From Single Compounds to Ambient Aerosols: A Machine-Learning-Based Estimation of Organic Hygroscopicity

Deshmukh, Shravan; Poulain, Laurent; Wehner, Birgit; Henning, Silvia; Herrmann, Hartmut; Pöhlker, Mira

doi:10.5194/egusphere-2026-1992

Preprints

https://doi.org/10.5194/egusphere-2026-1992

Preprints

30 Apr 2026

| 30 Apr 2026

Status: this preprint is open for discussion and under review for Atmospheric Chemistry and Physics (ACP).

From Single Compounds to Ambient Aerosols: A Machine-Learning-Based Estimation of Organic Hygroscopicity

Shravan Deshmukh, Laurent Poulain, Birgit Wehner, Silvia Henning, Hartmut Herrmann, and Mira Pöhlker

Abstract. Aerosol hygroscopicity strongly governs particle size, mixing state, and radiative effects, yet remains poorly constrained for organic aerosols due to their chemical complexity and limited observations. Here, we present laboratory-measured size-segregated hygroscopic properties of 22 organic compounds, including carboxylic acids, amino acids, sugars, and alcohols, using a hygroscopic tandem differential mobility analyzer (HTDMA) combined with chemical characterization by Aerosol Mass Spectrometry (AMS). Our results extend previous studies by resolving hygroscopic behaviour across the submicrometer size range most relevant to atmospheric processes and by systematically linking organic hygroscopicity (κ_org) across functional groups, as measured by AMS, with physicochemical properties. Structurally similar compounds may exhibit markedly different hygroscopic behavior, underscoring the role of molecular interactions. Similar to carbon chains, increased functionalization generally enhances hygroscopicity and induces a pronounced size dependence. Functional-group-based classifications from the AMS provide a useful approximation for estimating κ_org, but may not capture this complexity. Leveraging these laboratory constraints, we use a simple but extensible machine-learning framework that integrates laboratory-derived κ_org with ambient aerosol observations. The application of this hybrid approach to urban and rural environments demonstrates substantial improvements in predicting ambient hygroscopicity, with R² values increasing from 0.82 to 0.96 at the Paris suburban site SIRTA (France) and from 0.60 to 0.94 at the rural background site Goldlauter (Germany), compared to conventional composition-based models. By bridging controlled laboratory measurements with data-driven ambient analysis, this study provides a rigorous pathway to improve the representation of the direct aerosol radiative effect in atmospheric and climate models.

Received: 08 Apr 2026 – Discussion started: 30 Apr 2026

Competing interests: One co-author (B. Wehner) is a member of the editorial board for the Journal of ACP. The contact author has declared that none of the other authors has any competing interests.

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.

Download & links

Preprint (PDF, 1824 KB)

Supplement (1665 KB)

Download & links

Shravan Deshmukh, Laurent Poulain, Birgit Wehner, Silvia Henning, Hartmut Herrmann, and Mira Pöhlker

Status: open (until 19 Sep 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-1992', Anonymous Referee #1, 12 Jun 2026 reply
The manuscript "From Single Compounds to Ambient Aerosols: A Machine-Learning-Based Estimation of Organic Hygroscopicity" by Deshumkh et al. provides estimates of the hygroscopicity parameter kappa for single-component organic aerosols, and uses these hygroscopicity parameters to build a model to predict the hygroscopicity of ambient aerosol using functional groups that can be measured by aerosol mass spectrometry. Overall, this seems like a promising approach to predict the hygroscopicity of ambient aerosol. However, the manuscript text and figures need substantial revision to improve its readability.
Presentation of Figures:
Fig 2 - I believe the x-axis is dry diameter? Can the authors please clarify this.

Fig 2 - It is not clear why i) the error bars for the literature data are so large compared to this work, ii) where the literature data came from

Fig 2 - the caption says straight-cahin dicarboxylic acid particles but there are molecules that do not seem to fit this description in the panes.

Fig 2- CHOgt has never been defined.

Fig 3/4 /5/6 use k rather than kappa for the hygroscopicity parameter. These also do not look consistent between them, sometimes a capital and sometimes not?

Fig S1 and 2S both say straight-chained dicarboxylic acid particles in the caption, while none of these molecules fit into that category.

Between Fig 2 S1 and S2 there are more hygroscopicity plots than molecules in table S1. what is sucrose_gly mix in Fig S2? I do not think it is mentioned anywhere in the manuscript.

Fig s3 - the axis labels and text in the figure is all far too small to see. The x-axis label of diameter on the a-f sub plots is not clear as the label is only in the middle panel while the bottom row of the figure has every panel labeled.

Fig S4 - there are no x-axis labels included

Fig S5 - some of the x-axis labels are overlapping making them impossible to read.

Fig S6/S7 - please indicate the meaning of the pink regions in the figure captions.

Fig S8 - only one of these subplots is referred to in the text. Is the table part of the figure? Please only report the significant values of kappa and its standard deviation. The standard deviation should never have greater precision than the reported value.

Fig S9 - the text in panels b and d is too small to read.

Clarity of the text:
line 16 - what do the authors mean by 'similar to carbon chains, increased functionalization generally enhances hygroscopicity'.

line 123 - is the (Description -- Across, n.d.) text accurate? The other campaigns have site details.

The names and formation of the functional groups is not consistent throughout the text, figures, and SI. Please select one way to describe them and use consistent naming, including subscript and charge indication, if included, everywhere.

It is not clear in the modeling sections why everything is time-dependent. How is this time component handled for the laboratory-measured parameter analysis?

line 251 - should the bar be on top of the y, or is there a typo in the equation above?

line 290 - it is not clear why the hygroscopicity parameter would grow with dry particle diameter. is there previous data or theory that suggests this should be the case? What previous studies does this agree with? Fig 2- Malonic only shows one literature datapoint for 200 nm

The paragraph starting at line 285 is confusing. I think it would be helpful to have a paragraph discussing the linear dicarboxylic acids and then another paragraph discussing molecules with different functionalities, and then somehow comparing them. The way it reads now almost implies all the molecules are linear dicarboxylic acids.

hypothesis starting at line 300 - is there any previous work or modeling to back up this hypothesis, is is not clear why dry particle size alone would change the 'hygroscopicity of the organic compounds themselves.'

The authors may want to comment on the known differences in physical properties of odd and even carbon dicarboxylic acids.

the paragraph at the beginning of section 3.2 is confusing to read. A reader is unlikely to know off hand what functional groups are in each of these molecules. It would be helpful to clarify why you are comparing certain things and how those molecules differ.

The references in the SI are in author (year) but there is no bibliography included with the SI and the bibliography in the main text is numbered, not alphabetical.

a main theme in this work is that the hygroscopicity of a single-component particle is a function of size, but it is really not clear why this should be the case. how many repeated trials were performed of the hygroscopicity experiments? what kind of validation has been done to confirm this is real and significant?

How well is the model trained on data from two field campaigns likely to predict all ambient aerosol? It would be nice to see a real validation of the model using a dataset that is not used to train the model.

Other comments
is there a reason the authors have chosen to reference the 2013 IPCC report rather than the newest edition?

please check that abbreviations are defined at the first instance and then used throughout the text, there are instances of the same abbreviation being defined multiple times.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-1992-RC1
RC2:
'Comment on egusphere-2026-1992', Anonymous Referee #2, 26 Aug 2026 reply
The manuscript “From Single Compounds to Ambient Aerosols: A Machine-Learning-Based Estimation of Organic Hygroscopicity” presents a machine-learning (ML) framework for predicting aerosol hygroscopicity using laboratory hygroscopic tandem differential mobility analyzer (HTDMA) measurements combined with aerosol mass spectrometry (AMS) chemical characterization. The work falls within the scope of Atmospheric Chemistry and Physics, addresses a scientifically relevant topic, and proposes a novel approach for predicting organic aerosol hygroscopicity from chemical descriptors. The manuscript demonstrates that the ML framework improves the prediction of ambient aerosol hygroscopicity compared with Zdanovskii–Stokes–Robinson and Monte Carlo approaches.
However, the practical application of the proposed framework remains somewhat unclear, and several scientific and methodological issues should be addressed before publication.
Major Comments

The manuscript repeatedly highlights a size dependence in organic aerosol hygroscopicity. However, the physical origin of this dependence is insufficiently discussed. Could the observed size dependence be related to Kelvin-effect corrections used in the derivation of κ from HTDMA measurements? In that case, the effective water activity at the particle surface may differ from the ambient relative humidity. However, this explanation appears unlikely because the growth factor does not exhibit a consistent dependence on particle size across different compounds. Some compounds show increasing hygroscopicity with size, others decreasing hygroscopicity, and some show little dependence at all.

The Introduction omits discussion of established thermodynamic frameworks used for predicting hygroscopic growth in inorganic-organic aerosol mixtures, such as AIOMFAC, E-AIM, and UManSysProp. Why is an ML approach needed? What limitations of existing thermodynamic models does it address? Does the proposed framework improve predictive skill, computational efficiency, applicability, or interpretability? At minimum, the authors should discuss these existing approaches in the Introduction and clearly position their work relative to them. Ideally, some comparison against one or more thermodynamic models should be included.

The manuscript repeatedly claims that the framework can improve the representation of aerosol hygroscopicity in atmospheric and climate models (e.g. lines 25, 102, and 496). However, no concrete pathway for incorporating the framework into such models is presented. The ML framework relies on AMS-derived family-group information and O:C ratios, whereas global climate models typically do not simulate aerosol composition at this level of chemical detail. The authors should clarify, how the proposed framework could realistically be implemented in large-scale atmospheric models. Which model variables would be used as inputs?

In the conclusion (line 495), the authors state: “Together, this work provides a scalable, physically grounded pathway from molecular-scale organic composition to predictions of aerosol growth, cloud interactions, ...” This statement appears too strong given the evidence presented. The framework is trained and evaluated using HTDMA-derived hygroscopicity under subsaturated conditions. No cloud condensation nuclei (CCN) measurements are included. It is therefore unclear whether the predicted κ values can be directly applied to cloud activation studies. Previous studies show difference in hygroscopicities in sub-saturated conditions and those based on CCN measurements (e.g., Petters, M. D. and Kreidenweis, S. M. (2007), A single parameter representation of hygroscopic growth and cloud condensation nucleus activity, Atmos. Chem. Phys., 7, 1961-1971.)

Specific Comments
Line 172: I do not understand the following sentence: “Changing κ_i in calculations doesn’t significantly change κ_chem prediction, as lower soluble compounds like organics generally show only slight hygroscopic growth at relative humidities less than 98% (Petters et al., 2009; Wex et al., 2009)”. Please clarify the reasoning. The statement appears inconsistent with the demonstrated importance of organic hygroscopicity throughout the manuscript.

Line 231: Monte Carlo is not a machine-learning model. The sentence should be revised accordingly. There is also a typo in the beginning of the sentence. “Four-machine learning models” should be “Four machine learning models”

Technical Comments
The manuscript would benefit from a thorough language check. For example: Missing articles (“a”, “an”, “the”). Inconsistent verb tenses. Occasional grammatical errors. Several overly long and difficult-to-follow sentences.

Figure 1: The figure contains text that is difficult to read (e.g. “Decision Tree (1)” and “Regression time”). Because the figure is a bitmap rather than vector graphics, readability does not improve when zooming. I recommend increasing font sizes and providing a higher-resolution or vector version of the figure.

Recommendation
The manuscript addresses an important topic and presents an interesting and potentially useful ML-based framework. However, the scientific motivation and practical applicability of the approach require further clarification, particularly with respect to existing thermodynamic models, the physical interpretation of the observed size dependence, and the pathway toward atmospheric-model applications. I therefore recommend major revisions.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-1992-RC2

Shravan Deshmukh, Laurent Poulain, Birgit Wehner, Silvia Henning, Hartmut Herrmann, and Mira Pöhlker

Supplement

https://doi.org/10.5194/egusphere-2026-1992-supplement

Shravan Deshmukh, Laurent Poulain, Birgit Wehner, Silvia Henning, Hartmut Herrmann, and Mira Pöhlker

Viewed

Total article views: 518 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
377	96	45	518	46	40	41

HTML: 377
PDF: 96
XML: 45
Total: 518
Supplement: 46
BibTeX: 40
EndNote: 41

Views and downloads (calculated since 30 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	49	11	3	63
May 2026	251	39	13	303
Jun 2026	37	14	7	58
Jul 2026	28	22	6	56
Aug 2026	12	10	16	38

Cumulative views and downloads (calculated since 30 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	49	11	3	63
May 2026	251	39	13	303
Jun 2026	37	14	7	58
Jul 2026	28	22	6	56
Aug 2026	12	10	16	38

Viewed (geographical distribution)

Total article views: 494 (including HTML, PDF, and XML) Thereof 494 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 30 Aug 2026

Short summary

Improved representation of aerosol hygroscopicity is critical for understanding how atmospheric particles influence cloud formation, climate, air quality, and human health. This work advances understanding of aerosol hygroscopicity by connecting laboratory measurements of organic aerosol properties with ambient observations through an explainable machine learning framework.


Total:	0
HTML:	0
PDF:	0
XML:	0