the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 31 Aug 2026
Development and tests of machine-learned cloud droplet number concentration in a next-generation bulk microphysics scheme for operational applications
Abstract. A machine-learning (ML) model was developed to predict cloud droplet number concentrations. This ML model was trained on data from simulations using the Model for Prediction Across Scales–Atmosphere (MPAS-A) coupled to a next-generation version of the Thompson-Eidhammer bulk microphysics scheme. The ML model was tested for two applications. First, the ML model was used to predict the number concentration, and hence the effective radii of subgrid-scale clouds from the MYNN boundary-layer scheme. These subgrid-scale clouds, necessary for accurate predictions of radiation, have historically been parameterized using constant values for their effective radii. The ML model predicted cloud droplet number concentrations for subgrid-scale clouds so that their effective radii could be calculated using a method physically consistent with those of fully resolved clouds predicted by the bulk microphysics scheme. These results were compared with observations using a novel approach to show that the ML model provided reasonable estimates of effective radii. Second, cloud droplet number concentrations were replaced in the bulk microphysics parameterization with values predicted by the ML model. Results from this test were compared with aircraft and surface observations from two IOPs during the WINTRE-MIX field campaign. Overall, the simulation with the ML prediction of cloud droplet number concentrations reproduced the 24-h total precipitation for the case and reasonably represented the microphysical details. The ML model performed better for relatively short forecasts, which was attributed to the lack of model spin-up required by physical parameterizations. The biases in the ML model for this case study resulted in reduced drizzle production, which slightly degraded the forecast of precipitation phase. Benefits of the ML method include lower computational cost, better coupling between fully resolved and unresolved (subgrid-scale) clouds, and more rapid development of realistic forecasts when initial conditions are not fully spun up.
- Preprint
(88973 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 16 Nov 2026)
-
CEC1: 'Comment on egusphere-2026-4026 - No compliance with the policy of the journal', Juan Antonio Añel, 22 Sep 2026
reply
-
AC1: 'Reply on CEC1', Anders Jensen, 26 Sep 2026
reply
Thank you for your comment; we have uploaded more data to be in compliance with the data policy. Please see the new version of the dataset: https://zenodo.org/records/22964695
- We added the data used to train the ML model:
hrrrv5.history.2024-08-06_01.00.00_extract.nc - ML training data hurricane Debby case
hrrrv5.history.2024-08-06_07.00.00_extract.nc - ML training data hurricane Debby case
hrrrv5.history.2025-05-15_22.00.00_extract.nc - ML training data springtime convection case
hrrrv5.history.2025-05-16_03.00.00_extract.nc - ML training data springtime convection case
ww.history.2023-01-09_12.00.00_extract.nc - ML training data atmospheric river case
ww.history.2023-01-09_18.00.00_extract.nc - ML training data atmospheric river case
- We added the MRMS composite reflectivity to remove the need to point to AWS
composite_refc_2023-01-09-1800.nc - MRMS composite reflectivity atmospheric river case
composite_refc_2024-08-06-0700.nc - MRMS composite reflectivity hurricane Debby case
composite_refc_2025-05-15-2200.nc - MRMS composite reflectivity springtime convection case
- We added MPAS code to remove the need to point to github repositories
MPAS-Model-ml-training.zip - MPAS model code for ML training
MPAS-Model-pub-ml_nc.zip - MPAS model code for ML-NC simulation
MPAS-Model-pub-ml_re.zip - MPAS model code for ML-RE simulation
TEMPO-exp-ml_nc.zip - TEMPO code for ML-NC simulation
TEMPO-exp-ml_re.zip - TEMPO code for ML-RE simulation
- We added missing data needed to run the MPAS simulations
hrrrv5.static.nc - static file for MPAS simulations
sounding_locations.txt - Sounding locations
- We addded intermediate RAP data for MPAS simulations to avoid pointing to AWS
RAP:2022-03-06_12
RAP:2022-03-06_18
RAP:2022-03-07_00
RAP:2022-03-07_06
RAP:2022-03-07_12
RAP:2022-03-07_18
We plan on updating the Code and data availability statement and the references in the paper after the discussion period, but please let us know if you would like an updated version of the manuscript sooner.
Thanks,
Anders Jensen
Citation: https://doi.org/10.5194/egusphere-2026-4026-AC1 -
CEC2: 'Reply on AC1', Juan Antonio Añel, 26 Sep 2026
reply
Dear authors,
Thanks for addressing this issue so quickly. I have checked the repositories and we can consider now the current version of your manuscript in compliance with the code policy of the journal.
Juan A. Añel
Geosci. Model Dev. Executive Editor
Citation: https://doi.org/10.5194/egusphere-2026-4026-CEC2
-
AC1: 'Reply on CEC1', Anders Jensen, 26 Sep 2026
reply
-
RC1: 'Comment on egusphere-2026-4026', Anonymous Referee #1, 05 Oct 2026
reply
General assessment
This manuscript uses seven model state variables to diagnose cloud droplet number concentration and applies the resulting estimates to the calculation of effective radii for subgrid-scale clouds and to the microphysics of fully resolved clouds. The study combines offline training, online simulations, and comparisons with observations to demonstrate the potential of the approach, while also reporting overestimation of cloud droplet number concentration and reduced drizzle production. The main contribution is the development and coupling of the parameterization, which provides a useful example of simplifying cloud microphysics in weather models. I recommend major revision, as detailed below.
Major comments
1. The network does not use aerosol or CCN information (lines 135–139), although similar cloud water and meteorological conditions can occur under different aerosol conditions and with different histories of cloud droplet evolution. Please explain how omitting aerosol information affects the diagnosis of cloud droplet number concentration, and distinguish statistical prediction of the reference model output from the ability to respond to changes in aerosols. Lines 209–210 state that cloud droplet number concentration is diagnosed both before and after the microphysics calculation. Please also explain how the aerosol variables are updated during this sequence. The authors could use the existing data to compare diagnostic errors across aerosol concentration ranges to help establish the conditions under which the method is applicable.
2. Lines 188–193 define the bias-correction range as 1–1000 cm⁻³. The offline script ml_paper_nc_parameter_space.py (code lines 109–113 and 133) uses number concentration for the correction, whereas module_mp_tempo_ml.F90 (code lines 214–221) applies the range 0–3 to the logarithm of number mixing ratio. In relation to the online implementation described in lines 202–211, please verify the unit conversions and order of correction operations in the versions used for this study. Please also compare the outputs before and after correction in the two implementations using identical inputs, weights, and preprocessing parameters. I suggest providing these comparisons and a minimal working example with the code.
Minor comments
1. The abstract (lines 13–17) attributes the improved performance at shorter forecast lead times to reduced spin-up requirements and presents lower computational cost as an established benefit. The cloud conditions also differ between the two observation periods (lines 260–263). Please therefore describe spin-up as a possible explanation and revise lines 403–405 accordingly. Please provide a brief timing comparison under otherwise identical model configurations to quantify the computational benefit.
2. Please clarify in lines 179–193 which samples were used for training, model selection, testing, and quantile mapping, including the relationship between the correction and test datasets. For the correlation coefficient of 0.92 reported in lines 186–187, please specify the scale on which it was calculated, the samples included, and their relationship to the samples shown in Fig. 3. Offline performance based on the data split should be described separately from the evaluation in the independent online case study.
3. Appendix A (lines 432–438) already describes the vertical weighting used to calculate effective radius. Please expand lines 270–279 to clarify the relationship between grid-mean and in-cloud quantities and the common selection criteria applied to the model and observational samples. Lines 227–230 specify that the CDP measures droplets in the 2–50 μm size range. The aircraft comparison should explain the difference between measurements within this range and the total quantities represented in the model, and define the bias statistics used in lines 337–340. The summary of improvements in effective radius in lines 397–400 should also reflect the differences among sites described in lines 293–305.
4. Lines 389–391 discuss both the similarity between the two simulations and their agreement with observations. Please make the basis for these two assessments clear. In Fig. 15 and Appendix B (lines 440–444), please specify the 1 mm selection threshold and spatial masking rules used in the analysis code, and explain whether the error distributions include areas where precipitation is missed or falsely predicted.
Technical corrections
1. Please apply the unit conventions for mixing ratios and concentrations in lines 99–102 consistently throughout the manuscript. The caption of Fig. 4 identifies the blue and grey data as training and testing data, whereas lines 194–200 describe a comparison between predicted and reference values. Please check the legend, caption, and identification of results before and after correction.
2. Lines 251–254 give a liquid water path range of 5–45 g m⁻², whereas Appendix A (lines 436–438) uses 5–40 g m⁻². Please distinguish the valid retrieval range from the range used for sample selection, explain the choice of the 40 g m⁻² upper limit, and state the thresholds used in the figures and tables.
3. Please complete the definitions of the symbols, error bars, and bins in Figs. 9 and 11, and specify the sample sizes and sampling criteria. For Fig. 13, please clarify the units used for cloud droplet and raindrop number concentrations and the time window covered by the observations. For Fig. 14, please identify the data sources for the observational panels. For Fig. 15, please clarify whether white areas represent low precipitation amounts or missing data.
4. Please correct the citation format of “Han and et al.” and “Girouard and et al.” in lines 231–232, and remove duplicated DOI prefixes in the references, for example in lines 471 and 480. In the introduction (lines 46–48), I suggest replacing descriptions such as “remarkably accurate” with specific statements consistent with the findings of the cited studies.Citation: https://doi.org/10.5194/egusphere-2026-4026-RC1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 266 | 197 | 59 | 522 | 51 | 47 |
- HTML: 266
- PDF: 197
- XML: 59
- Total: 522
- BibTeX: 51
- EndNote: 47
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
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
To get access to part of your code you cite GitHub. However, GitHub is not a suitable repository for scientific publication. GitHub itself instructs authors to use other long-term archival and publishing alternatives, such as Zenodo. In addition, to get access to the data used for training and validation and produced in your work you cite several websites, and none of them fulfil GMD’s requirements for a persistent data archive because:
- It does not appear to have a published policy for data preservation over many years or decades (some flexibility exists over the precise length of preservation, but the policy must exist).
- It does not appear to have a published mechanism for preventing authors from unilaterally removing material. Archives must have a policy which makes removal of materials only possible in exceptional circumstances and subject to an independent curatorial decision,
- It does not appear to issue a persistent identifier such as a DOI or Handle for each precise dataset.
If we have missed a published policy which does in fact address this matter satisfactorily, please post a response linking to it. If you have any questions about this issue, please post them in a reply.
The GMD review and publication process depends on reviewers and community commentators being able to access, during the discussion phase, the code and data on which a manuscript depends, and on ensuring the provenance of replicability of the published papers for years after their publication. Please, therefore, publish your code and data in one of the appropriate repositories and reply to this comment with the relevant information (link and a permanent identifier for it (e.g. DOI)) as soon as possible. We cannot have manuscripts under discussion that do not comply with our policy.
Later, if the Topical Editor decides to continue with the review or publication process of your manuscript and you are requested to upload a new version of it, then The 'Code and Data Availability’ section of your manuscript must also be modified to cite the new repository locations, and corresponding references added to the bibliography.
I must note that if you do not fix this problem, we cannot continue with the peer-review process or accept your manuscript for publication in GMD.
Juan A. Añel
Geosci. Model Dev. Executive Editor