Following on from our chat in the AUS2200 meeting the other day, I’ve made this fun (though scientifically useless) visualisation tracking the maximum vertical wind across the AUS2200 domain for both the 30-second timestep and 75-second timestep model runs here: multiplot.mp4. Apologies in advance for my lack of design sense. Anyway, the green dot/line represents the 75-second model and the red dot/line is the 30-second. Your guess was spot on, it appears as if the ‘instability’ forms right over PNG on the northern boundary of the region. There is no green dot/line after 2:16:15 as the 75-second model crashes at that point.
For a bit of background, we (CLEX) encountered some model instability when running the AUS2200 2.2km regional UM model during the northern NSW floods of Feb-Mar 2022. We didn’t have a lot of time to get the data ready, so we shortened the timestep and hoped (which worked, fortunately). Now that the pressing deadline of the original WACI hackathon has passed, we have some time to drill down and figure out what went wrong. I’m no UM expert, so I bought it up at the last AUS2200 meeting hoping to learn what to look for and what can be done about model instability. “Vertical winds over PNG” was the first hint, and that does seem to be the case. The next hint was that the orography may need to be smoothed out, so I’m hoping to learn what that would entail.
In the regional ancillary suite, make sure rgNN_rsNN_orog_srtm=True is set (use shuttle orography)
Then a further option will appear rgNN_rsNN_srtm_smooth. Select “121” for this, click on the parameter name to get the help text on how to increase the smoothing.
Once it’s smoothed enough to start running you could combine the dataset with the original orography if you like, using the smoothed orography only over the maritime continent and the original orography over australia - you’d need to use mule for this.
It looks like the ancil suite is already producing a bunch of orographies with different smoothing factors applied, so that part is done. I don’t know how its picked which one to use though, it has linked qrparm.orog.mn to qrparm.orog.srtm.121x1, but SMOOTH_FACTORS="5 10 20 35 50 75 100". Does that mean it has done a single sweep of 121 smoothing on all points, or just over 1500m? In order to test this out, do I need to run the ancil suite again, or can I just link qrparm.orog.nm to a different file and run the forecast suite?
I think I need a crash course in how orography works in the UM because I’m struggling to determine exactly what orography the model is loading in. It must be doing some kind of transform on the input data, because even simple things like “the highest and lowest points” don’t match between the ancil, the model dump created by reconfiguration. Most notably, the lowest point on the input dump is apparently around -105m(?). There is also a 200m discrepancy between the highest points in the model dump vs the ancil file. Any help would be appreciated.
Edit: Here is a picture, this is the northern-most part of the region. The actual location of the -105m point (the arrow points to it) is 5 grid points (11km) in from the northern boundary. The eastern-most edge of the image is about 1100km from the eastern boundary of the domain. Maybe an interpolation/regridding gone wrong?
More edit: It took me far too long to realise that the instability occurs more or less at that point as well. And as if to confirm it, I ran the model with 100 sweeps of 121 smoothing and it still crashed.