Continued progress in numerical weather prediction requires sustained, collaborative research in weather observation processing, modelling and computing. Often research efforts in these different areas have to go hand in hand. For example, modelling the atmosphere and other Earth system components at higher resolution requires advances in numerical methods and computing.

Extra observations
Extra observations Example of additional observations assimilated in a single continuous data assimilation cycle compared to the current operational setup. They include satellite observations from a large number of instruments as well as in situ measurements.

ECMWF is well placed to achieve such coordination because its scientists carry out research and development across all areas of numerical weather prediction in close collaboration with partner organisations and fellow scientists across the world.

In 2018, several major research projects came to fruition or made decisive progress. Important advances were made in data assimilation and in our understanding of persistent biases in ECMWF’s medium-range forecasts of near-surface weather parameters. Progress was also made in ensemble size research, atmospheric composition priorities for numerical weather prediction and work on an alternative dynamical core using the Finite Volume Module.

Continuous data assimilation

Accurate weather predictions rely on an accurate specification of the initial state of the Earth system. Data assimilation combines tens of millions of Earth system observations with model information to arrive at a full set of initial conditions called the analysis. In ECMWF’s configuration operational in 2018, by the time the analysis was complete, the most recent observations that had gone into producing it were almost two hours old.

Building on previous work at the Centre and elsewhere, ECMWF scientists developed a revised configuration of the 4D-Var data assimilation system that enables the analysis to benefit from more recent observations. In the new, more continuous framework, observations taken around one and a half hours later can be assimilated. The change also makes it possible to add another iteration to the data assimilation calculations, thus improving their accuracy.

The new data assimilation configuration was found to improve forecast quality significantly. It was scheduled for inclusion in the 2019 upgrade of ECMWF’s Integrated Forecasting System to IFS Cycle 46r1.

50-member Ensemble of Data Assimilations

Since 2010, ECMWF has run an Ensemble of Data Assimilations (EDA) to help determine the initial conditions for its forecasts. The EDA is an ensemble of 4D-Var data assimilations that reflects uncertainties in observations, atmospheric boundary conditions and the model physics. It contributes to the high-resolution analysis by providing flow-dependent estimates of the errors in the short-range forecasts used in the data assimilation system. It is also used to determine the perturbed initial conditions for ensemble forecasts: EDA-based perturbations are combined with singular vectors to construct perturbations to the high-resolution analysis.

The year 2018 saw the successful conclusion of work on a new, optimised 50-member EDA configuration that has a comparable computational footprint to the previous 25-member configuration. The increase in ensemble size improves both the high-resolution analysis and the EDA-based perturbations to the initial conditions for the 50-member ensemble forecast. The change also means that ensemble forecast members are exchangeable. Among other things, this will increase the efficiency of research experimentation involving ensemble forecasts. The change was scheduled for inclusion in the 2019 upgrade of ECMWF’s Integrated Forecasting System to IFS Cycle 46r1.

Headline score
Exchangeable ensemble members The ensemble members produced using the 50-member EDA configuration are exchangeable, as illustrated by these charts of the mean absolute difference of 500 hPa geopotential in the northern extratropics between consecutive pairs of ensemble forecast members and non-consecutive pairs. The plots show the results of an experiment with plus–minus symmetry of the initial perturbations based on a 25-member EDA (left), and of an experiment without such symmetry based on the new 50-member EDA (right). The differences are averaged over a period of 41 days.

Ensemble size in research

An ensemble forecast is a set of forecasts that represent the range of possible future weather evolutions. The larger the ensemble, the better a probability distribution can be estimated. But research experiments to test new ideas are computationally expensive if the number of ensemble members is high. This can slow down progress: the larger the ensemble, the longer scientists have to wait for the results of their tests.

Work carried out in 2018 showed that some of the most relevant forecast performance scores obtained with a small number of ensemble members can be valid for large ensembles too. This can be achieved by making suitable statistical corrections in the score computation.

An important condition is that ensemble members are exchangeable. In that case, meaningful results for the operational 50-member ensemble can be achieved in tests involving as few as eight ensemble members. These findings can help to speed up progress in numerical weather prediction under the constraint of limited computing resources.

Probability densities for different ensemble sizes
Probability densities for different ensemble sizes The probability densities shown here, estimated for ensembles with 20, 50, 200, 1,000 and 4,000 members, illustrate the value of a large ensemble. The underlying true distribution is also shown (bottom right). For each ensemble size, 16 different ensembles were generated from the underlying distribution. The variations among the densities reflect the sampling uncertainty due to the finite ensemble size.

Reducing forecast biases

Investigations carried out in 2018 showed that persistent biases in ECMWF’s medium-range forecasts of near-surface weather parameters are closely related to the coupling between the atmosphere and the land surface in the Integrated Forecasting System. They are also related to other processes, such as turbulent mixing, radiation and clouds.

The investigations were part of an ECMWF initiative entitled ‘Understanding uncertainties in surface–atmosphere exchange’ (USURF), which started in November 2017. Key to making progress was the availability of suitable in situ data from ECMWF’s Member and Co-operating States and the development of a conditional verification methodology, which helped to identify specific processes as likely causes of some of the biases. Work in 2018 focused on 2-metre temperature biases in Europe. However, because of the physical links between 2-metre temperature, 2-metre humidity and 10-metre wind speed, investigations have also included some aspects of humidity- and wind-related processes.

The results of the initiative will feed into the operational implementation of a multi-layer snow scheme and an improved land scheme as well as a revised framework for moist processes relying on consistent assumptions and improved coupling between the turbulent diffusion, convection and cloud schemes and the dynamics.

Temperature bias variability
Temperature bias variability Two-metre temperature biases in the IFS vary according to the time of day as well as by region, as shown in these charts showing the mean error (bias) of the forecast for day 3 in winter 2017/18 (December –January–February) at 00 UTC (left) and 12 UTC (right). Verification was performed against a subset of SYNOP weather station observations.
Experimental snow scheme
Experimental snow scheme Tests showed that a substantial reduction in the 2-metre temperature warm bias in northern Scandinavia can be achieved when using different versions of an experimental multi-layer snow scheme. The tests were carried out for the box 64–70°N, 15–30°E and the period 17 February – 1 March 2018.

Atmospheric composition priorities

One of ECMWF’s strategic goals for 2025 is to develop an integrated global model of the Earth system to produce forecasts with increasing fidelity on time ranges up to a year ahead. This will be achieved through an increased level of complexity of physical and chemical processes as well as of Earth system interactions in the model. Atmospheric composition has the potential to be one of the sources of predictability at different timescales.

A cross-departmental working group on atmospheric composition was set up at ECMWF in spring 2018. It has worked to identify a set of priority developments that have the potential to improve forecasts at all scales, from a few days to seasons ahead.

The group made recommendations concerning modelling and data assimilation aspects for ozone, CO2 and aerosols; improving the atmospheric composition numerical infrastructure and code efficiency; and modifying the model performance evaluation process to account for atmospheric composition. In pursuing these objectives, the Centre will be able to leverage some of the capabilities of the Copernicus Atmosphere Monitoring Service (CAMS) implemented by ECMWF on behalf of the EU.

Aerosol optical depth

© ECMWF/CAMS

Aerosol optical depth At any one time, the atmosphere contains a mix of aerosols, such as dust, sea salt, biomass burning particles and sulphate, as shown in this example of a CAMS map for aerosol optical depth at 550 nm.

Finite Volume Module improvements

Over the next decade, many aspects of ECMWF’s Integrated Forecasting System (IFS) may need to change to enable the production of the computationally demanding higher-resolution global forecasts called for by ECMWF’s long-term Strategy. The dynamical core lies at the heart of the model infrastructure. It numerically solves the governing equations describing the resolved atmospheric dynamics.

The dynamical core is coupled to parametrizations of small-scale physical atmospheric processes and to models of other Earth system components. The current IFS dynamical core uses the spectral transform method to solve the governing equations. ECMWF is continuing to develop this dynamical core, which also includes a nonhydrostatic option, to make it as computationally efficient as possible.

Minutes per forecast day comparison
Minutes per forecast day comparison Elapsed time to run one day of the dry baroclinic instability benchmark test similar to the current high-resolution forecast (HRES) configuration using 350 nodes of ECMWF’s Cray XC40 supercomputer. Results are for the nonhydrostatic finite- volume (FV) method and the hydrostatic spectral transform (ST) method.

For added flexibility, it is also developing a new, nonhydrostatic dynamical core which uses the finite-volume method. In 2018, this ‘Finite-Volume Module’ (FVM) was shown to perform well compared to the current dynamical core in benchmark tests, and its computational cost was reduced substantially. Tests have shown that it holds great promise in terms of computational efficiency for global nonhydrostatic forecasts at high resolution run on future exascale high-performance computing facilities.

Benchmark tests
Benchmark tests One of the tests comparing results obtained using the FVM and the operational spectral transform (ST) method was for baroclinic instability at forecast day 10, showing pressure at the lowest full level for different grid spacings.