You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 18 Next »

Horizontal resolution upgrade

The horizontal resolution upgrade is being developed with a trade-off in mind between resolution and computational costs. A number of options of how to produce the most effective combination of horizontal resolutions between 4D-Var, EDA, HRES and ENS have been tested to establish computing costs and to derive possible efficiency gains.

A viable choice was found to employ the so-called cubic Gaussian grid (TC) instead of the current linear Gaussian grid (TL) where the shortest wave is represented by four rather than two grid points. By keeping the spectral truncation the same, more resolution is added in grid-point space to more accurately represent diabatic forcings and advection, which is then controlled through truncation in spectral space. In the current operational configuration the erroneous build-up of energy at the shortest scales is filtered by a lower-than-nominal resolution of the orography, strong horizontal diffusion and a de-aliasing filter. In the future this filtering will be able to be much reduced. The TC option also substantially improves mass conservation.

In order to reduce the computational cost further, a grid modification has been investigated, the cubic, spectral octahedral grid (TCO). The octahedral grid applies a new rule for computing the number of points per latitude circle. It is based on a new mesh that also allows for future implementations of a hybrid spectral – grid point model. The computational cost is reduced by about 25% compared to the cubic grid as fewer grid point calculations are needed and this new grid will also be implemented in the coming high resolution model cycle.

In summary, the anticipated upgrade will have a horizontal resolution that translates to about 9 km in the outer loop of 4D-Var as well as the high-resolution forecast and to about 16 km for the ensemble up to day 10.

Some background

Spectral representation of the IFS

The IFS uses a spectral transform method to solve numerically the equations governing the spatial and temporal evolution of the atmosphere.  The idea is to fit a discrete representation of a field on a grid by a continuous function.  This is achieved by expressing the function as a truncated series of spherical harmonics:

\[ \begin{equation*} A(\lambda,\mu,\eta,t) = \sum_{l=0}^{\mbox{T}} \sum_{|m|\leq l}^{\mbox{T}} \psi_{lm}(\eta,t) Y_{lm}(\lambda,\mu) \end{equation*} \]

where μ = sinθ with λ  the longitude and θ the latitude of the grid point, T is the spectral truncation number and Ylm are the spherical harmonic functions.

The spectral coefficients ψlm are computed from the discrete values known at each point of a Gaussian grid on the sphere by

  • a Fast Fourier Transform in the zonal direction followed by
  • a slow/fast Legendre transform in the meridional direction.

At each time step in the IFS:

  • derivatives, semi-implicit correction and horizontal diffusion are computed in spectral space;
  • explicit dynamics, semi-Lagrangian advection and physical parametrizations are computed in grid point space.

The representation in grid point space is on the Gaussian grid.  The grid point resolution is determined by the spectral truncation number, T.

Gaussian grids

Naming convention

The Gaussian grids are defined by the quadrature points used to facilitate the accurate numerical computation of the integrals involved in the Fourier and Legendre transforms. The grids are labelled by N where N is the number of latitude lines between the pole and the equator.  For example, for the N640 Gaussian grid, there are 640 lines of latitude between the pole and the equator giving 1280 latitude lines in total.

The grid points in latitude, θk, are given by the zeros of the Legendre polynomial or order N:  PN0(μk=sinθk) = 0.  A consequence of this is that a Gaussian grid has:

  • latitude lines which are not equally spaced;
  • no latitude points at the poles;
  • no line of latitude at the equator.

Regular (or full) Gaussian grid

A regular Gaussian grid has the following characteristics:

  • there are 4N longitude points along each latitude circle;
  • the longitudinal resolution in latitude-longitude is given by 90o/N;
  • the points get closer together (i.e. more crowded) as the latitude increases towards the poles;
  • the total number of grid points is 8N2.

Reduced (or quasi-regular) Gaussian grid

A reduced Gaussian grid:

  • has the same number of latitude lines (4N) as the corresponding regular Gaussian grid;
  • has a decreasing number of longitude points towards the poles;
  • has a quasi-regular grid spacing in distance at each latitude;

  • provides a uniform CFL (Courant–Friedrichs–Lewy) condition.

Up to and including IFS cycle 41r1, ECMWF has used a standard reduced Gaussian grid.  This has 4N longitude points at the latitude nearest to the equator, with the number of longitude points reducing in blocks as the latitudes approach the poles.

With the horizontal resolution increase at IFS cycle 42r1, ECMWF introduces a slightly different form of the reduced Gaussian grid which is referred to as the octahedral reduced Gaussian grid or, more simply, the octahedral grid.

N48 regular Gaussian grid

N48 standard reduced Gaussian grid

Relationship between spectral truncation and grid point resolution - linear, quadratic and cubic grids

The relationship between the spectral resolution, governed by the truncation number T, and the grid resolution depends on the number of points on the grid at the equator, 4N, at which the shortest wavelength field is sampled:

linear:  each wavelength is sampled by 2 grid points → 4N = 2(TL + 1)

quadratic: each wavelength is sampled by 3 grid points → 4N = 3(TQ + 1)

cubic: each wavelength is sampled by 4 grid points → 4N = 4(TC + 1)

Until the implementation of IFS cycle 18r5 on 1 April 1998, the IFS used a quadratic grid. The introduction of the two-time level semi-Lagrangian numerical scheme at IFS cycle 18r5 made possible the use of a Linear Gaussian Grid reflected by the TL notation.  The linear grid has been used since then, up to and including IFS cycle 41r1. At IFS cycle 42r1, the cubic grid will be used and will be indicated by the TC.

By changing the number of points at which the shortest wavelength is sampled, the effective grid point resolution can be increased while keeping the spectral truncation number constant.

Grid point computations, which are often non-linear, benefit from the higher resolution of the grid produced with cubic sampling.  Such a grid has no aliasing, less numerical diffusion and more realistic surface fields.

Increasing the horizontal resolution

The relationship between the spectral truncation, T, the sampling (linear, quadratic or cubic) and the grid point resolution, N, allows for three possible approaches for increasing the horizontal resolution:

  1. increase the number of wave numbers in the spectral representation but keep the same Gaussian grid by reducing the sampling frequency (i.e., increase T with TQ →TL );
  2. increase the number of wave numbers in the spectral representation and the Gaussian grid resolution but keep the sampling frequency constant (i.e. increase T and increase N).
  3. keep the number of wave numbers in the spectral representation constant and increase the Gaussian grid resolution by increasing the sampling frequency (i.e., TL →TC, and increase N).

The horizontal resolution upgrade at IFS cycle 42r1 is achieved by:

  • increasing the Gaussian grid resolution by keeping the spectral truncation constant but using cubic sampling;
  • introducing a new form of the reduced Gaussian grid - the octahedral grid.

This new grid will be referred to simply as a cubic octahedral grid.  This will be indicated with the notation TCO.

What is the octahedral grid ?

The octahedral grid has been inspired by the Collignon Projection of the sphere onto an octahedron. It is a form of reduced Gaussian grid with the same number of latitude lines located at the same latitude values as those of a standard Gaussian grid but with the number of longitude points at each latitude circle computed according to the formula:

\[ \begin{eqnarray*} \mbox{N}_{lat}(lat_N) & = & 20 \\ \mbox{N}_{lat}(lat_i) & = & \mbox{N}_{lat}(lat_{i+1}) + 4, \mbox{ for } i=\mbox{N} - 1,\ldots,1 \end{eqnarray*} \]

In other words, there are 20 longitude points at the latitude circle closest to the poles with the number of points increasing continuously by 4 at each latitude towards the equator. This is in contrast to the standard reduced Gaussian grid where there are 'jumps' between blocks of latitudes with a constant number of longitude points (a restriction imposed by the Fast Fourier Transform routines being used prior to IFS cycle 41r2).

As a consequence, the zonal resolution of the octahedral grid varies more with latitude than for the standard reduced Gaussian grid. This can be seen the figures to the right.

Note in particular that the octahedral grid has 4N + 16 longitude points at the latitude circle closest to the equator whereas the standard reduced Gaussian grid has 4N longitude points at the latitude circle closest to the equator.

There are also fewer total grid points in the octahedral grid compared to the standard reduced Gaussian grid.For example, the N1280 octahedral grid has about 20% fewer grid points than the equivalent N1280 linear grid. Generally, there are 4xNx(N+8) grid points in the octahedral grid or resolution N.

The octahedral grid has been shown to improve the calculation of local derivatives in grid point space.

Comparison of the resolution for the reduced Gaussian grids with latitude

Comparison of the resolution for the reduced Gaussian grids with latitude.  The full red and yellow curves show the resolution for the standard reduced Gaussian grids at N1024 and N1280.  Note that the resolution remains more or less constant at 10km and 8km, respectively, as the latitude varies.  The corresponding curves for the N1024 and N1280 octahedral grids are shown by the dashed and full blue lines, respectively.  Note that the resolution for the N1280 octahedral grid varies from about 8 km at the equator, increasing to about 10 km at 70oN and 70oS before decreasing again towards the poles. The resolution of the N640 standard reduced Gaussian grid used at IFS cycle 41r1 is at about 16 km. Also shown are the regular Gaussian grids at N1024 (black dashed line) and N1280 (black full line).

Comparison of the zonal variation in resolution between standard reduced and octahedral grids

Comparison of the zonal variation in resolution for the N1280 standard reduced Gaussian grid (left) with the corresponding octahedral grid (right) on the surface of the sphere

 

 

Frequently asked questions

Will the change to the cubic octahedral grid affect me if I use regular lat-lon data ?

No, users of regular lat-lon data will be unaffected by the change of model grid.  They will, however, benefit from the increase of the model horizontal resolution.

Will ECMWF make data available on the standard regular and reduced Gaussian grids ?

Yes, users will still be able to request data, both in dissemination and MARS, of the standard regular and reduced Gaussian grids. Note, however, that this will be interpolated from the values on the original model grid.See also I make computations involving flux parameters or accumulated fields (for example, to de-accumulate precipitation) and am advised to work on the model grid: which grid should I use ?

I use point data (e.g., for meteograms, vertical profiles, etc) - what do I need to do ?

Users of point data should not that the coordinates of the nearest grid point will have changed.  Users should take particular care for coastal points for which the nearest grid point may have changed from a land point to a sea point or vise versa.

Is the new land-sea mask and orography for the cubic octahedral grid available ?

Yes, the new land-sea masks and orography fields for HRES at TCO1279 (N1280), ENS Leg A at TCO639 (N640) and ENS Leg B TCO319 (N320) can be downloaded from ...

Do I need to upgrade the version of GRIB API I use in order to decode data on the cubic octahedral grid ?

Version 1.12.3 of grib_api can decoded fields on the cubic octahedral grid correctly.  At grib_api 1.14.0, a new computed key isOctahedral is introduced which allows users to query the grid type. For the cubic octahedral grid, isOctahedral=1; otherwise, isOctahedral=0.

Can GRIBEX decode data on the cubic octahedral grid ?

To be checked. But you should not use GRIBEX.

I make computations involving flux parameters or accumulated fields (for example, to de-accumulate precipitation) and am advised to work on the model grid:  which grid should I use ?

For performing computations with accumulated fields, users are advised to request data on the cubic octahedral grid.

Is there any change to the vertical resolution at IFS cycle 42r1 ?

No, only the horizontal resolution is increased.  The vertical resolution remains at L137 for HRES and L91 for ENS.

What will happen if I retrieve IFS cycle 42r1 data from MARS using grid=av ?

Users retrieving data from MARS with the keyword, grid=av ("archived value") will retrieve data on the model grid.  For data from IFS cycle 42r1 this will be the cubic octahedral grid.

What will happen if I retrieve IFS cycle 42r1 data from MARS using grid=1280 ?

This behaviour is unchanged. By default, users retrieving data from MARS with the keyword, grid=1280 will retrieve data on the regular N=1280 Gaussian grid. 

Will ERA-Interim fields also use the cubic octahedral grid ?

No, the horizontal resolution upgrade applies only to ECMWF HRES and ENS operational forecasts, including the monthly extension.

Will the ECMWF System 4 Seasonal Forecasts (SEAS) also use the cubic octahedral grid ?

No, the horizontal resolution upgrade applies only to ECMWF HRES and ENS operational forecasts, including the monthly extension.

 

  • No labels