A short in-person meeting was held at the 2022 International Radiation Symposium to discuss progress and priorities for ways forward in our collaboration.

• Robin Hogan's talk on results of the CKDMIP intercomparison so far is here
• As of July 2022, the entire CKDMIP dataset including the Evaluation-2 spectra (previously withheld) are available from https://dissemination.ecmwf.int/ecpds/home/ckdmip/
• Results from MSTRN, PSLACKD, RRTMGP and PyKdis have also recently been added to the CKDMIP results pages

In terms of priorities for then next steps, we agreed that the results so far raised important questions about how best to generate gas-optics schemes, and that we should address these before, say, starting to look at cloudy conditions. Specifically:

• We should list the steps involved in creating CKD models, and study in detail the impact of different methodologies in each of these steps.  The How CKD Tools Work page has a brief summary of how our tools work, and could form a basis for this. It would be useful if models who have not yet submitted information for this page could do so.
• What is the best way to spectrally average absorption coefficients to k terms, and what is the impact of different averaging techniques for the various tools?
  • A quick study of the impact on longwave ecCKD models is shown here, implying that the best method amongst those tried is to average the transmission across an equivalent atmospheric layer over which pressure changes by a factor of 3 (although ecCKD's subsequent optimization step was found to be crucial).
  • A recent paper by Webb, Solovjov and André suggests that the optimal approach is the geometric mean of the absorption coefficients at the bounds of the spectral interval in g space, i.e. k_opt = sqrt(k₁ x k₂) = exp((ln k₁ + ln k₂)/2).  This is similar to logarithmic averaging, but only considers the bounds of the spectral interval.
• Is there a fundamental limit to the accuracy of correlated-k models even with large numbers of k terms?  This is implied by Slide 11 or Robin's ecCKD talk, and in some other talks, and is likely due to the inability to represent the imperfect rank correlation of spectra at different heights. It would be useful if CKDMIP participants could submit models with quite large numbers of k terms (ideally all using the same "narrow" band structure) so that we can see if this fundamental limit is reached in all models, and whether the limiting accuracy is similar between models.  What can we do about it? SOCRATES has a method to treat imperfect correlation in height - can this be shared and applied by other tools?
• Other crucial aspects are the way spectra are sorted (either separately per pressure/temperature/gas or with a single sorting per gas) and the treatment of gas overlap.  Can we test different approaches using the same tool?
• It would be great to have some more submissions from different tools, ideally using at least the "narrow" band structure.

Other items raised at the meeting were:

• Can we extend the CKDMIP line-by-line database to more profiles?  We currently only have 100 base profiles (nominally 50 for training and 50 for evaluation) of temperature, water vapour and ozone, and this is not enough to fill parameter space in the training. Extending to more profiles is possible but somewhat laborious.  There might be a case to use the RFMIP method to select the profiles according to some criteria such as global coverage/representativity, or covering extreme values of each variable.
• What is the impact of vertical resolution?  The LBL datasets typically use 10 points per decade of pressure - is this enough?
• Can we extend the dataset spectrally to the far UV, e.g. the 100-200 nm range?  We would only need O2, O3 and Rayleigh, and this would cover a region important for photolysis.