the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 30 Jul 2026
A Geometry-Aware AI Emulator for the Coupled Whole Atmosphere from Earth’s Surface to the Ionosphere and Thermosphere
Abstract. Whole-atmosphere models such as WACCM-X resolve coupling from the Earth surface to the Mesosphere-Lower-Thermosphere (MLT), and Ionosphere-Thermosphere (IT) systems with expensive computational costs. Here we introduce CAM-NET, a geometry-aware Spherical Fourier Neural Operator (SFNO) surrogate for emulating WACCM-X variability from Earth’s surface to IT region. CAM-NET is trained on 3-hourly WACCM-X simulations and predicts neutral winds, temperature, pressure-coordinate vertical velocity, electron density, and zonal ion drift. The framework combines a Spherical Fourier Neural Operator (SFNO) backbone with a newly developed lightweight module that extends the frozen atmospheric representation to plasma variables. Compared with a planar Adaptive Fourier Neural Operator (AFNO) baseline, CAM-NET backbone achieves lower errors for neutral variables and higher anomaly correlations across variables and altitudes. For the held-out simulation, CAM-NET preserves the dominant IT morphology and remains stable during multi-day autoregressive rollouts. Spherical-harmonic diagnostics show that the model retains low-degree variability while damping high-wavenumber mesospheric structures, especially near 90 km where gravity wave breaks. CAM-NET is intended as a computationally efficient emulator of WACCM-X, rather than an operational forecasting system. These results demonstrate its potential for rapid ensemble experiments, uncertainty quantification, and sensitivity studies of large-scale coupled whole atmospheric variability.
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-4362', Anonymous Referee #1, 18 Sep 2026
-
RC2: 'Comment on egusphere-2026-4362', Anonymous Referee #2, 29 Sep 2026
Recommendation: Minor Revision / Technical Correction
This manuscript presents a promising proof-of-concept for CAM-NET as a fast, computationally efficient surrogate for WACCM-X. The use of SFNO to address poleward distortions is well-motivated, and the stability demonstrated over multi-day rollouts is a notable achievement. However, the manuscript requires adjustments to its overall framing, expanded discussion of its physical limitations, and modifications to its presentation to clarify the model's actual predictive skill. Because these issues can largely be addressed through text revisions, expanded discussion, and figure reorganization rather than requiring fundamental architectural changes or retraining, I recommend a minor revision.
Major Comments (Methodology & Framing)
- The authors made a substantial compromise by extracting only every fifth vertical level from the native WACCM-X grid to reduce dimensionality. I suspect that this structural choice may restrict the model's ability to resolve vertical wave propagation, which might contribute to the poor performance for gravity wave structures near 90 km. The authors should discuss the impact of this vertical downsampling choice on vertical wave momentum deposition and overall vertical wave propagation.
- Table 1 reports ACC only at the 120-h endpoint, which provides limited information about how predictive skill evolves during the autoregressive rollout. The lead-time-dependent ACC curves in Figure S1 provide useful information on error growth and model stability, although these comparisons are currently limited to the common 30-h interval because the AFNO baseline becomes unstable thereafter. I recommend moving Figure S1 to the main text, or otherwise providing CAM-NET lead-time-dependent ACC through the full 120-h rollout, so that the temporal degradation of predictive skill can be evaluated directly.
- To my understanding, the model utilizes a purely statistical geometric L2 loss function. The authors should comment on the consequences of this approach regarding the production of non-physically consistent states. (I.e., the model might find a state which minimizes statistical error but violates the primitive equations.) Additionally, the "frozen" atmospheric representation means the ionosphere is only one-way coupled to the neutral atmosphere. Because the atmospheric backbone is frozen and the ionospheric module acts downstream of the atmospheric representation, the model does not appear to permit explicit ionosphere-to-neutral feedback during inference. The authors should clarify which plasma–neutral coupling effects are represented implicitly through the WACCM-X training data and which cannot be dynamically represented by the CAM-NET architecture, particularly ion drag. Finally, the authors should briefly justify the specific selection of zonal ion drift over vertical ion drift as a primary target, given the importance of vertical plasma transport to ionospheric dynamics.
- The introduction identifies rapid ensemble generation and uncertainty quantification as important motivations for CAM-NET. However, the manuscript does not evaluate sensitivity to perturbed initial conditions (or perturbed boundary conditions, which are often used to generate ensemble spread in highly forced systems like the IT) or whether autoregressive trajectories develop physically meaningful ensemble spread. The authors should clarify whether such behavior has been tested and, if not, identify it as an important direction for future evaluation.
- The held-out inference period is limited to 2014, while the training period spans 2001–2011 and the model uses F10.7 and Kp as external drivers. The manuscript demonstrates an additional March 2015 geomagnetic-storm stress test in Appendix B, but this primarily addresses disturbed geomagnetic conditions rather than solar-cycle dependence. The authors should discuss whether CAM-NET is expected to generalize similarly across different phases of the solar cycle and acknowledge that multi-year validation spanning solar minimum and maximum conditions is needed to establish this. If the authors have already performed such an analysis, it would be useful to include or reference it explicitly.
Minor Comments
- The abstract would benefit from some quantification, for example quoting some of the reported 120-h ACC so that readers have an idea of how well the emulator performs from the abstract.
- The methodology section references training the model for K successive steps in the second multi-step fine-tuning stage. The value of K is never stated and should be explicitly provided.
- Briefly define "geometry-aware" and "frozen" within the introduction to ensure the concepts are immediately clear to readers unfamiliar with standard neural network architectural jargon.
- The first paragraph of the introduction currently contains no citations. In general, this paper could benefit significantly from more citations. The first paragraph could benefit from citations supporting the statements concerning whole-atmosphere coupling, computational cost, and the downstream impacts of thermospheric/ionospheric uncertainty.
- The second paragraph of the introduction heavily relies on AI weather models from 2023. Given the manuscript's July 2026 version date, the introduction should be updated to reflect major developments in AI-based weather prediction since the cited 2023 literature, including recent generative/physics-informed forecasting approaches and relevant reviews. (as an example, Selz & Craig 2026 includes a good summary of state-of-the-art ML emulators in the lower atmosphere).
- The first sentence of the second paragraph in the introduction cuts off abruptly ("Data-driven weather models have recently demonstrated strong foreca.").
Citation: https://doi.org/10.5194/egusphere-2026-4362-RC2
Viewed
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 165 | 0 | 1 | 166 | 0 | 0 |
- HTML: 165
- PDF: 0
- XML: 1
- Total: 166
- BibTeX: 0
- EndNote: 0
Viewed (geographical distribution)
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This paper demonstrates CAM-NET, a novel way to apply spherical Fourier neural operator to WACCM-X output to emulate WACCM-X in a computationally efficient way, and which addresses the limitations of other deep-learning methods.
Main comments:
In general, there is a lot of assumed knowledge on the part of the reader, both from the atmospheric modelling and deep-learning perspectives. I do not think the author currently bridges that gap in the text or references enough to allow both communities to follow all aspects of the paper.
It is not entirely clear how CAM-NET is run, i.e. what kind of WACCM-X state or file it is initialized from or how often/when it is re-initialized through the validation/test period. Also what is output from CAM-NET? Is that output on pressure levels?
The analysis performed in the results section needs to be expanded. The current analysis is explained in a way where it’s not entirely clear what assumptions have been made (for example if a global average has been taken without cosine latitude weighting). Additionally, absolute metrics are provided which makes comparison across altitudes and variables difficult. Some further analysis could be done at the discretion of the authors to explore the results pass global means (if this is what has been done).
The ionospheric module appears to be trained on every 5th level of the WACCM-X dataset. This could be skipping the ionospheric peaks such as the F2 peak and causing significant deviation from WACCM-X.
Appendix A could happily be moved into the main text
Small comments are included in the attached PDF.