the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 17 Mar 2025
Rapid near-term warming in a carbon neutral future attributed to substantial aerosol decline
Abstract. Greenhouse gases (GHGs) and aerosols have been the main contributors to climate change. Here, future aerosols and GHGs impacts on global and regional surface air temperature (SAT) are assessed using machine learning. We show that, following a carbon-neutral pathway, global SAT rapidly increases by 0.8 °C from 2015 to 2050, with 0.6 °C attributed to the substantial decline in aerosols. Considering both the varying aerosols and GHGs, temperatures during 2015–2035 under the carbon-neutral scenario are even higher than those under the high-emission scenario, demonstrating that the near-term warming related to aerosol reduction is noteworthy, although carbon neutral scenario is beneficial to slow down the warming. If CO2 is reduced to mitigate the warming caused by aerosol reduction, it has to decline from 400 ppm to 340–350 ppm. This study emphasizes the importance of anthropogenic forcings in regulating climate change and reveals the dominant role of aerosols in modulating climate in the near-term carbon neutral future.
Competing interests: At least one of the (co-)authors is a member of the editorial board of ACP.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(2464 KB) - Metadata XML
-
Supplement
(2599 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-840', Anonymous Referee #1, 25 Mar 2025
This work deploys a decision tree based ML to generate an emulator for CMIP6-like climate model to attribute the contribution of anthropogenic aerosol and GHGs in global and regional warming. The ML-attribution is done for SSP119 and SSP585 scenarios. Although the attribution has been explicitly simulated in CMIP6-DAMIP for SSP245, the SSP119 (Net-Zero) and SSP585 (high fossil fuel) scenarios have no multiple model detection & attribution modelling experiments, therefore is of particular interest to the community. While interesting, I feel there are major concerns that this study should be addressed accordingly, before can be accept for publishing. I will provide details of my concerns as below, in addition, I feel the language presentation would also need further polishing.
Specific comments:
- I doubt about the “perfect” validation of the ML model, eg. Fig.2a. I think the approach for model validation is not a fair way to do it. If I understand the method section correctly, the authors randomly leave-out 10% data from a bunch of CMIP6 models outputs for validation and use the rest 90% data for training. Therefore, the ML, in the majority cases, is only doing an easy interpolation job, rather than the challenging predicting job which we expected ML to do. Because, for example, CESM: 2010.01 – 2010.07 + 2010.09 – 2010.12 are used for training, but 2010.08 is used for validation (just an example, similar cases can happen in most of situations). In addition, your training dataset could still have the 2010.08 data from ECHAM, MRI-ESM, MPI-EMS, etc. Therefore, your ML model only learns the relationship between different models and interpolation between different months (or from neighbor grid cells). However, for attribution, we do need the capability for prediction in ML model, which is not trained in the design of current training approach.
- The description of ML is not clear enough in the method. Eg. What is import, what is output, how to train the ML etc.
- What is the representative of the ML model? Does it represent a single CMIP6 model (which one?), or it represent the multiple model average?
- As model-based attribution tech. pioneered by Prof Klaus Hasselmann, a key element for attribution is the inter-variability between climate models. Because this helps us understand the uncertainty and allows us to say that if the contribution of a climate forcer is significant enough to be detected, or not. I wonder how the multi-GCMs variability is represented in the ML model, and how is this been used to convince that the attributed aerosol/GHGs forcing is a significant fingerprint? Note that this multi-GCMs variability (stem from parameterization/structure/etc. uncertainties) is different from the shading area shown in Fig.5 (and many other figures), which only provides the uncertainty of ML training.
- I think Fig.S4 is worth more interpretation. I cannot read the message (L236) from Fig.S4. In addition, could you please help me understand why in some cases that CO2/CH4/N2O can contribute cooling in global and some regional scales (see Fig.S4 of the negative impacts on model temperature)?
Citation: https://doi.org/10.5194/egusphere-2025-840-RC1 - AC1: 'Reply on RC1', Li You, 26 May 2025
-
RC2: 'Comment on egusphere-2025-840', Anonymous Referee #2, 09 May 2025
The authors present a machine learning based approach for disentangling the effects of different anthropogenic forcing agents on historical temperatures by training on an opportunistic ensemble of climate model simulations. They find a significant cooling contribution due to aerosol, which leads to enhanced future warming as aerosol emissions decline. The study is framed well and, if true, would represent an interesting and useful approach for attributing observed warming trends without having to run dedicated single (or all-but-one) forcing simulations.
That said, I have strong concerns about the methodology, and in particular the approach for training and testing the model. Because the scenarios used for training the model are all strong correlated, the use of a random sub-sampling of test data leads to a serious risk of overfitting. That is, the randomly sampled test data provides no real validation that the model is able to extrapolate to the scenarios the authors then use the model to explore, undermining the presented results. A more appropriate approach to select test data would be to hold back a whole scenario, as in ClimateBench (Watson-Parris et al. 2022), which also specifically tests emulators against ssp245-aero to perform aerosol attribution.
If invited for resubmission, the manuscript would benefit from proof-reading by a native English speaker as there are many grammatical and style aspects that could be improved.
Citation: https://doi.org/10.5194/egusphere-2025-840-RC2 - AC2: 'Reply on RC2', Li You, 26 May 2025
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 474 | 48 | 21 | 543 | 41 | 15 | 24 |
- HTML: 474
- PDF: 48
- XML: 21
- Total: 543
- Supplement: 41
- BibTeX: 15
- EndNote: 24
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1