Ensemble run for ACCESS-CM2: Perturbations have no effect

Earth System Coupled Model
1

Hi all,

I’m interested in running an ensemble for ACCESS-CM2 - I’ll be perturbing the initial air temperatures by about 10^-13 randomly. There are a few people @ars599 @zoegillett27 @spencerwong who have successfully done this 3 years ago here, @sebmckenna did it 2 years ago here, and most recently @huazhen here. These thread are closed so I can’t comment there. I have the perturbIC.py file that is commonly used for this purpose from here.

My issue is that whilst I can use the file to perturb the warm restart files, I see no difference in conditions after 1 month, below is the daily mean air temperature at 1.5 m between 4 members.

Screenshot 2026-03-11 at 2.33.55 pm
Screenshot 2026-03-11 at 2.33.55 pm1752×714 118 KB

I would greatly appreciate any advice from those who’ve done this before or anyone with an idea!

Cheers,

Joel

2

Have you confirmed the perturbed files are actually different, e.g. by running diff FILEA FILEB or by plotting the perturbed field in each file?

3

I have no clue if this is important here and I have never done any of these before, but 1e-13 is close to 64-bit floating-point machine precision for temperatures in Kelvin with values ~270:

julia> eps(270.0)
5.684341886080802e-14 # <- machine precision for 64-big float value of 270.0

So you might have such a small perturbation that whatever chaotic bifurcation you were expecting has not had time to be fully realized:

julia> 270 + 1e-13
270.0000000000001

And if by any chance you are adding 32-bit floats, then your perturbation may be actually not perturbing anything:

julia> 270f0 + 1f-13
270.0f0

Maybe that’s your issue here? (Note I’m using Julia but the floating-point arithmetic issues are essentially the same across languages AFAIU.)

For something to be “felt” I would recommend taking the square root of eps, so in your case, for T in Kelvins, something on the order of 1e-7 for 64-bit floats and 0.005 for 32-bit floats.

julia> sqrt(eps(270.0))
2.384185791015625e-7

julia> sqrt(eps(270f0))
0.0055242716f0
2 Likes
4

another potential approach for debugging, building on the good suggestions that have come before: maybe hit it with a bigger hammer (say, 10**-2, or even 1 for clarity about whether anything changed) and check if that has an impact on the model’s evolution

2 Likes
5

Hi Joel,

In the end I didn’t need an ensemble, so I’m not too sure how that code works. I believe @huazhen has done this recently with ACCESS-CM2 though, so hopefully she can provide more help. I’m interested to know as well.

Cheers,
Zoe

6

Thanks all for your replies!

@Scott the diff command returns nothing, these files are identical. I even did a histogram of the air temperature differences and got the min, mean, max difference at 15 decimals and still no difference. Is it something to do with the script @tammasloughran?

@Benoit @jemmajeffree I tried a difference of 0.01 and I got the same results, very odd!

@zoegillett27 Thanks for your input, and yes if you have any tips on how you got it working @huazhen please let me know

Cheers,

Joel

7

@MartinDix
Am I right in remembering that perturbIC.py is your script? any suggestions or guidelines on how it should be used?

1 Like
8

The default size perturbation is large enough that significant differences appear very quickly.

Using a copy of /data/access/access-cm2/cmip6_restarts/bi889/restart/atm/bi889a.da09500101_00

module use ~access/modules
module load pythonlib/umfile_utils/access_cm2
perturbIC.py -s 0 bi889a.da09500101_00

mule-cumf shows theta differences with the expected default amplitude of 0.01.

mule-cumf bi889a.da09500101_00 /g/data/access/access-cm2/cmip6_restarts/bi889/restart/atm/bi889a.da09500101_00 
...
%%%%%%%%%%%%%%%%%%%%
* Field 171/15610  *
%%%%%%%%%%%%%%%%%%%%
Lookup compares, data DOES NOT compare
Compared 64/64 lookup values.
File_1 lookup info:
  t1(0950/01/01 00:00:00)  lblev(9999)/blev(0.0)  lbproc(0)
Data differences:
  Number of point differences  : 27264/27648
  Maximum absolute difference  : 0.009999598438156454
  RMS difference               : 0.0057510709134938765
  RMS diff as % of file_1 data : 0.0020506374334412678
  RMS diff as % of file_2 data : 0.0020506375586943374

Running with the perturbed file shows solver differences from the first step, e.g.
Original

********************************************
*    Linear solve for Helmholtz problem    *
* Outer Inner Iterations InitialError      *
*    1     1       16     0.100000E+01     *
*    1     2        7     0.108265E-01     *
********************************************

********************************************
*    Linear solve for Helmholtz problem    *
* Outer Inner Iterations InitialError      *
*    2     1        5     0.117254E+00     *
*    2     2        3     0.121000E-02     *
********************************************

Perturbed

********************************************
*    Linear solve for Helmholtz problem    *
* Outer Inner Iterations InitialError      *
*    1     1       16     0.100000E+01     *
*    1     2       11     0.108264E-01     *
********************************************

********************************************
*    Linear solve for Helmholtz problem    *
* Outer Inner Iterations InitialError      *
*    2     1        5     0.117142E+00     *
*    2     2        1     0.120943E-02     *
********************************************

Is it possible that you specified the amplitude as zero rather than the seed as zero? I’ve now modified the script to make this an error.

2 Likes
9

This worked well, thanks @MartinDix and @jemmajeffree for the help. It seems to be the script I was using was not working correctly, your script accessible with the modules

module use ~access/modules
module load pythonlib/umfile_utils/access_cm2

works as expected.

2 Likes