I think it’s a combination but I’m guessing to some extent. The different perturbations have the starting position tweaked slightly but I had always assumed that the ‘tweaking’ was in the estimates, not the actual collected data, because why would you alter something that was a ‘known’ rather than an estimate? Logically if you start changing too many of the actual datapoints it means you’re not projecting forwards from reality at all.

You might find this interesting, a description of the horizontal and vertical resolution in the ECM model: https://confluence.ecmwf.int/display/FUG/Section+2.1.1.1+Grid+point+Resolution

You’ll understand my point about the huge number of datapoints: at the Equator there are over 5,000 grid points, reducing to just 20 near the Poles, and 137 vertical grid points.  Then multiply that by all the variables needed at each grid point.

Conversely I always thought that it was the starting values that were slightly perturbed. If I recall unless all the starting points are known with unachievable accuracy, the feedback loops cause the models to become chaotic after about 6-7 days on average. Also the perturbations and resulting ensembles can be used as a measure of how sensitive this particular synoptic situation is to the starting values and hence its reliability going forward

That was my point, but why vary some known data hardly at all and others by more?  For example, the SLP values are pretty constant, as are the 500hPa temperatures.  I accept that there must be a program that tweaks some variables by specified amounts since it can’t be done manually, but someone has decided which variables to tweak and by how much.

I think this is the point: we don’t have accurate data for every data point in the grid, so I have assumed that the starting point must contain some estimates and that the ensemble suite tweaks the estimates, but clearly it’s not like that, perhaps because it’s just not possible.

On the more general point, yes, the ensembles are testing for the stability of the evolution.  Looking at the output you can see changes starting to become noticeable very quickly, like a couple of days.

It's very interesting. I wonder though, if the models, as Retron points out, all were ran from the same starting point, would they end up with the exact same solution each time? Surely that flies in the face of the idea of 'fluid motion'.? In real world situations, I don't think that happens naturally. Spill a jar of say, 50 marbles, from the same spot, with the same force, multiple times on a wide hall floor, and each of those marbles will not take the exact same direction (or path) or stall in the same spot after each 'run'. That would be just impossible. 

All the ensemble runs would end in the same position. 

The same model run with the same data would produce the same results - it's just a complicated set of "if this, then that" calculations, and it's why they're called deterministic models. Even the "AI" ones are working on a similar basis, it's just that instead of purely "if this bit is 10C, then that bit will be 8C", it's "if this bit is high pressure and history shows it becomes low pressure, then it becomes low pressure".

The modification of starting values is absolutely required if you want ensembles that actually differ between runs.

EDIT: And I see Brian posted much the same as I was posting! 😁

The Met Office provide a much better worded explanation:

Ensemble forecast systems are designed so that each member should be equally likely. The initial differences between the ensemble members are small, and consistent with uncertainties in the observations. But when we look several days ahead the forecasts can be quite different.

....

An ensemble samples the uncertainty of the forecast, assuming that the forecast model is perfect. If the starting conditions were known accurately, and the model was perfect, then an accurate forecast could, in theory, be produced. However, because it is not possible to know the exact starting conditions we use our best guess and generate a forecast which can sometimes be inaccurate (shown in red in the schematic above). By sampling the uncertainty in the starting conditions, and running several ensemble members forward with the model (shown in blue in the schematic above), we produce an estimate of the forecast uncertainty and an indication of which weather events may occur.

See:

https://www.metoffice.gov.uk/research/weather/ensemble-forecasting 

https://www.metoffice.gov.uk/research/weather/ensemble-forecasting/what-is-an-ensemble-forecast 

The key points are that even tiny starting errors are quickly magnified after a few days and it is NOT possible (except in a theoretical sense) to know the current state of the atmosphere. Increasing the amount of data only improves the approximation.

I told you so!  Bet against the Merkelslug at your peril

...and of course we see this in action every winter, when one of the models "goes off on one" - perhaps having a ridge push a low in just the right track to bring a foot of snow in the south, only for subsequent runs to remove that low entirely. It's essentially a variant of the butterfly effect, one small modelling error snowballs such that everything after a surprisingly short time is, in effect, junk.

It's also worth mentioning that despite all the massive increases in computing power, grid resolution improvements and indeed greater data input (the aircraft data in ECM is around 10 times what it was a few years back, for example)... we haven't actually seen much improvement in verification in the past 10 years. It *is* better, but only by a tiny bit - a much, much smaller increase than over the preceding 10 years.

If you were able to exactly replicate the coordinates of all the marbles in the jar and the exact parameters such as height, force, angle etc etc of the emptying then they would but even tiny differences can have big outcomes

Your highlighted words are exactly what I have said many times; no matter how much higher the model resolution becomes it’s impossible to know the exact starting position so the inherent errors will always be there and get amplified as the model runs forward.

Nonetheless, the model accuracy has clearly - based on the stats - improved over time.

It will be interesting to see whether the machine learning / AI stuff has much impact on medium to long range forecast accuracy. IIRC, they are quite confident that it will reduce running costs because AI can be used to "short cut" some of the computations. However, there seems less confidence in how much accuracy will be improved. I suspect that most of the gains will be in short range forecasting.

https://www.wetterzentrale.de/maps/GFSOPUK18_198_48.png 

I think this highlights the point perfectly that aircraft data accounts for a small slice of a much wider base of data attributed to each run. 

I reckon there's a limit as to how far conventional forecasting can go... as the smaller the grid gets, the more likely a tiny feature will be "hallucinated" (to use the AI terminology) and end up scuppering the whole run. That said, the "AI" models - essentially glorified pattern-matching - did show a slight skill advantage last year, even out to day 10. However, when you're looking at an improvement from say 0.86 to 0.87 correlation over the course of a year... you really aren't going to notice it in practice.

Here's a chart showing the model stats over the past 25 or so years - the gradual decline in improvement rate can clearly be seen. (CMC=GEM, FNMOC=NOGAPS)

https://ukwct.org.uk/weather/stats.jpg 

In the immediate future it couldn’t be more different with 14°C or maybe even 15°C possible in places today.

Not yet.

Not just saying it as it shows mild but I really don't rate the UKMO in the 144/168h range. It's often very wrong.  

The golden rule IMO is cross model consistency. So yes, the chance of a cold spell of sorts is probably increasing, but now is not the time to start counting down.