An online meeting was held at 13.00-15.30 UTC with around 25 participants, covering five agenda items interspersed with discussion. (1) Robin Hogan explained what was required of CKDMIP participants in terms of data that should be provided and scenarios that should be covered – more technical details can be found in section 4 of the Technical Guide, as well as in the GMD paper (links on the CKDMIP home page). (2) Robin summarized the results so far, taking figures from the CKDMIP results page to show the improvement of RRTMGP compared to RRTMG, and the accuracy-efficiency trade-off with ecCKD. (3) James Manners summarized the workings of the SOCRATES CKD tool, and how the concept of equivalent extinction was used to treat spectral overlap. (4) Robert Pincus described his project to build a tool chain including user-friendly line-by-line and CKD software tools. (5) This was followed by a general discussion on the ways forward.

  • Robin Hogan: What is required of participants; results so far; questions for discussion (PPT, PDF)
  • James Manners: SOCRATES: correlated-k methods (PPT, PDF)
  • Robert Pincus: Towards community tools for LBL modeling and CKD generation (PDF)
  • Video of meeting (only available to CKDMIP participants with an FTP password - around 1.2 GB): MP4

Questions arising:

  1. What is the meaning of g-point fraction required from participants in the spectral definition file? The meaning of this matrix was given on slide 11 of Robin’s talk – see also section 4 of the Technical Guide.  The purpose of recording it (rather than simply which g point lies in which band) is for the second phase of CKDMIP in which the same data submissions are used to quantify errors in cloudy conditions.  We will investigate errors both when cloud optical properties are averaged across a band (as in almost all current models) and averaged per g point, potentially taking into account the spectral correlation of cloud and water-vapour absorption features. More details are given in section 4.4 of the GMD paper.
  2. What is the surface albedo and does it have to be spectrally flat? The value that will be used in the first phase of CKDMIP is 0.15, an approximate global-mean value (Wild et al. 2013) – see section 3.5 of the GMD paper. In the second phase of CKDMIP, the accuracy of CKD models in the presence of a spectrally varying surface albedo will be tested. This should be possible without further data submissions from participants.
  3. How are continua treated, and can they be removed from the line-by-line datasets (especially for water vapour)? Can we accommodate different models for the water vapour continuum? The line-by-line datasets for water vapour are available on the FTP site both with and without the continuum (using MT_CKD 3.2). The CO2 continuum and the collision-induced bands of N2 and O2 are included in the line-by-line datasets for those gases.  The evaluation is of course performed including the water vapour continuum.  Some CKD models are unable to use the same continuum model as the line-by-line reference dataset, so we should take that into account when studying the results. If it is a problem, then one possibility would be to perform a parallel evaluation for those models using reference calculations in which the water vapour continuum has not been included, although this is then not representative of the real world.
  4. My CKD tool was developed primarily for radiance modelling and remote sensing – how does that differ to broadband fluxes for weather and climate modelling? In an atmospheric model the broadband fluxes are needed not only for the surface and top-of-atmosphere energy budget, but for the radiative heating profile. Since the heating rate is essentially the vertical divergence of the flux divided by air density, higher in the atmosphere one can get large heating rate errors associated with only small errors in the flux profile.  Therefore, it will be important to take the heating-rate profile into account when training your CKD scheme in a way that would not be necessary in training a radiance model.
  5. I can train my radiance model differently for clear and cloudy conditions – which approach should I use in CKDMIP? This is a very interesting question. It is possible that when you are modelling not just the top-of-atmosphere radiation but also the heating-rate profile and the surface fluxes, whether or not clouds are present is less important in the training.  It is suggested that for the first (clear-sky) phase of CKDMIP you train your model for clear skies, and then when clouds are added in the second phase you investigate whether separate training could be beneficial. One reason clouds may affect the specification of a CKD model in the shortwave is at wavelengths where cloud absorption is very weak but gas absorption is not; here, multiple scattering by clouds increases the pathlength of light rays and therefore increases the probability of absorption by gases, potentially putting different constraints on the required accuracy of the gas optics model.
  6. How can CKDMIP accommodate radiation codes such as SOCRATES that intertwine the gas optics and radiative transfer components? For SOCRATES, the proposed approach is outlined in section 4.2 of the GMD paper. For models that mix gas optics and radiative transfer in a different way, please discuss with Robin.
  7. James described in detail the workings of the SOCRATES CKD tool, and Robin stressed the need for this information to be gathered for all CKD tools as it will likely be important when comparing the performance of the various tools when more data have been submitted.
  8. Are the scripts used for plotting the results available to participants? Yes, they are in the ckdmip-0.9.tar.gz software package on the CKDMIP home page (in Matlab). However, since submissions have been received, these scripts have been modified to cope with some of the slight differences between the various model output. A revised software package will be released soon.
  9. Should CKDMIP investigate the impact of uncertain spectroscopy, as suggested by one of the reviewers of the GMD paper? A brief discussion established that the team believe it should not: previous studies have already highlighted errors in older radiation codes based on outdated spectroscopy; CKDMIP should focus on the errors due to the way that CKD models are formulated given a common underlying spectroscopy. Robert pointed out that the improvement in accuracy moving from RRTMG to RRTMGP is primarily due to the improved spectroscopy, rather than the changed numbers of g point.
  10. Should we worry about non-LTE effects given that we are quantifying errors up to the mid-mesosphere where they become important? The team thought not, since non-LTE models suitable for atmospheric modelling (thinking of those by Victor Fomichev and Manuel Lopez-Puertas) tend to be formulated in terms of an additional source of heating or cooling in the upper atmosphere. These models can, in principle, be added to any CKD model for fluxes, so can safely be decoupled from the concerns of CKDMIP.
  11. How do we address the needs of those who require spectral surface fluxes for needs such as vegetation modelling, ocean biology and solar energy? In the second phase of CKDMIP, the existing submissions will be used to investigate how well spectral surface fluxes can be predicted making use of the information in the spectral-distribution files. At this point we will re-engage with participants to decide whether it is necessary to test additional band structures targeting surface spectral fluxes.
  12. How could the ability to generate CKD models targeting specific applications be made available to non-expert users? This currently requires either obtaining large volumes of data (hundreds of GB) or running LBL models (currently not very user-friendly).  Robert suggested that this could be provided by a remote computing resource.  Jérôme Riedi has since generously suggested that this could be ICARE in Lille.
  13. What is a realistic timeline for gathering contributions from the various model, writing a CKDMIP results paper and the next meeting? Most groups are unable to commit to a timescale due to the fact that this is an unfunded activity, the COVID situation and so on. Robert said that in six months it ought to be able to produce versions of RRTMGP with reduced numbers of g-points.  We decided to target the next meeting for the spring when hopefully more submissions would have been received. Bill Collins also suggested a meeting to discuss the details of how to optimally discretize k distributions, but this might perhaps be best timed after more submissions have been received so that we have more evidence of what approaches work best. Robin stressed that so long as he remained employed at ECMWF he would be happy to help people who wanted to use the CKDMIP dataset to evaluate their models.


Write a comment…