acceptodds
Under review as a conference paper at ICLR 2027

AEGIS: Differentiable Mars Climate Model with Neural Closures

Abstract

General circulation models (GCMs) are the primary tool for simulating planetary atmospheres. They play a vital role in understanding Mars's atmosphere, as forecasting its unique weather is mission-critical for operations such as entry, descent, and landing. Mars poses unusual challenges for these models, as observations are sparse compared to Earth. In addition, a thin atmosphere alongside a radiatively active dust cycle creates a volatile atmosphere with large diurnal temperature swings and no true terrestrial analog for validation. Existing Mars GCMs, including the LMD PCM, the NASA Ames Mars GCM, and PlanetWRF, are mature and physically detailed but are implemented in legacy Fortran with finite-difference or finite-volume solvers, and they do not expose gradients for calibration or machine learning. Here we present AEGIS, a modular differentiable Mars climate model that couples Mars's unique atmospheric physics to the Dinosaur dynamical core, with interfaces for neural closures. We showcase stable ten-Mars-year simulations that reproduce the seasonal cycle while conserving the total inventory, capture realistic large-scale surface-temperature structure, and produce surface pressure that follows Mars Orbiter Laser Altimeter (MOLA) topography. Gradients through coupled trajectories agree with finite differences and support physical calibration and neural training. We compare with conventional GCMs, highlighting the framework's computational efficiency and differentiability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Differentiability, Mars, Physics, GCM, Climate Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NbUAOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2(−1)
Read review
Paper summary

The paper introduces Aegis, a framework to achieve both physically differentiable modeling and neural-physical hybrid modelling of planetary atmosphere, detail its application on Mars atmosphere, and explain its performance compared to reference datasets.

Critical strengths

The proposed method (hybrid modeling with differentiable programming and neural network) is aligned with the main current direction of the ML community for atmospheric and climate science. The work is hence promising. Can be used for both physical parametrisation and coupled with a learned neural model.

Critical weaknesses

The paper does not allow reproducible results as part of the paper are hard to read, details are missing and some claims are not fully supported by the evidence. In addition, it does not indicate in which practical context the use of the model output is beneficial.

Questions for authors

A. In table 2, it is not clear if the calibration gain is significant or in the uncertainty estimates (is there any?). Also, the neural heating gain is null and is not discussed. Does it make sense to keep it?

B. In table 3, is pressure comparison relevant if very mostly correlated to terrain? Could CO₂ pressure cycle be compared with observations?

C. In table 4, why does trajectory training lose here, and is there a setting where it wins?

D. Can parameters tuned during the 5 sols gradient study be valid for the 10 years stable simulation? The gap between the two horizons seems significant

E. The paper implies Aegis generalises the modeling of planet atmosphere and then apply Mars planetary constants, but this is untested. Would have been nice to confirm the approach by exposing the results of applications on other documented planet, such as the Earth. Otherwise better to just mention it in future work if intented?

F. The writing is very heterogeneous, some paragraphs are really high level, while some others are diving directly into very technical details without introduction (especially the math related explanations), forcing the reader to trust the authors about the correctness of the implementation, because lost with numerous technical details (while not even exhaustive, c.f. hardware below). It is surprising to jump from understandable claims, to complex maths, then to very generic inline codes. For example, section 2.1 jumps from a high-level description to Eqs. 1–2 without defining dI/dt or the projection step. Another is mention of MY32 and ARCO-MACDA without introduction. This is one of the main concern of the reviewer, the gap between the clear description of the objective and the technical explainations. Could this be improve?

G. Hardware setup not detailed enough ("single GPU" or "one GPU" mentioned several times) for teams that would want to reproduce the results. What card/how much VRAM? Moreover, this is critically related to the performance claims of the approach (the execution time in comparison with reference model in Fortran).

H. Could you indicate how the output resolution, levels and frequencies of the model are useful for downstream usecase such as the operations mentioned in abstract

I. Could you harmonised the way you reference figures and tables (fullname or abbreviations) and make sure all of them are referenced? For example, Figure 1 not referenced anywhere, but useful to understand the overall architecture.

J. Section 3.2 gives the full-year CO₂ drift as 2.8×10⁻⁹. Section 3.4 gives 4.28×10⁻⁹ for "the full-physics year". Aren't these the same run?

1 comment

Thank you. The framework supports updated condensation and cloud physics.

W1. Reproducibility and practical value

We provide inputs, training and averaging rules, and confirmed hardware.

We have expanded methods in Section 2.1 and Appendices B, C, and D, and explained uses in Appendix C.2.

A.1. Calibration gain and uncertainty

Calibration lowers test temperature error, a 7.84% reduction. The estimated gain interval is 0.109 to 0.314 K. It is based on only two held-out years.

We have reported this in Section 3.1 and Appendix B.2.

A.2. Null neural-heating gain

We present the neural components as a way for researchers to implement neural experiments using the AEGIS framework. We have further qualified our writing to support this view and reported the experiment values without any claims of efficiency or accuracy.

We have addressed this in Section 3.5 and Appendix B.3.

B.1. Terrain-driven pressure correlation

Shared terrain explains much of the raw agreement. Removing the elevation component reduces the given correlations compared with raw values.

We have added the adjusted correlations and method in Section 3.3 and Appendix C.1.

B.2. Observed pressure cycle

We now compare with Viking station records. Seasonal anomaly correlations are 0.959 and 0.950, while pressure errors are 56.8 and 76.0 Pa.

We have added this comparison in Section 3.3 and Appendix C.2.

C.1. Why trajectory training loses

We present the neural components and the experiments as a way for researchers to implement neural experiments using the AEGIS framework. We have further qualified our writing to support this view and reported the experiment values without any claims of efficiency or accuracy.

We have stated the budgets and results in Section 3.5.1 and Appendix B.4.

C.2. A setting where it wins

Trajectory fitting has a lower mean error on the follow-up interpolation test. Continued local fitting performs better under higher dust. The ranking depends on test conditions, and longer coupled runs also fail rollout checks. We have added the repeated tests and failures in Appendix B.4.

D. Short gradient study versus long simulation

The short study checks gradients at the default opacity; it does not tune parameters. Calibration uses shorter windows. The long integration tests whether the simulation remains finite.

We have clarified these experiments in Section 3.1 and Appendix B.5.

E. Other planets

Validation is limited to Mars. Other planets require their own physical models and tests.

We have stated the Mars scope and placed other-planet validation and terraforming in future work in Section 4.

F.1. Equations and projection

Inventory means total carbon dioxide across the tracked reservoirs. The projection corrects numerical errors in this total after each step. We have defined the budget in Section 2.1 and explained the correction and its effects in Appendices A.3 and D.2.

F.2. Dataset names and writing

The revision introduces Mars-year labels and ARCO-MACDA with their sources and averaging.

We have expanded the physical explanation in Section 2.1 and dataset descriptions in Section 3.1 and Appendices B.1 and C.1.

G.1. Hardware

Historical timing and sensitivity runs use an NVIDIA A100. New calibration and projection runs use an NVIDIA A100 40 GB VRAM and double precision. We have added confirmed hardware and timing scope in Appendix D.3.

G.2. Speed comparison

Hardware and configuration differences prevent a fair Ames speedup claim. We have removed it and report absolute throughput in Section 3.2 and Appendix D.3.

H. Resolution and practical use

The model uses a 64 by 32 transform grid with 12 vertical layers; comparison fields are saved about hourly. This supports coarse global and seasonal studies. Operational forecasts need local validation. We have specified resolution in Section 2.1, sampling in Appendix C.1, and uses in Appendix C.2 and Section 5.

I. Figures and references

We have cited the architecture figure in Section 2.2, checked cross-references, and corrected duplicate material and the seasonal count in Appendix C.

J. Conflicting annual drifts

These values come from different runs. The drift of 2.8 parts per billion belongs to the cold-start timing run with a 450-second step; 4.28 parts per billion belongs to the restart-based ablation with a 300-second step. Both measure the largest relative inventory drift at saved times.

We have distinguished the runs in Sections 3.2 and 3.4 and Appendix D.2.

Reviewer_cTvCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces AEGIS, a differentiable Martian climate simulation framework built upon the Dinosaur dynamical core. The authors incorporate Martian radiative processes, heat exchange between the surface and subsurface, CO₂ condensation/sublimation, and atmospheric mass coupling, while providing an interface for neural parameterization. The experiments cover gradient verification, long-term stable integration, comparisons with reference climate fields, and preliminary parameter calibration.

Critical strengths
  1. Integrating the atmosphere, surface, frost layers, and subsurface thermal states into a unified, differentiable system offers significant value for both engineering and scientific applications.

  2. The framework’s correctness is systematically validated through comparisons between automatic differentiation and finite-difference approximations, multi-year integrations, and ablation studies of physical processes.

  3. The neural radiation interface is well-designed, ensuring energy budget closure for discrete atmospheric columns through the construction of radiative fluxes.

Critical weaknesses
  1. Methodological innovation is limited. The core dynamics, automatic differentiation, and neural parameterization approaches are inherited from existing frameworks; The paper fails to clearly articulate any novel, generalizable computational or learning methods that emerge from this coupling process.
  2. Physical validation remains inadequate. The minimal drift in the CO₂ inventory is partly a result of successive projection; the magnitude of these corrections and their subsequent impact should be reported. Correlations in atmospheric pressure largely reflect shared topography, and sampling discrepancies between reference datasets further complicate the assessment of climate fidelity.
  3. Efficiency comparisons lack a unified basis. Comparisons with runtime data from legacy Ames documentation involve different hardware, resolutions, and physical configurations, rendering any quantification of speedup advantages unreliable.
Questions for authors
  1. Beyond the differentiable implementation of existing physical processes, what is the most critical novel technique introduced in this paper? What specific challenges—which existing frameworks cannot directly address—does it resolve?

  2. Could the practical value of gradient-based optimization be demonstrated—and compared fairly against gradient-free methods—through calibration or inverse-problem solving over longer time horizons and across multiple multi-year windows?

1 comment

Thank you for the review. We have updated the carbon dioxide condensation and cloud physics.

W1; Q1. Novel method and coupling challenge

We agree that the dynamical core, automatic differentiation, and the general approach to neural parameterization are inherited, and the revision says so plainly. What existing Dinosaur-based frameworks do not address is an atmosphere whose main constituent condenses. On Mars, the bulk gas itself moves between atmosphere and surface, and global-mean surface pressure swings by 30% over a year in our runs. Earth configurations have no analogous exchange. Handling this inside a differentiable model required three things:

  • Mass exchange in the prognostic state. CO₂ phase change feeds directly into the prognostic log surface pressure (Eq. 2), and frost and regolith temperature are carried in the differentiated state, so the mass exchange is part of the gradient.
  • A conservation projection that is itself differentiated. The gradients inherit the conservation law: the sensitivities of mean surface pressure and mean frost are equal and opposite at every horizon (Section 3.1).
  • A flux-form neural radiation interface. The network predicts interface fluxes, and heating is computed from their convergence, so the column energy budget closes for any weights.

Gradient checks confirm that the conservation correction preserves the atmospheric and frost balance. The learned radiation component closes the column radiation budget by predicting fluxes. This gives a reusable way to preserve physical balances while learning through a coupled simulator.

We have explained the contributions and credited inherited methods further in Appendix A.

W2.1. Projection corrections and impact

The small reported drift is enforced by the correction. We now measure the correction and compare runs with and without it, retaining physical phase change. Cumulative absolute corrections total 1.388% of initial inventory over the first year. Removing the correction produces 0.883% inventory growth and changes global pressure by 2.63 Pa RMS, with a maximum difference of 6.50 Pa.

We have added correction sizes and effects in Section 3.2 and Appendices A.3 and D.2.

W2.2. Topography and observed pressure cycles

Shared terrain inflates the raw pressure correlations, which range from 0.978 to 0.990. Against Viking station records, seasonal anomaly correlations are 0.959 and 0.950, with pressure errors of 56.8 and 76.0 Pa. Model peaks occur too early, and dust is unmatched to the observed years.

We have added these comparisons in Section 3.3 and Appendices C.1 and C.2.

W2.3. Temporal sampling

Using the same time averaging as the reference data lowers temperature errors in every comparison. We have updated the results in Section 3.3 and explained the averaging in Appendix C.1. These comparisons use the revised model, which also includes physics changes.

W3. Efficiency comparison

We have removed that claim of any speedup with comparable frameworks and report our own throughput with hardware and timing scope in Section 3.2 and Appendix D.3.

Q2. Optimization over longer and multiple multiyear windows

The current experiments demonstrate short-window calibration. In the spatial-albedo experiment, a uniform change performs better than gradient optimization with less computation. We have narrowed claims and clarified the tested horizons in Appendix B.5. The longer-horizon comparison remains a priority next step for improving the framework.

Reviewer_MHJVOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce a differentiable global circulation model for Mars' atmosphere with an interface for closure models.

Critical strengths
  • they introduce a differentiable GCM for Mars' atmosphere. I am sure that this is something the geophysics/planetary science community will appreciate.
  • the authors demonstrate that this allows for coupling and optimizing neural closures.
  • the paper is mostly well written
Critical weaknesses
  • my main concern is venue fit. Given that the paper develops a differentiable solver for Mars atmosphere, I think a domain-specific venue might be more appropriate.
  • the paper mainly reads like a documentation of a software integration project with little information about what the methods are that its implements with the GCM and what associated challenges are.
Questions for authors
  • why the choice of a lat-on-grid, why finite difference methods?
  • what measures are taken to ensure stability
  • what is the CFL condition and is it prohibitive in training?
2 comments

We thank you for the review.

W1. Venue fit

AEGIS supports learning through a differentiable simulator. ICLR explicitly accepts papers under the category "application to physical sciences". In the past, ICLR has published physical-science simulation methods such as "Learning to Control PDEs with Differentiable Physics" and "Learning Mesh-Based Simulation with Graph Networks" which resemble the work this paper presents. AEGIS brings this approach to an atmosphere whose main gas freezes onto the surface. Its time-stamped trajectories also support representation learning and neural-surrogate training alongside reanalysis data.

We have clarified the contribution in Section 1 and believe this is the right venue for this work.

W2. Methods and coupling challenges

We agree that the submission spent too much space on interface details and too little on methods and their difficulties. The revision removes the interface-level description and adds a numerical formulation subsection covering the discretization, time stepping, diffusion and the Mars-specific, we significantly revised each paragraph to improve coherence and clarify the methods used and the challenges presented.

The main challenge is tracking atmospheric gas as it freezes and returns from the surface while keeping that exchange differentiable. The inherited core does not provide this Mars coupling. Gradient checks confirm that the conservation correction preserves the balance between atmospheric and frost sensitivities. The learned radiation component also preserves the column radiation budget through its flux predictions.

We have explained these contributions and credited the inherited methods in Sections 1, 2.1 and 2.2 and Appendix A.

Q1. Grid and finite differences

AEGIS uses a spectral solver. The longitude-latitude grid evaluates nonlinear terms and local physics. Finite differences independently check gradients. We chose Dinosaur because it provides a differentiable atmospheric core with established use in NeuralGCM. We have clarified the solver and grid roles in Section 2.1 and Appendix D.1.

Q2. Stability measures

The solver handles fast gravity waves implicitly, damps small-scale noise, and corrects the carbon-dioxide inventory after each step. Learned heating and radiation outputs have explicit bounds. The long physical-model integrations remain finite at saved states. Longer learned-radiation tests also reveal failures, which we report.

We have explicitly documented the stability safeguards in Section 2.1 and Appendix D.1, the integration evidence in Section 3.2, and the learned-model failures in Appendix B.4.

Q3. CFL condition and training cost

Gravity waves are handled implicitly; transport and other explicit processes still limit the time step. Halving the step changes temperature by only 0.0042 K at a 44,400 s horizon, and it is not prohibitive in training. It uses 25 steps of about 296 seconds and recomputes intermediate states to save memory.

We have added additional step-size evidence in Appendix D.1 and training details in Section 2.2 and Appendices B and D.3.

Thank you for the clarifications. I acknowledge that the dynamical core is spectral and that my Q1 was based on a misreading of the original text - revising Section 2.1 would help. I don't think you can expect many readers in this community to be aware of the Dinosaur dynamical core.

My remaining concern is the strength of the ML contribution:

  • how is the condensation/sublimation threshold made differentiable, and how well-behaved are gradients near the frost point?
  • Learned closures are trained on ~2 h unrolls (25 × 296 s), and the step-size test covers only ~0.5 sol, while Appendix B.4 reports failures in longer learned-radiation runs. Could you quantify when these failures occur and whether longer training horizons help?
  • How does the physical model compare against an established Mars GCM or the Mars Climate Database, and what are the per-step runtime and memory costs?

I am still mainly concerned about venue fit but I recognize that my other concerns are mostly addressed and I have decided to raise my score to a 2. My justification for this is the strength of the ML contribution and the venue fit.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.