the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 Jul 2026
A 21-year global dataset of particle number concentrations for aerosol-cloud interaction studies
Abstract. Aerosol particles larger than roughly 50–100 nm in diameter are climatically important because they can act as cloud condensation nuclei (CCN), making their global number concentrations essential for understanding aerosol–cloud interactions. However, observationally constrained, long-term global datasets of particle number concentrations in this size range remain scarce. In this investigation, we present a global dataset of ground-level particle number concentrations for the period 2003–2024, produced by combining in situ observations with a machine-learning approach. The dataset includes two variables: the number concentrations for particles larger than 100 nm (N100) and larger than 50 nm (N50), provided at 0.75° × 0.75° spatial resolution and daily temporal resolution. To generate this dataset, we trained an eXtreme Gradient Boosting (XGB) model using measurements from 62 in situ stations as targets and reanalysis variables as predictors, enabling a data-driven representation of particle number concentrations at the global scale. We evaluated the dataset against independent observations from 12 additional stations. At 2/3 of these stations, the dataset shows good performance, capturing the median concentrations within a factor of 1.5 from the observations. Furthermore, we describe the main characteristics of the dataset in terms of global spatial patterns, temporal variability, and seasonal cycles, and demonstrate its ability to capture long-term trends in particle number concentrations, including both increasing and decreasing tendencies reported in the literature. This work provides the first observation-constrained, machine-learning-based global dataset of N50 and N100 at daily resolution over two decades, bridging the gap between sparse measurements and computationally expensive process-based models. The dataset, publicly available at https://doi.org/10.5281/zenodo.20202080, offers a valuable resource for evaluating model simulations, improving CCN-related parameterizations, and supporting weather and climate studies without the need for explicit knowledge of the aerosol particle microphysics.
- Preprint
(8353 KB) - Metadata XML
-
Supplement
(6857 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on essd-2026-415', Anonymous Referee #1, 28 Aug 2026
-
RC2: 'Comment on essd-2026-415', Anonymous Referee #2, 20 Oct 2026
This study applies a machine learning approach to derive a global gridded dataset of aerosol number concentrations across two size ranges (N₅₀ and N₁₀₀). The machine learning model ingests surface reanalysis data (ERA5 and CAMS) and is trained and validated using a large selection of ground-based measurement stations.
The results demonstrate good model performance at many of the independent holdout stations against which it is evaluated.
In summary, this study is well written, employs appropriate methodologies, and has the potential to be of great value to the scientific community for the evaluation of atmospheric models. The dataset is significant, but requires more rigorous uncertainty quantification and analysis for its applicability as a global dataset for model evaluation.
Major comments
1: Effective Size Limits of the ML Model
The authors state that the upper size limit was defined as the maximum particle diameter measured by the instrument (Lines 113–115). The effective maximum size limit of the resulting ML model dataset require clarification, as this is critical for its utility as a model evaluation tool. Specifically, what particle diameter range does each predicted data point correspond to? Constraining the maximum diameter to a fixed value is important for the dataset to serve as a robust model evaluation resource, since most climate models employ modal aerosol schemes. For any evaluation of a climate model against this dataset to be rigorous, the size limits over which aerosol number concentration should be integrated must be clearly defined. The authors should also consider providing guidance on a recommended workflow for climate modellers using modal aerosol schemes who wish to use this dataset for model evaluation.
2: Hyperparameter Selection and Sensitivity
The study states that the ML model employs a similar setup to Ovaska et al. (2025) (Lines 153–154). Given that this study incorporates a substantially different observational training dataset, it is unclear why the same hyperparameters are retained without retuning. The authors should justify this decision and demonstrate how model performance and the resulting dataset would change if hyperparameters were re-optimised given the updated training data.
3: Quantification of Dataset Uncertainty and Confidence Measures
One of the stated aims of the study is to provide a dataset that can be used to evaluate other models. Whilst a continuous global gridded dataset of N₅₀ and N₁₀₀ is highly advantageous for this purpose, a more rigorous characterisation of the dataset’s limitations would strengthen the study, along with an explicit mapping of those limitations onto the structural uncertainties of the ML framework. At present, the discussion of the drivers of model weaknesses and associated biases is quite speculative in nature in places. The dataset would benefit substantially from the inclusion of a quantitative confidence measure as an additional data field to support users, and to demonstrate consistency of the data at different locations to those tested. Furthermore, the authors are encouraged to include Figure S2 in the main body of the paper, as these results are central to a transparent assessment of biases.
4: Sensitivity Analysis to Identify Bias Drivers
A rigorous characterisation of the spatio-temporal limitations of the dataset requires a more in-depth analysis to reveal any weaknesses inherent in the ML framework. This would benefit from a sensitivity analysis to identify the meteorological and aerosol predictor regimes associated with the largest biases relative to in situ observations. The extensive in situ observational dataset could be leveraged to identify such regimes. For example, do the largest biases in predicted N₅₀ or N₁₀₀ correspond to aerosol particle number size distributions with particular characteristics, such as those influenced by new particle formation or cloud processing? Identifying regimes, either through the predictor variables or through the observations that are associated with the largest biases, would clarify whether specific processes drive model deficiencies. Importantly, this would better inform users about limitations of the dataset if using for model evaluation. For instance, does the systematic under-representation of peak concentration values (Line 213) reflect an inability of the datasets representation of new particle formation events? At what specific concentration thresholds does ML model performance degrade (Lines 203–208)?
5: Extrapolation and CAMS Limitations
The ML dataset addresses the sparsity of the current spatio-temporal record of in situ surface observations by generating a global gridded product spanning 2003–2024. However, tree-based methods such as XGBoost are known to struggle when predicting beyond the bounds of the training data. Several of the training stations have significant gaps in their observational records (Figure S1), which raises the question of extrapolation and its potential impact on model performance, particularly for trend analysis. Discussion of the dataset performance should be expanded to acknowledge known extrapolation limitations as well as any known limitations associated with the CAMS input data, specifically highlighting any regions and time periods that any previous studies have identified the CAMS performance to be weakest. For example, how do spatial and temporal uncertainties in the CAMS data used — including diurnal and seasonal biases — propagate into this dataset, and what implications does this have for its use in trend analysis and model evaluation?
6: Performance During Aerosol Perturbation Events
How does the dataset reproduce known large aerosol perturbation events? Several key natural perturbation events have previously been exploited to investigate the capacity of models to represent aerosol-cloud interactions, including volcanic eruptions (Jordan et al., 2025; Duncan et al., 2026). A persistent challenge in such studies has been the spatial sparsity of in situ observations during these events. It would be of significant scientific value to demonstrate the ability of the dataset to capture known large aerosol perturbation events previously employed to investigate aerosol-cloud interactions to highlight the usefulness of this dataset as an additional model constraint for such studies in the future. Does the dataset contain peaks associated with such events?
7: Altitude Representation of CAMS Predictors at Mountain Stations
The authors acknowledge that complex terrain and local topography may affect ML model performance (Line 266). Several stations in the training dataset are located in complex terrain or at high altitude (Table 1). The authors should clarify how the CAMS aerosol predictor fields are sampled in the vertical dimension relative to these high-altitude stations, and provide an assessment of whether the vertical representativeness of the CAMS data employed influenced model performance at these locations and, if so, in what way.
8: Antarctic Station Selection and Extrapolation Concerns
For an XGBoost based model to represent global conditions reliably, it is essential that the training dataset adequately spans the observed concentration range. Dome-C (DMC) was selected as a holdout station despite exhibiting extremely low concentrations that fall outside the range of the training data (Line 224). The authors should clarify why DMC was retained as a holdout station rather than used as a training station to mitigate the extrapolation challenges reported. It would also be informative to demonstrate how the dataset would change if the model were trained using DMC data.
An extension in this study compared to Ovaska et al., 2025 is to incorporate additional measurement sites. Given the acknowledged data sparsity in Antarctica (Line 250), the authors should address why no further Antarctic measurement stations, such as those described in Lachlan-Cope et al. (2020), were incorporated into the training dataset to improve coverage in this climatically sensitive region.
9: Attribution of Model Biases and Surface Reanalysis Limitations
The authors postulate that systematic overestimation at the EPC station (Line 256) is attributable to its heterogeneous environment. This attribution is rather speculative given that many other stations in the dataset are similarly located in complex, heterogeneous environments. A more rigorous investigation of ML model performance, as recommended in earlier comments, is required to substantiate such assertions. For the ISP station, the authors note that the model underestimates concentrations (Line 263); it should be clarified whether this bias is unique to ISP or whether nearby stations exhibit comparable underestimation. More generally, the discussion should more clearly acknowledge the limitations of using surface ERA5 reanalysis variables across all regions, recognising that for certain environments other atmospheric processes can play an important role for surface aerosol properties, for example transport from aloft, e.g. McCoy et al., 2021.
10: Physical Consistency of Predicted N₅₀ and N₁₀₀
The authors correctly identify instances where the predicted N₁₀₀ exceeds N₅₀, is physically inconsistent given that N₁₀₀ is a subset of N₅₀ (Lines 284–295). Quantitative statistics on how frequently the N₅₀/N₁₀₀ ratio approaches or falls below 1 should be provided, along with a discussion of how this ratio is affected by the upper particle diameter size limit (see Comment 1). The authors should also briefly comment on whether joint training of the model on both N₅₀ and N₁₀₀ is feasible and whether it could reduce these physical inconsistencies. Additionally, the assertion that N₅₀ generally exhibits lower temporal variability than N₁₀₀ (Line 311) appears counterintuitive, given that sub-100 nm particle concentrations are typically highly variable due to processes such as new particle formation. The authors should elaborate on the physical basis for this finding. Finally, the claim that peak concentrations in polar regions occur during December–February (Lines 324–327) should be discussed in the context of previous literature on Arctic haze, some of which have identified the peak occurrence between March and May (e.g. Tunved et al., 2013).
Minor comments + technical comments
Please clarify the provenance of the aerosol size distribution data used, including any version numbers / download dates, to facilitate reproducibility where possible. A dataset reference list is provided in Table 1; however, observational datasets are updated over time as new quality-control frameworks are introduced.
The interpretation of the global concentration estimates (Section 4.1.2) would benefit from additional detail. The estimates presented in this study appear systematically lower than those reported in the prior literature. Can the authors provide any insight as to whether this discrepancy is attributable to differences in the underlying ML framework assumptions or to differences in the definition of the predicted variables?
The coefficient of determination (R²) is a commonly used statistical measure. Did the authors consider reporting it, alongside the coefficient of variation in Table 2?
The colour used in Figure 4 for the observed N₁₀₀ time series is difficult to distinguish. The authors should revise the colour scheme to improve clarity and accessibility.
Line 55: Some key references relating to previous studies that evaluate models against measurements are missing, for example, Johnson et al., 2020 could be considered to provide context to the wider literature on global model evaluation strategies.
Line 60: The authors should consider acknowledging the EarthCARE satellite mission in this context.
Line 69: Consider acknowledging the most recent version of this dataset (Anders et al., 2026).
Line 78: The key differences between the Georgiades et al. (2025) model and the present study should be stated explicitly.
Line 80: Would benefit from acknowledging other recent machine learning approaches that apply XGBoost to predict the same or similar variables, such as Duncan et al. (2025) and highlight any differences in the machine learning approach.
Line 105: The authors should provide justification for the exclusion of other polar measurement stations for which data is available from the training dataset.
Line 119: The fraction of data removed by the three standard deviation outlier filter should be quantified (as a percentage).
Line 136: For reproducibility, please specify the interpolation method used (e.g. bilinear interpolation).
Line 138: Please specify how the derived meteorological variables (relative humidity, wind speed) were calculated. This is relevant since different models use different schemes for deriving relative humidity (e.g. with respect to ice or liquid water).
Line 143: The selection of a minimum value of 10⁻¹⁷ appears somewhat arbitrary. The authors should provide justification for this threshold and discuss its sensitivity.
In places, it is unclear whether "average" refers to the mean or the median. The authors should use consistent and unambiguous terminology throughout, including in the description of the any averaging procedure.
Line 225: The term "otherwise" is slightly ambiguous in this context and should be clarified.
Line 254: The use of N₅₀ and N₁₀₀ as proxies for CCN may overstate their utility, as cloud condensation nucleus activation depends on aerosol composition and dynamics in addition to particle number size.
Line 256: "the model" should be replaced with "the ML model" for clarity and check other instances throughout.
Figure 5: Consider overlaying station locations as circles to contextualise the spatial patterns relative to the observational network.
Line 314: The discussion of variability should acknowledge the potential contribution of new particle formation events.
Figure 7: The colour bar tick labels would be clearer if reported in more readable format, e.g. 10 cm-3, 100 cm-3.
Figure 8: Consider providing regional latitude/longitude bounds in the figure caption or supplementary material for reproducibility.
Line 374: Phrasing: “Possibly driven” > “It is possible that … “
Line 387: Please clarify how this pattern is discernible from the presented results.
References
Anders, Y. E., Block, K., Pöhlker, M., and Quaas, J.: Evaluation and improvement of CAMS-derived CCN number concentrations using in-situ measurements, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2026-1404, 2026.
Duncan, E. K., Fieldsend, J. E., Sellar, A., Tovazzi, E., Kim, P., Haywood, J. M., and Partridge, D. G.: A framework to holistically investigate processes controlling the aerosol lifecycle using explainable AI techniques, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2025-4298, 2025.
Duncan, E. K., Jordan, G., Kim, P., Haywood, J. M., Watson-Parris, D., Johnson, B., Sellar, A., Kipling, Z., Teixeira, J., Malavelle, F., and Partridge, D. G.: Lagrangian investigation of GCMs during the 2014–15 Holuhraun eruption reveals large differences in the representation of aerosol size distribution, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2026-1043, 2026.
Johnson, J. S., Regayre, L. A., Yoshioka, M., Pringle, K. J., Turnock, S. T., Browse, J., Sexton, D. M. H., Rostron, J. W., Schutgens, N. A. J., Partridge, D. G., Liu, D., Allan, J. D., Coe, H., Ding, A., Cohen, D. D., Atanacio, A., Vakkari, V., Asmi, E., and Carslaw, K. S.: Robust observational constraint of uncertain aerosol processes and emissions in a climate model and the effect on aerosol radiative forcing, Atmos. Chem. Phys., 20, 9491–9524, https://doi.org/10.5194/acp-20-9491-2020, 2020.
Jordan, G., Malavelle, F., Haywood, J., Chen, Y., Johnson, B., Partridge, D., Peace, A., Duncan, E., Watson-Parris, D., Neubauer, D., Laakso, A., Michou, M., and Nabat, P.: How well are aerosol–cloud interactions represented in climate models? – Part 2: Isolating the aerosol impact on clouds following the 2014–2015 Holuhraun eruption, Atmos. Chem. Phys., 25, 13393–13428, https://doi.org/10.5194/acp-25-13393-2025, 2025.
Lachlan-Cope, T., Beddows, D. C. S., Brough, N., Jones, A. E., Harrison, R. M., Lupi, A., Yoon, Y. J., Virkkula, A., and Dall'Osto, M.: On the annual variability of Antarctic aerosol size distributions at Halley Research Station, Atmos. Chem. Phys., 20, 4461–4476, https://doi.org/10.5194/acp-20-4461-2020, 2020.
McCoy, I. L., Bretherton, C. S., Wood, R., Twohy, C. H., Gettelman, A., Bardeen, C. G., & Toohey, D. W. (2021). Influences of recent particle formation on Southern Ocean aerosol variability and low cloud properties. Journal of Geophysical Research: Atmospheres, 126, e2020JD033529. https://doi.org/10.1029/2020JD033529
Tunved, P., Ström, J., and Krejci, R.: Arctic aerosol life cycle: linking aerosol size distributions observed between 2000 and 2010 with air mass transport and precipitation at Zeppelin station, Ny-Ålesund, Svalbard, Atmos. Chem. Phys., 13, 3643–3660, https://doi.org/10.5194/acp-13-3643-2013, 2013.
Citation: https://doi.org/10.5194/essd-2026-415-RC2
Data sets
A 21-year global dataset of particle number concentrations (N100 and N50) for aerosol-cloud interaction studies Aino Ovaska, Daniel Holmberg, Elio Rauth, Alghamdi Mansour, Paulo Artaxo, Eija Asmi, John Backman, Benjamin Bergmans, Matthew Boyer, Liezl Bredenkamp, Maurizio Busetto, Dleano Campos De Oliveir, Juan Andres Casquero Vera, Darius Ceburnis, Tak Chan, Tommy Chan, Sebastien Conil, Daniele Contini, Suzanne Crumeyrolle, Valentin Duflot, Konstantinos Eleftheriadis, Johan Esveld, Ekaterina Ezhova, Markus Fiebig, Shahzad Gani, Olga Garmash, Francisco J. Gómez-Moreno, Roy M. Harrison, András Hoffer, Rakesh K. Hooda, Antti Hyvärinen, Tareq Hussein, Jorma Joutsensaari, Nikos Kalivitis, Heinz Kaminski, Jutta Kesti, Radovan Krejci, Adam Kristensson, Chongai Kuang, Markku Kulmala, Lauri Laakso, Ari Leskinen, Heikki Lihavainen, Andreas Massling, Maik Merkel, Steffen Noe, Jakub Ondracek, Noemi Perez, Jean-Eudes Petit, Tuukka Petäjä, Michael Pikridas, Christopher Pöhlker, Mira Pöhlker, Jean Philippe Putaud, Ximeng Qi, Cristina Reche, Sergio Rodríguez, Petr Roztocil, Jean Sciare, Karine Sellegri, Dongjie Shang, Ashish Singh, Mikko Sipilä, Henrik Skov, Mar Sorribas, Tamanna Subba, Junying Sun, Peter Tunved, Ville Vakkari, Pieter van Zyl, Aki Virkkula, Jens Voigtländer, Kay Weinhold, Alfred Wiedensohler, Hee-Jung Yoo, Putian Zhou, Kai Puolamäki, Tuomo Nieminen, Veli-Matti Kerminen, Victoria Sinclair, and Pauli Paasonen https://doi.org/10.5281/zenodo.20202080
Daily Measurements of Number Concentrations (N100 and N50) at 74 stations (2003-2024) Aino Ovaska, Daniel Holmberg, Elio Rauth, Alghamdi Mansour, Paulo Artaxo, Eija Asmi, John Backman, Benjamin Bergmans, Matthew Boyer, Liezl Bredenkamp, Maurizio Busetto, Dleano Campos De Oliveir, Juan Andres Casquero Vera, Darius Ceburnis, Tak Chan, Tommy Chan, Sebastien Conil, Daniele Contini, Suzanne Crumeyrolle, Valentin Duflot, Konstantinos Eleftheriadis, Johan Esveld, Ekaterina Ezhova, Markus Fiebig, Shahzad Gani, Olga Garmash, Francisco J. Gómez-Moreno, Roy M. Harrison, András Hoffer, Rakesh K. Hooda, Antti Hyvärinen, Tareq Hussein, Jorma Joutsensaari, Nikos Kalivitis, Heinz Kaminski, Jutta Kesti, Radovan Krejci, Adam Kristensson, Chongai Kuang, Markku Kulmala, Lauri Laakso, Ari Leskinen, Heikki Lihavainen, Andreas Massling, Maik Merkel, Steffen Noe, Jakub Ondracek, Noemi Perez, Jean-Eudes Petit, Tuukka Petäjä, Michael Pikridas, Christopher Pöhlker, Mira Pöhlker, Jean Philippe Putaud, Ximeng Qi, Cristina Reche, Sergio Rodríguez, Petr Roztocil, Jean Sciare, Karine Sellegri, Dongjie Shang, Ashish Singh, Mikko Sipilä, Henrik Skov, Mar Sorribas, Tamanna Subba, Junying Sun, Peter Tunved, Ville Vakkari, Pieter van Zyl, Aki Virkkula, Jens Voigtländer, Kay Weinhold, Alfred Wiedensohler, Hee-Jung Yoo, Putian Zhou, Kai Puolamäki, Tuomo Nieminen, Veli-Matti Kerminen, Victoria Sinclair, and Pauli Paasonen https://doi.org/10.5281/zenodo.20200333
Model code and software
global_nconc_estimates Aino Ovaska https://version.helsinki.fi/oaino/global_nconc_estimates/-/tree/e9f752af3ccd69cc2b2137f478b3386ce496fb37/
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 574 | 385 | 177 | 1,136 | 145 | 188 | 178 |
- HTML: 574
- PDF: 385
- XML: 177
- Total: 1,136
- Supplement: 145
- BibTeX: 188
- EndNote: 178
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This work presents a global dataset of ground-level particle number concentrations for the period 2003–2024 generated by training an extreme gradient boosting machine learning regression model. It attempts to provide a validation and climatology analysis using their dataset.
The methods presented are being developed by the authors with their previous work. As such, the data are new but the method relies on their previous work for explanation. Their data seems like it would be useful for hematological studies in the future. The references and citations appear to be represented accurately. As such, the article is appropriate to support the publication.
The dataset and training data are accessible at the given identifiers. There were no error estimates provided or discussed in the article. There is much effort to validate the dataset using holdout stations, there is no discussion as to what error estimates should be used with the data. Additionally, their methods for constraining the models are not effective as the final dataset contains non physical values of N_50/N_100. This calls into question the reliability of this data set. The dataset is significant and unique, but it needs to have physically consistent values and some error estimates to be useful.
While the authors report this as a physically constrained method, it is really just a model trained with ground station data. There is no demonstration that we can expect the data at various locations to be consistent with those tested. As mentioned, the authors report nonphysical measurements in their data. I would be able to understand and use the dataset.
I recommend minor revision. I have also provided a series of specific comments and technical corrections in the attached PDF file.