Review of "Assessment of Climate Biases in OpenIFS Version 43R3 across Model Horizontal Resolutions and Time Steps"

This study evaluates the sensitivity of an atmospheric general circulation model (OpenIFS cycle 43R3) to different combinations of time-step and horizontal resolution. The authors evaluate several aspects of the mean climate and variability in simulations with prescribed sea surface temperatures (SSTs) and sea-ice over the the period 1979-2019. The authors identify several regions where reducing model time-step from 60 minutes to 15 minutes can have positive impacts on systematic biases that are comparable to the impact of increasing resolution. The manuscript is clear and concise and the topic is within the scope of GMD. The results of this study will likely be of interest to the many users of the IFS model in weather and climate sciences. In particular, these results raise interesting questions about model development strategy when it is often necessary to work with cheaper and/or reduced resolution surrogates of more expensive operational/production configurations. I believe this manuscript can be suitable for publication in GMD but I have several comments that I think would improve and clarify the current manuscript.  

Main comments:

(1) In section 3.1 the authors focus on biases of near-surface fields, and how these are alleviated with reduced time-step and/or increased resolution. I encourage the authors to extend their analysis to other levels in the atmosphere (e.g. zonal means of temperature/wind against model/pressure levels). Given the changes in convection and vertical mixing identified later in the paper, I think it is possible that the authors will find similar sensitivities in the troposphere. I also think it is possible that changes at other levels may result in increased rather than reduced biases. This is fine, as the most interesting aspect is the sensitivity to time-step and how this varies with region (e.g. is it limited to near-surface/troposphere). I think it is unlikely that reducing time-step will improve biases in all regions/levels, so it would also be interesting to discuss and interpret any regions of increased bias (e.g. whether they might indicate a role for compensating errors). 

(2) The abstract concludes with the general statement that "reducing the time step in the OpenIFS model, one can alleviate some climate biases at a lower cost than by increasing the horizontal resolution." I would like the authors to to add some discussion of whether they expect their results to generalise to resolutions and/or time-steps not tested in this manuscript. For example, how far is the LR configuration from converging? Would reducing time-step in a much higher resolution model (e.g. 9km) bring similar benefits? Depending on these additions, the authors may wish to qualify the concluding line of the abstract. 

(3) What is the impact time-step/resolution on the representation of extremes? It is plausible that changes in time-step that improve the mean state have a limited impact on extremes that are more sensitive to horizontal resolution (e.g. orographic precipitation or tropical cyclones). As cited by the authors, the mean climate of the 25km and 50 km HighResMIP configurations of IFS are very similar. However, the differences in horizontal resolution are evident in the representation of extremes (examples below):  

https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2019JD032184

https://journals.ametsoc.org/view/journals/clim/33/7/jcli-d-19-0639.1.xml

Minor comments:

Introduction: This section would benefit from an overview of the "expected" impacts of reducing time-step in simple models in terms of truncation error and how this might not always hold true in a more complex system. For example, in simple finite difference models, solutions converge as grid-spacing and time-step are decreased due to reduced truncation errors. The choice of time-step and grid-spacing may also be constrained by stability criteria. However, this intuition does not always hold in complex models due to the coupling between many different elements. For instance, it is plausible that the unconditionally stable semi-implicit semi-Lagrangian scheme used in the IFS allows a user to configure the model with a long time step to reduce the cost. Later developments on top of this configuration introduce compensating errors in other aspects of the physics that reduce biases. Reducing the time step at a later stage may then leads to increased biases as the model configuration has been implicitly tuned for a particular combination of time-step and resolution. 

Lines 110-119. The authors state that their "detailed analysis of phase speed in such a manner is novel in literature". This may be the case, but I would like the authors to provide a more detailed summary of the method used to diagnose the amplitude and phase speed of extratropical Rossby waves (e.g. a bullet point list of the main processing steps). The current description is insufficient for reproduction of the analysis. In particular, it is not clear from the text how wave packets and associated phase speeds are diagnosed.

Line 169. Typo? Should be "Roberts et al. (2018)" as in intro?

Line 198. Is it correct to include Coriolis? Work done by Coriolis term should be zero since it acts perpendicular to motion of air parcels. 

Lines 226-228: It is possible that the lower precipitation RMSE in OIFS-LRA-1h is due to a "double penalty" effect that penalises higher resolution models, which have more structure in the precipitation fields. Is the precipitation in the LR-1h experiment notably smoother? Other metrics (e.g. fractions skill score) may provide a different ranking of models. More details on double-penalty effects and fraction skill score here:

https://www.ecmwf.int/en/about/media-centre/science-blog/2023/verifying-high-resolution-forecasts 

Lines 237-238 and figure 4: Why do the authors standardise the wave amplitude biases in figure 4 instead of showing the absolute values? This standardisation emphasises errors in regions the authors argue are unimportant, which complicates interpretation of the plots. Specifically, the  authors focus their analysis of Rossby waves on the "region where the wave amplitude is larger than 5 ms-1 is termed core region, which mostly covers the area that is occupied by the thick black contours in Fig. 4". However, biases are presented "relative to ERA5 (model – ERA5), normalized by the ERA5 detrended  variability expressed by the standard deviation", which highlights errors in the high-latitude high-wavenumber waves that are dismissed by the authors as "unimportant as these waves have a small amplitude and little effect on variability". 

Section 3.2. How do the authors interpret the impact on Rossby wave amplitude/phase speed biases? For example, is it related to the representation of tropospheric jets and associated wave guides and their biases?

Section 3.3 and figure 7. What is the sampling uncertainty in these composites and estimates of pattern correlation (e.g. estimated using bootstrap resampling of available dates)? Are the differences between configurations significant? 

Table 1. What is the HPC cost of the different configurations (e.g. core hours per model year)? Do they scale as expected from changes in time-step and number of grid points? 

Figure 2. What is the sampling uncertainty in these estimates of RMSE? Are the differences in RMSE between configurations significant?

Overall the manuscript is clear, concise, to the point and relevant. I appreciated the final recommendations on the recommended resolutions and timestep settings.

One thing I did miss is a discussion on the wiggles in the surface fields in the LRA and MRA configurations.  This is in fact a known issue that one of the co-authors complained about in the OpenIFS CONFLUENCE page https://confluence.ecmwf.int/pages/viewpage.action?pageId=188034913. In this page it is described that the Tco (octahedral reduced Gaussian grid ) grid is a poor choice for the low resolution configurations, instead I would encourage the use of compatible TL (reduced Gaussian grid) such as TL159 and TL399.

Nevertheless, the findings of this study are valid and I do not recommend repeating the exercise using the TL grids, but I would expect some discussion on this in the introduction, why the authors did not use the TL grids for the LRA and MRA configurations, despite the known issues? Also this should be pointed out in the results section.

Following are some other suggestions:

- p. 2 line 38 correct "have been widely used"

- p. 2 line 41 correct "which lead to"

- p. 2 line 60 correct "while increasing from"

- p. 2 line 63 correct "when increasing from"

- p. 2 line 66 correct "IFS models"

- p. 3 line 76 cite Döscher et al (2022) https://doi.org/10.5194/gmd-15-2973-2022 in which the timestep of various resolutions is given, instead of Van Noije et al (2021).

- p. 4 line 103, explain which CMIP6 forcings are used ? mole fraction of CO2? authors only specify that aerosol and ozone concentrations are from climatology

- p. 4 line 105, which scenario exactly ? SSP5-8.5 ?

- p. 4 line 111, correct "OIFS simulation datasets"

- p. 4 line 130, please confirm that the Pearson's correlation is computed

- p. 5 line 156, authors should justify in the intro. why they did not run OIFS-HRA at 1h timestep (i.e. for numerical stability) in order to evaluate if the improvements are due to resolution or timestep

- p. 5 line 162, correct "large difference"

- p. 6 line 194, black lines are not visible, authors should give their values

- p. 7 line 205, correct "lower stratosphere and troposphere"

- p. 7 line 215, is there any plausible explanation for this? predominance of ocean surface over continental ones? 

 - p. 7 line 220, The is an abrupt transition from analysis of winds to that of temperature, a proper sentence to indicate this change of focus is needed. 

 - p. 7 line 224, Fig. 2b indicates there is no notable improvement in RMSE  from the shortened timestep.

 - p. 11 line 345, correct "accounting for most variability"

 - p. 12 line 369, cite cite Doscher (et al 2022) instead or, or in addition to, Haarsma et al 2020