In 2020, ECMWF scientists continued to push for advances in numerical weather prediction (NWP) despite the obstacles set up by the COVID-19 pandemic. The impact of the pandemic on the aircraft data used in weather prediction is a case in point: as the availability of the data decreased, efforts were stepped up to find alternatives. The case is also a good example of working with other agencies, especially European ones, to address the shortfall in weather data.

The year was also marked by attempts to exploit machine learning in NWP; the finding that reduced North Atlantic sea-surface temperature biases led to significantly improved sub-seasonal to seasonal forecasts; progress towards reducing near-surface forecast biases as part of a three-year project; and, in the case of intense tropical cyclones, advances in the estimation of maximum wind speed, which previously tended to be severely underestimated even given the correct central pressure.

We also report on the final stages of refinements and verification of major changes to the moist physics in ECMWF forecasts. This work will be reflected in a future update of the Integrated Forecasting System (IFS). Finally, in a historic first in NWP, wind data from the new Aeolus satellite began to be assimilated into forecasts.

Coordinated response to loss of aircraft weather data

A coordinated response involving EUMETNET (a network of 31 European national meteorological services), national meteorological services and private companies helped to mitigate any adverse effects of the COVID-19-related loss of aircraft-based observations on weather forecasts.

In March and April 2020, there was a sharp drop in flights and thus in aircraft-based observations available to ECMWF. The continued availability of complete sets of satellite observations ensured that there was no severe impact from the loss of aircraft observations. Aircraft reports include temperature and wind and in some cases humidity and turbulence. They are used together with many other observations to help estimate the state of the Earth system at the start of forecasts.

Responses to the drop in observations included the use of previously untapped aircraft-based observations; an increase in the number of radiosonde launches from some locations; and the assimilation of additional satellite data. In an example of successful collaboration with the private sector, the companies FLYHT and Spire stepped in to provide additional aircraft-based observations and radio occultation satellite data, respectively.

A warmer stratosphere in forecasts
Numbers of global aircraft reports received at ECMWF per dayThere is some thinning and a small proportion of rejections in the aircraft reports normally received at ECMWF, so that the number assimilated (blue) is less than the number received (black). Most of these reports are received as part of the World Meteorological Organization’s Aircraft Meteorological Data Relay (AMDAR) programme. In addition, the European Meteorological Aircraft Derived Data Center (EMADDC) at the Dutch national meteorological service (KNMI) has been processing ‘Mode-S’ air traffic control signals to derive wind and temperature information. ECMWF started using the data (green) in July.

Machine learning at ECMWF

In 2020, ECMWF made a significant effort to see how applications of artificial intelligence and machine learning may improve numerical weather prediction at the Centre. There has recently been a surge in new methods which have the potential to bring significant changes to the work of operational weather prediction centres. Such methods include the use of deep neural networks, which can learn the behaviour of very complex non-linear systems from data.

Take the example of Earth system data assimilation. This combines the latest observations with a short-range forecast constrained by previous observations to obtain the best possible estimate of the current state of the Earth system. Patterns of bias in observations can be learnt through machine learning where this is hard to understand by purely physical reasoning.

The example shown here uses an offline neural network bias correction applied to the solar-dependent bias of the Special Sensor Microwave Imager/Sounder (SSMIS). This varies through the orbit and through the year in a complex pattern. The red arches near the bottom of the plot are the most difficult part. The difficulties of applying a neural network are a lack of training data and the fact that it would need to be continually re‑trained to keep up with the evolving bias. Other development areas for machine learning in data assimilation that are equally promising are targeting the problem of model error estimation and corrections.

Bias diagramsBias diagrams showing daily binned bias in Kelvin between SSMIS F-17 channel 11 observations and the simulations from the IFS, stratified by the angle around the orbit (left) and bias learnt by an evolving neural network that is then used as a bias prediction model for the next day (right).

Gulf Stream errors and sub-seasonal forecasts

ECMWF has established that reducing North Atlantic sea-surface temperature (SST) biases leads to significantly improved sub-seasonal to seasonal (S2S) forecasts of atmospheric circulation anomalies over Europe. The results from sensitivity experiments suggest that higher-resolution ocean models, with a grid spacing of less than 10 km rather than the 25 km currently used, would be beneficial. In particular, they would be able to better resolve the position of the Gulf Stream. Such models will be investigated at ECMWF during the next few years.

The results were achieved by evaluating the impact of SST errors on S2S forecasts using an SST-bias correction methodology. The applied bias correction effectively reduces SST biases in the North Atlantic region. The resulting southward shift of the Gulf Stream has consequences for atmospheric predictability that extend beyond the North Atlantic. For example, the image shows that reducing North Atlantic SST biases leads to significantly improved S2S forecasts of meridional wind at 200 hPa in Europe and elsewhere.

Impact of SST bias correction on forecast skill of meridional wind at 200 hPaChanges in the continuous ranked probability skill score (CRPSS; shading) and anomaly correlation (grey contour spacing of 0.1) shown for weekly mean anomalies derived from forecast days 26 to 32.

Progress towards reducing near-surface forecast biases

In 2020, ECMWF brought to a close a three-year project on ‘Understanding uncertainties in surface–atmosphere exchange’ (USURF). The aim was to investigate the systematic forecast biases in near-surface weather parameters, like temperature or winds, disentangle their sources and identify ways to reduce them in the future. These biases limit predictive skill from hours to seasons ahead.

These systematic errors often have complicated geographical patterns and temporal structure and result from an interplay between atmospheric and land surface processes. USURF defined a plan to address some of these issues. This will include an improved representation of snow and a revision of the land cover and vegetation maps, accompanied by a retuning of uncertain parameters in the surface–atmosphere coupling.

USURF has also provided further evidence that increases in near-surface forecast skill not only depend on an improved representation of physical processes. They also rely on the availability of comprehensive observations and in-depth studies using process-based diagnostics that can correctly attribute model error. Ongoing improvements to the diagnostic and verification tools used at ECMWF are therefore an important contribution towards further enhancements of forecast skill, alongside model developments.

Temperature biasesThe first two figures show spatial maps of March–April 2014 daily minimum and maximum temperature biases for the ECMWF operational system at a lead time of 2 days. The third panel shows the March–April 2014 mean diurnal cycle of 2-metre temperature at Sodankylä (Finland) in observations and in the ECMWF forecasting system with single-layer and multi-layer snow. It shows that the underestimation of the amplitude of the diurnal cycle of 2-metre temperature is due to a lack of sensitivity to changes in radiation, which is partly the result of using a single-layer snow model.

Moist physics upgrade

In 2020, the final stages of refinements and verification of major changes to the moist physics in ECMWF forecasts took place. One of the main drivers of this project was ECMWF’s strategic decision to move towards an ensemble forecast horizontal grid spacing of about 5 km, down from 18 km in 2020. The goal was thus to make it possible for the IFS to be run across a broader range of horizontal resolutions, including convection-permitting resolutions.

To this end, the complicated interactions between turbulence in the lowest part of the atmosphere, convective motions and the cloud physics should be described as simply, efficiently, accurately and scale-independently as possible. Longstanding systematic model errors in clouds, precipitation and radiation across all resolutions and forecast lead times should also be addressed.

For example, moist processes in the IFS are represented with physically based parametrizations for turbulent mixing, convection, subgrid clouds and microphysics. Each parametrization is called sequentially during every model time step. Individual developments over many years have led to some complications and inconsistencies in the way these schemes work together. These inconsistencies have been resolved in the moist physics upgrade.

IFS physics parametrization call sequenceThe two flow charts highlight the differences in saturation adjustment between moist physics in the IFS in 2020 and the moist physics upgrade planned for IFS Cycle 48r1.

Maximum wind speed in tropical cyclones

Prior to the summer of 2020, ECMWF forecasts generally severely underestimated maximum wind speed for intense tropical cyclones even given the correct central pressure. This was found to be closely linked to the modelling of momentum exchange at the ocean surface.

The momentum exchange is generally expressed in terms of the drag coefficient, which connects the magnitude of the surface stress to the square of the wind speed at a certain height above the surface. In the IFS, there is an active two-way coupling between the atmosphere and ocean waves, which results in an extra dependency of the drag coefficient on the sea state (waves).

Over the last decade, it has been suggested that the drag coefficient should tail off for strong winds. Recent wave model developments tried to address this issue, but not sufficiently for hurricane-force winds. A further reduction for such winds was tested at a horizontal grid spacing down to about 5 km to probe its limits.

For the operational high-resolution forecast with a grid spacing of about 9 km, it was found that, generally, the new system yields a much better maximum wind speed – minimum pressure relation. It was therefore implemented in June 2020 in IFS Cycle 47r1. However, as the figure indicates, forecasts continue to underestimate some of the most intense cases compared to observational estimates.

Relationship between maximum 10 m wind speed and corresponding minimum mean sea level pressureScatter plots for maximum 10 m wind speed and corresponding minimum mean sea level pressure for all 10-day forecasts at TCo1279 resolution (corresponding to a grid spacing of about 9 km) from 00 UTC for the period 25 August 2019 to 1 January 2020 (coloured squares; the dashed line indicates mean central pressure for a given wind speed), and corresponding reported values (6-hourly Best Track data: purple circles; the dotted line indicates mean central pressure for a given wind speed), for 20 tropical cyclones, showing results for IFS Cycle 46r1 (left) and IFS Cycle 47r1 (right).

Assimilating Aeolus

ECMWF started to assimilate wind observations from the European Space Agency’s (ESA) pioneering Doppler wind lidar mission Aeolus on 9 January 2020, 16 months after the first wind profiles became available. It was the first numerical weather prediction centre to go operational with Aeolus, followed by the German national meteorological service (DWD), Météo-France and the UK Met Office later in 2020. ECMWF was able to benefit from the data quickly because of its strong involvement with the mission since its conception in the 1980s and its selection as an ESA Earth Explorer Mission in 1999.

Aeolus is carrying the world’s first functioning space-based Doppler wind lidar and Europe’s first space-based lidar. The aim of the mission is to demonstrate this new technology in space for the benefit of weather forecasting and to improve the understanding of atmospheric dynamics, especially in the tropics.

Thanks to careful investigations, in which ECMWF model winds were used as a reference, the Aeolus wind biases were found to be strongly correlated with the instrument’s main telescope temperature, which varies slightly with the Earth’s top of atmosphere radiation. The telescope temperatures are fortunately available in real time. A bias correction scheme using these temperatures as predictors went operational on 20 April 2020.

Observing system experiments have shown that Aeolus winds provided a statistically significant positive impact, especially in the tropics and in the polar regions.

Aeolus wind profilesThe plot shows the Aeolus Level-2B Rayleigh-clear and Mie-cloudy horizontal line-of-sight wind speed profiles on 14 February 2020 over the North Atlantic. The blue area highlights the location of a jet streak with wind speeds up to about 100 m/s. Around this period, some records were broken for the speed of transatlantic flights due to the strength of the jet stream.

Observing system experiments have shown that Aeolus winds provided a statistically significant positive impact.