New science upgrades to the Met Office's high resolution models over the UK
November 2016 - Significant science upgrades have been made this month to the UK high resolution Numerical Weather Prediction (NWP) models run at the Met Office.
These latest upgrades have taken advantage of the new supercomputer, which had its first phase implemented earlier this year.
Parallel testing of new science changes
Upgrades to configurations of the Unified Model (UM) run at the Met Office are tested in parallel to the operational version of our weather forecasting models in what are known as parallel suites. The latest upgrades, which were tested in parallel suite 38 (PS38), were made operational this month.
PS38 has seen a series of improvements to the UK High Resolution model (UKV) and our UK High Resolution UK ensemble model (MOGREPS-UK). The major changes to these UK high resolution models are described in this article.
The Met Office has for several years been producing forecasts from UM simulations over the UK at resolutions that are fine enough to resolve the local dynamics of many smaller-scale processes that contribute dramatically to the local weather (the UKV has a 1.5 km grid size over the UK).
A key example is the vertical motion generated in deep convective storms and hence the name “convection-permitting” is often applied to models at these resolutions.
These convection-permitting UK models have shown much improved rainfall characteristics, as well as benefits for local low-level winds and temperatures compared to the significantly coarser resolution global model.
It is well recognised, however, that a 1.5km grid is still too coarse to resolve accurately most of the scales relevant even in deep convection. Furthermore, turbulent motions occur across a range of scales. The model must therefore be able to represent (parametrize) the effects of the smaller scales whilst resolving the largest ones where it can.
Extended forecast lead times
The two UK models (UKV and MOGREPS-UK) have benefited from an extension of forecast lead times from T+36 hours to T+54 hours for all runs, and T+120 hours for two of the UKV model runs per day.
MOGREPS-UK is similar to the UKV model although the calculations are performed at a coarser horizontal resolution. The inner, fixed resolution domain of MOGREPS-UK is at 2.2km resolution compared to 1.5km in the UKV model. MOGREPS-UK uses 12 runs of the model to provide sampling of uncertainty in the forecasts. This allows us to better assess the risks of high impact weather.
The outer, variable resolution domain of the high resolution models has been extended in all directions, something that has been made possible by the increased computational power of our new supercomputer.
The new domain is represented by the full extent of the image on the right. The old extent of the domain is shown by the solid black line. The darker green shaded box shows the fixed resolution inner domain, which has not changed.
This larger domain will allow a better spin up of convection over the UK, especially showers moving in from surrounding waters and continental Europe. The initial benefit of this domain increase is to improve the forecast in the fixed resolution inner domain, over the UK.
We will be verifying forecast data from the outer domain so that the extended region can be used for forecasting and in any downstream products in future.
The challenge in developing high resolution forecasts
The "grey-zone" challenge
Accurately maintaining the separation between small and large scale convection as the dominate scale of motion crosses from subgrid to resolved, or vice versa (a regime now known as the “grey zone”) is a key challenge in developing more accurate high resolution forecasts.
One negative characteristic observed in the early days of the UKV was a tendency to break up solid stratocumulus cloud sheets too readily and this was at least partly attributed to grey zone issues. These issues can be described as a combination of errors in the representation of poorly resolved (near grid-scale) turbulent eddies and partial double-counting of the turbulent fluxes.
“Blended” boundary-layer parametrization
The “blended” boundary-layer parametrization scheme (Boutle et al 2014a) was introduced to the UK high resolution models in February 2015 as an initial step towards improving the separation between small and large scale convection.
In this parametrization, the turbulent mixing is defined to be a weighted combination of the one-dimensional planetary boundary layer (1D PBL) scheme and a three-dimensional Smagorinsky sub-grid turbulence parametrization, with a weighting that depends on the ratio of eddy size to grid length (i.e., on how well the model should be able to resolve the largest scales of the turbulence).
Cloud microphysics scheme
At the same time, a similarly motivated scale-aware cloud microphysics scheme with an explicit representation of sub-grid variability (Boutle et al 2014b), was introduced that primarily improved forecasts of light rain. The combined effect was significant improvement to boundary layer cloud cover, especially stratocumulus, and consequently to screen-level temperatures.
Improved forecasting of showers
In PS38, we have revised how subgrid turbulent mixing into developing cumulus clouds is represented, which allows the UK models to initiate small scale showers more readily.
This means PS38 models are less likely to miss showers in more marginal situations. It also improves the tendency of the current model to be slow to develop showers in the morning.
PS38 has also seen a change to the stochastic forcing that helps improve the spin up of showers, particularly as they come in through the boundaries. These improvements to convection have not degraded the representation of stratocumulus, something which has been difficult to achieve up until now.
PS38 example of a convective rainfall outbreak
The images above show results from a convective rainfall outbreak over the UK, which occurred at 1400 UTC on 27th August 2015.
The top left image shows the visible satellite image and the top right image shows the radar derived surface precipitation rate.
The two images on the bottom row show forecasts from a 0300 UTC analysis from PS37 (left) and PS38 (right).
This PS38 example shows more shower activity over Ireland and SW England, as well as more broken cloud over central England that agrees better with the observations.
Summary of UK model upgrades