Decoupling the ACCESS ESM for long spin-up

We are planning to run some glacial experiments with the ACCESS ESM. Since the boundary conditions are quite different from pre-industrial/interglacial conditions, we will need to run the model for more than 2000 years.
We were thus wondering if there were some ways the spin-up could be “accelerated”? ie the atmospheric model is the slow and heavy part of the ACCESS-ESM but it is the ocean that takes a long time to equilibrate. Would there be a way to increase the length of the atm. time step or keep the atm. constant for a while to speed up the spin-up?


Maybe @rmholmes or @aekiss have some ideas? Could an ACCESS-OM2 model be used to spin-up the model? Could you save 10 years of atmospheric forcing and run a repeat-decadal forcing @mauricehuguenin?

To be clear, do you just want it to go faster in real-time, or do you want that and to use less compute time?

Thanks Aidan. While we can probably run the coupled model for 2000yrs, it feels like a waste of resources. We want to save time, compute and storage space. We could run the model for 100 years in the coupled state, then save the last 10 years and run a repeat-decadal forcing, and re-run it in a coupled state. The ACCESS ESM includes CICE4.1 and not CICE5.1, so that might cause difficulties.

1 Like

An issue here is that the offline ocean will not necessarily converge towards the same state as the coupled ocean since it doesn’t have all the feedback in. So it’s very hard to know if running part of the spin-up offline would help or hinder the process. You could have to re-run the same number of years again in the coupled model to get to an equilibrium in the coupled model. Even if you start from an ocean at equilibrium in the offline model.


Running the ocean-only model (ACCESS-OM2) with repeat decadal forcing stemming from the coupled model should work.

I assume one issue during glacial times would be the ocean circulation and bathymetry is quite different, for example there is no Bass Strait and the North Sea is land. So it could be tricky to prepare the ocean-only model first for these glacial conditions.

I have not used ACCESS-ESM yet. Would a glacial run with the ESM model just mean:

  • Force the model with glacial c(CO2)?
  • Remove the upper 13 grid cells of the ocean and set them as land? So that the ocean starts way below in -100 m compared to present times?
1 Like

Yeah, we encountered an issue with the CAFE coupled model where simply changing the atmospheric timestep drove the system off towards an entirely new equilibrium state.

1 Like

I think this would be technically feasible. The ESM and OM2 ocean code has been harmonised, although if the CICE versions are different that is an issue. There would be some work involved in setting up the forcing fields correctly from the atmospheric model (and there would be some approximations to do with time averaging).

I actually think this is a really good application of “repeat decade” forcing, although you might want to consider cycling over an even longer time period depending on what kind of decadal variability the ESM is capturing.

I agree that @clairecarouge’s point is an important one. I don’t have a feel for how big the numerics-related drift could be compared to the variations you’re expecting on glacial time-scales (associated with the atmospheric coupling)? We have a few 1-degree ACCESS-OM2 runs lying around that could be used to make that kind of comparison.

1 Like

Thank you all for your thoughts!
To put the model in a glacial state, we were thinking about doing it in steps (also because it is interesting scientifically):

  • changing the orbital parameters and GHG
  • including glacial ice-sheets (i.e. modifying high lat. vegetation, albedo, topography), & modifying the main changes in land/ocean mask related to the lower sea-level
  • modifying ocean salinity and nutrient concentrations (would work well offline).

The issue of the different equilibrium is important. This is why I was thinking that there would be several rounds of coupled/offline/coupled. I am not too worried about numerical drift, but more the high latitude processes related to changes in sea-ice and oceanic circulation. I suppose it depends on i) whether it is technically heavy to save the atm. fields and implement them in the ACCESS-OM2, and ii) whether the restarts between the two different configurations of CICE are compatible. I suppose it also depends on how much cheaper/faster is the 1deg ACCESM-OM2 compared to the ACCESS-ESM (I don’t have the nb for the ACCESS-OM2).

I am not sure whether such a design will go forward for these glacial expt, but I think it is worth keeping in mind, as it would be a very useful tool for BGC studies (the C cycle takes a long time to equilibrate in the deep ocean) under different boundary conditions.


Off-topic, but this is worrying.

1 Like

The approximate costs are:

ACCESS-OM2 1-degree: Model cost: 0.15kSU/year, Walltime: 95 years/day
ACCESS-ESM1.5: Model cost 1.1kSU/year, Walltime: 16 years/day

So you’re looking at about a factor of 10 less in cost, and a factor of 5 in speed.

It doesn’t seem like this would be too difficult, but I don’t know for sure. @Dhruv_Bhagtani’s project creating a flux-forced version of MOM5 using an ACCESS-OM2 control run is the closest thing I can think of that has been done in this space.

I’m not sure about the CICE question.

1 Like

The fluxes between the atmosphere and MOM all go via CICE, and (if I recall correctly) this is done differently for ACCESS-CM2/ESM and ACCESS-OM2.

You might not get the same results with flux forcing of the ocean-ice run it could well be different its been a while since we tried this back in CSIRO MK3 days with MOM2.2 ocean and as @Dougie said for the CAFE case it can produce different equilibrium. The difference between CICE4.1 and CICE5.1 codes might introduce some round off errors that could grow but would not be my major concern, its the deep ocean that you are trying to equilibrate.

Yes @aekiss, also the OASIS-MCT set up is different in ACCESS-ESM1.5 v ACCESS-CM2 v ACCESS-OM2 so there is a fair bit of engineering to handle. The ACCESS-OM2 is only doing basic fluxes in a simplified boundary layer whilst in the coupled model we have full fluxes coming from the UM implicit boundary layer. I don’t think this is viable @LaurieM its a nice idea, I think one of the Japanese groups did some speed ups for one of the paleo-type runs involving ice sheets, I will see if I can dig out the paper for you.

1 Like

One question you ask here @LaureM is about whether restarts between CICE4.1 and CICE5.1 are compatible, I have never used them this way, we do now use CICE5.1 restarts with some offline editing in CICE6. runs one of the improvements which were made both in ACCESS-M2 and separately in ACCESS-CM2 was in the calendar/clock handling in the restarts so that I expect will be a problem again. I will cross check the rest of the data inputs if you want me too.

1 Like

Sorry I wasn’t suggesting you try flux forcing. I was just saying that Dhruv’s project is the most recent project I know using forcing from one model run in ACCESS-OM2 to force another one. I agree that bulk forcing is better.

Coming in late to the discussion - we are running our coupled ice-sheet/ocean model with some accelerated forcing approaches, applied at the time evolving melt rate. It works well. We have a manuscript close to being submitted on this. Please get in touch and we can chat and perhaps I can share some details. It is lead from Norway so I’d need to ask to share. Ben

1 Like

Thanks all for the suggestions. An accelerated approach would be the best. Siobhan, if you have some ideas please let me know. Ben, I’ll contact you. Another option is to use the “fast” version of the ESM with lower atm. resolution. Is this operational? Happy to test if needed.

Hi all,
So, this is also something I am interested in. I previously tried doing an iterative coupling procedure to spin up a deep time paleo run. My method was to use a 50-year coupled simulation. Then use the final 10 years to force a 500-year ocean-only simulation. You can see the results I got in Figure 2 of this paper: CP - Climate sensitivity and meridional overturning circulation in the late Eocene using GFDL CM2.1

Although we saved time in one way, i.e. running the ocean-only model was faster, we also slowed down the spinup in terms of model years. This was because it tended to “lock” the SST boundary condition in a non-equilibrated state, which meant that the deep ocean (and in fact all layers of the ocean) equilibrated more slowly.

I would add that our initial conditions were not actually very good. I.e. they were a very long way from equilibrium, and we could have chosen a smarter starting point. It’s also possible that my method of saving out the boundary conditions could be improved… Nevertheless, my conclusion in that case was that iterative coupling did not save us time overall.