the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 19 Aug 2026
Strong Cloud-Mediated Aerosol Cooling in TaiESM1 Diagnosed Using Cloud Radiative Kernels and APRP Decomposition
Abstract. Aerosol–cloud interactions remain a major uncertainty in estimating human influence on climate. Here, we diagnose how aerosols, clouds, radiation, and ocean coupling interact in the Taiwan Earth System Model version 1 (TaiESM1). We use two complementary diagnostic methods to separate cloud radiative responses, aerosol-mediated effects, and the roles of aerosol–cloud and aerosol–radiation interactions. The contrast between the historical simulation and the pre-industrial control simulation produces a strongly negative total cloud radiative response (−2.08 W m⁻²), dominated by an aerosol-mediated component (−2.24 W m⁻²), despite a positive global-mean cloud feedback (+0.84 W m⁻² K⁻¹). This cooling is concentrated over the North Pacific and North Atlantic, where changes in low- and middle-level clouds are associated with stronger reflection of solar radiation. The shortwave aerosol effective radiative forcing is −2.34 W m⁻², mainly from aerosol–cloud interactions (−2.00 W m⁻²), while aerosol–radiation interactions are weaker (−0.34 W m⁻²). Within the aerosol–cloud component, cloud scattering dominates (−2.14 W m⁻²), whereas cloud-amount changes are small and positive (+0.12 W m⁻²). Compared with selected Coupled Model Intercomparison Project Phase 6 models, TaiESM1 lies near the strong aerosol–cloud cooling end of the sample. Fully coupled and prescribed-sea-surface-temperature simulations indicate that ocean coupling mainly alters regional patterns rather than the global-mean magnitude. These results identify cloud scattering of sunlight as the dominant driver of TaiESM1’s aerosol-related cloud cooling and highlight aerosol activation, cloud droplet number, and cloud optical depth as priorities for model development.
- Preprint
(7767 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-4942', Johannes Mülmenstädt, 21 Sep 2026
-
RC2: 'Comment on egusphere-2026-4942', Anonymous Referee #2, 25 Sep 2026
This study uses well-established radiative kernel and APRP technices to diagnose anthropogenic aerosol radiative cooling and its cloud-mediated components in the TaiESM1 model, primarily in the context of historical minus pre-industrial conditions. It is a very readable paper, and I appreciated learning about a fairly unique model that has not yet featured prominently in aerosol forcing diagnosis work of the past. This paper will be relevant to ACP readers as the community prepares for similar analysis expected in CMIP7, and will be useful as a reference for future TaiESM developers. That being said, at times the paper lacked sufficient motivation for the methodological choices and analysis provided, and at times a lack of methodological description left me confused. At points I also felt the analysis was a bit surface level given that this is a single-model paper. My comments below reflect that.
General: The kernel method used in this study is useful for a multi-model evaluation, because it uses very common CMIP experiments. But this is a single-model study, so the authors could have just run fixed-SST timeslice experiments designed for radiative forcing analysis, like the piClim RFMIP protocol. This is a more direct way to diagnose aerosol foricng and the simulations are cheaper. Why then did the authors rely on the kernel method and the accompanying coupled simulations? Some motivation should be provided. If the motivation for this choice was to also be able to capture the corresponding temperature-driven responses, it would be useful to say so in the introduction as a motivator. For instance, was the lack of warming in this model an impetus for this analysis?
Line 95-103: These lines make the argument for why we can’t just look at AOD to understand ACI forcings. I agree, but I don’t think anyone views ACI so narrowly in that way anymore. I wonder if there is another way to frame this? Maybe pointing out that ACI is driven by many factors and there is a benefit to decomposing the relevant radiation changes?Section 2.1: Some additional, relevant detail on the TaiESM1 aerosol and cloud schemes would be helpful here. The authors could describe the activation scheme, how aerosol size distribution is handled (bulk? Modal?) Have the cloud microphysics been modified since CESM1?
Section 2.1 and Table 1: piNTCF is an “all but one” experiment where everything is set to historical while NTCF is returned back to pre-industrial conditions, while in contrast AerosolOnly is a “one but all” experiment where the aerosols are perturbed historically and everything else is set back to pre-industrial. Does this protocol discrepancy matter for TaiESM and thus complicate the combined analysis of these experiments? Some comment on this in the text would be helpful. Simpson et al. (doi.org/10.1175/JCLI-D-22-0666.1) suggest it can matter, but not for all models.
Line 153: The authors should provide more information about histSST. Are time-evolving SST’s prescribed? If so, over what time period? Or is a climatology of SST’s prescribed? Where do the SST’s come from?
Equation 2 and related text: An important point with this method is that because alpha is derived from 1pctCO2, it is not just a sensitivity of cloud feedback, but also include the contributions of CO2 cloud and other rapid adjustments, which can be large. This is why removing the alpha_1pctCO2 * dT term isolates aerosol adjustments so well. Without including the CO2-induced rapid adjustments (i.e. if alpha from abrupt4xCO2 was used), the residual calculation would still include potentially large CO2 adjustment terms. This should be explained.
Line 176-182: I see mention of good agreement with aerosol cooling from fixed-SST experiments. I presume this means the author’s “aerosol-mediated cloud radiative response” calculation is essentially an estimate of aerosol forcing “cloud rapid adjustment”? If so, it would be helpful to mention the connection to the term “rapid adjustment”, since it is widely used in the radiative forcing literature now.Line 166: The authors should clarify whether “cloud radiative response” differs from “cloud radiative effect” and how. For instance, is cloud masking removed?
Lines 215-218 and Figure 1a: I am curious why the AOD change is positive in the southern ocean for historical-piControl when the aerosol-dominated and greenhouse gas-dominated AOD changes in that region appear to be equal and opposite in sign. Is there another component contributing AOD? Nonlinearities? Or am I mistaken about the magnitudes? If the latter, it may be because the colorbar is too broad and thus hard to decipher the magitude of the background AOD changes.
Line 222-234: Similar to two comments above, this is the first mention of cloud radiative effect, but the authors should define it in the methods section and clearly state how it differs from cloud radiative response. Also, I would prefer they use the abbreviations LWCRE and SWCRE instead of LWCF and SWCF, because they call it cloud radiative effect and not cloud forcing.
Line 233-235: The authors argue the LW cloud forcing in the aerosol-related case is not spatially coherent, but I’m not sure what they mean by spatially coherent. I see a clear tropical Pacific signature that appears ENSO-like and what one may expect from a cloud feedback/circulation process. It is important to clarify this, because they further argue in the next sentence that this lack of coherency suggests aerosol-cloud changes is dominated by SW not LW cooling. I agree that’s true in a global-mean sense, but not necessarily regionally.
Figure 3 Caption: I think there is some confusion here about “row” vs. “column”. The text description of figure position does not align with that is actually shown.
Figure 3: It would be helpful to print the global-mean values for each subplot like has been done for other figures in the paper.Line 259-260: I am surprised the global-mean total cloud response and aerosol-mediative cloud response are so similar. Likewise, the spatial maps of these terms in Fig 3a look almost identical. Even the AerosolOnly-piControl has a larger difference between total and aerosol-mediated response. Does it suggest the historical simulation has very little contribution from temperature-driven radiative changes or from GHG radiative forcing? Is warming (seems yes, as shown later) and greenhouse gas radiative forcing small in this model? Some context would be helpful.
Line 271-273: It’s not clear why the authors have started using the AerosolOnly – piControl contrast to diagnose aerosol response after using historical-piNCTF previously. Is it because they want to highlight the effects of the different base states? If so, that should be stated at the beginning of this paragraph.
Line 287-291: Since this is a single-model study, there is opportuntity to provide more detail about the aerosol-mediated cloud response from piNCTF-piControl. Could dust or sea-salt changes be playing a role? Does this model have interactive chemistry so that interplay between CH4, O3, and OH could be having an impact on background sulfur dioxide or organic aerosols? Without more detail, the evaluation of piNCTF-piControl doesn’t add much to the story, as we know anthropogenic aerosol changes are not the main culprit based on experimental design.
Line 367: Define TS here and again in the relevant figure captions.
Section 4.1 generally: I appreciate the authors put the TaiESM1 results into the context of the CMIP ensemble. That was helpful. But since this is a single model study, I feel the authors could go deeper in their discussion of the results. Why is ACI scattering so large, for example? Is it related to model tuning choices made? Is there something unique about the TaiESM1 aerosol activation or microphysics schemes that may be the culprit? I don’t expect the authors to find all of the answers, but a more model-specific discussion of the causes would really elevate this paper and its novelty relative to the many multi-model papers these methods are usually used for.
Line 378-385: This paragraph is written as if the aerosol-cloud coupling is wrong or deserve additional development. Is it? Some evidence (e.g. citation of observational constraints) would be useful.
Section 4.2 generally: I don’t really understand the point of this analysis. Some additional motivation context would be helpful. Anthropogenic aerosol forcing is largely unrelated to ocean dynamics or SSTs, and since the SST fields are the same in the historical and histSST experiments (are they?) isn’t it obvious the temperature-mediated cloud radiative changes would be extremely similar too? The authors point to some regional differences in cloud fraction between historical and histSST, but could it just be noise? In summary, I can appreciate that ocean coupling dynamics, separate from SST change, can have an impact on the climate, but it just doesn’t seem particularly relevant to a study about aerosol radiative forcing and I think the results here prove that.
Citation: https://doi.org/10.5194/egusphere-2026-4942-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 95 | 68 | 76 | 239 | 66 | 64 |
- HTML: 95
- PDF: 68
- XML: 76
- Total: 239
- BibTeX: 66
- EndNote: 64
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
I have reviewed "Strong Cloud-Mediated Aerosol Cooling in TaiESM1 Diagnosed Using Cloud Radiative Kernels and APRP Decomposition" by Tsai et al., who used these methods to study the response of the climate to anthropogenic aerosols in the TaiESM Earth system model.
The manuscript is engagingly written (with a few rough spots due to length or content noted below). It is nice to see application of these state-of-the-art methods to diagnosing aerosol-cloud interactions in the model climate. A large portion of the results are necessarily TaiESM-specific, but the authors do present interesting findings that may generalize to the way we diagnose ERFaci in other ESMs (or in general). These findings concern the importance of the climate base state used in the ERFaci diagnosis, the importance of ERFaer to the historical temperature evolution across models, and the relative unimportance of atmosphere-ocean coupling. On the plus side, this makes the study interesting for the wider community. On the minus side, I have some methodological concerns specifically for these findings, which I would like to see addressed (or rebutted, since it is possible I misunderstood the method) before I am comfortable recommending publication.
Scientific results:
- If this study is intended to guide further model development (l. 46, l. 378-385), it would be nice to know which part of the cloud scattering response is large (Nd or LWP) since very different processes are implicated depending on the answer to this question. This shortcoming could be addressed by a figure (between Figs. 1 and 2) showing Delta LWP (or even better Delta log LWP) and a figure showing Delta CDNC (or Delta log CDNC) because (a) then the reader can see all the changes entering into Delta CRE and (b) it also give the reader an idea about the partitioning between droplet activation and LWP adjustment, which APRP doesn't provide.
For the same reason (forcing and adjustment processes are not separated by either method employed), I would dispute the use of "process-oriented" in the Conclusions (l. 458).
- l. 85: probably good to cite some of the extensive pre-Chen2021 literature on the feedback-forcing inverse correlation (e.g., Kiehl 2007 https://doi.org/10.1029/2007GL031383, Forster et al. 2013 https://doi.org/10.1002/jgrd.50174).
l. 365-377: Thinking further about the previous point in response to the discussion here. It is a bit hard to evaluate the statements about forcing-feedback compensation between models when the feedbacks are not shown. But, accepting the premise, doesn't the fact that the Delta T_s varies by a factor of 6 (by eyeball, since the actual values are not given) among the models shown argue against the "compensation mechanism proposed by Wang et al."? I would expect the Delta T_s differences between models to be suppressed if the forcing and feedbacks compensated each other. If Fig. 6 included isolines of total ERFaer, it seems to me (again by eyeball, since the values are only shown by color scale) that Delta T_s follows ERFaer quite closely, as would conventionally be expected (e.g., Rotstayn et al. 2015, https://doi.org/10.1175/JCLI-D-14-00712.1).
It would be interesting to include a time series of the diagnosed forcing and coupled-model SST to support the statements made in this paragraph. If the historical time evolutions were presented, the presence or absence of "pothole cooling" in the 1970s and "catch-up warming" in the more recent past, Zhang et al 2021 (https://doi.org/10.5194/acp-21-18609-2021) would break the degeneracy between the strong-forcing/strong-feedback and weak-forcing/weak-feedback model temperature evolution to some extent. Otherwise, the conclusion that "the realized warming amplitude also depends on greenhouse-gas forcing, cloud feedback strength, ERFari, ocean heat uptake, and circulation adjustment. Therefore, TS should be interpreted as an integrated climate response" is obviously true but not obviously connected to what is seen in TaiESM.
- l. 271-280: I find it hard to believe that the difference between the diagnosed aerosol cloud response in the historical vs AerosolOnly experiments is physical. Rather, I think it must be an artifact of the method. I was very confused by this paragraph, however, so correct me as needed. I think AerosolOnly uses a coupled ocean; if so, the climate cools relative to piControl. Trying to use the 1%CO2 cloud feedback parameter may not work very well here because the cloud feedbacks are different between warming and cooling whenever there is a component due to cloud thermodynamic phase or due to liquid vs ice sea surface. I think the bright red bands in the high-latitude oceans in the AerosolOnly aerosol-mediated cloud response (Fig. 3b) are highly suggestive that something like this is going wrong in the method, since there is no large positive ERFaci in the Southern Ocean in the actual climate. But if this is the problem, it will affect most of the rest of the globe as well (anywhere the clouds aren't pure liquid in both climate states), so it could explain why the global means are so different. I am not sure what "preindustrial background state" on l. 278 means ("background" is too imprecise), but whether it means temperature or atmospheric composition, I don't think it can explain a factor of 2 in diagnosed ERFaci (assuming that -2.08 W m-2 and -1.15 W m-2 are the numbers I'm meant to compare here).
- l. 335-344: You have invested a significant amount of work into implementing both APRP and the kernel-based method. I think you let yourself down a bit in this paragraph by simply saying "these are different quantities", and in general by mostly discussing these results separately rather than connecting them with each other. If I understand correctly, the difference is that the aerosol cooling of the coupled SST induces circulation changes that modulate the cloud responses to aerosols, which by definition (it involves Delta SST) is not part of the aerosol ERF. You already have the results to tell an interesting story along the lines of Soden and Chung (2017), (I think) simply by plotting the difference between ERFaci and the Soden/Chung aerosol-mediated cloud response. I think you should!
- Fig. 6: Where do the values for the CMIP6 models come from? Did you compute them yourself using the Zelinka et al. (2023) method? The values are quite different from the values for the same models in Zelinka2023 Table 2. Is the difference because you used different experiments in your calculation than Zelinka et al. used in theirs? But in that case, isn't the fact that the ERF components can be so different a troubling reflection on the method?
Also, note that the modeling center name alone is not always enough to identify a model (e.g., IPSL).
Presentation:
l. 98-103: This paragraph is a bit choppy. I don't think it is necessary nowadays to underscore not "relying solely on aerosol optical depth". The statements about the relative importance of RF and adjustments to ERF and about the activation coefficient reflect a narrow sliver of the breadth of work being done in ERFaci world, and more to the point not questions that APRP can address, since activation and LWP adjustment are not separated. Why not simply point out that decomposing the aerosol effects on climate allows us to understand different mechanisms (ERFaer components; the Soden/Chung aerocol-mediated cloud response) and that comparing ERFaer component by component to other lines of evidence makes it more likely to catch compensating errors in the ERFaer?
l. 140: would be good to list the ACI-relevant parameterizations (i.e., does TaiESM inherit the Park and Bretherton cloud/boundary layer scheme and MG microphysics?) and whether the new convection scheme is coupled to cloud microphysics
l. 169: "combined influence of ..." add "and the ERF of the other forcing agents"? Since there is no mention of, e.g., the cloud adjustment to the CO2 forcing, does that mean Wang2021 conclude this component is negligible? If so, it would be good to state that here so the reader understands why the CO2 ERF is absent from eq. (1). Is this also the reason you say "... historical – piNTCF as the aerosol-related component ..." on l. 208 although historical – piNTCF technically includes the CO2 ERF?
l. 190-193: APRP ERFaci and ERFari definitions differ from the IPCC definitions (because cloud adjustments to ARI are part of ERFari in the IPCC definition and part of ERFaci in the APRP definition), see the discussion after eqs. (6-7) in Zelinka2023. It would be good to mention this here.
l. 218: I think it is a mathematical identity that line 1 = line 2 + line 3 for all the plots in Fig. 1, correct? Maybe a few remarks about why the piNTCF - piControl Delta AOD is so large (~30% of the historical - piControl) or equivalent why the difference between historical - piControl and piNTCF - piControl Delta AOD is so large?
l. 236-243: In general, I find Figs. 1 and 2 a bit confusing because the second row of plots doesn't include the cloud feedback (both historical and piNTCF use the present-day SST) whereas the first and third rows do (piControl uses the preindustrial SST). I guess the reader will just have to live with doing a lot mental switching back and forth here. But shouldn't l. 242 then say "greenhouse-gas driven cloud adjustments and cloud feedback"? (And similarly for l. 245?) And shouldn't "greenhouse gas" be replaced by CO2, since NTCF includes the other greenhouse gases (CH4 and O3)?
l. 271-281: It would be easier to follow this paragraph if the "aerosol-only" experiment were described a little more clearly, specifically whether this is a coupled or fixed-SST run (maybe Table 1 could have columns specifying which forcers are PI and which are time-varying and whether the ocean is coupled or fixed, and if fixed, which SST is being used).
l. 354-364: The purpose of the least-squares fits in Figs. 6 and 7 appears to be to document the correspondence between the climate responses on the x and y axes and to show that TaiESM lies somewhere close to the line. It looks like the fit is x ~ y rather than (the more usual, though of course there is no real constraint on choice of axes in this case) y ~ x. If the fit were done the other way around, the line in Fig. 6 would probably be close to horizontal and TaiESM would look like an outlier. I say this not to quibble with the paper's conclusions but rather to say that I find the straight-line fit more confusing than illuminating.
l. 362: As noted above, I am having trouble seeing where this statement about IPSL comes from. In Zelinka2023, all IPSL models have ERFaci < 0.
l. 365: TS has not been defined
l. 387, though I should have noticed in Sec. 2: is the prescribed SST the one from the TaiESM historical run or an observed one? I assume the former, but please state. If the latter, that would make the historical - histsst plots difficult to interpret.
All of Sec. 4.2: This section is 3 multipanel figures and two pages of dense text describing many fields that change hardly at all between prescribed and interactive SST. I think this could be made substantially shorter and reader-friendlier by focusing on what's actually different between historical and histsst and then focusing on why the interactive ocean is important for those differences (through fast feedbacks in surface exchange etc.).
Figures, especially the many-panel ones: the font is *so* tiny
Fig. 3: would be nice to have the global mean values indicated in each panel
Fig. 4b: would be nice to have a narrower color scale here so the spatial patterns are actually visible
Fig. 6: an isoline of total ERFaer would be quite useful to guide the eye and make the connection to the Delta T_s argument (l. 365), given how different the scales on the axes are; probably more useful than the fit line