«Medium-Range Weather Prediction Austin Woods Medium-Range Weather Prediction The European Approach The story of the European Centre for Medium-Range ...»
Concerns with the envelope orography were being felt. Short-range forecast errors had increased, if only slightly. The envelope behaved differently in differing weather regimes, especially in summer. There were differences between the levels at which the weather observations were reported, and the model heights. Masao Kanamitsu, a scientist visiting from Japan, joined Jarraud and Simmons in reassessing the impact of the envelope orography at various resolutions, in preparation for implementation of a higher-resolution model in May 1985. While concluding that the envelope was, on the whole, satisfactory, it was becoming clear that a more sophisticated approach to modelling the effects of mountains was required.
In May 1986, three additional levels were introduced in the model stratosphere, giving 19 levels in total, with the top level now at 30 km.
Research by Martin Miller, Tim Palmer and others, in parallel with work in other major forecasting centres, was showing the importance of considering “gravity wave drag”. Waves are generated as the air flows across large mountain ranges like the Rockies. The high-level wind was slowed by the waves breaking at high levels, thus extracting momentum from the flow.
Incorporating the effect in the model in 1986 reduced the systematic overprediction of the speed of the westerlies, and improved modelling of the ultra-long waves around the Hemisphere.
Prediction of surface temperatures and other weather elements when snow was lying on the model surface was being investigated. A canopy of vegetation can mask snow on the ground. Research was under way into a scheme to describe the interaction between snow and canopy.
For the first decade or so, the development of the analysis was to a large extent independent of model development; this followed common practice The Medium-Range Model 107 at all major forecasting centres. For example, the spectral model code was separate to that of the analysis, leading to some duplication of work and the risk of inconsistency between the codes. A researcher in the analysis section, Jan van Maanen, was devoting virtually all his time to analysis-related aspects of keeping the spectral model going in operations.
We have referred elsewhere to the substantial co-operation between scientists at the Centre and those in the Member States. We will see now that a fortuitous accident of timing and personal contact led to many years of collobaration in model development between the Centre and France, with a level of co-operation almost unique in meteorology. The development of a new forecasting system began in 1987. It lead to integration in a single consistent Fortran code of the world’s biggest set of forecasting models, analysis code and other numerical tools, the so-called “Integrated Forecast System” or IFS.
Development began from the advances in computer hardware. Computer memories were becoming bigger. Simmons recalled how the prospect of keeping the model in the computer’s central memory, as an “in-core” model, was an attractive possibility that could soon be realised. Coding for the repeated in-out transfers and the associated problems could be avoided. A re-coding of the ECMWF model was required.
We saw in the previous Chapter that Philippe Courtier of Météo France in Toulouse had been investigating variational data assimilation. Courtier, now an ECMWF staff member, and Simmons were discussing recent research, in Courtier’s case the need to code the “adjoint” of the Centre’s operational model, which would be required for this kind of assimilation technique, over coffee in the Centre’s restaurant. They agreed that the paths of their research were very close. They decided jointly that a new global spectral model should be coded, together with its “tangent linear” version.
This was a necessary step to coding the adjoint. The model and its equations are at the core of the data assimilation algorithm in variational assimilation;
the assimilation is in fact built around the model. The model code had to be integrated into the assimilation code if the Centre was to be able to use the promising very powerful technique of variational assimilation.
Discussions between Simmons, Courtier and Geleyn referred to in the previous Chapter evolved naturally into an informal and fruitful collaboration. Formally, there was no “management” agreement or decision, either on the part of Lennart Bengtsson or of the management of Météo France.
The collaboration evolved naturally over the years, with communication scientist-to-scientist, programmer-to-programmer, group-to-group. It was a stunningly successful example of co-operation between scientists of many
108 Chapter 9nationalities and backgrounds, male and female, some experienced, some recent graduates, working (most of the time!) in harmony to improve the two different but complementary systems. The exchange of scientists between France and the Centre was a key factor.
In the following years, Météo France in Toulouse developed its “Action de Recherche Petite Echelle Grande Echelle” or ARPEGE system in parallel with the Centre’s development of the IFS. In the literature, the terms IFS/ARPEGE or ARPEGE/IFS are used. Scientists at the Centre in Reading and those in Toulouse developed and maintained in common a single major code. Both the scientific and technical aspects needed for research experiments and operational forecasts were kept consistent. Mats Hamrud had with Courtier written the first lines of the IFS code. The new system integrated most of the applications, from analysis to initialisation to modelling, into this single code. At the time of writing, Hamrud continued to manage the truly vast code of the entire IFS system; both Simmons and Miller remarked on his invaluable knowledge and expertise.
Model development began quickly in Paris until 1991, thereafter in Toulouse, and at the Centre. The Centre adapted its existing model physics;
Toulouse developed a new physics package. The first operational ARPEGE model was operating in Toulouse by September 1992, two years ahead of the Centre’s operational IFS. The stretched grid became operational in Toulouse in October 1995. The code was robust; it survived several changes of computer systems in Toulouse and Reading.
Soon after his arrival back at the Centre in 1990 as Head of the Operations Department, Michel Jarraud noted that there was a need for more systematic, perhaps even formal, interaction between the scientists working in the Research Department and those in the Meteorological Division of the Operations Department, to communicate better the monitoring results from the Meteorological Operations Room.
He instituted regular so-called “OD/RD meetings”, held four times a year, at which useful scientific and technical information was exchanged and actions followed up. Meteorological Operations staff presented results of operational monitoring of data and verification scores of the forecasts, and research staff presented their diagnoses of the assimilation and model.
Questions and issues arising from these presentations were then aired. The meetings were restricted to Centre staff, allowing opinions to be freely expressed and discussed. Some “Special Topics” were included, for example performance of tropical cyclone predictions, or the behaviour on the model in polar regions. Over the years these meetings proved themselves to be surprisingly useful. Major issues were identified and addressed, some The Medium-Range Model 109 not even having been recognised by the scientist whose presentation had raised the issues!
September 1991 saw the next major change in model resolution at the Centre. Following a programme of research that had stretched over five years, a T213L31 model, able to define 213 waves around the globe, and with 31 levels, was introduced.
This new higher-resolution system depended crucially on a major improvement to the numerical scheme of the model: the “semi-Lagrangian” scheme. With this, the time step can be made relatively long, without falling foul of the mathematical criterion leading to computational instability: the numerical collapse of the forecast. Hal Ritchie, a visiting scientist from the Meteorological Research Branch of Environment Canada worked on this scheme. Ritchie, with Mariano Hortal, Clive Temperton and Adrian Simmons, implemented a significant new dynamical core for the model, providing the basis for model development in the future. Tests on the original version of the new model showed that a three-minute time-step was required; increasing this to four minutes led to computational instability.
Use of the “semi-Lagrangian” scheme allowed a 20-minute step, which together with the reduced Gaussian grid enabled completion of a ten-day forecast in four hours rather than 24!
At this resolution, waves in the atmosphere with a wavelength of 190 km and above could be followed. There were now 4,154,868 points in the model at which wind, temperature and humidity were predicted, almost ten times as many as in the 1979 model. The grid became a “reduced” Gaussian grid.
The number of grid points along a latitude circle decreased towards the poles, so the grid point spacing was about 60 km on the whole globe. In
• Three surface and sub-surface levels took into account vegetation cover, gravitational drainage, capillarity exchange, surface and subsurface runoff, deep-layer soil temperature and moisture.
• High, medium, low and convective clouds were all modelled, as were stratiform and convective precipitation.
• Carbon dioxide was fixed at 345 parts per million by volume.
• Aerosols, ozone, solar angle, diffusion, ground & sea roughness, ground and sea-surface temperature, ground humidity, snow-fall, snow-cover & snow melt, radiation (incoming short-wave and out-going long-wave), friction (at surface and in free atmosphere), gravity wave drag, evaporation, sensible and latent heat flux were all included.
In 1992, model low-level cloud was changed to reduce errors in prediction of near-surface temperatures near the Baltic and North Sea coasts, and
110 Chapter 9reduce over-prediction of low-level clouds over the Mediterranean in summer, and over snowfields in winter.
Improvements to the cloud and radiation parametrization were made in
1993. Experiments on soil surface, including hydrology, and very low level (boundary layer) processes lead to many improvements to the operational model in August 1993. However, further experiments on envelope orography gave an unexpected result - its continued use improved the forecasts significantly. This was despite the fact that the mismatch between model level and the height of observations over hills and mountains meant a significant loss of low-level data, there was over-prediction of convective rain and snow, and heavy rain related to orographic lifting was incorrectly widened and intensified. It appeared that the benefits of envelope orography could still be realised by further work on planetary boundary layer.
The benefits of gravity wave drag in the model were confirmed by the same set of experiments. Advantage was taken of field experiments over the Pyrenees to compare the model drag with that in the real world. The model was found to underestimate the mountain torque; flow separation in the lee of the Pyrenees had been underestimated. Development of a new representation of orography began - but the envelope orography had served the Centre well for more than 10 years, even though most of the staff of the Research Department had never been completely comfortable with its use in the model.
In March 1994, after the major rewrite of the forecast model, the Integrated Forecast System became the operational system. The research team, in collaboration with the GMD National Research Center for Information Technology in Bonn and Météo France, also developed a portable version of the IFS code to be used as a “benchmark” code for testing and comparing parallel distributed-memory or Massively Parallel Processing computers.
In April 1995 the envelope orography was — at last — replaced by a smoothed mean orography together with a scheme to parameterise the effects of sub-grid-scale orography. Model mountains were now correctly blocking low-level wind flow, and drag on the wind due to flow separation caused by this sub-grid-scale orography was better modelled - these were novel features. In addition, a new and unique scheme developed by Michael Tiedtke to model the main processes associated with clouds consistently was introduced into the model. Both cloud fraction, and the ice and water content of clouds, were being predicted as model variables.
In the years following, research continued at an accelerating rate on improving the numerical and physical aspects of the model, including much The Medium-Range Model 111 more efficient use of the two-time-level version of the semi-Lagrangian scheme. In September 1996, the operational suite was implemented on the new Fujitsu VPP700.
Research started in the mid-1990s to improve the stratospheric resolution and to raise the top level of the model. The higher levels were needed to assimilate new kinds of data collected by satellite from the mesosphere, 50 to 80 km above the surface, well above the existing model top of about 30 km. Agathe Untch, newly-arrived at the Centre, quickly found herself fully occupied with the task.
In April 1998, the model resolution was increased from T213 to T319 on a linear grid; now waves down to 125 km were predicted. By March 1999, Untch had succeeded in the difficult task of raising the model ceiling;
according to Hortal, this was “a remarkable achievement”. There were now 50 levels, with the highest close to 65 km. Stratospheric ozone data could now be assimilated and modelled, and - another triumph for 4D-Var - wind information could now be gleaned from the ozone measurements in the stratosphere.
In October 1999, ten more levels were added close to the ground. The grid-point total had now reached 8,300,760, with in addition 553,384 in surface and sub-surface layers.
In June 2000, a new scheme for parameterizing surface fluxes and processes was implemented. A grid-box was separated into fractions, called “tiles”, with six over land: bare soil or ground, high or low vegetation, high vegetation with snow under, snow on low vegetation, and two over oceans, one for water, the other for ice. Separate calculations were made for each tile.