioning, the ‘sink’ skin temperature control variable needed for radiance assimilation, and correlated observation errors. These features are needed to assimilate the wide variety of observations required to produce high-quality initial conditions. Without them OOPS can only run with a limited set of observations.
Initial tests encouraging
In order to test OOPS-IFS, it has been compared to both the full IFS and a simplified IFS, where data assimilation components and observations not yet available under OOPS are switched off. Here, only conventional observations (excluding aircraft) and the important satellite radiances from AMSU-A and ATMS (mostly giving temperature information) are assimilated. Comparing OOPS to the simplified IFS allows us to assess whether its current capability is working well and how close we are to an operational level of skill.
The chart shows that OOPS produces a model short-range forecast that does not fit observations from the ATMS satellite instrument as well as the full IFS. This is to be expected as not all components of the full system are yet in place. However, it fits ATMS slightly better than the simplified IFS. Other observation fits and forecast impacts have also been examined, albeit for short periods. They lead to the same conclusion. Other data assimilation diagnostics are giving good results with OOPS, including the adjoint test and the minimisation and convergence of 4D-Var. The conclusion is that technically OOPS-IFS is performing similarly to the IFS when run in the same configuration, but that the remaining components that are still lacking in OOPS are critical to its pre-operational readiness. It is planned to implement these remaining components into a test branch of the IFS in 2017 with the intention of inclusion in IFS Cycle 46r1.
The target is that OOPS-IFS will be operational in 2019.