Grid feedback and convective parameterisations - why they matter!

So if you've been following ManUniCast over the first week or so of its operation, you've probably seen something like this (as we discussed in the previous blog entry).  All plots that follow are from a 54-hour precipitation forecast, effective at 0000 UTC on 9 November, 2013, with the model being initialised at 1800 UTC on 6 November 2013.

As we've discussed before, this is caused by employing two-way grid feedback.  The precipitation that is formed by the convective parameterisation on the large-scale model grid does not translate over into the small-scale grid due to this interface between the two.  In fact, if we concentrate solely on the inner grid, we get a very different pattern of precipitation then what is shown on the large scale.

What this shows is that there is light precipitation falling on Northern Ireland, with the maximum precipitation falling in both the English Channel and northern France as well as areas of western Scotland.  You can see the impact of the coastline on the precipitation fields, with rain occurring in northern Ireland.  

If we remove this grid feedback from the process, then the model topography and small-scale interactions that occur on the inner grid do not impact the larger grid, as shown below.

Without grid feedback, the large-scale model precipitation is now continuous from across the ocean until it hits the Irish and Scottish coast, where the land surface starts impacting how the cumulus parameterisation operates.

 If we look at the inner domain precipitation, we see that it hasn't changed much compared to when grid feedback is turned on, but there are slight differences in the highest amount of precipitation in Scotland and the English Channel.

So, by turning off grid feedback, we can explore some of the sensitivity in the model to various parameters that we have set as options for our model operation (a full list of these parameterisations is listed on the ManUniCast website).  By turning off grid feedback, we can make a better looking large-scale domain precipitation figure, but we may lose some increased accuracy in coastal regions.  This is because the inner domain can resolve the interactions on the coast while the outer domain cannot- and this can cause the difference in prediction of precipitation in Northern Ireland in the small scale domain but not in the large scale domain in these examples.  

So, how can we find out if there are changes we can make to the model setup that would resolve in us getting a more accurate precipitation field on both domains?   There are a lot of different options that we can select for the convective/cumulus parameterisation, and they all impact the model forecast in different ways.  The WRF-ARW team have a great description of the different options present from a lecture that they've given at WRF workshops in the past.  

For the images above, the 20 km large-scale domain uses the Kain-Fritsch cumulus parameterisation scheme.  The Kain-Fritsch scheme is most likely the most used and the most-well known scheme, with the three main papers that discuss the parameterisation scheme being cited almost 3500 times in the 23 years since the first paper (which if you really want to read, is available here).  However, the Kain-Fritsch scheme is not always going to be the most valid, and other parameterisation schemes may create "better looking" weather patterns- or certainly different looking weather patterns.  

A lot of other modelling simulations use different cumulus schemes, such as the Tiedtke scheme that is also present for use in WRF.  The Tiedtke scheme is used by the ECMWF for their model simulations, and treats the problem of small-scale convective processes on larger grid spacing somewhat differently then the Kain-Fritsch scheme.  If we keep model grid feedback on, but use the Tiedtke scheme instead of Kain-Fritsch, we produce this graph for our large-scale total precipitation field:  

The Tiedtke cumulus parameterisation scheme with grid feedback turned on still causes the sharp boundary from our coarse to fine scale domain.  However, the precipitation of the Atlantic Ocean looks more realistic and much more homogenous then using the Kain-Fritsch scene above.  

If we look at our inner domain over the UK, use of the Tiedtke scheme on the large scale domain produces quite a different picture then the Kain-Fritsch scheme does, even though the inner domain does not use a cumulus parameterisation!

 The Tiedtke scheme still has the problem of cutting off the precipitation at grid boundaries on the large-scale domain.  However, there are increased amounts of coastal precipitation that may be more realistic - which is what was missing in the large-scale precipitation fields produced using the Kain-Fritsch cumulus parameterisation and turning off grid feedback.

So, with that being said, if we turn off grid feedback and use the Tiedtke parameterisation, we may have the "best looking" option for production of large-scale precipitation, and these final two plots do just that.

With grid feedback turned off and using the Tiedtke parameterisation, we still have a continuous precipitation field that shows many of the same features as the previous plots, but with more coastal precipitation in Scotland and Ireland then in the Kain-Fritsch scheme with grid feedback turned off.  

If we look at the inner grid, we can see that not much has changed from the simulation we ran that used grid feedback and the Tiedtke scheme.

Which is exactly what we want to happen.

So, this post can show how making 2 small changes in grid feedback and cumulus parameterisation impact the model precipitation field.  ManUniCast began operations using the Kain-Fritsch cumulus parameterisation scheme with grid feedback at the beginning, but has now switched to using the Tiedtke scheme without grid feedback- and the differences in these precipitation patterns and the physical reasoning behind it is why.