In this paper the authors couple a global emulator (MAGICC) with a regional ESM emulator (MESMER) to create simulated fields of regional surface temperature anomalies from pre-industrial under different future climate scenarios. The combined emulator can be run either to simulate the response to particular CMIP6 models, or in an "observationally constrained" manner in which the global mean response ensemble from MAGICC has been pre-selected to be consistent with observed temperature and global mean uptake. This framework will be very useful to provide regional projections of climate change to future scenarios. Most importantly, and possibly a point that the authors undersell, is that the regional climate projections from any scenario can be produced, not just those that were run by ESMs (e.g. the SSPs). This would allow specific scientific questions to be answered such as the regional responses to different pre-defined levels of global mean warming, forcing, or total carbon budgets, and regional climate change commitments. The cheap computational framework, allowing millions of ensemble members to be run (memory is suggested as a limiting factor, but not processing speed), allows for statistically robust projections of climate risk in different regions and can aid adaptation planning. The modular, open-source framework appears to be flexible and extensible: while only annual mean regional temperature anomalies are included at the moment, the opportuntity to include precipitation and other climate variables (provided robust predictors are found), or higher simulated time resolution, should be possible without redesigning the whole framework.

This paper does something important and useful, and does it well, so my comments are limited to being quite minor.

lines 19-20: I assume IPCC 2021 also says this - though given the increased roles of emulators in the Sixth Assessment, perhaps ESMs are not our primary tools any more!

figure 2: It took me a few minutes to fully decipher what the right side was showing here. Faint horizontal lines that separate the four rows would make it clearer that the spatial plots correspond to the same rows as the time series plots. In the first map in row (b) I think is taken that the pink, purple and orange lines are very similar and the forced GSAT plot from the blue line is a bit warmer. I would also take from this that in (b) and (c), dividing any of the four scenarios by any of the others would give a constant pattern scaling ratio of local to global GSAT that is the same everywhere. 

lines 139-140: more for my interest, but what regression coefficient did you determine for the stratospheric aerosol optical depth? Also, if you wanted to be totally CMIP6-consistent you could use the CMIP6 SAOD time series (ftp://iacftp.ethz.ch/pub_read/luo/CMIP6/CMIP_1850_2014_extinction_550nm_strat_only_v3.nc) rather than NASA-GISS one.

lines 193-194: of course, more predictors will reduce error. I was satisfied that the possibility of overfitting is addressed later on (lines 216-218). Still, I think the flow of the paper is a bit disjointed: we have a standard MESMER config (Beusch et al. 2020), then we introduce a relationship with more predictors and shows it performs slightly better (fig 3 and 4), but actually we don't use it for the main MESMER-MAGICC coupling in section 3.3 (explained in lines 249-255) that is the main subject of the paper. Perhaps a reordering to put the "additional predictors" section later on could be explored.

line 247: I'm in general agreement with the authors' opinion on the implausibility of SSP5-8.5 but I think Hausfather & Peters 2020 gets abused a bit - particularly as we're running everything concentration driven here and high-end climate responses to a SSP3-7.0 (or even SSP4-6.0) like emissions pathway can't be ruled out.

line 268: MESMER (typo)

lines 285-286: "however, most CMIP6 ESMs perform in an observationally-consistent manner in most regions (Beusch et al. 2020b)." For the ignorant such as me, you might want to briefly explain the discrepancy between poorly performing global GSAT and well-performing regional GSAT in ESMs. I don't know if it is true, or covered in Beusch et al. 2020b, but is it because the regions are mostly land regions and cover 30% of the surface, so much of the poorly performing regions with respect to observations are over the ocean?

line 301: minor style typo: 360,000

lines 304-306 (remark only, no need to respond): I believe this sparse sampling would be sufficient actually. We did some pattern scaling with FaIR where the global delta T from FaIR was combined with one of 10 ESM simulatons chosen at random where the ESM performed well over the UK domain. Indeed, the full span of uncertatiny was well sampled.

figure 5: it could be a PDF rendering issue, but it would look nice if the alpha (1 - transparency) value for the shaded regions was < 1 to see the overlapping regions between blue and orange.

line 311: SSP1-1.9

section 4.2: is it worth saying where these 600 constrained ensemble members came from? Is it the MAGICC AR6 WG1 config, or is it from one of the RCMIP papers - Nicholls et al. 2020?

lines 336-337: optinally, you could hammer this home by giving the mean and range of ECS of these 6 models, compared to the full CMIP6 ensemble and/or the AR6 assessed range.

line 338: "is clearly incompatible ... (Fig. 6)." Well, only for 2 out of the 6 models - the other 4 look reasonable to me

lines 358-360: Indeed, there was a whole unofficial MIP (PDRMIP) devoted to this behaviour in ESMs. Tom Richardson derived a global 

precipitation emulator based on emissions of different GHGs, aerosols and GSAT: https://doi.org/10.1175/JCLI-D-17-0240.1. He had a regional one somewhere too but don't think it ever made it into a publication. Definitely something to explore.

Beusch et al discuss an emulation framework for the prediction of climate change impacts as a function of emissions scenarios. The framework is structured around the coupling of the MAGICC global simple climate model, and the MESMER regional pattern scaling model.



This paper presents the overall framework, including the methodology for the calibrations of the model components. It discusses the evaluation of regional results (as assessed through the metric of annual mean temperature), and the comparison of scenario projections with MAGICC-MESMER as compared with simulations in the CMIP6 archive.



The paper is clear, and well written and provides a novel and useful tool for the impacts community. I have no major issues barring publication, just some minor revisions are required to discuss some of the structural limitations which are implicit in the approach, and some comments which might help motivate further study.

Minor Issues

1 -In the discussion of limitations and future developments, it should be noted that the model structure allows for no dependency of climate variability on warming level - though there are probably elements of internal variability which are themselves dependent on warming or forcing level (Zheng 2018, Pendergrass 2017, Dorr 2021). These limitations, and potential for future developments, should be discussed a little more.



2- Furthermore, the internal variability is represented as a pattern related to global mean temperature deviations from a forced trajectory, plus a random term with imposed regional correlations through kriging. It is unclear from the present study whether this approach adequately reproduces (in a stationary climate) the noise-covariance structure of the original ESM which is being emulated, and the tests employed here - which focus on point-level errors - do not assess the skill of the emulator in producing realistic modes of natural variability. Though it's way beyond the scope here, and not necessary for this model overview, a future review could consider the relative performance of noise representation in the current scheme and other approaches (e.g. Perkins 2020, Alexeeff 2018, Holden 2010)



3 - the use of a lowess time filter to distinguish forced and variable components may exclude low frequency elements of natural variability which would ultimately be excluded from the model. The authors could test this in cases where large initial condition ensembles are available by combining ensemble members to produce an improved estimate of the underlying forced signal, smoothing (or not, if the ensemble is very large) and using the residual to estimate the noise component of the timeseries.



4 - The introduction of the additional predictors (a quadratic dependency of regional temperatures and a term dependent on global ocean heat uptake) are interesting extensions to the model. My concern is whether there is sufficient data to unambiguously fit these additional degrees of freedom, and whether there is spatial coherence in the relative role of the non-linear term and the ocean heat uptake term over the gridded field. As the authors move towards a larger number of predictors, they might want to consider an EOF prefilter for spatial fields - allowing the model parameters to be fitted in a lower dimensional space which enforces covariance structure. This framework might also aid ultimately in the probabilistic calibration of the MESMER component.



5 - The ocean heat uptake term is a very useful extension to MESMER for representing temperatures in deep mitigation scenarios. In future versions, the authors might find it useful to partition the heat uptake by depth, to represent the pattern effects of heat stored in the oceanic mixed layer as distinct from the deep ocean.



6 - There a brief discussion of the role of non-GHG forcers in the current study, and how this might be incorporated in the future. A brief note on how the simplified current framework might therefore introduce bias would be useful. i.e. to what degree are aerosol/ghg pathway co-dependencies 'baked into' the MESMER configration, and is this evident by looking at scenario outliers like SSP3-RCP7?









References:



Zheng, Xiao-Tong, Chang Hui, and Sang-Wook Yeh. "Response of ENSO amplitude to global warming in CESM large ensemble: uncertainty due to internal variability." Climate Dynamics 50.11 (2018): 4019-4035.



Pendergrass, A. G., Knutti, R., Lehner, F., Deser, C., & Sanderson, B. M. (2017). Precipitation variability increases in a warmer climate. Scientific reports, 7(1), 1-9.



Dörr, J., Årthun, M., Eldevik, T. and Madonna, E., 2021. Mechanisms of regional winter sea-ice variability in a warming Arctic. Journal of Climate, 34(21), pp.8635-8653.



Perkins, W. Andre, and Greg Hakim. "Linear inverse modeling for coupled atmosphere‐ocean ensemble climate prediction." Journal of Advances in Modeling Earth Systems 12.1 (2020): e2019MS001778.



Alexeeff, Stacey E., et al. "Emulating mean patterns and variability of temperature across and within scenarios in anthropogenic climate change experiments." Climatic Change 146.3 (2018): 319-333.



Holden, P. B., and N. R. Edwards. "Dimensionally reduced emulation of an AOGCM for application to integrated assessment modelling." Geophysical Research Letters 37.21 (2010).