Possibly this has been discussed many times in the past but definitely not in this forum…
The question is how does one initialize a 1/4-degree simulations given 1-degree temperature/salinity. What sort of interpolations we need to do to ensure that the 1/4-degree initial conditions are smooth (smooth enough?) and also that the total amount of tracers is conserved? (Or should we even enforce column-wise tracer conservation?)
(We are trying to set up global configurations with Oceananigans.jl)
I don’t know of an interpolation method that provides smoothness, conservation and monotonicity. AFAIK you can have any two of these, but not all three.
ESMF offers 1st order conservative interpolation but this is piecewise-constant, i.e. no smoother than the original data.
ESMF also has 2nd order conservative interpolation which is smoother but can generate new extrema (including extrema of the opposite sign to the original data, which is why we stopped using it in ACCESS-OM2 to interpolate surface forcing fields so we wouldn’t get negative rainfall etc).
A point for discussion: In the context of global, bulk-forced ocean sea-ice models, is this worth worrying about in much detail? These models have to be spun-up to equilibriate things like ocean heat content anyway, and so I wouldn’t have thought how the initial condition is treated to this detail has a big impact (providing you’re not kicking off instability that make the model unstable)? From memory, the ACCESS-OM2 1-degree and 1/4-degree versions drift in different directions (in terms of global ocean heat content), so starting from a spun-up 1-degree solution (as opposed to initial conditions from observations) gets you further from where you want to be.
@aekiss That’s useful information about the regridding options in ESMF! Especially useful (and makes sense) that high-order conservative schemes can violate monotonicity and that this can be undesirable.
I think it’s a great idea to try initializing from ACCESS-OM2 output… !
More generally than finding a particular initial condition that works well, I also want to get up to speed / implement good strategies for initialization, since I think we’ll continue to have the need to find initial conditions (today’s problem is 1/4 degree, but we’ll be running at many other resolutions and probably initializing with various products). As Navid says we also would like to be able to initialize from ECCO (to any resolution), and probably plenty of other products in the future.
@rmholmes we’ve been working on numerics that are as minimally dissipative as possible — which seems to work well in realistic eddying states, but perhaps is more susceptible to instabilities associated with initialization. So the short answer is yes, one of the main issues is that it appears a poor choice of initial conditions can cause our simulation to go unstable. I agree with you we shouldn’t be over-concerned with fine details; the main issue is finding stable initial conditions that can be spun up in a reasonable amount of time, without having to resort to very small time-steps to resolve unphysical stuff.
@glwagner ok that makes sense. If you’re worried about instabilities right at the beginning of the run then I can see why you need to worry about this.
Re:
I think you would want to think carefully about doing this. As mentioned above, the drift in ACCESS-OM2 could be very different to what Oceananigans wants to give you, meaning that you could easily increase rather than reduce the time it takes to spin-up. And you’re starting further from observations.
However, one strategy could be to start from a very short ACCESS-OM2 run, long enough that it has smoothed out any initial shocks (and developed a velocity field that is consistent with the density field - I know we start from zero velocity in ACCESS-OM2) but not long enough to have any significant drift?
on cold start initialisation and instabilities - parallel with coastal modelling
Global model are often initialised from climatologies, but indeed nothing prevents us from starting from another products: Initializing with ECCO state estimate or reanalysis products (like BRAN or GLORYS)
Using these products is what is used in regional and coastal modelling
Initialisation and stability have been debated for a while in coastal modelling, especially to initiate from a physically consistent state and limit drift, spin up, unstable model…etc.
I have not looked into this for a long time, but here is an old paper which is probably quite on topic
While looking for the ref, I also came across this other paper (from the same people /team), but looking into extrapolation (when LR model does not have canyons, but the HR does). Might not be on topic for a 1/4 deg model but might be of interest. https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2006GL026099
on vertical interpolation:
Trevor and Paul wrote the following paper showing how we can create or destruct water mass through interpolation from one vertical grid to another.
I always thought I should look into this for initialisation of my regional models, but have not find the time yet…
If you are looking into good initialisation strategy, this is definitely worth a read!