the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Aug 2026
Machine learning based radiation emulation: development and performance evaluation in operational reforecast experiments
Abstract. Radiation is typically the most time-consuming physical process in numerical models. One solution is to use machine learning methods to simulate the radiation process to improve computational efficiency. From a practical application standpoint, this study investigates the key issues for hybrid modelling: coupling compatibility and long-term integration stability. A residual convolutional neural network is employed to approximate the Rapid Radiative Transfer Model for General Circulation Models (RRTMG) within the global operational system of China Meteorological Administration. We adopted an offline training and online coupling approach. First, a comprehensive dataset is generated through model simulations, encompassing all atmospheric columns. To ensure the stability of the hybrid model, the dataset is enhanced via experience replay, and additional output constraints based on physical significance are imposed. Meanwhile, a LibTorch-based coupling method is utilized, which is more suitable for real-time operational computations. The hybrid model is capable of performing ten-day integrated forecasts as required. A two-month operational reforecast experiment demonstrates that the machine learning emulator achieves accuracy comparable to that of the traditional physical scheme, while accelerating the computation speed by approximately eightfold.
- Preprint
(5251 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-1735', Anonymous Referee #1, 29 Sep 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-1735/egusphere-2026-1735-RC1-supplement.pdfCitation: https://doi.org/
10.5194/egusphere-2026-1735-RC1 - AC1: 'Reply on RC1', Hao Jing, 30 Oct 2026
-
RC2: 'Comment on egusphere-2026-1735', Anonymous Referee #2, 30 Oct 2026
General Comments:
This paper describes the development of a machine-learning radiation emulator and its integration and testing in the CMA-GFS prediction model. While improving the computational efficiency of radiative transfer is a worthy objective and neural networks have some potential to achieve this goal, it is also critically important that fluxes and heating rates provided by radiation codes be physically based and strictly validated with measurements. The authors have reported significant progress in developing and integrating their ML radiation emulator into the weather prediction model, but their case for its success in replicating the operational model could be more rigorous. For example, the authors present mean absolute errors between the operational radiation code and the neural network (NN) that are averaged globally and over long periods of time to document broad scale biases, but this approach does not document any errors on the regional and temporal time scales that are most relevant to numerical weather prediction. The authors then show differences in synoptic fields and point out areas where the NN “outperforms” the operational model. If the intent of the NN at this stage is to simply duplicate the result of the operational radiation model, then there is no physical basis (i.e. improvement in physics) that justifies such a claim. In fact, it could be argued that the altered outcome suggests that the NN has not achieved its objective. The documented results may represent a situation in which randomly induced errors produce better results in some circumstances but for the wrong reasons, which to this reviewer does not represent an advancement in numerical weather prediction. The authors have made important advances in stabilizing the ML emulator in this context, but the consequences of the stabilization adjustments on the simulation and how frequently they are required should be documented. Furthermore, it’s not clear how the NN performs near the limits of its training when its output may be in error, but not to a degree that causes the prediction model to fail. While the NN development described by the authors represents progress toward this objective, a stronger case must be made that the NN approach is replicating accurate radiative fluxes and heating rates at spatial and temporal scales relevant to weather prediction. Reconsidering the manuscript for publication after major revisions is recommended.
Specific Comments:
Abstract
1. Introduction
2. RRTMG – Original radiation scheme in CMA-GFS
Page 4, Line 101: The authors may want to consider revising this sentence to read “…serves as the foundational code for the development and validation of many related radiation models”.
Page 4, Line 103: Recommend rewording to read “…while accounting for the spectroscopic effects of individual absorption lines”.
3. Dataset
Page 5, Line 149: What is the subset of input and output variables that were selected for use by the neutral network, and how were those variables selected? Were tests performed using other variable subsets, or with all variables, to be able to assess the relative impact of the subset selected on neutral network accuracy?
Page 5, Line 156: For model training and validation, was the dataset was randomly sampled globally or over some limited area? Were land and ocean areas appropriately sampled, and over land are both high and low elevation areas represented? Were clouds and aerosols adequately sampled? How much of the calendar year 2022 was represented in the data sets?
Page 5, Line 162: The data normalization procedure may support the neural network training, but is it known whether this step inadvertently over-emphasizes or under-emphasizes specific radiative processes in ways that may contribute to neutral network instability?
4. Machine learning model and training
Page 8, Line 195: Please clarify the spatial and temporal extent of the validation dataset used to generate the results in Figures 2 and 3? Presenting only the mean absolute error averaged over the entire dataset seems intended to show only mean bias and to average over any regional or short-term errors. What is the magnitude of any errors on the smaller spatial and temporal scales that are most relevant to numerical weather prediction?
Page 8, Line 196: The intent of examining the mean absolute error by model level over large spatial and temporal scales may be to quantify the presence of any bias on such large scales. However, any errors introduced by the ML emulator can be of either sign, and looking only at global MAE made disguise biases on smaller scales. To make a convincing case for the NN, the authors should present a more comprehensive comparison to the reference model that also indicates the largest errors that occur on scales most relevant to the diversity of radiative processes being simulated.
5. Coupling the ML emulator and CMA-GFS
6. Enhance the stability of the hybrid model
Page 12, Line 268: Please explain the justification for the use of “nearly eightfold” in describing the difference in performance for each model on different numbers of CPU cores listed in Table 2. The average difference appears to be 6-7 times.
Page 12, Line 281: What will happen in the circumstance when the ML emulator is operating in atmospheric conditions that are close enough to the limits of its training data to produce relatively large errors, but those errors are not large enough to cause the simulation to fail catastrophically over time? Will this situation be detected?
Page 13, Line 291: Define the symbol
Page 14, Line 324: Constraining the ML predicted fluxes from exceeding pre-defined “physical” limits is unlikely to produce the correct flux and heating rate for the associated atmospheric state. How often does this situation occur in the evaluation experiments?
Page 14, Line 340: Clarify that 500 hPa geopotential anomaly correlation and differences are shown in Figure 5 to represent “performance in synoptic field forecasting”
7. Performance in the operational reforecast experiments
Page 14, Line 343: Provide some detail on the derivation of the Threat Score presented and what its value means.
Page 14, Line 344: If the intent of the neutral network is to reproduce the operational radiation code as closely as possible, then there is no physical basis for the NN to “outperform” the operational model or for such apparent results to be described that way. Any discrepancies in the simulated fields are simply random errors introduced by the NN, some of which will result in benefit to the simulation and some of which will not. The differences shown in Figures 5-9 appear to indicate that the NN is not reproducing the fluxes and heating rates of the operational model on synoptic and smaller scales.
Page 14, Line 345: “Excelled” seems like an exaggeration give the threat score numbers presented.
Page 14, Line 347: Please review the discussion of precipitation bias in the text, which does not appear to match the results shown in Figure 7. While the hybrid model shows lower bias for the first two forecast days for light rain in January, the plots show it has higher biases for all other precipitation types and forecast times.
Page 17, Line 372: The discussion of Figure 10 seems inconsistent with the figure itself. January differences in 2m temperature absolute error appear to be up to 0.2 K in the figure, not “generally comparable” as described in the text. July differences in 2m temperature absolute error appear lower for the hybrid model at 12 hours and negligible at 24 and 36 hours, not “…the hybrid model outperformed the operational model within the first 1.5 days…” as stated in the text. Also, the July absolute error appears to increase with time for the hybrid model during the five days shown. What happens at longer forecast times? Describe what is meant by “relative error” in Figures 10 and 11.
Page 19, Line 405: The precipitation plots for the operational model in Figure 14a appear to show precipitation only over land, while the comparable plots for the hybrid model and for the observations show precipitation over both land and ocean. Redraw the panels in Figure 14a to show all three sources the same way (either land only or land and ocean) to make the comparison more insightful.
Page 20, Line 416: Training the NN directly from LBLRTM is possible, though it would require a considerable computational commitment to obtain a sufficiently comprehensive training dataset. Also, the issue of training for cloudy radiative transfer with LBLRTM would further complicate such an activity, especially in the shortwave spectral region, since a multiple scattering capability would also have to be considered.
8. Discussion and future work
Code and data availability
Page 21, Line 435: The two Zenodo data repositories listed here should be included in the references.
Acknowledgments
Figures/Tables
Technical Corrections:
1. Introduction
Page 2, Line 45: Replace “integrating” with “integrates”
Page 2, Line 51: Spell out the acronym CAM2 as “Community Atmosphere Model”
Page 3, Line 78: If CAM is defined earlier, then the acronym can be used here
Page 3, Line 86: Replace “adopts” with “adopt”
2. RRTMG – Original radiation scheme in CMA-GFS
Page 4, Line 104: Reword as “radiative fluxes and heating rates” or “radiative flux and heating rate”
3. Dataset
4. Machine learning model and training
5. Coupling the ML emulator and CMA-GFS
6. Enhance the stability of the hybrid model
Page 12, Line 284: Replace “form” with “from”
Page 13, Line 312: Replace “networks” with “network”
Page 13, Line 313: Replace “becomes” with “become”
7. Performance in the operational reforecast experiments
8. Discussion and future work
Figures/Tables
Page 6: In Table 1, in the row for “ch4vmr”, replace “Mathane” with “Methane”
Page 9: In Figure 2a, change “SW” to “LW” in the top left panel y-axis label. Add units to the y-axis labels in all panels.
Page 15: In the Figure 5 caption, replace “fields” with “field”. Revise the caption and the text to mention that 500 hPa geopotential anomaly correlation and differences are shown in the figure. Explain the meaning of the bars in the difference plots.
Citation: https://doi.org/10.5194/egusphere-2026-1735-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 253 | 108 | 36 | 397 | 38 | 38 |
- HTML: 253
- PDF: 108
- XML: 36
- Total: 397
- BibTeX: 38
- EndNote: 38
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
