Liu et al. (2025) uses 6 different statistical models to bias correct surface UKESM1 ozone comparing to CAMS reanalysis. A weighted approach is shown to improve performance over any single model. This bias correction is then applied to future scenarios, specifically SSP3-7.0 and SSP3-7.0-lowNTCF. The manuscript is well written and the figures are broadly of good quality.

Major concerns

My major concern with this study is the assumptions behind it and the validity of the bias correction for future scenarios. The authors state "We assume that UKESM1 exhibits systematic biases that are associated with other self-generated variables" (L97), and UKESM biases are then corrected by comparing UKESM1 to CAMS. However, there was only a limited discussion of CAMS, and while CAMS shows reduced biases compared to TOAR observations there were no detailed comments on whether CAMS uses the same emissions as UKESM or how emissions and how these are represented in models could influence biases against observations. How much is CAMS constrained by data assimilation compared to a free-running model such as UKESM, and is correcting ozone for e.g. biases in temperature a valid and fair approach? I feel that more reasoning is required here, along with clear caveats on the approach taken. 

I am currently also unconvinced that bias-correcting to CAMS for the present day, then applying this bias correction to future scenarios is also valid and fair. How sure are we that both UKESM and CAMS have the correct internal relationships in the present day simulations to be sure that projecting this bias correction into the future, with a different climate state and different emissions, would lead to that bias correction still being valid? Again, I feel that a greater discussion here of the validity of this method and the caveats in doing so should be made, especially as quite strong statements are made assuming that this is an entirely valid approach. The authors state that "This indicates that the UKESM1 has a greater sensitivity of seasonal O3 changes due to unknown reasons" (L149) in the discussion of future surface O3 changes, but these unknown reasons may make this approach less certain to succeed. How sure are the authors that they have the correct assumptions in the calculation of the biases in the present day, and is UKESM the correct model to consider regional air pollution in the context of their wider questions? 

Although at the end of the paper, the authors do state that "we acknowledge that uncertainties remain, particularly regarding the use of CAMS data as a reference for model training" (L262-3) I would have preferred a much more detailed discussion of the assumptions and limitations of this study, as from the current manuscript I am not convinced that this is a valid approach. In terms of a technical peice of work it is well formulated and presented, but scientifically I am nervous about the strength of the scientific statements that the authors make, particularly around the size of the biases UKESM may simulate when considering future climates.

Specific issues

Figure 3 - why do there seem to be discontinuities at 0.01, 0.1, and 1? The behaviour of the curve seems to jump at each of these values of sigma.

Figure 4 - I needed to zoom in quite a lot to be able to see the detail of the hatching. I would recommend making this plot bigger, perhaps a full-page 6x2 rather than the 3x4 currently presented. 

Figure 6 - given that the error bars (only 1 standard deviation) in winter mostly all straddle 0, can it be said that there are any biases in that season at all from this method?

Surface ozone (O3) as simulated in CMIP6 models, like UKCA/UKESM, tend to show large biases against observations (e.g., Turnock et al., 2020). In particular these models tend to simulate too much O3 in the Northern Hemisphere (NH) summer months (JJA) and too little in the NH winter months (DJF). The reasons behind these are not completely understood but likely reflect biases in the models climate, chemistry and emissions (obviously). 



Liu et al use a range of techniques -- from very simple statistical techniques to increasingly complex machine learning (ML) ones -- to bias correct UKCA modelled surface O3 during the historic period using "observations" from the Copernicus Atmosphere Monitoring Service (CAMS) reanalysis. CAMS itself is not perfect in this regard as it is model data but does provide globally gridded data at a comparable resolution to the UKCA outputs.



They find that the Transformer (an attention based ML model) performs best and multiple linear regression performs worst. They pool the results of their bias correction models using an ensembling technique and then then use their ensemble bias correction model to bias correct future surface O3 projections. Using this they find that the UKCA global mean surface O3 in JJA under the SSP-370 is overestimated by more than 12 ppb. 



I think the results are interesting, but not surprising, the methodology is clear to follow, but not state-of-the-art, and overall this is a well written and thought provoking study. I have my reservations on the reliability of the results but I think this is a publishable study. 



Main concerns: 

1) Nudging -- as I understand it the model simulations are not nudged and as a result one source of the O3 disagreement in UKCA is from the incorrect simulation of climate. Is this true? I think it would be worth making this clearer. Temperature pops out as an important feature and presumably bias correcting for temperature still does nothing for the O3 bias (as shown before in Archibald et al. 2020). 

2) Links to bias correction literature -- on my reading it seems like there is little to no attempt to discuss and place this work in the context of the very broad literature on bias correction, especially for climate models. I would like to know how techniques such as those by Ayar et al. (2021), Vrac and Friederichs (2015) or Nivron et al. (2025) complement, challenge or reinforce the choices of the approach of the task of bias correction taken here. 

3) CAMS data -- I think the caveat of the use of these data needs to be made more clearly and centrally. 



More general: 

It is stated that you use 20 features in training your models but you don't articulate exactly what these are. Please can you add a table. 

Figure 7 is interesting and the central point about the shift in NOx/VOC emissions regimes changing is clear but could you map out the areas in space where these points are coming from is to provide some spatial context? I am struggling to see how there is only ever one O3 value for a given NOx/VOC ratio (and is this concentrations or emissions? If so be clear on if it's g(N)/g(C) or what the units are). 



References: 

Vaittinada Ayar, P., Vrac, M. & Mailhot, A. Ensemble bias correction of climate simulations: preserving internal variability. Sci Rep 11, 3098 (2021). https://doi.org/10.1038/s41598-021-82715-1

Nivron et al. 2025. A temporal stochastic bias correction using a machine learning attention model.  doi:10.1017/eds.2024.42

Vrac, M., and P. Friederichs, 2015: Multivariate—Intervariable, Spatial, and Temporal—Bias Correction. J. Climate, 28, 218–237, https://doi.org/10.1175/JCLI-D-14-00059.1.