Machine learning significantly improves the simulation of hourly-to-yearly scale cloud nuclei concentration and radiative forcing in polluted atmosphere

Ren, Jingye; Zou, Songjian; Xu, Honghao; Liu, Guiquan; Wang, Zhe; Zhang, Anran; Zhao, Chuanfeng; Hu, Min; Shang, Dongjie; Tang, Lizi; Huang, Ru-Jin; Sun, Yele; Zhang, Fang

doi:https://doi.org/10.5194/egusphere-2025-1483

Preprints

https://doi.org/10.5194/egusphere-2025-1483

Preprints

03 Jun 2025

| 03 Jun 2025

Machine learning significantly improves the simulation of hourly-to-yearly scale cloud nuclei concentration and radiative forcing in polluted atmosphere

Jingye Ren, Songjian Zou, Honghao Xu, Guiquan Liu, Zhe Wang, Anran Zhang, Chuanfeng Zhao, Min Hu, Dongjie Shang, Lizi Tang, Ru-Jin Huang, Yele Sun, and Fang Zhang

Abstract. The accurate prediction of cloud condensation nuclei (CCN) number concentration (N_CCN) on a large spatiotemporal scale is challenging but critical to evaluate the aerosol cloud interaction (ACI) effect. Combining multi-source dataset and the N_CCN simulated by the Weather Research and Forecasting coupled with Chemistry (WRF-Chem) model, we have developed a new machine learning-based model which predicts well both regional and hourly-to-yearly scale N_CCN at typical supersaturations in the North China Plain (NCP). We show that the prediction bias of N_CCN compared to observations is reduced from -39 % with the WRF-Chem model to approximately -8 % with the new model. The greatest improvement is seen in polluted cases. The new model captures well the spatial variation and better describes long-term trends of N_CCN than the WRF-Chem. More importantly, the study reveals a significant long-term decreasing trend of N_CCN in NCP due to a rapid reduction in aerosol concentrations from 2014 to 2018, during which a series of strict emission reduction measures were implemented by the Chinese government. This reflects the climate benefit of pollution control. Our study further illustrates that the new model reduces the uncertainty in simulating cloud radiative forcing from an overestimation of 1.07±0.76 W m^-2 to only 0.18±0.65 W m^-2, illustrating the high sensitivity of climate forcing to changes in N_CCN. This work offers a new modeling framework that has the potential to greatly improve the assessment of the ACI effect in current models, and guides the way to simulate CCN in other regions around the world.

Received: 28 Mar 2025 – Discussion started: 03 Jun 2025

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, 2398 KB)

Supplement (2763 KB)

Download & links

Jingye Ren, Songjian Zou, Honghao Xu, Guiquan Liu, Zhe Wang, Anran Zhang, Chuanfeng Zhao, Min Hu, Dongjie Shang, Lizi Tang, Ru-Jin Huang, Yele Sun, and Fang Zhang

Status: final response (author comments only)

CEC1:
'Comment on egusphere-2025-1483 - No compliance with the policy of the journal', Juan Antonio Añel, 23 Jun 2025

Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".

https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
Beyond one Zenodo repository which contains the machine learning model and the meteorological input variables hosted in the RDA-NCAR, none of the other sites that you cite to get access to the code (e.g. WRF) or data, are valid repositories for scientific publication, and they do not comply with the requirements exposed in the policy of the journal.
Therefore, the current situation with your manuscript is irregular, as we can not accept manuscripts in Discussions that do not comply with our policy. Please, publish your code and data in one of the appropriate repositories according to our policy and reply as soon as possible to this comment with a modified 'Code and Data Availability' section for your manuscript, which must include the relevant information (link and permanent identifier (e.g., handle, DOI)) of the new repositories, and which you must include in a potentially reviewed manuscript.
I must note that if you do not fix this problem, we will have to reject your manuscript for publication in our journal.
Juan A. Añel

Geosci. Model Dev. Executive Editor

Citation: https://doi.org/10.5194/egusphere-2025-1483-CEC1
- AC1:
  'Reply on CEC1', Fang Zhang, 23 Jun 2025
  
  Re: Thank you for your efforts and time on handling the paper. The source codes of WRF-Chem, Python and the Scikit-Learn machine learning library have been revised in the Code and Data availability. See as follows: “
  Code and Data availability
  The data and code are publicly accessible at https://zenodo.org/records/15523200 (Ren et al., 2025). This includes the machine learning code, the corresponding training and testing dataset (chemical compositions, gaseous pollutants, meteorological datasets and simulated CCN concentration from WRF-Chem) and the observation CCN concentrations, the script and namelist file used in WRF-Chem and the scripts used for plotting, supporting the findings of this study. The release version of WRF-Chem source code is archived on GitHub (https://github.com/wrf-model/WRF, last access: May, 2025). The release version of Python and the Scikit-Learn machine learning library are open source from https://github.com/python and https://github.com/scikit-learn.”
  
  Citation: https://doi.org/10.5194/egusphere-2025-1483-AC1
  - CEC2: 'Reply on AC1', Juan Antonio Añel, 24 Jun 2025
    
    Dear authors,
    Unfortunately, your reply does not solve some of the issues regarding the compliance with our policy. First, you have stored part of your assets in a GitHub site. However, GitHub is not a suitable repository for scientific publication. GitHub itself instructs authors to use other long-term archival and publishing alternatives, so you must store the assets that you have linked to GitHub in one of the suitable repositories according to our policy.
    A similar issue happens with the data. You have hosted them in sites that do not comply with our policy (e.g. acom.ucar.edu or meicmodel.org.cn). You must store them in one of the suitable repositories.
    Thererefore, please, address these issues and reply to this comment with the information for the new repositories.
    Juan A. Añel
    Geosci. Model Dev. Executive Editor
    
    Citation: https://doi.org/10.5194/egusphere-2025-1483-CEC2
    
    AC2: 'Reply on CEC2', Fang Zhang, 25 Jun 2025
    
    Re: Thank you for your efforts and time on handling the paper. We have updated the section of Code and Data availability and see as follows: “
    Code and Data availability
    The data and code are publicly accessible at https://zenodo.org/records/15737652 (Ren et al., 2025). This includes the WRF-Chem model version 4.1.5 used in this study, the machine learning code, the corresponding training, testing datasets and the CCN observation datasets, the emissions inventory and scripts used in WRF-Chem and the scripts used for plotting, supporting the findings of this study. The release version of WRF-Chem is also open-access and can be publicly available at NCAR https://www2.mmm.ucar.edu/wrf/users/download/get_source.html (Skamarock et al., 2019, last access: 10 May 2025). The initial meteorological variables are from the National Center for Environmental Prediction's Final Operational Global (NCEP/FNL) and available at https://doi.org/10.5065/D6M043C6 (NCEP, 2000).
    References:
    Ren, J., Zou, S., Xu, H., Liu, G., Wang, Z., Zhang, A., Zhao, C., Hu, M., Shang, D., Tang, L., Huang, R.-J., Sun, Y., & Zhang, F.: Machine learning significantly improves the simulation of hourly-to-yearly scale cloud nuclei concentration and radiative forcing in polluted atmosphere [Data set]. Zenodo. https://zenodo.org/records/15737652, 2025.
    Skamarock, W., Klemp, J., Dudhia, J., Gill, D. O., Liu, Z., Berner, J., Wang, W., Powers, J. G., Duda, M. G., Barker, D., and Huang, X.-Y.: A Description of the Advanced Research WRF Model Version 4.1, UCAR/NCAR, https://doi.org/10.5065/1dfh-6p97, 2019 (code available at https://www2.mmm.ucar.edu/wrf/users/download/ get_source.html, last access: 10 May 2025).
    NCEP: NCEP FNL Operational Model Global Tropospheric Analyses, continuing from July 1999, National Centers for Environ mental Prediction [Data set], https://doi.org/10.5065/D6M043C6, 2000 (last access: 10 May 2025).“
    
    Citation: https://doi.org/10.5194/egusphere-2025-1483-AC2
    
    CEC3: 'Reply on AC2', Juan Antonio Añel, 26 Jun 2025
    
    Dear authors,
    Many thanks for your reply. We can consider now the current version of your manuscript in compliance with our code and data policy.
    Juan A. Añel
    Geosci. Model Dev. Executive Editor
    
    Citation: https://doi.org/10.5194/egusphere-2025-1483-CEC3
RC1:
'Comment on egusphere-2025-1483', Anonymous Referee #1, 01 Jul 2025
Ren et al. used the Random Forest Regression (RFR) model to predict cloud condensation nuclei (CCN) concentrations in the North China Plain. Their results showed a reduced prediction bias compared to WRF-Chem. Moreover, by incorporating observational data such as PM2.5, NO2, SO2, the model captured the long-term decreasing trend in aerosol concentration from 2014 to 2018.
While the approach and the results are interesting, I have several concerns regarding the current manuscript:
The methodology particularly and the manuscript generally lacks critical information for readers to follow. Why the authors decided to put information such as study domain, details about WRF-Chem, RFR configurations/training and validating and more in the Supplemental Information (SI)? I find it’s very hard to understand and follow the methodology session. For example, it is unclear whether nitrate was the output of the WRF-Chem? A minor point: “new model” is not a preferred terminology in the result figures.

While the relative importance of the input parameters is informative, it does not adequately explain the substantial improvement in CCN predictions. With a large bias in WRF-Chem simulations (~39%), it is unclear why the authors included WRF-Chem outputs as predictors? On the other hand, if the authors only used observational data to train the ML model, then the current results would not be apple to apple comparison. A more appropriate benchmark would be a comparison between the RFR and other machine learning methods used in previous studies.
Citation: https://doi.org/10.5194/egusphere-2025-1483-RC1
- AC3:
  'Reply on RC1', Fang Zhang, 05 Aug 2025
  Referee 1
  Ren et al. used the Random Forest Regression (RFR) model to predict cloud condensation nuclei (CCN) concentrations in the North China Plain. Their results showed a reduced prediction bias compared to WRF-Chem. Moreover, by incorporating observational data such as PM_2.5, NO₂, SO₂, the model captured the long-term decreasing trend in aerosol concentration from 2014 to 2018.
  While the approach and the results are interesting, I have several concerns regarding the current manuscript:
  The methodology particularly and the manuscript generally lacks critical information for readers to follow. Why the authors decided to put information such as study domain, details about WRF-Chem, RFR configurations/training and validating and more in the Supplemental Information (SI)? I find it’s very hard to understand and follow the methodology session.
  
  Re: Thank you for your efforts and time on handling the paper. We have put more information in the main text of the paper and updated the section of Methods and see as follows or Lines 106-225: “
  Methods
  
  2.1 Study area
  In this work, we select the North China Plain (NCP) (32°-40°N and 114°-121°E) as the study area. Being one of the most polluted areas in China, the aerosol particles in NCP are with more complex composition and mixing state, which leads to great challenge in accurate prediction of cloud concentration nuclei (CCN) concentrations. In recent years, emissions of gas pollutants and fine particles have shown a significant downward trend year by year (Wei et al., 2023) due to the implementation of the vigorous emission reduction in China (Zheng et al., 2018). This also makes changes in aerosols CCN activity in the study area from the point of view in assessment of the climate effect of aerosols.
  2.2 Model construction and validation
  Here we develop the ML-based N_CCN prediction model by employing the Random Forest Regression method (RFRM) that has been demonstrated and can solve multivariate and nonlinear regression problems (Nair and Yu, 2020; Liang et al., 2022).
  The diagram of the model construction and the N_CCN prediction is shown in Figure 1. Due to lack of a large spatial scale observed N_CCN data, we use simulated N_CCN by WRF-Chem model as the targeted variable that can basically capture the ambient temporal variability of CCN concentration despite a deviation of ~40% by comparing with our six field observations (Figure 4). The input parameters include the chemical components of PM_2.5 (organic, sulfate, nitrate, ammonium, black carbon) from the Tsinghua University Tracking Air Pollution in China dataset (Liu et al., 2022) and gas and particulate pollutants (nitrogen dioxide (NO₂), sulfur dioxide (SO₂), carbon monoxide (CO), ozone (O₃) and PM_2.5) collected from the China National Environmental Monitoring Centre network. Meteorological parameters are from the European Centre for Medium-range Weather Forecasts Reanalysis version 5 (ERA-5) and include temperature, relative humidity (RH), total precipitation (TP), wind speed (WS), wind direction (WD), planetary boundary layer height (BLH), surface pressure (SP) and surface net solar radiation (SNSR). These datasets have undergone validation and been proven highly suitable for developing machine learning models for atmospheric applications (Nair and Yu, 2020; Wei et al., 2023). Cartesian coordinates were also added as input due to the spatiotemporal nature of the input data (Yang et al., 2022). Supplemental Table 1 provides more details about the input parameters.
  When constructing the model, all the aforementioned species data (predictor and target variables) were processed to match the same spatiotemporal resolution, and then split into 7:3 ratio for model training and testing respectively. As a result, a total of 274365 samples is included in the training datasets. In order to assure a stronger generalization ability of the N_CCN prediction model, the 10-fold cross-validation is adopted (Wei et al., 2023). The optimization parameters of RF model were examined by varying hyperparameters (Fig. S1). In addition, cross-validation (CV) is applied to select the hyperparameters during the data preprocessing (Yang et al., 2022). The CV results showed that when the number of trees (n_estimators) was less than 200, the prediction accuracy increased rapidly with the increase of the number of trees, and then gradually stabilized. According to the CV score and the number of data sample, the number of trees was set to 500 in this study. The impact of max depth on the CV score showed that, with the increase of depth, the complexity of the model increases. Thus, the max depth is set to 28. Also, the model generalization error was larger when the minimum sample number of the leaf and branch node are large, indicating that the model itself is close to the optimal model complexity level. Therefore, a higher value was set given the large sample size in this case. The influence of the maximum selection feature number on CV score showed a trend of increasing first and then decreasing, so the maximum value of CV curve was set to 16.
  
  Figure 2. Comparison of RFRM retrieval and WRF-Chem simulated N_CCN at S=0.2%. (a and b) Density plots of retrieval N_CCN at S=0.2% as a function of the simulations from WRF-Chem on the testing dataset.
  Fig. 2 shows the performance of our developed RFRM models by comparing the testing dataset (sample size: N=117585) of CCN simulated by WRF-Chem with the predicted CCN. Here the quality metrics for model performance are based on the correlation coefficient (R²), root mean square error (RMSE) and the slope of the RFRM predicted and the WRF-Chem simulated CCN concentrations. It showed that the estimated N_CCN at S=0.2% are highly correlated with the values from WRF-Chem, with the correlation R² of ~0.89-0.91 and slopes of 0.83 and 0.86 in our developed RFRM model. This suggests that our model works well in estimating N_CCN with a high aerosol loading environment. We also found that the accuracy of the CCN prediction will deteriorate slightly if not including the information of chemical compositions (Fig. S2), or if using XGBoost algorithm (Fig.S3) when constructing the model.
  Given that the ultimate goal of model development is to more accurately predict the actual atmospheric concentration of CCN, we further conducted a comparative analysis of the prediction results against the in-situ observation data in this region, spanning from the hourly scale to the interannual scale (see Figure 3-6), which will be presented and discussed in detail in Section 3.2 and 3.3.
  2.3 Data and other details in the model construction
  2.3.1 N_CCN simulated by WRF-Chem model
  The WRF-Chem version 4.1.5 is used to simulate N_CCN in this study, which nested a domain in 10 km×10 km covering the entire NCP (Fig. S4) and contained 181×170 grids. The simulation in WRF-Chem is conducted from 1 January 2014 to 31 December 2018 with an hourly resolution. In the WRF-Chem modeling system, the sectional Model for Simulating Aerosol Interactions and Chemistry (MOSAIC), the Morrison two-moment scheme (Morrison et al., 2009) and the Carbon Bond Mechanism Z photochemical mechanism (Zaveri et al., 1999) are employed. We also compared the simulation using the Regional Acid Deposition Model (Stockwell et al., 1990) and the Lin microphysics scheme (Lin et al., 1983). Considering the calculation efficiency and accuracy with the measurements, the CBMZ-MOSAIC and Morrison 2-monment scheme were finally applied to simulate the long-term CCN concentration. More details about the other parameterizations used for WRF-Chem simulation were given in SI.
  2.3.2 Gound-based measurements and datasets
  Ground measurements of atmospheric gaseous precursors, fine particles chemical compositions, and CCN number concentration (at supersaturations of 0.2% and 0.4%) were collected during six field campaigns at three sites in the NCP (Fig. 4), used to assess the performance of the developed ML-based model in predicting N_CCN. The six campaigns were conducted as follows: at the Beijing (BJ) site from 8–30 November 2014, 20 August to 6 October 2015, 16 November to 20 December 2016, and 28 May to 27 June 2017; at the Xingtai (XT) site from 17 May to 14 June 2016; and at the Gucheng (GC) site from 23 January to 3 February 2018. They are accordingly named BJ2014_WIN, BJ2015_AUT, BJ2016_WIN, BJ2017_SUM, XT2016_SUM, and GC2018_WIN (Fig. 4a).
  The BJ site (Longitude: 116.37° E; Latitude: 39.97° N) is located at the meteorological tower station of the Institute of Atmospheric Physics, Chinese Academy of Sciences. It is representative of the general emission conditions in urban areas of the northern NCP. The primary pollution sources here are surrounding traffic and residential emissions. The XT site (Longitude: 114.37° E; Latitude: 37.18° N) is situated at a national weather station. It is primarily influenced by emissions from surrounding towns and factories (e.g., coal-fired power plants, coking, steel, cement, and chemical industries) and thus reflects polluted suburban conditions in the southern NCP. The GC site (Longitude: 115.74° E; Latitude: 39.15° N) is located at the Integrated Ecological-Meteorological Observation and Experiment Station of the Chinese Academy of Meteorological Sciences. Surrounded mainly by nearby villages, farmland, and transportation networks, this site represents the regional background pollution in the northern NCP.
  The CCN number concentrations were measured by using the Droplet Measurement Technologies CCN counter (model CCNC-100, DMT Inc. Lance et al., 2006) at BJ and XT site. The supersaturation (S) levels set for each CCN measurement cycle were 0.1%, 0.2%, 0.4%, and 0.8%, respectively. Another measurement at GC site was referred from Zhang et al. (2020). In this study, the comparisons between the measured and predicted N_CCN were mostly based on the value at S=0.2% and S=0.4%. The observed N_CCN varies from a few hundred to tens of thousands at these sites, and the campaign mean mass concentration of PM_2.5 ranges from 35.6 to 160 μg m^-3 (Fig. 4b), indicating that the observations can represent various atmospheric conditions, spanning from clean to polluted in the region. More details about the observations could be found in Fan et al. (2020), Ren et al. (2018), and Zhang et al. (2019). In addition, the long-term measurement of particle number size distribution (PNSD) at a field site in Beijing (Fig. S5, Shang et al., 2022) is also used for deriving the long-term trend of yearly averaged N_CCN.
  For example, it is unclear whether nitrate was the output of the WRF-Chem?
  Re: The nitrate was from the Tsinghua University Tracking Air Pollution in China dataset (Liu et al., 2022). We have provided a detailed introduction to the data in the methodology section and see as follows or Lines 125-127:
  “…The input parameters include the chemical components of PM_2.5 (organic, sulfate, nitrate, ammonium, black carbon) from the Tsinghua University Tracking Air Pollution in China dataset (Liu et al., 2022) …”
  A minor point: “new model” is not a preferred terminology in the result figures.
  Re: The “new model” has been replaced with “RFRM” in the result figures.
  While the relative importance of the input parameters is informative, it does not adequately explain the substantial improvement in CCN predictions. With a large bias in WRF-Chem simulations (~39%), it is unclear why the authors included WRF-Chem outputs as predictors?
  Re: Currently, N_CCN observation is still very lacking, mostly consisting of single field observations (Ren et al., 2018; Zhang et al., 2019). The N_CCN from satellite‐retrieved method can be as high as there are usually large deviations of -30%–+90% in the estimation accuracy (Shen et al., 2019). Numerical model has been demonstrated that it can capture the relative amplitude of the variability of the aerosol particle number concentration and CCN number concentration (Fanourgakis et al., 2019; Nair and Yu, 2020). To develop a spatiotemporal-scale model for predicting CCN concentrations, the CCN concentration output by WRF-Chem is used as the target variable. We have added some explanations as follows or see Lines 122-125:
  “…Due to lack of a large spatial scale observed N_CCN data, we use simulated N_CCN by WRF-Chem model as the targeted variable that can basically capture the ambient temporal variability of CCN concentration despite a deviation of ~40% by comparing with our six field observations (Figure 4) …”
  On the other hand, if the authors only used observational data to train the ML model, then the current results would not be apple to apple comparison. A more appropriate benchmark would be a comparison between the RFRM and other machine learning methods used in previous studies.
  Re: Thank you for your suggestion. We have added detailed description regarding the RFRM model construction, and also compared with XGBoost model as follows or see Lines 117-173:
  “… 2.2 Model construction and validation
  Here we develop the ML-based N_CCN prediction model by employing the Random Forest Regression method (RFRM) that has been demonstrated and can solve multivariate and nonlinear regression problems (Nair and Yu, 2020; Liang et al., 2022).
  The diagram of the model construction and the N_CCN prediction is shown in Figure 1. Due to lack of a large spatial scale observed N_CCN data, we use simulated N_CCN by WRF-Chem model as the targeted variable that can basically capture the ambient temporal variability of CCN concentration despite a deviation of ~40% by comparing with our six field observations (Figure 4). The input parameters include the chemical components of PM_2.5 (organic, sulfate, nitrate, ammonium, black carbon) from the Tsinghua University Tracking Air Pollution in China dataset (Liu et al., 2022) and gas and particulate pollutants (nitrogen dioxide (NO₂), sulfur dioxide (SO₂), carbon monoxide (CO), ozone (O₃) and PM_2.5) collected from the China National Environmental Monitoring Centre network. Meteorological parameters are from the European Centre for Medium-range Weather Forecasts Reanalysis version 5 (ERA-5) and include temperature, relative humidity (RH), total precipitation (TP), wind speed (WS), wind direction (WD), planetary boundary layer height (BLH), surface pressure (SP) and surface net solar radiation (SNSR). These datasets have undergone validation and been proven highly suitable for developing machine learning models for atmospheric applications (Nair and Yu, 2020; Wei et al., 2023). Cartesian coordinates were also added as input due to the spatiotemporal nature of the input data (Yang et al., 2022). Supplemental Table 1 provides more details about the input parameters.
  When constructing the model, all the aforementioned species data (predictor and target variables) were processed to match the same spatiotemporal resolution, and then split into 7:3 ratio for model training and testing respectively. As a result, a total of 274365 samples is included in the training datasets. In order to assure a stronger generalization ability of the N_CCN prediction model, the 10-fold cross-validation is adopted (Wei et al., 2023). The optimization parameters of RF model were examined by varying hyperparameters (Fig. S1). In addition, cross-validation (CV) is applied to select the hyperparameters during the data preprocessing (Yang et al., 2022). The CV results showed that when the number of trees (n_estimators) was less than 200, the prediction accuracy increased rapidly with the increase of the number of trees, and then gradually stabilized. According to the CV score and the number of data sample, the number of trees was set to 500 in this study. The impact of max depth on the CV score showed that, with the increase of depth, the complexity of the model increases. Thus, the max depth is set to 28. Also, the model generalization error was larger when the minimum sample number of the leaf and branch node are large, indicating that the model itself is close to the optimal model complexity level. Therefore, a higher value was set given the large sample size in this case. The influence of the maximum selection feature number on CV score showed a trend of increasing first and then decreasing, so the maximum value of CV curve was set to 16.
  
  Figure 2. Comparison of RFRM retrieval and WRF-Chem simulated N_CCN at S=0.2%. (a and b) Density plots of retrieval N_CCN at S=0.2% as a function of the simulations from WRF-Chem on the testing dataset.
  Fig. 2 shows the performance of our developed RFRM models by comparing the testing dataset (sample size: N=117585) of CCN simulated by WRF-Chem with the predicted CCN. Here the quality metrics for model performance are based on the correlation coefficient (R²), root mean square error (RMSE) and the slope of the RFRM predicted and the WRF-Chem simulated CCN concentrations. It showed that the estimated N_CCN at S=0.2% are highly correlated with the values from WRF-Chem, with the correlation R² of ~0.89-0.91 and slopes of 0.83 and 0.86 in our developed RFRM model. This suggests that our model works well in estimating N_CCN with a high aerosol loading environment. We also found that the accuracy of the CCN prediction will deteriorate slightly if not including the information of chemical compositions (Fig. S2), or if using XGBoost algorithm (Fig. S3) when constructing the model.
  
  Figure S2. Comparison of RFRM-ShortVars model retrieval and WRF-Chem simulated NCCN at S=0.2%. (a and b) Density plots of retrieval NCCN at S=0.2% as a function of the simulations from WRF-Chem on the testing dataset.
  Figure S3. Same as Figure S2 but from XGBoost model.
  Given that the ultimate goal of model development is to more accurately predict the actual atmospheric concentration of CCN, we further conducted a comparative analysis of the prediction results against the in-situ observation data in this region, spanning from the hourly scale to the interannual scale (see Figure 3-6), which will be presented and discussed in detail in Section 3.2 and 3.3 …”
  References:
  Morrison, H., Thompson, G., Tatarskii, V.: Impact of cloud microphysics on the development of trailing stratiform precipitation in a simulated squall line: Comparison of one-and two-moment schemes, Monthly Weather Review, 137(3), 991–1007, https://doi.org/10.1175/2008MWR2556.1, 2009.
  Zaveri, R. A., Peters, L. K.: A new lumped structure photochemical mechanism for large-scale applications, Journal of Geophysical Research: Atmospheres, 104, 30387–30415, https://doi.org/10.1029/1999JD900876, 1999.
  Stockwell, W. R., Middleton, P., Chang, J. S., et al.: The second-generation regional acid deposition model chemical mechanism for regional air quality modeling, Journal of Geophysical Research: Atmospheres, 95(D10): 16343-16367, https://doi.org/10.1029/JD095iD10p16343, 1990.
  Lin, Y., Farley, R., Orville, H.: Bulk parameterization of the snow field in a cloud model, Journal of Applied Meteorology and Climatology, 22(6): 1065-1092, https://doi.org/10.1175/1520-0450(1983)022<1065:BPOTSF>2.0.CO;2, 1983.
  Lance, S., Nenes, A., Medina, J. et al.: Mapping the operation of the DMT continuous flow CCN counter, Aerosol Science and Technology, 40(4), 242–254, https://doi.org/10.1080/02786820500543290, 2006.
  Liang, M., Tao, J., Ma, N. et al.: Prediction of CCN spectra parameters in the North China Plain using a random forest model, Atmospheric Environment, 289, 119323, https://doi.org/10.1016/j.atmosenv.2022.119323, 2022.
  Liu, S., Geng, G., Xiao, Q., Zheng, Y., Liu, X., Cheng, J., & Zhang, Q.: Tracking daily concentrations of PM2.5 chemical composition in China since 2000, Environ Sci Technol, 56, 16517–16527, https://doi.org/10.1021/acs.est.2c06510, 2022.
  Fan, X., Liu, J., Zhang, F. et al.: Contrasting size-resolved hygroscopicity of fine particles derived by HTDMA and HR-ToF-AMS measurements between summer and winter in Beijing: the impacts of aerosol aging and local emissions, Atmos. Chem. Phys., 20, 915–929, https://doi.org/10.5194/acp-20-915-2020, 2020.
  Ren, J., Zhang, F., Wang, Y. et al.: Using different assumptions of aerosol mixing state and chemical composition to predict CCN concentrations based on field measurements in urban Beijing, Atmos. Chem. Phys., 18(9), 6907–6921, https://doi.org/10.5194/acp-18-6907-2018, 2018.
  Yang, N., Shi, H., Tang, H. et al.: Geographical and temporal encoding for improving the estimation of PM_2.5 concentrations in China using end-to-end gradient boosting, Remote Sensing of Environment, 269, 112828, https://doi.org/10.1016/j.rse.2021.112828, 2022.
  Zhang, Y., Tao, J., Ma, N. et al.: Predicting cloud condensation nuclei number concentration based on conventional measurements of aerosol properties in the North China Plain, Science of The Total Environment, 719, 137473, https://doi.org/10.1016/j.scitotenv.2020.137473, 2020.
  Wei, J., Li, Z., Wang, J. et al.: Ground-level gaseous pollutants (NO₂, SO₂, and CO) in China: daily seamless mapping and spatiotemporal variations, Atmos. Chem. Phys., 23, 1511–1532, https://doi.org/10.5194/acp-23-1511-2023, 2023.
  Zhang, F., Ren, J., Fan, T. et al.: Significantly enhanced aerosol CCN activity and number concentrations by nucleation-initiated haze events: A case study in urban Beijing, Journal of Geophysical Research: Atmospheres, 124(24), 14102–14113, https://doi.org/10.1029/2019JD031457, 2019.
  Shen, Y., Virkkula, A., Ding, A. et al.: Estimating cloud condensation nuclei number concentrations using aerosol optical properties: role of particle number size distribution and parameterization, Atmos. Chem. Phys., 19(24), 15483–15502, https://doi.org/10.5194/acp-19-15483-2019, 2019.
  Nair, A. A., Yu, F.: Using machine learning to derive cloud condensation nuclei number concentrations from commonly available measurements, Atmos. Chem. Phys., 20(21), 12853–12869, https://doi.org/10.5194/acp-20-12853-2020, 2020.
  Fanourgakis, G., Kanakidou, M., Nenes, A. et al.: Evaluation of global simulations of aerosol particle and cloud condensation nuclei number, with implications for cloud droplet formation, Atmos. Chem. Phys., 19(13), 8591–8617, https://doi.org/ 10.5194/acp-19-8591-2019, 2019.
  Zheng, B., Tong, D., Li, M., Liu, F. et al.: Trends in China’s anthropogenic emissions since 2010 as the consequence of clean air actions, Atmos. Chem. Phys., 18, 14095-14111, https://doi.org/10.5194/acp-18-14095-2018, 2018.
  Shang, D., Tang, L., Fang, X. et al.: Variations in source contributions of particle number concentration under long-term emission control in winter of urban Beijing, Environmental Pollution, 304, 119072, https://doi.org/10.1016/j.envpol.2022.119072, 2022.
  
  Citation: https://doi.org/10.5194/egusphere-2025-1483-AC3
RC2:
'Comment on egusphere-2025-1483', Anonymous Referee #2, 02 Jul 2025
The manuscript by Ren et al. presents a machine learning (ML) model based on Random Forest Regression (RFRM) to predict cloud condensation nuclei number concentration (N_CCN) at typical supersaturations in the North China Plain (NCP), which they then demonstrate that their model reduces the prediction bias from 39% (WRF-Chem) to 8%. The study also analyzes the importance of different input factors, evaluates the model’s performance in simulating spatiotemporal variability, and quantifies the reduction in cloud radiative forcing uncertainty achieved by mitigating N_CCN simulation biases. While the topic is highly relevant to GMD and represents a potential advancement in climate modeling, I have significant concerns regarding the methodological approach and validation logic, as well as the clarity of the manuscript. I recommend returning the manuscript to the authors with encouragement to revise and resubmit after addressing the following concerns.

Major Concerns
The training and testing datasets use WRF-Chem simulations as the output variable (target), but model performance is evaluated against observations. From a machine learning perspective, if the target variable (output) in both training and testing is WRF-Chem simulations, the model’s objective should be to approximate WRF-Chem’s output—not observations. Thus, model performance should be assessed based on how well it replicates WRF-Chem’s results, not observations. The authors report that WRF-Chem exhibits a significant bias (~39%) compared to observations, yet their ML model, trained to emulate WRF-Chem, shows a much smaller bias (~8%). This discrepancy is counterintuitive and requires a thorough explanation, which is currently missing.

Use multisource datasets as input and WRF-Chem simulations as output to make training and test sets. Table S1 indicates that all input variables used in the ML model are also available from WRF-Chem outputs. However, the author uses datasets from different sources and different spatiotemporal resolutions to construct the data set through interpolation. The authors provide no rationale and advantage of this approach. Conventionally, one would train ML models using consistent model outputs and later test with observational inputs to assess potential gains.

The model is trained on data from 378 monitoring sites across the NCP (2014–2018) but evaluated using three sites (Beijing, Xingtai, Gucheng) within the same region and time period. This does not test generalizability. True validation requires spatiotemporal extrapolation: e.g., evaluation outside the training period or region. Performance on training-era/training-region data is expected to be favorable and does not demonstrate robust predictive skill.

Other Major Concerns:
The manuscript positions itself as unique by focusing on polluted regions (Lines 88–95), yet only 6 field campaigns are used for evaluation, with no dedicated analysis of heavy-pollution events. Additionally, heavy reliance on "mean prediction bias" is misleading: if RMSE/MSE is the training loss (unstated in the text), ML models inherently bias predictions toward the mean. Therefore, the improvement of the "mean prediction bias" cannot fully prove the performance of the ML model in real scenarios, and it is more meaningful to conduct a detailed evaluation and analysis of a single severe pollution event.

A common but concerning trend in ML applications is showcasing successes while neglecting failures. This paper follows that pattern. There is no discussion of scenarios where the model underperforms, its limitations, or potential pitfalls. For instance, an eager graduate student might misuse this model for policy analysis without realizing its constraints (e.g., lack of generalizability), leading to significant wasted effort. A rigorous journal paper must present a balanced view of model capabilities and weaknesses.

Minor Concerns:
Clarity and Presentation Issues.
The methods section is placed in the Supplement, making the manuscript harder to follow.

Numerous ambiguities and errors hinder comprehension. For example:
Lines 109–110 describe the study domain as 32°–40°N and 114°–121°E, but Figure S1 shows a different region.

Lines 117–119 incorrectly state that simulated N_CCN is an input to the RFRM model (it should be the output).

Lines 154–155: The phrase "more to the model’s output" is unclear.

Figure 3c: The frequency unit appears to be 1e-8, but this is not explicitly stated.

Figure 6e shows N_CCN uncertainties within 150%, while Figures 6a–d display uncertainties exceeding 500%.

Similar problems appear in many places throughout the article, accompanied by punctuation errors, improper use of terms, etc., making reading extremely difficult.
Insufficient Explanation of Counterintuitive Results.

For instance, Figure 2a shows that sulfate has low permutation importance but high R-Square. The authors do not adequately explain or validate this finding, leaving readers to speculate.

Overstated Claims About Model Applicability.

Lines 382–384 suggest that integrating this framework into traditional climate models could reduce aerosol indirect effect uncertainties. However, since the model is only validated within its training spatiotemporal domain, such claims about generalizability are premature. The authors should temper these statements or provide evidence of the model’s robustness beyond the tested conditions.
Citation: https://doi.org/10.5194/egusphere-2025-1483-RC2
- AC4:
  'Reply on RC2', Fang Zhang, 05 Aug 2025
  The manuscript by Ren et al. presents a machine learning (ML) model based on Random Forest Regression (RFRM) to predict cloud condensation nuclei number concentration (N_CCN) at typical supersaturations in the North China Plain (NCP), which they then demonstrate that their model reduces the prediction bias from 39% (WRF-Chem) to 8%. The study also analyzes the importance of different input factors, evaluates the model’s performance in simulating spatiotemporal variability, and quantifies the reduction in cloud radiative forcing uncertainty achieved by mitigating N_CCN simulation biases. While the topic is highly relevant to GMD and represents a potential advancement in climate modeling, I have significant concerns regarding the methodological approach and validation logic, as well as the clarity of the manuscript. I recommend returning the manuscript to the authors with encouragement to revise and resubmit after addressing the following concerns.
  Major Concerns
  The training and testing datasets use WRF-Chem simulations as the output variable (target), but model performance is evaluated against observations. From a machine learning perspective, if the target variable (output) in both training and testing is WRF-Chem simulations, the model’s objective should be to approximate WRF-Chem’s output—not observations. Thus, model performance should be assessed based on how well it replicates WRF-Chem’s results, not observations. The authors report that WRF-Chem exhibits a significant bias (~39%) compared to observations, yet their ML model, trained to emulate WRF-Chem, shows a much smaller bias (~8%). This discrepancy is counterintuitive and requires a thorough explanation, which is currently missing.
  Re: The reviewer put forward an insightful comment. However, the reviewer may misunderstand the part where we compared the CCN predicted by the RFRM model with both the observed CCN and that simulated by WRF-Chem. This could be attributed to that we do not elaborate on this clearly in the methodology section. Due to the scarcity of a large spatial scale CCN observations (regional and global), in this study, the RFRM model was trained using simulated CCN concentrations from WRF-Chem as the target variable, which can capture general temporal variations in ambient CCN concentrations (Fig. 4) despite some biases.
  When constructing the model, all the aforementioned species data (predictor and target variables) were processed to match the same spatiotemporal resolution, and then split into 7:3 ratio for model training and testing respectively. Fig. 2 shows the performance of our developed RFRM models by comparing the testing dataset (sample size: N=117585) of CCN simulated by WRF-Chem with the predicted CCN. It showed that the estimated N_CCN at S=0.2% are highly correlated with the values from WRF-Chem, with the correlation R² of ~0.89-0.91 and slopes of 0.83 and 0.86 in our developed RFRM model. This suggests that our model works well in estimating N_CCN with a high aerosol loading environment
  However, given that the ultimate goal of model development is to more accurately predict the actual atmospheric concentration of CCN, we further conducted a comparative analysis of the prediction results against the in-situ observation data in this region, spanning from the hourly scale to the interannual scale (see Figure 3-6). For this comparison, we also plotted and included the WRF-Chem simulated CCN in the figures. As a result, we can get how the developed model has improved the WRF-Chem simulations.
  We have added detailed descriptions in the revised text or see follows (Lines 117-173):
  “… 2.2 Model construction and validation
  Here we develop the ML-based N_CCN prediction model by employing the Random Forest Regression method (RFRM) that has been demonstrated and can solve multivariate and nonlinear regression problems (Nair and Yu, 2020; Liang et al., 2022).
  The diagram of the model construction and the N_CCN prediction is shown in Figure 1. Due to lack of a large spatial scale observed N_CCN data, we use simulated N_CCN by WRF-Chem model as the targeted variable that can basically capture the ambient temporal variability of CCN concentration despite a deviation of ~40% by comparing with our six field observations (Figure 4). The input parameters include the chemical components of PM_2.5 (organic, sulfate, nitrate, ammonium, black carbon) from the Tsinghua University Tracking Air Pollution in China dataset (Liu et al., 2022) and gas and particulate pollutants (nitrogen dioxide (NO₂), sulfur dioxide (SO₂), carbon monoxide (CO), ozone (O₃) and PM_2.5) collected from the China National Environmental Monitoring Centre network. Meteorological parameters are from the European Centre for Medium-range Weather Forecasts Reanalysis version 5 (ERA-5) and include temperature, relative humidity (RH), total precipitation (TP), wind speed (WS), wind direction (WD), planetary boundary layer height (BLH), surface pressure (SP) and surface net solar radiation (SNSR). These datasets have undergone validation and been proven highly suitable for developing machine learning models for atmospheric applications (Nair and Yu, 2020; Wei et al., 2023). Cartesian coordinates were also added as input due to the spatiotemporal nature of the input data (Yang et al., 2022). Supplemental Table 1 provides more details about the input parameters.
  When constructing the model, all the aforementioned species data (predictor and target variables) were processed to match the same spatiotemporal resolution, and then split into 7:3 ratio for model training and testing respectively. As a result, a total of 274365 samples is included in the training datasets. In order to assure a stronger generalization ability of the N_CCN prediction model, the 10-fold cross-validation is adopted (Wei et al., 2023). The optimization parameters of RF model were examined by varying hyperparameters (Fig. S1). In addition, cross-validation (CV) is applied to select the hyperparameters during the data preprocessing (Yang et al., 2022). The CV results showed that when the number of trees (n_estimators) was less than 200, the prediction accuracy increased rapidly with the increase of the number of trees, and then gradually stabilized. According to the CV score and the number of data sample, the number of trees was set to 500 in this study. The impact of max depth on the CV score showed that, with the increase of depth, the complexity of the model increases. Thus, the max depth is set to 28. Also, the model generalization error was larger when the minimum sample number of the leaf and branch node are large, indicating that the model itself is close to the optimal model complexity level. Therefore, a higher value was set given the large sample size in this case. The influence of the maximum selection feature number on CV score showed a trend of increasing first and then decreasing, so the maximum value of CV curve was set to 16.
  
  Figure 2. Comparison of RFRM retrieval and WRF-Chem simulated N_CCN at S=0.2%. (a and b) Density plots of retrieval N_CCN at S=0.2% as a function of the simulations from WRF-Chem on the testing dataset.
  Fig. 2 shows the performance of our developed RFRM models by comparing the testing dataset (sample size: N=117585) of CCN simulated by WRF-Chem with the predicted CCN. Here the quality metrics for model performance are based on the correlation coefficient (R²), root mean square error (RMSE) and the slope of the RFRM predicted and the WRF-Chem simulated CCN concentrations. It showed that the estimated N_CCN at S=0.2% are highly correlated with the values from WRF-Chem, with the correlation R² of ~0.89-0.91 and slopes of 0.83 and 0.86 in our developed RFRM model. This suggests that our model works well in estimating N_CCN with a high aerosol loading environment. We also found that the accuracy of the CCN prediction will deteriorate slightly if not including the information of chemical compositions (Fig. S2), or if using XGBoost algorithm (Fig. S3) when constructing the model.
  
  Figure S2. Comparison of RFRM-ShortVars model retrieval and WRF-Chem simulated N_CCN at S=0.2%. (a and b) Density plots of retrieval N_CCN at S=0.2% as a function of the simulations from WRF-Chem on the testing dataset.
  Figure S3. Same as Figure S2 but from XGBoost model.
  Given that the ultimate goal of model development is to more accurately predict the actual atmospheric concentration of CCN, we further conducted a comparative analysis of the prediction results against the in-situ observation data in this region, spanning from the hourly scale to the interannual scale (see Figure 3-6), which will be presented and discussed in detail in Section 3.2 and 3.3 …”
  Use multisource datasets as input and WRF-Chem simulations as output to make training and test sets. Table S1 indicates that all input variables used in the ML model are also available from WRF-Chem outputs. However, the author uses datasets from different sources and different spatiotemporal resolutions to construct the data set through interpolation. The authors provide no rationale and advantage of this approach. Conventionally, one would train ML models using consistent model outputs and later test with observational inputs to assess potential gains.
  Re: Although the input feature variables used in model training can indeed be derived from the output of WRF Chem. However, they exhibit varying degrees of bias. Assuming the non-linear relationship between predictor features and target variables is accurate, train the RFRM model. So, choosing more accurate feature factors to train machine learning models is crucial for improving model accuracy. It is also consistent with the standard practice of environmental data modeling. In machine learning modeling, we used PM_2.5 chemical composition from the Tsinghua University Tracking China Air Pollution Dataset (Liu et al., 2022), gas and particulate pollutants from the China National Environmental Monitoring Center network, and meteorological parameters from the European Centre for Medium Range Weather Forecast Reanalysis 5th edition (ERA-5) data as input variables. These are commonly used datasets for building atmospheric application machine learning models (Nair and Yu, 2020; Wei et al., 2023). The results indicate that the model trained on carefully selected predictive factors can reliably predict CCN. We have revised as follows and see Lines 125-138:
  “…The input parameters include the chemical components of PM_2.5 (organic, sulfate, nitrate, ammonium, black carbon) from the Tsinghua University Tracking Air Pollution in China dataset (Liu et al., 2022) and gas and particulate pollutants (nitrogen dioxide (NO₂), sulfur dioxide (SO₂), carbon monoxide (CO), ozone (O₃) and PM_2.5) collected from the China National Environmental Monitoring Centre network. Meteorological parameters are from the European Centre for Medium-range Weather Forecasts Reanalysis version 5 (ERA-5) and include temperature, relative humidity (RH), total precipitation (TP), wind speed (WS), wind direction (WD), planetary boundary layer height (BLH), surface pressure (SP) and surface net solar radiation (SNSR). These datasets have undergone validation and been proven highly suitable for developing machine learning models for atmospheric applications (Nair and Yu, 2020; Wei et al., 2023). Cartesian coordinates were also added as input due to the spatiotemporal nature of the input data (Yang et al., 2022). Supplemental Table 1 provides more details about the input parameters …”
  The model is trained on data from 378 monitoring sites across the NCP (2014–2018) but evaluated using three sites (Beijing, Xingtai, Gucheng) within the same region and time period. This does not test generalizability. True validation requires spatiotemporal extrapolation: e.g., evaluation outside the training period or region. Performance on training-era/training-region data is expected to be favorable and does not demonstrate robust predictive skill.
  Re: Thank you. In this study, the RFRM model is trained on data from 378 sites across in NCP from 2014 to 2018. The six observations were excluded from the set of 378 monitoring sites across NCP, and the corresponding time periods were eliminated in the model training process. Therefore, the evaluation is conducted outside of the training period or region.
  Besides, observations at these sites could represent the average polluted and background conditions in the NCP (eg., PM_2.5 shown in Fig. 4b). Compared with the observed values, the better performance of the model can indicate its good predictive ability. See Lines 188-225:
  “…2.3.2 Gound-based measurements and datasets
  Ground measurements of atmospheric gaseous precursors, fine particles chemical compositions, and CCN number concentration (at supersaturations of 0.2% and 0.4%) were collected during six field campaigns at three sites in the NCP (Fig. 4), used to assess the performance of the developed ML-based model in predicting N_CCN. The six campaigns were conducted as follows: at the Beijing (BJ) site from 8–30 November 2014, 20 August to 6 October 2015, 16 November to 20 December 2016, and 28 May to 27 June 2017; at the Xingtai (XT) site from 17 May to 14 June 2016; and at the Gucheng (GC) site from 23 January to 3 February 2018. They are accordingly named BJ2014_WIN, BJ2015_AUT, BJ2016_WIN, BJ2017_SUM, XT2016_SUM, and GC2018_WIN (Fig. 4a).
  The BJ site (Longitude: 116.37° E; Latitude: 39.97° N) is located at the meteorological tower station of the Institute of Atmospheric Physics, Chinese Academy of Sciences. It is representative of the general emission conditions in urban areas of the northern NCP. The primary pollution sources here are surrounding traffic and residential emissions. The XT site (Longitude: 114.37° E; Latitude: 37.18° N) is situated at a national weather station. It is primarily influenced by emissions from surrounding towns and factories (e.g., coal-fired power plants, coking, steel, cement, and chemical industries) and thus reflects polluted suburban conditions in the southern NCP. The GC site (Longitude: 115.74° E; Latitude: 39.15° N) is located at the Integrated Ecological-Meteorological Observation and Experiment Station of the Chinese Academy of Meteorological Sciences. Surrounded mainly by nearby villages, farmland, and transportation networks, this site represents the regional background pollution in the northern NCP.
  The CCN number concentrations were measured by using the Droplet Measurement Technologies CCN counter (model CCNC-100, DMT Inc. Lance et al., 2006) at BJ and XT site. The supersaturation (S) levels set for each CCN measurement cycle were 0.1%, 0.2%, 0.4%, and 0.8%, respectively. Another measurement at GC site was referred from Zhang et al. (2020). In this study, the comparisons between the measured and predicted N_CCN were mostly based on the value at S=0.2% and S=0.4%. The observed N_CCN varies from a few hundred to tens of thousands at these sites, and the campaign mean mass concentration of PM_2.5 ranges from 35.6 to 160 μg m^-3 (Fig. 4b), indicating that the observations can represent various atmospheric conditions, spanning from clean to polluted in the region. More details about the observations could be found in Fan et al. (2020), Ren et al. (2018), and Zhang et al. (2019). In addition, the long-term measurement of particle number size distribution (PNSD) at a field site in Beijing (Fig. S5, Shang et al., 2022) is also used for deriving the long-term trend of yearly averaged N_CCN …”
  Other Major Concerns:
  The manuscript positions itself as unique by focusing on polluted regions (Lines 88–95), yet only 6 field campaigns are used for evaluation, with no dedicated analysis of heavy-pollution events. Additionally, heavy reliance on "mean prediction bias" is misleading: if RMSE/MSE is the training loss (unstated in the text), ML models inherently bias predictions toward the mean. Therefore, the improvement of the "mean prediction bias" cannot fully prove the performance of the ML model in real scenarios, and it is more meaningful to conduct a detailed evaluation and analysis of a single severe pollution event.
  Re: Thanks for your suggestion, we have added some discussion about haze events in the revised text, see as follows or Lines 290-311:
  “…Compared to WRF-Chem simulations, the RFRM model showed the greatest improvement during the winter campaigns when PM_2.5 concentrations were usually higher. For example, during the GC2018_WIN campaign, the observed N_CCN is underestimated as large as 61% by the WRF-Chem (Fig. S8), while the underestimation is largely improved with the predicted bias of only 3% in the RFRM model (Fig. S8). WRF-Chem simulations for warm seasons noticeably improved, e.g., the uncertainty decreased to 8% during the BJ2015_AUT campaign (Fig. S8). Overall, the RFRM model still performs better than the WRF-Chem model and is with averaged predicted bias of 18% during summer campaigns. Occasionally, the WRF-Chem model overestimated the N_CCN apparently, e.g., the episodes of September 21 to 24 during the BJ2015_AUT campaign, and May 28 to 31 during the BJ2017_SUM campaign. Here four pollution events from different seasons have been selected to further examine the capability of RFRM model to predict CCN concentrations (Figure 5). Fig. 5a presents a case from 14th to 18th September, 2015, during which PM_2.5 levels increased from 50 to 315 µg m^⁻3. As pollution intensified, CCN concentrations also rose. Compared to observations, the RFRM model exhibited lower relative bias. Fig. 5b–d display three additional individual pollution episodes of varying severity with PM_2.5 ranging from 10 to 660 µg m^⁻3. In all cases, the RFRM model more accurately captures the peak CCN concentrations during pollution events, exhibiting consistently lower relative bias. Especially for the case of 2nd to 5th December in 2016, the RFRM model can better capture the peak N_CCN of high pollution, while the WRF-Chem did not simulate the peak on December 4th very well …”
  
  Fig. 5 Performance of the RFRM model in predicting N_CCN during haze events. (a) Case of 24 to 18 September in 2015, (b) case of 6 to 18 May in 2016, (c) case of 2 to 5 December in 2016, (d) case of 2 to 8 June in 2017.
  A common but concerning trend in ML applications is showcasing successes while neglecting failures. This paper follows that pattern. There is no discussion of scenarios where the model underperforms, its limitations, or potential pitfalls. For instance, an eager graduate student might misuse this model for policy analysis without realizing its constraints (e.g., lack of generalizability), leading to significant wasted effort. A rigorous journal paper must present a balanced view of model capabilities and weaknesses.
  Re: Thanks for the suggestion, some discussions about the model limitations were revised in the section of 4.2 Limitations and outlook or see Lines 481-500:
  “4.2 Limitations and outlook
  In this study, the RFRM model was trained using simulated CCN concentrations from WRF-Chem as the target variable, assuming that the nonlinear relationships between the predictor features and the target variable are accurate. However, as noted earlier, even though WRF-Chem simulations can capture the variation of N_CCN, they carry an uncertainty of ~20–40% compared to observations (Fanourgakis et al., 2019). This contributes directly to uncertainty in the RFRM model’s predictions. Additionally, note that in this study, observational data from six campaigns at three sites are analyzed. Validating the simulated N_CCN through comparisons with observations at more ground sites is thus warranted. In the future, it is crucial to obtain comprehensive monitoring data of CCN and other key aerosol properties (e.g., particle size distribution, chemical compositions) in different environments.
  The RFRM framework presented here relies on readily available atmospheric state variables (eg., chemical compositions, gas pollutants, and meteorology elements) and significantly improves the accuracy of N_CCN prediction, thereby helping to bridge observational gaps. Our modeling framework could then be used to simulate ground-level CCN data in other regions around the world and even on a global scale. Moreover, this approach may guide the development of machine‑learning–based models to predict CCN vertical profiles, which are critical for accurately assessing aerosol–cloud interactions…”
  Minor Concerns:
  Clarity and Presentation Issues.
  The methods section is placed in the Supplement, making the manuscript harder to follow.
  Re: Thank you for your efforts and time on handling the paper. We have updated the section of Methods and see as follows or Lines 106-225:
  “2. Methods
  2.1 Study area
  In this work, we select the North China Plain (NCP) (32°-40°N and 114°-121°E) as the study area. Being one of the most polluted areas in China, the aerosol particles in NCP are with more complex composition and mixing state, which leads to great challenge in accurate prediction of cloud concentration nuclei (CCN) concentrations. In recent years, emissions of gas pollutants and fine particles have shown a significant downward trend year by year (Wei et al., 2023) due to the implementation of the vigorous emission reduction in China (Zheng et al., 2018). This also makes changes in aerosols CCN activity in the study area from the point of view in assessment of the climate effect of aerosols.
  2.2 Model construction and validation
  Here we develop the ML-based N_CCN prediction model by employing the Random Forest Regression method (RFRM) that has been demonstrated and can solve multivariate and nonlinear regression problems (Nair and Yu, 2020; Liang et al., 2022).
  The diagram of the model construction and the N_CCN prediction is shown in Figure 1. Due to lack of a large spatial scale observed N_CCN data, we use simulated N_CCN by WRF-Chem model as the targeted variable that can basically capture the ambient temporal variability of CCN concentration despite a deviation of ~40% by comparing with our six field observations (Figure 4). The input parameters include the chemical components of PM_2.5 (organic, sulfate, nitrate, ammonium, black carbon) from the Tsinghua University Tracking Air Pollution in China dataset (Liu et al., 2022) and gas and particulate pollutants (nitrogen dioxide (NO₂), sulfur dioxide (SO₂), carbon monoxide (CO), ozone (O₃) and PM_2.5) collected from the China National Environmental Monitoring Centre network. Meteorological parameters are from the European Centre for Medium-range Weather Forecasts Reanalysis version 5 (ERA-5) and include temperature, relative humidity (RH), total precipitation (TP), wind speed (WS), wind direction (WD), planetary boundary layer height (BLH), surface pressure (SP) and surface net solar radiation (SNSR). These datasets have undergone validation and been proven highly suitable for developing machine learning models for atmospheric applications (Nair and Yu, 2020; Wei et al., 2023). Cartesian coordinates were also added as input due to the spatiotemporal nature of the input data (Yang et al., 2022). Supplemental Table 1 provides more details about the input parameters.
  When constructing the model, all the aforementioned species data (predictor and target variables) were processed to match the same spatiotemporal resolution, and then split into 7:3 ratio for model training and testing respectively. As a result, a total of 274365 samples is included in the training datasets. In order to assure a stronger generalization ability of the N_CCN prediction model, the 10-fold cross-validation is adopted (Wei et al., 2023). The optimization parameters of RF model were examined by varying hyperparameters (Fig. S1). In addition, cross-validation (CV) is applied to select the hyperparameters during the data preprocessing (Yang et al., 2022). The CV results showed that when the number of trees (n_estimators) was less than 200, the prediction accuracy increased rapidly with the increase of the number of trees, and then gradually stabilized. According to the CV score and the number of data sample, the number of trees was set to 500 in this study. The impact of max depth on the CV score showed that, with the increase of depth, the complexity of the model increases. Thus, the max depth is set to 28. Also, the model generalization error was larger when the minimum sample number of the leaf and branch node are large, indicating that the model itself is close to the optimal model complexity level. Therefore, a higher value was set given the large sample size in this case. The influence of the maximum selection feature number on CV score showed a trend of increasing first and then decreasing, so the maximum value of CV curve was set to 16.
  
  Figure 2. Comparison of RFRM retrieval and WRF-Chem simulated N_CCN at S=0.2%. (a and b) Density plots of retrieval N_CCN at S=0.2% as a function of the simulations from WRF-Chem on the testing dataset.
  Fig. 2 shows the performance of our developed RFRM models by comparing the testing dataset (sample size: N=117585) of CCN simulated by WRF-Chem with the predicted CCN. Here the quality metrics for model performance are based on the correlation coefficient (R²), root mean square error (RMSE) and the slope of the RFRM predicted and the WRF-Chem simulated CCN concentrations. It showed that the estimated N_CCN at S=0.2% are highly correlated with the values from WRF-Chem, with the correlation R² of ~0.89-0.91 and slopes of 0.83 and 0.86 in our developed RFRM model. This suggests that our model works well in estimating N_CCN with a high aerosol loading environment. We also found that the accuracy of the CCN prediction will deteriorate slightly if not including the information of chemical compositions (Fig. S2), or if using XGBoost algorithm (Fig. S3) when constructing the model.
  Given that the ultimate goal of model development is to more accurately predict the actual atmospheric concentration of CCN, we further conducted a comparative analysis of the prediction results against the in-situ observation data in this region, spanning from the hourly scale to the interannual scale (see Figure 3-6), which will be presented and discussed in detail in Section 3.2 and 3.3.
  2.3 Data and other details in the model construction
  2.3.1 N_CCN simulated by WRF-Chem model
  The WRF-Chem version 4.1.5 is used to simulate N_CCN in this study, which nested a domain in 10 km×10 km covering the entire NCP (Fig. S4) and contained 181×170 grids. The simulation in WRF-Chem is conducted from 1 January 2014 to 31 December 2018 with an hourly resolution. In the WRF-Chem modeling system, the sectional Model for Simulating Aerosol Interactions and Chemistry (MOSAIC), the Morrison two-moment scheme (Morrison et al., 2009) and the Carbon Bond Mechanism Z photochemical mechanism (Zaveri et al., 1999) are employed. We also compared the simulation using the Regional Acid Deposition Model (Stockwell et al., 1990) and the Lin microphysics scheme (Lin et al., 1983). Considering the calculation efficiency and accuracy with the measurements, the CBMZ-MOSAIC and Morrison 2-monment scheme were finally applied to simulate the long-term CCN concentration. More details about the other parameterizations used for WRF-Chem simulation were given in SI.
  2.3.2 Gound-based measurements and datasets
  Ground measurements of atmospheric gaseous precursors, fine particles chemical compositions, and CCN number concentration (at supersaturations of 0.2% and 0.4%) were collected during six field campaigns at three sites in the NCP (Fig. 4), used to assess the performance of the developed ML-based model in predicting N_CCN. The six campaigns were conducted as follows: at the Beijing (BJ) site from 8–30 November 2014, 20 August to 6 October 2015, 16 November to 20 December 2016, and 28 May to 27 June 2017; at the Xingtai (XT) site from 17 May to 14 June 2016; and at the Gucheng (GC) site from 23 January to 3 February 2018. They are accordingly named BJ2014_WIN, BJ2015_AUT, BJ2016_WIN, BJ2017_SUM, XT2016_SUM, and GC2018_WIN (Fig. 4a).
  The BJ site (Longitude: 116.37° E; Latitude: 39.97° N) is located at the meteorological tower station of the Institute of Atmospheric Physics, Chinese Academy of Sciences. It is representative of the general emission conditions in urban areas of the northern NCP. The primary pollution sources here are surrounding traffic and residential emissions. The XT site (Longitude: 114.37° E; Latitude: 37.18° N) is situated at a national weather station. It is primarily influenced by emissions from surrounding towns and factories (e.g., coal-fired power plants, coking, steel, cement, and chemical industries) and thus reflects polluted suburban conditions in the southern NCP. The GC site (Longitude: 115.74° E; Latitude: 39.15° N) is located at the Integrated Ecological-Meteorological Observation and Experiment Station of the Chinese Academy of Meteorological Sciences. Surrounded mainly by nearby villages, farmland, and transportation networks, this site represents the regional background pollution in the northern NCP.
  The CCN number concentrations were measured by using the Droplet Measurement Technologies CCN counter (model CCNC-100, DMT Inc. Lance et al., 2006) at BJ and XT site. The supersaturation (S) levels set for each CCN measurement cycle were 0.1%, 0.2%, 0.4%, and 0.8%, respectively. Another measurement at GC site was referred from Zhang et al. (2020). In this study, the comparisons between the measured and predicted N_CCN were mostly based on the value at S=0.2% and S=0.4%. The observed N_CCN varies from a few hundred to tens of thousands at these sites, and the campaign mean mass concentration of PM_2.5 ranges from 35.6 to 160 μg m^-3 (Fig. 4b), indicating that the observations can represent various atmospheric conditions, spanning from clean to polluted in the region. More details about the observations could be found in Fan et al. (2020), Ren et al. (2018), and Zhang et al. (2019). In addition, the long-term measurement of particle number size distribution (PNSD) at a field site in Beijing (Fig. S5, Shang et al., 2022) is also used for deriving the long-term trend of yearly averaged N_CCN…”
  Numerous ambiguities and errors hinder comprehension.
  For example:
  Lines 109–110 describe the study domain as 32°–40°N and 114°–121°E, but Figure S1 shows a different region.
  Re: Here Figure S1 has been revised as Figure S4. It shows the simulation domain of WRF-Chem, which nested a domain in 10 km×10 km covering the entire North China Plain and contained 181×170 grids. A region within 32°-40°N and 114°-121°E in the NCP is chosen as the study area. The distance between the study area and the boundary of the simulation domain must be greater than 10 times of the resolution. Our study area is within the range of Fig. S4. The sentence has been revised as follows or see Lines 108-109 and 176-178:
  “…In this work, we select the North China Plain (NCP) (32°-40°N and 114°-121°E) as the study area …”
  “…The WRF-Chem version 4.1.5 is used to simulate N_CCN in this study, which nested a domain in 10 km×10 km covering the entire NCP (Fig. S4) and contained 181×170 grids …”
  Lines 117–119 incorrectly state that simulated N_CCN is an input to the RFRM model (it should be the output).
  Re: The sentence has been revised as follows or see Lines 122-125:
  “…Due to lack of a large spatial scale observed N_CCN data, we use simulated N_CCN by WRF-Chem model as the targeted variable that can basically capture the ambient temporal variability of CCN concentration despite a deviation of ~40% by comparing with our six field observations (Figure 4) …”
  Lines 154–155: The phrase "more to the model’s output" is unclear.
  Re: The sentence has been revised as follows or see Lines 241-243:
  “…During the winter, changes in BLH contribute more to CCN predictions than PM_2.5 (Fig. 3b) and the model’s output changes more significantly with this factor (Fig. 3c) …”
  Figure 3c: The frequency unit appears to be 1e-8, but this is not explicitly stated.
  Re: Note that figure 3 has been revised as figure 4. The figure has been revised. See follows:
  
  Fig. 4 Performance of the RFRM model in predicting N_CCN at field sites in NCP. (a) Time series of the observed and predicted CCN number concentrations at S=0.2% for the six campaigns (BJ2015_AUT, BJ2017_SUM, XT2016_SUM, BJ2014_WIN, BJ2016_WIN, GC2018_WIN) in the North China Plain; (b) Map for average mass concentration of PM2.5 of 2014 from TAP dataset in NCP (http://tapdata.org.cn/) and field observed average mass concentration of PM_2.5 during the six field campaigns (see embedded histogram); (c) Scatter plots of the observed N_CCN at S=0.2% with the RFRM model predicted (top) and WRF-Chem simulated (bottom) respectively.
  Figure 6e shows N_CCN uncertainties within 150%, while Figures 6a–d display uncertainties exceeding 500%.
  Re: Note that figure 6 has been revised as figure 8. And here figure 8a–d present all-sites data points of the six observation campaigns. The statistical results show that the Random Forest Regression Model (RFRM) errors range from –90 to +600%, whereas the WRF‑Chem model exhibits a broader error span of –100 to +1800% when compared with the observations. Figure 8e summarizes the mean values across these campaigns with the N_CCN uncertainties within 150%. Some descriptions have been added as follows or see Lines 446-448:
  “…While, the mean uncertainties for all these parameters are largely reduced when the mean underestimation of ~8±38% in N_CCN at S=0.2% that is caused by RFRM model is applied (Fig. 8e) …”
  Similar problems appear in many places throughout the article, accompanied by punctuation errors, improper use of terms, etc., making reading extremely difficult.
  Insufficient Explanation of Counterintuitive Results.
  
  For instance, Figure 2a shows that sulfate has low permutation importance but high R-Square. The authors do not adequately explain or validate this finding, leaving readers to speculate.
  Re: Despite the high correlation between sulfate features and the target variable, their importance scores within the RFRM model remain low. Two main factors explain this:
  Nonlinear model behavior
  
  Random Forest is a nonlinear algorithm that constructs an ensemble of decision trees; it captures complex, non-additive interactions between predictors and response variables. As a result, even a feature with a strong linear correlation to the outcome may not play a pivotal role in the trees’ local splits. Thus, sulfate may exhibit high correlation with CCN concentration but contribute little to the actual partitioning decisions made by the model.
  Collinearity with other predictors.
  
  Strong inter-feature correlations (e.g. sulfate with nitrate/ammonium at 0.84-0.92/0.92-0.95) lead the model to favor one predictor (e.g. nitrate) over others when building decision splits. Because Random Forest often uses only one variable from a set of highly correlated candidates to optimally partition the data, sulfate’s importance score can be artificially diminished, despite sharing information with the target.
  Considering that the high hygroscopicity of sulfates is an effective seed for CCN, sulfate features were not removed during model training in our study.
  Some explanation has been added in the revised text, see as follows or Lines 250-259:
  
  Figure S7. Heatmap of the feature variables in the winter half of year (a) and summer half of year (b).
  “… Note that the impact of sulfate aerosols on N_CCN prediction is much less important in both summer and winter seasons compared to nitrate particles, with a permutation importance score ranging from ~0.02 to 0.03 but with higher correlation of ~0.31-0.49., This is mainly because the collinearity with nitrate features (~0.84-0.92) as seen in Fig. S7. In general, the machine learning algorithm often chooses one variable from a set of highly correlated candidates to optimally partition the data. Here sulfate’s importance score can be artificially diminished, largely due to its decreased proportion in PM_2.5 in recent years (Liang et al., 2022; Li et al., 2020). As a note, due to the high hygroscopicity of sulfates is an effective seed for CCN, it was not removed in RFRM model…”
  Overstated Claims About Model Applicability.
  Lines 382–384 suggest that integrating this framework into traditional climate models could reduce aerosol indirect effect uncertainties. However, since the model is only validated within its training spatiotemporal domain, such claims about generalizability are premature. The authors should temper these statements or provide evidence of the model’s robustness beyond the tested conditions.
  Re: The sentence has been revised in the text, see as follows or Lines 477-480:
  “…Given the simplified setting in current climate models, this work emphasizes the necessity and urgency to obtain the precise N_CCN values, offering a new framework for predicting CCN concentrations based on machine learning algorithms and effectively filling the observation gap of CCN concentrations…”
  References:
  Fanourgakis, G., Kanakidou, M., Nenes, A. et al.: Evaluation of global simulations of aerosol particle and cloud condensation nuclei number, with implications for cloud droplet formation, Atmos. Chem. Phys., 19(13), 8591–8617, https://doi.org/ 10.5194/acp-19-8591-2019, 2019.
  Nair, A. A., Yu, F.: Using machine learning to derive cloud condensation nuclei number concentrations from commonly available measurements, Atmos. Chem. Phys., 20(21), 12853–12869, https://doi.org/10.5194/acp-20-12853-2020, 2020.
  Yang, N., Shi, H., Tang, H. et al.: Geographical and temporal encoding for improving the estimation of PM_2.5 concentrations in China using end-to-end gradient boosting, Remote Sensing of Environment, 269, 112828, https://doi.org/10.1016/j.rse.2021.112828, 2022.
  Wei, J., Li, Z., Wang, J. et al.: Ground-level gaseous pollutants (NO₂, SO₂, and CO) in China: daily seamless mapping and spatiotemporal variations, Atmos. Chem. Phys., 23, 1511–1532, https://doi.org/10.5194/acp-23-1511-2023, 2023.
  Morrison, H., Thompson, G., Tatarskii, V.: Impact of cloud microphysics on the development of trailing stratiform precipitation in a simulated squall line: Comparison of one-and two-moment schemes, Monthly Weather Review, 137(3), 991–1007, https://doi.org/10.1175/2008MWR2556.1, 2009.
  Zaveri, R. A., Peters, L. K.: A new lumped structure photochemical mechanism for large-scale applications, Journal of Geophysical Research: Atmospheres, 104, 30387–30415, https://doi.org/10.1029/1999JD900876, 1999.
  Stockwell, W. R., Middleton, P., Chang, J. S., et al.: The second generation regional acid deposition model chemical mechanism for regional air quality modeling, Journal of Geophysical Research: Atmospheres, 95(D10): 16343-16367, https://doi.org/10.1029/JD095iD10p16343, 1990.
  Lin, Y., Farley, R., Orville, H.: Bulk parameterization of the snow field in a cloud model, Journal of Applied Meteorology and Climatology, 22(6): 1065-1092, https://doi.org/10.1175/1520-0450(1983)022<1065:BPOTSF>2.0.CO;2, 1983.
  Fan, X., Liu, J., Zhang, F. et al.: Contrasting size-resolved hygroscopicity of fine particles derived by HTDMA and HR-ToF-AMS measurements between summer and winter in Beijing: the impacts of aerosol aging and local emissions, Atmos. Chem. Phys., 20, 915–929, https://doi.org/10.5194/acp-20-915-2020, 2020.
  Ren, J., Zhang, F., Wang, Y. et al.: Using different assumptions of aerosol mixing state and chemical composition to predict CCN concentrations based on field measurements in urban Beijing, Atmos. Chem. Phys., 18(9), 6907–6921, https://doi.org/10.5194/acp-18-6907-2018, 2018.
  Zhang, F., Ren, J., Fan, T. et al.: Significantly enhanced aerosol CCN activity and number concentrations by nucleation-initiated haze events: A case study in urban Beijing, Journal of Geophysical Research: Atmospheres, 124(24), 14102–14113, https://doi.org/10.1029/2019JD031457, 2019.
  Lance, S., Nenes, A., Medina, J. et al.: Mapping the operation of the DMT continuous flow CCN counter, Aerosol Science and Technology, 40(4), 242–254, https://doi.org/10.1080/02786820500543290, 2006.
  Zhang, Y., Tao, J., Ma, N. et al.: Predicting cloud condensation nuclei number concentration based on conventional measurements of aerosol properties in the North China Plain, Science of The Total Environment, 719, 137473, https://doi.org/10.1016/j.scitotenv.2020.137473, 2020.
  Liang, M., Tao, J., Ma, N. et al.: Prediction of CCN spectra parameters in the North China Plain using a random forest model, Atmospheric Environment, 289, 119323, https://doi.org/10.1016/j.atmosenv.2022.119323, 2022.
  Li, S., Zhang, F., Jin, X. et al.: Characterizing the ratio of nitrate to sulfate in ambient fine particles of urban Beijing during 2018-2019, Atmospheric Environment, 237, https://doi.org/10.1016/j.atmosenv.2020.117662, 2020.
  Liu, S., Geng, G., Xiao, Q., Zheng, Y., Liu, X., Cheng, J., & Zhang, Q.: Tracking daily concentrations of PM2.5 chemical composition in China since 2000, Environ Sci Technol, 56, 16517–16527, https://doi.org/10.1021/acs.est.2c06510, 2022.
  Zheng, B., Tong, D., Li, M., Liu, F. et al.: Trends in China’s anthropogenic emissions since 2010 as the consequence of clean air actions, Atmos. Chem. Phys., 18, 14095-14111, https://doi.org/10.5194/acp-18-14095-2018, 2018.
  Shang, D., Tang, L., Fang, X. et al.: Variations in source contributions of particle number concentration under long-term emission control in winter of urban Beijing, Environmental Pollution, 304, 119072, https://doi.org/10.1016/j.envpol.2022.119072, 2022.
  
  Citation: https://doi.org/10.5194/egusphere-2025-1483-AC4

Jingye Ren, Songjian Zou, Honghao Xu, Guiquan Liu, Zhe Wang, Anran Zhang, Chuanfeng Zhao, Min Hu, Dongjie Shang, Lizi Tang, Ru-Jin Huang, Yele Sun, and Fang Zhang

Supplement

https://doi.org/10.5194/egusphere-2025-1483-supplement

Jingye Ren, Songjian Zou, Honghao Xu, Guiquan Liu, Zhe Wang, Anran Zhang, Chuanfeng Zhao, Min Hu, Dongjie Shang, Lizi Tang, Ru-Jin Huang, Yele Sun, and Fang Zhang

Viewed

Total article views: 1,161 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
1,062	70	29	1,161	38	9	33

HTML: 1,062
PDF: 70
XML: 29
Total: 1,161
Supplement: 38
BibTeX: 9
EndNote: 33

Views and downloads (calculated since 03 Jun 2025)

Month	HTML	PDF	XML	Total
Jun 2025	147	37	15	199
Jul 2025	84	19	9	112
Aug 2025	383	10	5	398
Sep 2025	448	4	0	452

Cumulative views and downloads (calculated since 03 Jun 2025)

Month	HTML	PDF	XML	Total
Jun 2025	147	37	15	199
Jul 2025	84	19	9	112
Aug 2025	383	10	5	398
Sep 2025	448	4	0	452

Viewed (geographical distribution)

Total article views: 1,160 (including HTML, PDF, and XML) Thereof 1,160 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 15 Sep 2025

Short summary

In this study, a new framework of cloud condensation nuclei (CCN) prediction in polluted region has been developed and it achieves well prediction of hourly-to-yearly scale across North China Plain. The study reveals a significant long-term decreasing trend of CCN concentration at typical supersaturations due to a rapid reduction in aerosol concentrations from 2014 to 2018. This improvement of our new model would be helpful to aerosols climate effect assessment in models.


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