MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres

G.Villanueva; J.A. Caraballo-Vega; J. Li; M.J. Frost; M.L. Carroll; S.D. Guzewich

REVIEW 4 major objections 5 minor 10 references

MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An Earth-trained AI weather model can be fine-tuned to forecast Martian temperature and wind, reproducing the diurnal cycle within 10 training epochs.

desk verdict Useful first demonstration of fine-tuning an Earth weather foundation model for Mars, but the evaluation stops short of proving forecast skill beyond climatology. read the letter →

arxiv 2608.05054 v1 pith:KKUIOPGG submitted 2026-08-05 astro-ph.EP cs.AIcs.CVcs.LG

M.L. Carroll , J. Li , S.D. Guzewich , G.Villanueva , J.A. Caraballo-Vega , M.J. Frost This is my paper

classification astro-ph.EPcs.AIcs.CVcs.LG

keywords GraphCastMarsatmospheretransferlearningweatherforecastingfoundationmodelClimateDatabasediurnalcyclefine-tuning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether a weather foundation model trained on Earth's atmosphere can be transferred to another planet. Taking the Earth-trained GraphCast model and the Mars Climate Database, the authors first show that zero-shot forecasts collapse toward climatological mean states and lose the day-night cycle. They then fine-tune GraphCast with Mars temperature and wind fields, top-of-atmosphere solar radiation, and constant humidity, producing a model they call MarsCast. Within about 10 training epochs, MarsCast begins to capture the diurnal cycle, and 10-day autoregressive forecasts reproduce seasonal and vertical temperature structure over most of the atmosphere. The paper's central claim is that Earth-trained AI weather models encode transferable atmospheric dynamics that can be adapted to a very different planet with modest data and compute.

What carries the argument

The load-bearing machinery is the GraphCast architecture, a graph neural network that represents atmospheric state on a geodesic grid and evolves it through an encoder-processor-decoder with message passing. The adaptation procedure surrounds it with Mars-specific choices: min-max rescaling of Mars fields into the ERA5 temperature range to avoid activation saturation, top-of-atmosphere solar radiation as forcing, six-hour sampling at four local times with sequential triplets, a land-sea mask set to unity, and humidity held at its global mean. Fine-tuning uses AdamW at learning rate $1 \times 10^{-6}$ with mean squared error loss and all encoder, processor, and decoder weights unfrozen, so the pretrained spatial representations are preserved while planetary thermal dynamics are learned.

What would settle it

Take dates with independent Mars observations, such as lander station temperature and pressure records or orbital retrievals, hold those dates out of training, and compare MarsCast forecasts to the observations rather than to the Mars Climate Database. If MarsCast reproduces the database fields but misses observed diurnal amplitudes or dust-storm temperature signals, the central claim of forecasting real Martian weather is not established.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that an Earth-trained weather foundation model can be adapted to forecast Martian weather through targeted fine-tuning. Zero-shot GraphCast, fed Martian initial conditions, produces plausible spatial patterns but no diurnal variability and decays rapidly to climatology. After fine-tuning with Mars Climate Database variables and solar forcing, the model learns Martian diurnal heating and cooling, reproduces the vertical temperature profile at lander sites, and keeps skill across four seasonal initializations and out to 10 days. The authors interpret this as evidence that the spatial and temporal propagation structures learned from Earth reanalysis are partly universal, while the radiative and diurnal forcing must be learned from planet-specific data.

Load-bearing premise

The load-bearing premise is that the Mars Climate Database faithfully represents the real Martian atmosphere; the paper trains and evaluates entirely against this simulation, and the authors themselves note that comparison with observations is still needed.

Editorial extensions

If this is right

MarsCast produces 10-day global forecasts on a 1-degree grid in roughly two GPU-minutes, orders of magnitude cheaper than a full Mars general circulation model, which makes ensemble forecasting practical.
Because the fine-tuned model inherits GraphCast's 0.25-degree internal training resolution, downscaling coarse Mars Climate Database input to 1 degree appears to retain finer-scale detail in the output, pointing to a cheap learned downscaling capability.
Prediction quality improves with the number of training days and varies with the season of initialization, so operational users could choose training and initialization windows to maximize skill.
The same fine-tuning recipe, with variable rescaling and a solar forcing term, is directly transferable to other planetary atmospheres for which a reanalysis-like dataset exists.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the Mars Climate Database misrepresents real events such as dust storms, MarsCast's apparent skill against held-out database dates will not carry over to actual Mars weather; the paper's own caveat that observational validation is still needed marks this as the next test.
The constant-humidity simplification could mask a real coupling: near-surface humidity on Mars varies with surface properties and time of day, and letting humidity be prognostic might change boundary-layer temperature evolution, a prediction that could be tested by un-freezing humidity during fine-tuning.
The observed downscaling behavior suggests part of the pretrained model acts as a learned super-resolution operator; a direct test would be training MarsCast on Mars Climate Database output at several native resolutions and measuring whether finer output skill improves beyond bilinear interpolation.
Read as a physics probe, the zero-shot collapse identifies which Earth-encoded processes are regime-specific, such as radiative diurnal forcing and boundary-layer thermal inertia, rather than universal, such as horizontal propagation and large-scale wave structure, offering a quantitative map of what a foundation model has actually learned.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, a useful proof-of-concept: the authors are the first to take a pretrained Earth weather foundation model (GraphCast) and fine-tune it for Mars. They release code and checkpoints, and the zero-shot-versus-fine-tuned comparison is informative. The qualitative evidence that the model learns diurnal and seasonal temperature structure is credible, and the claim about computational cost savings is reasonable.

The soft spot is the word 'forecast.' All evaluation is against MCD, the same simulation family that produced the training data, and the paper provides no quantitative skill scores against persistence, climatology, or the zero-shot model. Because the inputs include solar longitude and local time, a model that simply reproduces the diurnal/seasonal climatology would score well on held-out dates. The figures look good, but they do not separate climatology from actual predictive skill. This is the load-bearing gap, and it is fixable: report RMSE and anomaly correlation against persistence and climatology, and, ideally, a short validation against real Mars observations (for example, REMS or MCS data).

Smaller issues: the min-max scaling may have been computed over the full dataset, so it is worth checking for normalization leakage; three cited works (Roy et al. 2026, Uprety et al. 2025, Singh et al. 2026) are missing from the reference list; and the 'downscaling' observation in the discussion is based on visual inspection, not a metric.

Who is this for: planetary scientists and machine learning researchers working on transfer learning for atmospheres. It is a legitimate first step and deserves a serious referee. The revision needs to add quantitative baselines and ideally an observational check; without that, the central claim remains a plausible hypothesis rather than a demonstrated result.

Referee Report

4 major / 5 minor

Summary. The paper investigates whether an Earth-trained AI weather foundation model (GraphCast) can be adapted to Mars by fine-tuning on Mars Climate Database (MCD) fields. The authors evaluate zero-shot GraphCast, which they find decays to climatology and lacks diurnal variability, and then fine-tune on MCD temperature, wind, and solar-forcing inputs with humidity held constant. They report that after roughly ten epochs the fine-tuned model (MarsCast) begins to capture the diurnal cycle, that 10-day forecasts reproduce seasonal and vertical temperature structure, and that winds are predicted with qualitative accuracy. The paper releases training code and model checkpoints.

Significance. If the claims were quantitatively established, this would be a useful proof-of-concept for transferring terrestrial learned weather representations to other planets, with practical implications for mission support and ensemble forecasting. The central idea is novel and the computational-cost argument is compelling. The paper also has tangible reproducibility assets: a public GitHub repository, Hugging Face checkpoint release, and a compact fine-tuning configuration. However, the evidence presented is almost entirely qualitative, and the lack of any skill metric against persistence or climatology means the core claim of 'forecasting Martian weather' is currently unsupported even within the MCD framework.

major comments (4)

[§4, Figures 2–6] The paper reports no quantitative forecast-skill metrics (RMSE, MAE, anomaly correlation, vector-error statistics) and no comparison against baselines such as persistence, climatology, or the zero-shot GraphCast. Because the fine-tuning loss is MSE on temperature (§3.3) and the inputs include solar longitude and local time (§3.4), a model can minimize loss by deterministically predicting the seasonal/diurnal climatology while having no day-to-day weather skill. The conclusion in §6 that MarsCast 'can be adapted to forecast Martian weather' therefore needs quantitative support: domain-averaged and per-level skill scores versus persistence and climatology, at lead times from 6 h to 10 days.
[§3.5 and §4.2] The validation protocol is under-specified. The text states that checkpoints were selected by 'lowest validation RMSE' and that rollouts were compared with 'dates that were held out of the training,' but there is no description of how validation/test dates were partitioned (chronological or random), whether seasons used for validation overlap with training seasons, or how many distinct initialization dates were used for the reported figures. Also, the min–max rescaling in §2.1—if computed over the full MCD dataset—would allow test-set statistics to enter the inputs. Please specify the data split, the normalization bounds, and the exact checkpoint-selection rule.
[§5, last paragraph] The authors correctly note that 'comparison with observations is still needed,' but the abstract and conclusion claim the model predicts 'Martian weather' without that qualification. Since training and evaluation both use MCD, a single GCM simulation, the fine-tuned model may be learning simulator-specific behavior (e.g., parametrized dust, thermal tides, or interpolation artifacts) rather than real Mars dynamics. The load-bearing limitation should be addressed either by restricting all central claims to 'skill within the MCD framework' or by including validation against an independent MCD realization or against observations (e.g., REMS, MEDA, MCS).
[§4.5, Figure 6] Wind performance is reported only through a visual comparison and the qualitative bound 'within ±4 m/s for at least the first 48 hours.' No time series of wind RMSE, vector-error statistics, or comparison with persistence is provided, even though winds are a claimed output of MarsCast. The discrepancy figures are useful, but a quantitative error analysis is needed to support the claim that winds are 'successfully' predicted.

minor comments (5)

[§4.4 and Figure 4 caption] The text lists seasonal initializations as Ls = 60°, 150°, 240°, and 330°, while the Figure 4 caption lists Ls = 60°, 120°, 240°, and 330°; please reconcile the inconsistency.
[§4.5, first sentence] The sentence 'For the final experiment both meridional and zonal winds were included during fine-tuning, enabling MarsCast model to jointly predict meridional and zonal winds in addition to temperature (Figure 5)' appears to refer to Figure 6, not Figure 5; the figure callouts in this section need correction.
[Figure 1 caption] The caption states 'training loss shown in the middle,' but a zero-shot forecast involves no training; the middle panel likely shows a different quantity and should be relabeled.
[References] The Introduction cites Roy et al. (2026), Sagar Uprety et al. (2025), and Singh et al. (2026), but these references are missing from the reference list; all cited works should be included.
[§2.2, Table 1] The row 'Sea surface temperature Computed from temperature and pressure' is confusing for a land-only Mars configuration; please clarify how this quantity is used, given that the land–sea mask is set to unity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MarsCast's held-out MCD forecasts are genuine test-set outputs; the MCD-only evaluation is a data-fidelity limitation, not a circular derivation.

full rationale

No circularity was found. The fine-tuned MarsCast model is trained on MCD fields and evaluated autoregressively on dates held out from training (Section 3.5: 'Autoregressive rollouts were performed using the model checkpoints with those results compared to dates that were held out of the training'). These are genuine test-set outputs, not refitted values, so the supervised learning loop is not circular. No fitted parameter is renamed as a prediction, no load-bearing self-citation or imported uniqueness theorem appears, and the GraphCast architecture is cited to external work (Lam et al. 2023). The paper itself acknowledges in Section 5 that 'comparison with observations' and quantitative degradation analysis remain future work. The main scientific limitation—training and evaluation both use the MCD simulator, so skill against MCD may not transfer to real Mars—is a data-fidelity and external-validity concern, not a definitional reduction of the claimed result to its inputs. Similarly, the absence of RMSE/baseline comparisons (e.g., persistence or climatology) is an evaluation-strength issue, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central claim depends on data-normalization statistics, a fixed humidity value, and the choice of MCD as ground truth, which are the main free parameters and assumptions.

free parameters (4)

Min-max rescaling bounds per Mars variable = not stated
Temperature fields are rescaled to the ERA5 range using global Mars min and max (Section 2.1). If these statistics are computed over the whole dataset including validation dates, test information leaks into the normalization.
Global mean specific humidity = not stated
Humidity is held constant at the global mean z-score to isolate temperature variability (Section 2.2). This is a hand-chosen simplification.
Learning rate = 1e-6
AdamW learning rate chosen to preserve pretrained representations while allowing Mars adaptation (Section 3.3).
Training epochs and checkpoint selection = 1000 epochs, lowest validation RMSE
Early stopping was turned off and the checkpoint with the lowest validation RMSE was used for experiments (Section 4.2), a model selection choice.

assumptions (5)

domain assumption MCD is a faithful representation of the Martian atmosphere
The study treats MCD as ground truth for evaluation (Section 2.2). If MCD diverges from real Mars, the fine-tuned model's skill may not transfer to actual conditions.
ad hoc to paper Min-max rescaling to the ERA5 dynamic range preserves the physics needed for fine-tuning
The transformation in Section 2.1 is introduced to avoid activation saturation in the pretrained network; it assumes the rescaled fields retain the dynamical information relevant to Mars.
domain assumption GraphCast's pretrained representations are transferable to a different atmospheric regime
This is the central premise of the study (Section 1), tested empirically by fine-tuning; the paper assumes some universal spatiotemporal dynamics exist.
ad hoc to paper Holding humidity constant does not prevent learning thermal variability
The paper sets specific humidity to a constant global mean (Section 2.2) and acknowledges in Section 5 that this is unrealistic, yet treats it as an acceptable simplification.
domain assumption Land-sea mask set to unity is adequate for Mars
Mars has no liquid water bodies, so the land-sea mask is set to 1 (Table 1); this assumes the mask has no residual effect on the model's learned dynamics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres." pith.science (2026). https://pith.science/paper/KKUIOPGG

@misc{pith2026260805054,
  author       = {Pith},
  title        = {Pith review of: MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KKUIOPGG}},
  note         = {Machine review of arXiv:2608.05054}
}

read the original abstract

We investigate the transferability of Earth weather foundation models to planetary atmospheres by adapting the GraphCast graph neural weather forecasting model to Mars. While GraphCast achieves state-of-the-art performance for terrestrial forecasting, its applicability to non-Earth environments remains unexplored. Using the Mars Climate Database (MCD), which provides global atmospheric fields across vertical altitude levels (similar to Earth pressure levels), we evaluate zero-shot and fine-tuned GraphCast predictions of Martian temperature and wind fields. Zero-shot forecasts produce a surprisingly accurate depiction of current conditions but fail to reproduce diurnal variability and rapidly decay toward climatological mean states. To address this limitation, we fine-tune GraphCast using MCD variables and top-of-atmosphere solar radiation forcing while holding humidity constant. Fine-tuning enables rapid learning of Martian thermal variability. Within as few as 10 training epochs, the model begins to capture the diurnal cycle and forecasts up to 10 days reproduce seasonal and vertical temperature structure. Prediction quality improves with training sample size and exhibits sensitivity to seasonal initialization. These results demonstrate that Earth-trained AI weather models can be adapted to simulate Martian atmospheric dynamics, providing a pathway toward rapid planetary weather prediction to support mission operations, dust storm risk mitigation, and future human exploration.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 1 canonical work pages

[1]

The zonal‐mean circulation. J. Geophys. Res. 98, 3093–3123. https://doi.org/10.1029/92JE02946 Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., Battaglia, P.,

work page doi:10.1029/92je02946
[5]

Science 382, 1416–1421

Learning skillful medium-range global weather forecasting. Science 382, 1416–1421. https://doi.org/10.1126/science.adi2336 Lang, S., Alexe, M., Chantry, M., Dramsch, J., Pinault, F., Raoult, B., Clare, M.C.A., Lessig, C., Maier-Gerber, M., Magnusson, L., Bouallègue, Z.B., Nemesio, A.P., Dueben, P.D., Brown, A., Pappenberger, F., Rabier, F.,

doi:10.1126/science.adi2336
[6]

https://doi.org/10.5194/epsc2024-516 Price, I., Sanchez-Gonzalez, A., Alet, F., Andersson, T.R., El-Kadi, A., Masters, D., Ewalds, T., Stott, J., Mohamed, S., Battaglia, P., Lam, R., Willson, M.,

doi:10.5194/epsc2024-516
[8]

Nature 637, 84–90

Probabilistic weather forecasting with machine learning. Nature 637, 84–90. https://doi.org/10.1038/s41586-024- 08252-9 Richardson, M.I., Toigo, A.D., Newman, C.E.,

doi:10.1038/s41586-024-
[10]

Prithvi WxC: Foundation Model for Weather and Climate. https://doi.org/10.48550/ARXIV.2409.13598 Supplemental materials Figure S1: For 10-m zonal “u” wind component, comparison of (a) MarsCast predictions (b) MCD reference and (c) difference at for +06 hours to +48 hours, initialized for solar longitude (Ls =

arXiv doi:10.48550/arxiv.2409.13598
[1999]

Improved general circulation models of the Martian atmosphere from the surface to above 80 km. J. Geophys. Res. 104, 24155–24175. https://doi.org/10.1029/1999JE001025 Haberle, R.M., Pollack, J.B., Barnes, J.R., Zurek, R.W., Leovy, C.B., Murphy, J.R., Lee, H., Schaeffer, J.,

doi:10.1029/1999je001025
[2007]

PlanetWRF: A general purpose, local to global numerical model for planetary atmospheric and climate dynamics. J. Geophys. Res. 112, 2006JE002825. https://doi.org/10.1029/2006JE002825 Schmude, J., Roy, S., Trojak, W., Jakubik, J., Civitarese, D.S., Singh, S., Kuehnert, J., Ankur, K., Gupta, A., Phillips, C.E., Kienzler, R., Szwarcman, D., Gaur, V., Shinde,...

doi:10.1029/2006je002825
[2023]

Nature 619, 533–538

Accurate medium-range global weather forecasting with 3D neural networks. Nature 619, 533–538. https://doi.org/10.1038/s41586-023-06185-3 Bodnar, C., Bruinsma, W.P., Lucic, A., Stanley, M., Allen, A., Brandstetter, J., Garvan, P., Riechert, M., Weyn, J.A., Dong, H., Gupta, J.K., Thambiratnam, K., Archibald, A.T., Wu, C.-C., Heider, E., Welling, M., Turner...

doi:10.1038/s41586-023-06185-3

Show all 10 references

[2024]

AIFS -- ECMWF’s data-driven forecasting system. https://doi.org/10.48550/ARXIV.2406.01465 Millour, E., Forget, F., Spiga, A., Pierron, T., Bierjon, A., Montabone, L., Lefèvre, F., Montmessin, F., Chaufray, J.-Y., Lopez-Valverde, M., Gonzalez-Galindo, F., Desjean, M.-C., Cipria...

doi arXiv
[2025]

Nature 641, 1180–1187

A foundation model for the Earth system. Nature 641, 1180–1187. https://doi.org/10.1038/s41586-025-09005-y Forget, F., Hourdin, F., Fournier, R., Hourdin, C., Talagrand, O., Collins, M., Lewis, S.R., Read, P.L., Huot, J.,

doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

The zonal‐mean circulation. J. Geophys. Res. 98, 3093–3123. https://doi.org/10.1029/92JE02946 Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., Battaglia, P.,

work page doi:10.1029/92je02946

[2] [5]

Science 382, 1416–1421

Learning skillful medium-range global weather forecasting. Science 382, 1416–1421. https://doi.org/10.1126/science.adi2336 Lang, S., Alexe, M., Chantry, M., Dramsch, J., Pinault, F., Raoult, B., Clare, M.C.A., Lessig, C., Maier-Gerber, M., Magnusson, L., Bouallègue, Z.B., Nemesio, A.P., Dueben, P.D., Brown, A., Pappenberger, F., Rabier, F.,

doi:10.1126/science.adi2336

[3] [6]

https://doi.org/10.5194/epsc2024-516 Price, I., Sanchez-Gonzalez, A., Alet, F., Andersson, T.R., El-Kadi, A., Masters, D., Ewalds, T., Stott, J., Mohamed, S., Battaglia, P., Lam, R., Willson, M.,

doi:10.5194/epsc2024-516

[4] [8]

Nature 637, 84–90

Probabilistic weather forecasting with machine learning. Nature 637, 84–90. https://doi.org/10.1038/s41586-024- 08252-9 Richardson, M.I., Toigo, A.D., Newman, C.E.,

doi:10.1038/s41586-024-

[5] [10]

Prithvi WxC: Foundation Model for Weather and Climate. https://doi.org/10.48550/ARXIV.2409.13598 Supplemental materials Figure S1: For 10-m zonal “u” wind component, comparison of (a) MarsCast predictions (b) MCD reference and (c) difference at for +06 hours to +48 hours, initialized for solar longitude (Ls =

arXiv doi:10.48550/arxiv.2409.13598

[6] [1999]

Improved general circulation models of the Martian atmosphere from the surface to above 80 km. J. Geophys. Res. 104, 24155–24175. https://doi.org/10.1029/1999JE001025 Haberle, R.M., Pollack, J.B., Barnes, J.R., Zurek, R.W., Leovy, C.B., Murphy, J.R., Lee, H., Schaeffer, J.,

doi:10.1029/1999je001025

[7] [2007]

PlanetWRF: A general purpose, local to global numerical model for planetary atmospheric and climate dynamics. J. Geophys. Res. 112, 2006JE002825. https://doi.org/10.1029/2006JE002825 Schmude, J., Roy, S., Trojak, W., Jakubik, J., Civitarese, D.S., Singh, S., Kuehnert, J., Ankur, K., Gupta, A., Phillips, C.E., Kienzler, R., Szwarcman, D., Gaur, V., Shinde,...

doi:10.1029/2006je002825

[8] [2023]

Nature 619, 533–538

Accurate medium-range global weather forecasting with 3D neural networks. Nature 619, 533–538. https://doi.org/10.1038/s41586-023-06185-3 Bodnar, C., Bruinsma, W.P., Lucic, A., Stanley, M., Allen, A., Brandstetter, J., Garvan, P., Riechert, M., Weyn, J.A., Dong, H., Gupta, J.K., Thambiratnam, K., Archibald, A.T., Wu, C.-C., Heider, E., Welling, M., Turner...

doi:10.1038/s41586-023-06185-3