| 08 Jul 2026
Machine Learning-Based Observational Constraints on Cloud Condensation Nuclei Responses to Australian Wildfire Aerosols over Remote Oceans
Abstract. Australian “Black Summer” wildfires in 2019–2020 released large amounts of smoke that affected aerosols and clouds over the South Pacific. Here we quantified wildfire perturbation on aerosol loading and cloud condensation nuclei (CCN, particles that can act as seeds for cloud droplets) using a machine learning (ML) method. We trained ML models with meteorological datasets to represent counterfactual “no-wildfire” conditions, which were contrasted against satellite observations of “wildfire” conditions of aerosol optical depth (AOD), Aerosol Index, and CCN. We found strong and robust aerosol and CCN responses during the wildfire months over the South Pacific downwind plume region (140° E–100° W, 10° S–50° S). The wildfire perturbation substantially increased AOD by 36 % and 68 % in December 2019 and January 2020, respectively, and Aerosol Index by 21 % and 53 %, respectively. In comparison, CCN increased by 40 % in December 2019 and 20 % in January 2020. We find that the AOD and Aerosol Index enhancements form a broad band following the main smoke plume and extend across the full studied region over the South Pacific, while the CCN response is weaker and more localized, with enhancements mainly confined to the plume central-line along the main smoke transport pathway with a rapid decay further downwind. This difference suggests that transport and aging processes could reduce wildfire aerosol capability in enhancing CCN number concentration. Moreover, our results provide further observational evidence that Aerosol Index can provide a useful complement to AOD when interpreting smoke impacts on CCN.
Review for “Machine Learning-Based Observational Constraints on Cloud Condensation Nuclei Responses to Australian Wildfire Aerosols over Remote Oceans”
The study by Yu et al. employs a Random Forest machine learning (ML) approach to quantify changes in cloud condensation nuclei (CCN) concentrations associated with the Australian wildfires over remote oceans. The overall methodology of using ML to predict satellite-derived CCN proxies (MODIS AOD, OMI UV Aerosol Index, and CALIPSO-derived CCN) from reanalysis variables and then estimating these proxies under counterfactual "no-wildfire" conditions is reasonable. However, I have major concerns regarding the interpretation of two of the CCN proxies. First, the authors incorrectly use the OMI-derived UV Aerosol Index (UV AI), which is an indicator of the presence of UV-absorbing aerosols (e.g., smoke and dust), as a proxy for fine-mode aerosols that contribute to the CCN population (see my first comment for details). Second, the authors assume that the CALIPSO-derived CCN product captures changes in aerosol chemical composition and particle size distribution resulting from aerosol aging. In reality, the retrieval is based on predefined aerosol size distributions and does not account for such microphysical or chemical evolution (see comment 5 for more details).
These two issues substantially weaken the physical interpretation of the results. If these assumptions are removed, the manuscript offers limited novelty. Consequently, I do not believe the study meets the level of scientific novelty expected for publication in Atmospheric Chemistry and Physics in its current form, and I therefore recommend rejection.
Nevertheless, I provide the following detailed comments in the event that the manuscript is considered for publication in Atmospheric Chemistry and Physics or another journal.
1. Lines 76–86: The authors have confused the UV Aerosol Index (AI) derived from OMI with the aerosol index commonly derived from multispectral imagers such as MODIS. These are fundamentally different quantities. The OMI UV AI is an indicator of the presence of UV-absorbing aerosols, such as mineral dust, biomass-burning smoke, and volcanic ash, and is primarily sensitive to aerosol absorption and layer height rather than particle size. In contrast, the MODIS aerosol index, commonly defined as the product of the aerosol optical depth (AOD) and the Ångström exponent (AE), is used as a proxy for the abundance of fine-mode aerosols that are more relevant to cloud condensation nuclei (CCN). The AE is calculated from AOD retrieved at two wavelengths (typically 550 and 860 nm over ocean; e.g., Gryspeerdt et al. (2023)), based on the principle that fine particles exhibit a much stronger wavelength dependence of scattering than coarse particles.
Therefore, while the OMI UV AI is appropriate for identifying absorbing aerosol events (e.g., smoke or dust), it is not a valid proxy for CCN, whose concentration is more closely related to aerosol number and fine-mode aerosol loading. Consequently, the interpretation of the results based on the OMI UV AI as a surrogate for CCN is not justified and represents a significant limitation of the study.
Gryspeerdt, E., Povey, A. C., Grainger, R. G., Hasekamp, O., Hsu, N. C., Mulcahy, J. P., Sayer, A. M., and Sorooshian, A.: Uncertainty in aerosol–cloud radiative forcing is driven by clean conditions, Atmos. Chem. Phys., 23, 4115–4122, https://doi.org/10.5194/acp-23-4115-2023, 2023.
2. Line 174-176: The R² values for the three Random Forest models should be reported to assess their predictive performance. Reporting these statistics will enable readers to evaluate the fidelity and robustness of the models.
3. Line 237: Please refer the reader to the appropriate figure where this information is presented.
4. Lines 250–252: While the non-linear relationship between AOD and CCN is physically reasonable, this may not be the only explanation for the absence of high CCN values across the entire domain, unlike the spatial distribution of the UV AI. It is worth noting that, unlike the UV AI, the MODIS AOD exhibits a similar spatial pattern to the CCN product up to approximately 140°W.
Another important consideration is the large difference in sampling characteristics between MODIS and CALIOP. MODIS is a passive imaging radiometer with a swath width of approximately 2,330 km, whereas CALIOP is an active lidar with a laser footprint of only about 70 m along a narrow ground track. This substantial difference in spatial sampling may significantly influence the observed aerosol spatial variability and, consequently, the comparison with the CCN product.
To further investigate this issue, I recommend that the authors examine the CALIOP-derived aerosol extinction coefficient or aerosol optical depth (AOD) to determine whether these products exhibit spatial variations similar to those of the CCN product. These CALIOP aerosol products are available from NASA Earthdata: https://www.earthdata.nasa.gov/data/catalog/larc-cloud-cal-lid-l3-tropospheric-apro-allsky-standard-v5-00-v5-00.
5. Section 3.3: The interpretation presented in this section appears to assume that the CALIPSO-derived CCN product captures variations in smoke hygroscopicity and particle size distribution associated with aerosol aging. However, this is not the case.
The CALIPSO CCN product is derived by converting the CALIPSO aerosol extinction coefficient to aerosol number concentration using predefined aerosol size distributions (Choudhury and Tesche, 2022). For example, for smoke aerosols, the retrieval scales a prescribed smoke size distribution by adjusting the total particle volume until the calculated extinction matches the CALIPSO-derived extinction coefficient. The effect of aerosol hygroscopic growth is accounted for only by first converting the ambient extinction coefficient to dry extinction prior to the scaling procedure (Choudhury et al., 2022). Subsequently, CCN concentrations are estimated by integrating the retrieved dry size distribution for particles larger than 50 nm.
Therefore, the retrieval does not account for changes in the chemical composition or particle size distribution of smoke resulting from atmospheric aging during transport. Such information cannot be retrieved from an elastic backscatter lidar such as CALIOP without additional observational constraints.
I recommend that the authors revise this section to acknowledge the inherent limitations of the CALIPSO-derived CCN product and avoid overstating the conclusions regarding the effects of smoke aging on CCN variability.
Choudhury, G. and Tesche, M.: Estimating cloud condensation nuclei concentrations from CALIPSO lidar measurements, Atmos. Meas. Tech., 15, 639–654, https://doi.org/10.5194/amt-15-639-2022, 2022.
Choudhury, G., Ansmann, A., and Tesche, M.: Evaluation of aerosol number concentrations from CALIPSO with ATom airborne in situ measurements, Atmos. Chem. Phys., 22, 7143–7161, https://doi.org/10.5194/acp-22-7143-2022, 2022.
6. Lines 290–297: The discussion in this section should be revised because the UV Aerosol Index is an indicator of the presence of UV-absorbing aerosols (e.g., biomass-burning smoke and mineral dust), rather than fine-mode aerosol particles or aerosol number concentration. Therefore, the interpretation presented here is not fully supported by the quantity being analyzed. Please reframe the discussion to reflect what the UV AI physically represents and avoid interpreting it as a proxy for fine-mode particles or CCN.