Implementing extreme indices output module

[Johannesfjeldsaa]

What is the feature/what would you like to discuss?

To support studies of climate extremes I suggest a module for calculating extreme indices during a model run. The benefit of this enhancement is that it allows us to store less data (monthly and yearly) as opposed to daily data. The enhancement should be portable across model versions as it only needs to make use of temperature and precipitation fields from the model.

I suggest that a first implementation should focus on implementing indices described by Zhang, X et al. (2011, https://wires.onlinelibrary.wiley.com/doi/10.1002/wcc.147) these are frequently used in studies and are designed to support comparison with observations. A list of the 27 core indices can be found here - https://etccdi.pacificclimate.org/list_27_indices.shtml.

Is there anyone in particular you want to be part of this conversation?

@mvertens @gold2718 @adagj

Will this change (regression test) answers?

I Don't Know

Will you be implementing this enhancement yourself?

Yes

[gold2718]

Hi @Johannesfjeldsaa, I have a few questions to help with implementation:

• Are the temperature and precipitation fields you need all already contained in the state? state and pbuf?
• Can all the calculations go into a separate module (e.g., extreme_indices_diag)?
• Can all these indices be updated at the same point in the model time step?

[Johannesfjeldsaa]

Hi, @gold2718. I'll partly answer your questions and deliberately omit addressing the first. I have been spending some time to look at how this has been done offline until now. The motivation being twofold, a) having a way of verifying an online module, and b) perhaps it could be part of the online module. Turns out @maritsandstad did quite some work for the official CMIP6 release of these indices and so shes been very helpful.

About the indices - in a nutshell there are two types of indices; "normal" indices which are calculated using data that is from within some time slice and thresholded indices which require quantile calculation from a reference period from your specific model (e.g. by using the WMO base periods from a historical run). There seems to me that there has been quite some research put into optimizing the quantile estimation so it is not the most straight forward thing. Which for me suggests that it might be a better plan to call on a dependency (but let's discuss pros and cons).

For the base set of variables from Zhang, X et al. (2011) there are only four model fields needed - tasmin, tasmax, tas and pr:

• tasmax-based indices (daily max)
  • Simple indices: txx, txn, su, id
      * Percentile/threshold indices: tx10p, tx90p, wsdi
• tasmin-based indices (daily min)
    * Simple indices: tnx, tnn, fd, tr
    * Percentile/threshold indices: tn10p, tn90p, csdi
  *tas-based indices (daily average)
    * gsl
• Combined-temperature indices (tas + tasmin/tasmax)
  • dtr
• Precipitation-based indices (pr)
    * Simple indices: rx1day, rx5day, sdii, r10mm,
      r20mm, rnnmm, cdd, cwd, prcptot
    * Percentile/threshold indices: r95ptot, r99ptot
  Some will be calculated monthly, some annually:
• mon - txx, txn, tx10p, tx90p, tn10p, tn90p, dtr, rx1day, rx5day, r10mm, r20mm, rnnmm, r95ptot, r99ptot
• yr - su, id, wsdi, fd, tr, csdi, gsl, sdii, cdd, cwd, prcptot

I think it is quite common to calculate all the monthly indices annually as well but perhaps not the most important.

Technical solution - For cmip6 Cicero used an R package, which unfortunetlely does not seem to be maintained anymore. I have gotten to run by making a singularity image, I am currently attempting to recreate some of the cmip6 results. Today I also was tipped by Marit that a newer CDO-based library has been created, I have not yet tested it, but the developers have apparently tested it against the R package with success. For online coupling I think the R package is a no-go. Do you think a CDO based package can be coupled without a significant computational overhead compared to pure Fortran?