Graph-Based Learning for Multi-Horizon Martian Atmospheric Forecasting

Gary Myler; James Holmes; Manish Patel; Amel Bennaceur

cs.SESep 28, 2026

Graph-Based Learning for Multi-Horizon Martian Atmospheric Forecasting

Authors: Gary Myler, James Holmes, Manish Patel, Amel Bennaceur

Organizations: The Open University, UK.

Abstract

Martian weather forecasting is important for future exploration, but atmospheric behaviour on Mars combines spatial, temporal, vertical, and dust-driven processes in ways that challenge current modelling and forecasting approaches. This paper introduces MaGMA (Martian Graph-based Multi-horizon Atmospheric Forecasting), a graph-based data engineering framework that transforms OpenMARS reanalysis fields into structured learning objects for Martian atmospheric forecasting. Local atmospheric patches are represented as graph nodes and linked through spatial neighbourhoods, temporal continuity, longer temporal dependencies, and dynamically similar atmospheric states. The model integrates recent atmospheric history, engineered physical descriptors, and vertical atmospheric information to support forecasting across multiple horizons. We evaluate MaGMA across five unseen Martian years, including regular years and a global dust storm year. In regular years, the model achieves overall R^2 values of approximately 0.73-0.85. For dust-column forecasting, it outperforms classical and deep temporal baselines in most year-horizon comparisons. During the global dust storm year, dust-column prediction remains strong at shorter horizons, with R^2 above 0.8 for the first two horizons, while broader multivariate performance declines. The results show that graph-based data engineering can create reusable and diagnostically useful representations for planetary atmospheric forecasting, while highlighting the need for better learning under rare extreme regimes and improved use of vertical atmospheric structure.

Figures & tables

Martian Year	Experimental role	Earth-date coverage	Approx. sols
MY28	Training/validation	24 Sep 2006–8 Dec 2007	433
MY29	Evaluation	9 Dec 2007–25 Oct 2009	669
MY30	Evaluation	26 Oct 2009–12 Sep 2011	669
MY31	Evaluation	13 Sep 2011–30 Jul 2013	669
MY32	Evaluation	31 Jul 2013–17 Jun 2015	669
MY34	Evaluation	5 May 2017–22 Mar 2019	669

Table 1: Martian years used in the MaGMA experiments and their approximate Earth-date coverage. MY28 begins at sol 235 in the OpenMARS product used here; the remaining evaluation years span approximately complete Martian years.

Variable	Description	Unit
tsurf	Surface temperature	Kelvin
ps	Surface pressure	Pascals
u	Near-surface ( $\sim$ 4 m) zonal wind	m/s
v	Near-surface ( $\sim$ 4 m) meridional wind	m/s
dustcol	Dust column optical depth at visible wavelength	No unit
temp	Atmospheric temperature at $\sim$ 20 km altitude	Kelvin

Table 2: Variables used from the OpenMARS dataset.

Time	Tsurf	Psurf	Cloud	Vapour	u_wind	v_wind	Dust	Temp
1998-07-15 21:23:39	264.042	721.113	0.092	0.027	-7.451	8.604	0.428	179.686
1998-07-15 23:26:53	274.736	705.090	0.145	0.026	-7.053	4.934	0.427	174.502
1998-07-16 01:30:07	265.939	700.691	0.105	0.026	-6.825	-0.063	0.427	173.429
1998-07-16 03:33:21	238.624	697.252	0.134	0.025	-5.373	-4.048	0.426	173.556
1998-07-16 05:36:35	213.634	717.146	0.139	0.026	-3.899	-3.133	0.426	174.789

Table 3: Illustrative rows from the cleaned local time-series representation used in earlier work.

Figure 1: Overview of the MaGMA workflow. OpenMARS gridded atmospheric fields are transformed into graph-structured learning objects by partitioning the data into local spatiotemporal patches, constructing node features from recent temporal history, engineered descriptors, and vertical column information, and connecting nodes through spatial, temporal, multi-step temporal, and atmospheric-state similarity edges. During graph construction, each node is also assigned multi-horizon targets corresponding to H1, H3, and H6. The resulting graph object is processed by a neural forecasting architecture that combines temporal encoding, static and engineered features, vertical column encoding, relation-aware graph propagation, adaptive routing, and horizon-specific residual forecast heads. The model predicts surface pressure, surface temperature, dust column, zonal wind, meridional wind, atmospheric temperature, and uplift at each horizon.

Year	MAE	RMSE	$R^{2}$
MY29	0.271	0.522	0.728
MY30	0.156	0.303	0.851
MY31	0.271	0.511	0.738
MY32	0.275	0.524	0.726
MY34	0.458	0.753	0.363

Table 4: Overall generalisation performance

Year	Horizon	ps	tsurf	dust	u	v	temp	uplift
MY29	H1	0.052	0.333	0.079	0.174	0.482	0.238	0.497
MY29	H3	0.063	0.173	0.176	0.187	0.573	0.200	0.471
MY29	H6	0.070	0.184	0.244	0.212	0.579	0.215	0.489
MY30	H1	0.032	0.107	0.042	0.122	0.339	0.066	0.178
MY30	H3	0.047	0.095	0.086	0.151	0.434	0.082	0.232
MY30	H6	0.056	0.110	0.121	0.174	0.439	0.109	0.253

Table 5: MAE across horizons and variables

Year	Horizon	ps	tsurf	dust	u	v	temp	uplift
MY29	H1	0.070	0.436	0.163	0.237	0.679	0.316	0.929
MY29	H3	0.090	0.245	0.323	0.274	0.804	0.293	0.942
MY29	H6	0.101	0.282	0.448	0.322	0.824	0.338	0.982
MY30	H1	0.048	0.148	0.074	0.170	0.455	0.110	0.371
MY30	H3	0.077	0.152	0.129	0.222	0.597	0.158	0.457
MY30	H6	0.091	0.189	0.169	0.264	0.608	0.211	0.515

Table 6: RMSE across horizons and variables

Year	Horizon	ps	tsurf	dust	u	v	temp	uplift
MY29	H1	0.995	0.810	0.973	0.944	0.540	0.900	0.048
MY29	H3	0.992	0.940	0.896	0.925	0.355	0.914	0.125
MY29	H6	0.990	0.921	0.799	0.896	0.321	0.886	0.115
MY30	H1	0.998	0.978	0.959	0.972	0.722	0.986	0.734
MY30	H3	0.994	0.977	0.873	0.953	0.521	0.972	0.592
MY30	H6	0.992	0.964	0.783	0.934	0.505	0.949	0.515

Table 7:

R^{2}

across horizons and variables

Figure 2: Truth-versus-prediction scatter plots for MY30 and MY34 at H1, comparing dust-column and uplift forecasts. Dust-column predictions remain closely aligned with the one-to-one reference line, while uplift predictions show weaker agreement, particularly under the MY34 global dust storm regime.

Year	Acc	Prec	Recall	F1	AUC
MY29	0.998	0.994	0.999	0.996	1.000
MY30	1.000	1.000	1.000	1.000	1.000
MY31	0.995	0.986	0.998	0.992	1.000
MY32	0.997	0.992	0.998	0.995	1.000
MY34	0.627	0.001	0.000	0.000	0.200

Table 8: Auxiliary dust-regime classification performance

Figure 3: Regional prediction examples for MY34 dust-column forecasting at H3. The panels show representative best-case, median-case, and worst-case spatial regions. The best and median regions demonstrate that the model can track the main dust evolution in some locations, while the worst-case region illustrates a regional failure mode with a false positive late spike.

Variant	Mean $R^{2}$	$\Delta R^{2}$	Dust H3 $R^{2}$	Dust H3 $\Delta R^{2}$
No temporal sequence	-0.016	-0.698	-0.193	-1.059
No graph edges	0.618	-0.063	0.854	-0.012
No temporal graph edges	0.645	-0.036	0.871	0.005
No engineered/static features	0.656	-0.026	0.855	-0.011
No similarity edges	0.676	-0.005	0.868	0.002
No vertical column	0.680	-0.001	0.866	0.000

Table 9: Inference-time component sensitivity analysis. Negative

\Delta R^{2}

values indicate performance degradation relative to the full model.

Martian Year	Horizon	Linear Regression	Random Forest	GNN	GNN Gain
MY29	H1	0.920	0.532	0.973	0.053
MY29	H3	0.857	0.616	0.896	0.039
MY29	H6	0.772	0.556	0.799	0.027
MY30	H1	0.952	0.960	0.959	-0.001
MY30	H3	0.881	0.880	0.873	-0.008
MY30	H6	0.776	0.751	0.783	0.007

Table 10: Dust-column forecasting baseline comparison measured using

R^{2}

. Gain is computed relative to the strongest stable classical baseline for each year–horizon pair. Linear Regression produced unstable out-of-distribution extrapolations under MY34 and is omitted from those rows.

Martian Year	Horizon	LSTM	GRU	TCN	GNN
MY29	H1	0.943	0.947	0.938	0.973
MY29	H3	0.853	0.872	0.845	0.896
MY29	H6	0.725	0.766	0.684	0.799
MY30	H1	0.933	0.946	0.933	0.959
MY30	H3	0.849	0.868	0.820	0.873
MY30	H6	0.681	0.707	0.558	0.783

Table 11: Dust-column forecasting comparison against deep temporal baselines measured using

R^{2}

Explore similar work

MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres
Aug 5, 2026M. L. Carroll, J. Li, S. D. Guzewich +3Climate Exoplanet Atmospheres
1) NASA Goddard Space Flight Center · 2) ADNet Systems, Inc.
ARCO-Mars: A Unified Cloud-Optimized Archive of Mars Atmosphere Reanalysis
Jun 19, 2026Ananyo BhattacharyaEra5 Reanalysis Data Atmospheric Dynamics
University of Michigan, Department of Climate and Space Sciences and Engineering, Ann Arbor, USA
Towards a Foundation Model for the Martian Atmosphere
May 16, 2026Sujit Roy, Udayshankar Nair, Yuling Wu +16Atmospheric Dynamics Data Assimilation
Earth System Science Center, University of Alabama in Huntsville, AL, USA · NASA Marshall Space Flight Center, Huntsville, AL, USA · Department of Electrical & Computer Engineering, Colorado State University, Fort Collins, CO, USA +7

Aug 5, 2026, astro-ph.EPJ/K move · Enter open · S save

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

M. L. Carroll, J. Li, S. D. Guzewich, G. Villanueva+2

1) NASA Goddard Space Flight Center · 2) ADNet Systems, Inc.

Climate Exoplanet Atmospheres Transfer Learning Graph Representations

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

ARCO-Mars: A Unified Cloud-Optimized Archive of Mars Atmosphere Reanalysis

Towards a Foundation Model for the Martian Atmosphere