Outcomes of ECMWF/OceanPredict event on ocean data assimilation

Magdalena A. Balmaseda (ECMWF), Andrew Moore (University of California, Santa Cruz), Matthew Martin (UK Met Office)


More than 170 scientists from around the world discussed recent progress and challenges ahead in ocean data assimilation in a virtual event from 17 to 20 May 2021. The workshop was organised jointly by ECMWF and OceanPredict, a science programme for the coordination and improvement of global and regional ocean analysis and forecasting systems. It was designed to meet the ever-increasing requirements of marine, weather, environmental and climate services. It was very timely since 2021 marks the beginning of the UN Decade of Ocean Science for Sustainable Development.

During this event, experts from different domains addressed the multidisciplinary science underpinning climate and environmental monitoring and predictions, the exploitation of novel observations, and interactions in the ocean–atmosphere–sea-ice–biogeochemistry system at global and regional scales.

A combination of 36 plenary talks, 29 poster presentations, four working group discussions and several informal virtual breaks facilitated the exchange of information and seeding of new ideas. The working group discussions addressed questions common to all applications, including treatment of model error, the specification of short-range forecast and observation errors, the balancing of resolution and ensemble configurations, and the exploitation of machine learning. The discussions also covered the infrastructure needed to share developments among different domains, and between operations and research.

Outcomes of ECMWF/OceanPredict event on  ocean data assimilation


In the past few years, operational ocean activities have consolidated in different centres, and the number of ocean specialists has increased. There is now a critical mass to spark productive collaborations. Availability of a wider range of ocean observations greatly facilitates development efforts. It has given rise to activities on observing system impact and design as well as observation operators and treatment of observation error. Oral and poster sessions presented progress on variational and ensemble methodology and algorithms, opportunities for data assimilation arising from machine learning, theoretical considerations on the validity of the various assimilation methods for systems with different degrees of non-linearity, and practical considerations on balancing the competing demands of higher resolution and larger ensembles.

The workshop also touched on common developments for sharing data assimilation infrastructure, with major initiatives such as JEDI (Joint Effort on Data Assimilation Integration) and PDAF (Parallel Data Assimilation Framework) being presented.

More details on the workshop can be found here: A brief summary of the recommendations is presented below.

Overarching recommendations

Coupled data assimilation: Recommendations concern advancing the scientific foundations underpinning the coupling among Earth system components, which will require the productive engagement of all domain scientists. They also suggest exploring machine learning solutions and the use of targeted observations.

Resolution: Models should have sufficient resolution to resolve the relevant physical processes. Approaches for affordable solutions should be explored. Emerging observations with the ability to sample the ocean at finer spatial/temporal scales may become more useful. Observations that sample coarser scales might need different treatment in a high-resolution data assimilation system than before.

Methodology: Further development of methods for representing multi-scale flow-dependent background errors, which include the time dimension, and for balancing resolution/ensemble needs is recommended. The exploration of machine learning solutions for different aspects of data assimilation is encouraged. Efforts to develop methods targeting coupled data assimilation, treatment of model error and parameter estimation should continue and strengthen.

Ensembles: In principle, the quality of the background error statistics diagnosed from ensembles will improve with an increasing number of ensemble members. In practice, computational resources constrain the ensemble size, and it is important to invest in enhanced ensemble generation strategies to improve the reliability of the ensemble and control sampling issues.

Observing System (Simulation) Evaluation experiments (OSEs/ OSSEs): Comparison studies of coordinated OSSEs and OSEs will be helpful to learn from different systems. It is recommended to involve the observational community in the design of observation impact experiments, and to share the OSSE/OSE outputs with the wider community. Specific activities contributing to the UN decade of Ocean Science are encouraged.

Evaluation: As well as standardised metrics of fit to the assimilated observations, evaluation against independent data and error growth diagnostics are important. The reliability of the ensemble and the temporal consistency of the estimation should be considered.

Treatment of observations: Progress should be made to develop efficient methods to model observation error correlations in data assimilation systems. There is a need for developing and sharing methods for automatic quality control and observation bias correction. Historical observation repositories should be updated regularly and as promptly as possible.

Infrastructure developments: Modular and open source software infrastructure that facilitates exchange of developments and their application to different models are welcome and encouraged. Management of the complexity of the data assimilation infrastructure is crucial to facilitate its uptake. The need for data-sharing infrastructure to facilitate collaborations was also identified.

Links with the modelling community: Stronger links between the data assimilation and modelling communities will benefit the scientific and infrastructure developments in both domains. The capability to compare model fields with observations, the information from analysis increments, and the possibility of using data assimilation for parameter tuning are important assets for model development.

Training and recruitment: Investment in training of the next generation of data assimilation scientists is identified as critical. This should include specific training on the use of modern software development/collaboration techniques. Beyond training, sustained funding for data assimilation in the research community is required to maintain a sufficient pool of expertise to exploit new computer architectures and observing systems.

Next steps

The recommendations of the working groups will be used to inform the plans for ocean and coupled data assimilation developments at ECMWF and other forecasting centres around the world. They will also inform work within the Data Assimilation Task team of OceanPredict and will help to improve collaborative efforts being planned under the UN Decade of Ocean Science for Sustainable Development.

The full workshop report can be found here:

This article was written on behalf of the workshop’s organising committee, working group chairs and rapporteurs.