Set of initialised ACCESS-ESM1-5 PI ensembles to explore predictability

[jemmajeffree]

*If you would use this data, let me know what variables you are interested in*

Experiment title : ACCESS-ESM1-5 PI initialised ensembles

Summary :

Initialise 10 ensemble members, from 10 distinct initial conditions, and run these for 10 years (1000 years total). All run under pre-industrial forcing.

Scientific motivation:

To explore the predictability of internal variability in the ACCESS-ESM1-5 model.

I often use data from the Decadal Climate Prediction Project (DCPP, part of CMIP). These are hindcast ensembles initialised yearly from observations, and run with historical forcing. It’s nearly impossible to separate predictability from initial conditions vs that from climate change, so the global warming trend provides apparent ENSO forecast skill at 10 years lead time, which I suspect is unphysical. Model drift, due to the initialisation from observations, is also a confounding factor.

Therefore, I want to run a set of initialised perfect-model forecasts (i.e. model forecasts of the model itself) of a pre-industrial control run, to understand the predictability of ENSO in ACCESS-ESM1.5, when ignoring the effects of climate change and model drift.

I’d hope this experiment could also be utilised to understand the predictability of other aspects of the earth system (e.g. other climate modes, sea ice extent, drought, etc)

Experiment Name : init_pi
People : Jemma Jeffree
Configuration: ACCESS-ESM1-5 PI control:
Initial conditions: 10 distinct initial conditions, 5+ years apart, taken from a pre-industrial control simulation. Open to requests for particular initialisation years.
Run plan: 10 ensemble members, using micro-scale perturbations, for 10 years each
Simulation details:
Total KSUs required : 1,000
Total storage required : Depends on the variables saved. Personally, I only want monthly SST, 20 degree isotherm, surface winds (~10GB total), but it’d be great to keep other variables that the community is interested in. I’d be keen to keep only monthly variables, given this is an experiment for interannual predictability, and ideally these would nearly all be 2d.
[As of 20/8/25, everything except monthly 3d BGC is sitting on gdata/lg87 (5.5TB total)]
Storage lifetime : 1-2 years
Long term data plan : delete most of it and put the rest onto zenodo or tape
Outputs:
Restarts:
Related articles:

Analysis:

Conclusion:

[jemmajeffree]

@dkhutch @tiloz
Following on from the ESMWG meeting and putting spoken conversations into writing, may I have ~100kSU to finish off this ensemble? I ended up running nearly 20 initialisations with 10 ensemble members, but only out to 6 years, with various sources of compute at the end of last quarter

[dkhutch (David Hutchinson)]

Hi Jemma,
This plan sounds good to me. Please go ahead with your plan to use ~100 KSU.

[tiloz (Tilo Ziehn)]

Yes, agree. Please go ahead.

[jemmajeffree]

This data is now occupying 5.5 TB on lg87, under /g/data/lg87/experiment/init_pi. Naming scheme is r[perturbation_seed]i[restart_name_from_pi_control][bunch_of_garbage_that_payu_generated_but_I’m_too_scared_to_delete]. So r1irestart079-r1irestart079-2607b707 is the first ensemble member of the batch which branch off restart079, and is initialised from almost identical conditions to r2irestart079-r2irestart079-b570f055.

I’m aiming to get this into an intake catalogue soon, and then I’ll put a couple of example analysis scripts somewhere

I’m deleting the 3d monthly ocean bgc in 24h unless someone explicitly asks me not to, because it takes up a disproportionate amount of space and I can’t see it being useful.

As time goes by and/or data storage gets tight, I’ll delete more of this, with the eventual aim to have the handful of variables that go into publications on zenodo and the rest deleted.

[Aidan (Aidan Heerdegen, ACCESS-NRI Release Team Lead)]

@jemmajeffree

bunch_of_garbage_that_payu_generated_but_I’m_too_scared_to_delete

That is a shortened experiment_uuid

Usage — payu 1.1.7 documentation

It uniquely identifies the experiment. It is used to uniquely identify the data in an intake catalogue. It prevents namespace clashes in the archive directory, and is a good thing ™.

[dougiesquire (Dougie Squire, ACCESS-NRI Ocean/Sea-Ice Modelling Team)]

@jemmajeffree

I often use data from the Decadal Climate Prediction Project (DCPP, part of CMIP). These are hindcast ensembles initialised yearly from observations, and run with historical forcing. It’s nearly impossible to separate predictability from initial conditions vs that from climate change, so the global warming trend provides apparent ENSO forecast skill at 10 years lead time, which I suspect is unphysical.

Probably something you’re across already, and I’m a bit late to the party here, but a common approach is to use a corresponding set of uninitialised historical simulations to quantify the skill added by initialisation. There’s a whole world of skill metrics out there for doing this. You might find the docs (and associated code) I put together on the CSIRO CAFE decadal forecasts useful, in particular the section with the skill assessment.

[jemmajeffree]

Thanks Dougie. I have been using the CESM1 initialised and uninitialised experiments together, but I think I had some reason why I couldn’t just use the uninitialised experiment as a baseline for what I was trying to do. Maybe drift from initialising from observations?