PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph-Based Learning for Multi-Horizon Martian Atmospheric Forecasting

Gary Myler, James Holmes, Manish Patel, Amel Bennaceur

arXiv:2609.35042v1cs.SEcs.LG

TL;DR

Martian forecasting must address spatial, temporal, vertical, and dust-driven atmospheric processes, while flattening reanalysis fields can discard relationships relevant to prediction. MaGMA transforms OpenMARS into graph-structured learning objects for multi-horizon forecasting; performance is stable across regular unseen years but declines under the MY34 global dust storm regime.

  • Problem

    Martian atmospheric reanalysis contains spatial, temporal, vertical, and dynamically coupled relationships that flattening into isolated sequences may discard, despite forecasting's importance for science and exploration.

    Source: §1 · pp. 2–3

  • Method

    MaGMA transforms multidimensional OpenMARS fields into graph objects representing local patches, histories, vertical information, spatial neighbourhoods, temporal links, and dynamically similar states.

    Source: §1 · p. 3, §4 · p. 9

  • Results

    Across regular unseen Martian years MY29–MY32, overall R2 ranges from 0.726 to 0.851; MY34 multivariate performance drops, although dust-column prediction remains strong at H1 and H3.

    Source: §7.4 · p. 31

  • Takeaways & Limitations

    The graph-construction representation can be reused with atmospheric datasets beyond OpenMARS, as supplementary Venus and Titan demonstrations show.

    Source: §7.4 · p. 32, §8 · p. 32

    Generalisation under rare dust-storm regimes and use of vertical structure need improvement, and the temporal-only baselines do not exhaust comparisons with modern spatiotemporal forecasting systems.

    Source: §8 · p. 32

Abstract

from arXiv · show

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.

1 Introduction

Martian forecasting matters for science and exploration, but dust-driven dynamics and the spatial, temporal, and vertical structure of reanalysis fields challenge forecasting workflows that flatten data into isolated sequences. MaGMA represents OpenMARS fields as relational graph objects and evaluates multi-horizon prediction across unseen Martian years, including a global dust storm year.

1 Introduction

  • Martian forecasting supports climate and circulation research as well as mission operations such as observation planning, energy management, rover traversability, and instrument use.The motivation also emphasizes reusing increasingly rich atmospheric datasets while preserving their structure.Source: §1 · p. 2
  • Dust storms are difficult to forecast because their evolution involves nonlinear interactions among surface lifting, circulation, radiative feedback, and changing dust distributions.Global storms spread dust loading across much of the planet, making the distinction between routine variability and storm behaviour a central modelling challenge.Source: §1 · p. 2
  • Flattening multidimensional reanalysis into independent vectors or time series may discard forecast-relevant relationships across locations, vertical columns, and dynamically related regions.Relevant signals may lie in recent local history, neighbouring regions, or atmospheric columns.Source: §1 · pp. 2–3
  • MaGMA transforms OpenMARS fields into graph objects whose local-patch nodes and spatial, temporal, longer-range, and atmospheric-state edges preserve relationships for multi-horizon forecasting.The model combines temporal history, engineered descriptors, vertical information, and relation-aware propagation.Source: §1 · p. 3
  • The evaluation tests cross-year generalisation on complete unseen Martian years, including regular years and a separate global dust storm year.The design assesses whether the model captures dust-related atmospheric structure beyond common regimes.Source: §1 · p. 3
  • The paper also compares dust-column forecasts with linear, tree-based, recurrent, and temporal-convolutional baselines and analyses how information pathways contribute to prediction.The stated contributions include graph-based data engineering and diagnostic analysis of temporal, engineered, vertical, and graph components.Source: §1 · pp. 3–4

2 Related Work

Related work moves from local time-series prediction toward structured global atmospheric forecasting, while Mars-specific work remains constrained by sparse observations and limited data-driven forecasting studies. MaGMA is positioned as a focused effort to turn OpenMARS fields into reusable graph representations between single-site prediction and broader foundation-model agendas.

2 Related Work

  • The related-work framing identifies a representational gap: Martian reanalysis couples spatial grids, vertical structure, temporal evolution, and atmospheric dynamics that local time series do not capture together.MaGMA addresses this gap by building graph-structured learning objects from those fields.Source: §2 · p. 4, Weather forecasting · p. 5

Weather forecasting

  • Earth forecasting research spans local station models and global structured systems such as GraphCast [21], FourCastNet, and Pangu-Weather.The cited global systems preserve spatial or vertical structure in atmospheric data, motivating a related direction for OpenMARS.Source: Weather forecasting · pp. 4–5

Weather Forecasting for Mars

  • Mars forecasting draws on sparse and heterogeneous observations, while dust-storm reviews report that domain-knowledge and statistical methods remain insufficient for accurate, timely forecasts.Data assimilation has improved forecasts for specific quantities such as carbon monoxide; the cited Curiosity study remains a univariate, single-site prediction task.Source: Weather Forecasting for Mars · pp. 5–6
  • Mars foundation-model work [31] surveys datasets, architectures, and downstream tasks, whereas MaGMA focuses specifically on reusable graph-structured OpenMARS inputs for multi-horizon forecasting.The distinction places MaGMA between single-site machine learning and broader foundation-model ambitions.Source: Weather Forecasting for Mars · p. 6

3 OpenMARS Dataset

The study uses OpenMARS, a global multidimensional reanalysis product combining spacecraft observations with a Mars circulation model, and converts its year-specific fields into graph-learning inputs. The experiments use MY28 for training and validation and five unseen years for cross-year evaluation.

3 OpenMARS Dataset

  • OpenMARS combines spacecraft observations with a Mars Global Circulation Model and covers MY28–MY35, while the forecasting experiments use the MY28–MY34 subset.The MY28 record begins at sol 235 rather than the start of that Martian year.Source: §3 · pp. 6–7
  • The experiments train and validate on MY28 and evaluate on MY29, MY30, MY31, MY32, and MY34.The split separates the product's full coverage from the years used in this study.Source: §3 · pp. 6–7
  • OpenMARS contains 72 longitude points, 36 latitude points, 35 vertical levels, and 360 time outputs per file, with surface and atmospheric fields stored on their respective dimensions.The array-oriented NetCDF4 structure preserves coordinates, units, and metadata rather than flattening fields into tabular records.Source: §3 · p. 7
  • The selected inputs are surface pressure, surface temperature, dust-column optical depth, zonal and meridional wind, and atmospheric temperature.The fields represent near-surface state, dust loading, thermal structure, and horizontal circulation; surface CO2 ice is excluded because the task focuses on the evolving atmosphere.Source: §3 · pp. 7–8
  • The illustrative cleaned local time-series rows are not the year-specific graph inputs and do not retain the full spatial and vertical structure used by MaGMA.The sampled local series supports conventional time-series benchmarking, while graph inputs are constructed from year-specific OpenMARS fields.Source: §3 · p. 8, Table 3 · p. 8

4 MaGMA: Martian Graph-based Multi-horizon Atmospheric Forecasting

MaGMA first converts gridded OpenMARS fields into graph learning objects, then processes them with a multi-path forecasting architecture. The design combines local temporal histories and engineered or vertical features with relational information to predict multiple variables at three forecast horizons.

4 MaGMA: Martian Graph-based Multi-horizon Atmospheric Forecasting

  • The workflow assigns H1, H3, and H6 targets and uses horizon-specific heads to forecast seven atmospheric variables, including dust column, winds, temperature, pressure, and uplift.The architecture also supports diagnostics of the temporal, engineered, vertical, and graph-based information pathways.Source: Figure 1 · p. 9, §4 · p. 10
  • Graph construction segments atmospheric time series into windows and partitions the spatial grid so each node represents a local patch and its recent history.Nodes are enriched with engineered descriptors such as wind magnitude, transformed dust variables, geographic encodings, seasonal indicators, and vertical statistics.Source: §4 · pp. 9–10
  • The forecasting architecture encodes recent sequences, vertical columns, and static or engineered state features before fusing them for graph propagation.Propagation exchanges information across spatial, temporal, longer-range temporal, and dynamically similar-state relations.Source: §4 · p. 10
  • Adaptive routing selects how strongly multiple graph-propagation pathways contribute for each node, allowing the relational-information scale to vary by atmospheric state.The model uses multiple pathways rather than a single fixed propagation depth.Source: §4 · p. 10

5 Graph Construction

MaGMA converts gridded OpenMARS fields into supervised graph-learning objects whose nodes represent local atmospheric states and whose features, edges, and targets retain spatial, temporal, vertical, and atmospheric-state information.

5.1 From Gridded Fields to Local Atmospheric Nodes

  • Each node represents a 3 × 3 patch (15° × 15°) over an eight-step history, reducing each graph-time slice to 288 spatial nodes while retaining global coverage.Each node is paired with future atmospheric targets at multiple horizons.Source: §5.1 · p. 11

5.2 Node Feature Engineering

Full Vertical Column Tensor

5.3 Graph Edge Construction

  • Four edge types connect neighboring patches, the same patch across adjacent or up-to-two graph-time steps, and dynamically similar patches within a time slice.The two-step temporal limit and five-neighbor similarity relation are pragmatic locality and sparsity choices, not empirically optimized settings.Source: §5.3 · p. 13

5 Graph Construction

  • The graph object packages node features, vertical tensors, targets, connectivity, relation labels, a storm mask, and an auxiliary dust-regime target.Source: §5.3 · p. 13

5.4 Forecasting Targets

  • Targets cover H1, H3, and H6 offsets—8, 24, and 48 raw time indices—and each horizon includes six core variables plus a derived uplift proxy.Rows without an available future target window are excluded from evaluation through finite-value masking.Source: §5.4 · pp. 13–14

6 Graph Neural Architecture for Multi-Horizon Forecasting

MaGMA fuses local histories, engineered state descriptors, and vertical profiles before relation-aware graph propagation, adaptive routing, and separate multi-horizon forecast heads. It predicts residual changes from persistence and supports diagnostic analysis of its information pathways.

Model Overview

  • The architecture fuses temporal, static-feature, and vertical-column pathways into node representations, then applies graph propagation and horizon-specific residual forecast heads.The three heads predict H1, H3, and H6 outcomes for seven targets.Source: §6 · p. 14, Model Overview · pp. 14–15

Temporal Encoder

  • Two one-dimensional convolutional layers encode each node’s recent six-variable, eight-step history to capture short-term local atmospheric evolution.Adaptive average pooling and projection produce the shared temporal embedding.Source: Temporal Encoder · p. 15

Static and Engineered Feature Branch

  • The static branch combines engineered descriptors with the latest observed state and one-step trend, giving the model immediate local-state information for residual forecasting.The concatenated features pass through a feed-forward projection with GELU activation and layer normalization.Source: Static and Engineered Feature Branch · p. 15

Vertical Column Encoder

  • A vertical encoder processes selected full-column profiles as ordered level sequences using a GRU, residual MLP, and self-attention; a learned scalar gate scales this pathway during fusion.The fused representation is hbase = htemp + hstatic + γhvert, where γ is the learned vertical scaling factor.Source: Vertical Column Encoder · pp. 15–16

Auxiliary Dust-Regime Head

  • An auxiliary head predicts a weak dust-active label, and its predicted probability both supplies a training signal and conditions graph-expert routing.The signal is an auxiliary regime indicator, not an independently validated physical dust-storm classifier.Source: Auxiliary Dust-Regime Head · p. 16

Relational Graph Propagation

  • Relation-aware graph convolution lets nodes exchange information across spatial, temporal, multi-step temporal, and state-similarity links.Residual connections and normalization are used in propagation blocks to stabilize propagation and reduce excessive smoothing risk.Source: Relational Graph Propagation · pp. 16–17

Adaptive Routing over Propagation Depth

  • A learned router mixes outputs from shallow, medium, and deep graph pathways, conditioning its softmax weights on the fused node representation and predicted dust-regime probability.The different pathway depths provide different effective relational receptive fields.Source: Adaptive Routing over Propagation Depth · p. 17

Multi-Horizon Residual Forecast Heads

7 Experiments

The experiments train MaGMA on MY28, evaluate it on later Martian years including MY34, and compare it with classical and temporal forecasting baselines.

7 Experiments

  • The experiments train MaGMA on MY28, evaluate it on later Martian years including MY34 global dust storm year, and compare it with classical and temporal forecasting baselines.Source: §7 · p. 18

7.1 Training and Evaluation Strategy

The evaluation trains on MY28 and tests across later Martian years, comparing graph-based forecasting with classical and temporal baselines using multi-horizon metrics and diagnostic analyses.

Scalable Subgraph Training

  • Training uses time-batched induced subgraphs because full-graph GPU training is impractical within available accelerator memory.Dust-active nodes are oversampled toward a target storm ratio of 0.25.Source: Scalable Subgraph Training · pp. 18–19

Train–Validation Split and Cross-Year Evaluation

Training Objective

  • The objective combines weighted regression across H1, H3, and H6 with dust-regime classification, output-correlation, router-entropy, and smoothness terms.Dust column and uplift receive greater target weights than other variables.Source: Training Objective · pp. 19–20

Optimization Configuration

  • Optimization uses AdamW with a 2 × 10^-3 learning rate, 256 hidden dimensions, dropout of 0.05, and early stopping based on validation loss.The router temperature is annealed from 1.5 to a lower bound of 0.7 during training.Source: Optimization Configuration · p. 20

Baseline Models

  • Baseline comparisons include Linear Regression, Random Forest, LSTM, GRU, and TCN, trained on MY28 and evaluated on the same later-year set.The deep temporal baselines use raw node sequences without graph edges.Source: Baseline Models · pp. 20–21

Evaluation Metrics

  • Forecasting is assessed with MAE, RMSE, and R2 by year, horizon, and target, with dust-column R2 as the primary baseline comparison.The study also evaluates dust-regime classification, regional dust-column skill, and inference-time component sensitivity.Source: Evaluation Metrics · p. 21
  • Component sensitivity is an inference-time dependence diagnostic, not a full retrained ablation study.Source: Evaluation Metrics · p. 21

7.2 Results

MaGMA performs consistently across regular Martian years, but overall skill falls in the MY34 global dust storm regime. Dust-column forecasts remain comparatively strong, while other variables and the model’s auxiliary dust-regime classifier show important weaknesses.

Overall Multi-Year Forecasting Performance

  • Overall R2 ranges from 0.726 to 0.851 in MY29–MY32, but falls to 0.363 in MY34, showing that performance does not transfer uniformly to the global dust storm regime.The supplied overall-performance table reports performance across MY29, MY30, MY31, MY32, and MY34.Source: Table 4 · p. 22, Overall Multi-Year Forecasting… · p. 22

Variable- and Horizon-Specific Forecast Skill

  • Pressure, dust, temperature, and zonal wind generally have stronger R2 skill than meridional wind and uplift, with MY34 meridional-wind R2 negative at every horizon.Uplift is a derived diagnostic proxy rather than a directly observed core atmospheric variable.Source: Variable- and Horizon-Specific… · p. 22, Table 7 · p. 24

Global Dust Storm Regime Behaviour in MY34

  • MY34 dust-column R2 remains 0.908, 0.812, and 0.691 at H1, H3, and H6 despite overall R2 of 0.363; weaker results in other variables drive the broader decline.The model retains meaningful dust forecasting skill but struggles with the broader multivariate atmospheric regime.Source: Global Dust Storm Regime Behaviour in… · p. 24

Auxiliary Dust-Regime Classification

  • In regular evaluation years, the auxiliary dust-regime classifier reaches accuracy of 0.995–1.000 and AUC of 1.000, but MY34 accuracy is 0.627 and AUC is 0.200.The classifier uses a weak binary dust-active label derived from a dust-threshold flag and is not an independently validated storm detector.Source: Auxiliary Dust-Regime Classification · pp. 24–25, Table 8 · p. 26

Regional Forecast Heterogeneity

  • MY34 H3 regional examples range from tracking dust evolution with underestimated peaks to a worst-case late false-positive spike, illustrating spatially heterogeneous forecast skill.The examples show best-case, median-case, and worst-case regions.Source: Figure 3 · p. 26, Regional Forecast Heterogeneity · p. 26

Inference-Time Component Sensitivity

  • Inference-time sensitivity identifies recent temporal history as dominant: removing it lowers mean R2 from 0.681 to −0.016, while graph and engineered features provide secondary gains and the full vertical tensor adds little under this test.The authors note that vertical summaries already occur in engineered features and the test is not a fully retrained ablation; they call for retrained comparisons and alternative vertical representations.Source: Inference-Time Component Sensitivity · pp. 27–28

Dust-Column Baseline Comparison Against Classical Models; Dust-Column Baseline Comparison Against Deep Temporal Models

  • The GNN achieves the best dust-column R2 in 13 of 15 classical-baseline comparisons and 14 of 15 deep-temporal comparisons, with mean R2 of 0.863 versus 0.853 for the strongest stable classical baseline and 0.831 for GRU.The classical-baseline mean excludes unstable MY34 extrapolation rows.Source: Dust-Column Baseline Comparison… · pp. 28–29, Table 10 · p. 29

7.3 Cross-Planet Framework Transferability

MaGMA’s graph-construction and multi-horizon forecasting framework is demonstrated on Venus and Titan datasets beyond OpenMARS. These supplementary examples show portability, not complete atmospheric forecasting studies for either planet.

7.3 Cross-Planet Framework Transferability

  • The paper frames the additional-dataset experiments as supplementary tests of whether its graph-based approach can be reused with different atmospheric grids and variable sets.The passage also describes future experiments extending the demonstration to Titan’s polar-cloud simulation dataset.Source: §7.3 · pp. 29–30
  • The Venus experiment reused the Mars representation and architecture, including spatiotemporal nodes, vertical tensors, graph edges, relational propagation, adaptive routing, and multi-horizon heads.The Venus dataset is identified as Venus Planetary Climate Model output for age of air.Source: §7.3 · p. 29
  • The Venus H1 forecast for vertical wind velocity achieved R2 = 0.774, while the demonstration is not a complete Venus forecasting study.Figure 4 presents truth-versus-prediction forecasts at H1 and H3.Source: §7.3 · pp. 29–30, Figure 4 · p. 30
  • The Titan temperature demonstration tracks spatially averaged temporal evolution across H1, H3, and H6 using the same graph-construction and multi-horizon framework.Figure 5 is described as a portability demonstration rather than a full Titan forecasting study.Source: §7.3 · p. 31, Figure 5 · p. 31

7.4 Summary of Findings

MaGMA performs consistently across regular unseen Martian years and retains dust-column skill during MY34, though broader performance declines in that extreme regime. Sensitivity analyses identify recent history as the dominant information source, with graph structure contributing and the vertical pathway contributing less than expected.

7.4 Summary of Findings

  • R2 values of 0.726–0.851 across regular unseen Martian years MY29–MY32 indicate stable overall performance.Stronger results occur for surface pressure, dust column, atmospheric temperature, and zonal wind.Source: §7.4 · p. 31
  • During MY34, overall multivariate performance drops, but dust-column prediction remains strong at H1 and H3 while meridional wind, uplift, and surface temperature are harder to forecast.The results characterize dust-loading skill as useful while the broader coupled atmospheric response remains difficult.Source: §7.4 · p. 31
  • For dust-column forecasting, MaGMA outperforms classical baselines in most year–horizon comparisons and deep temporal baselines in nearly all comparisons.Its advantage is clearest at medium and longer horizons, especially under MY34.Source: §7.4 · p. 31
  • Removing raw temporal sequences causes the largest performance collapse; graph structure adds a smaller but useful contribution, while the vertical pathway contributes less than expected.The findings motivate stronger vertical encoders and more targeted use of column information, especially during dust-active periods.Source: §7.4 · pp. 31–32
  • Supplementary Venus and Titan demonstrations show that the graph-construction pipeline can be reused across datasets with different variables, physical regimes, and planetary contexts.These are portability demonstrations, not complete forecasting studies.Source: §7.4 · p. 32

8 Conclusion

MaGMA represents multidimensional atmospheric data as graph-structured learning objects and performs consistently across regular unseen Martian years, with strong short-horizon dust forecasting during MY34. The conclusion also identifies rare extreme regimes, vertical information use, and broader architecture comparisons as priorities for further work.

8 Conclusion

  • MaGMA transforms multidimensional reanalysis fields into graph-structured learning objects representing local patches, recent histories, vertical information, spatial neighbourhoods, temporal links, and dynamically similar states.The conclusion describes this as a shift from isolated sequence prediction to structured atmospheric learning.Source: §8 · p. 32
  • MaGMA performs consistently across regular unseen Martian years and provides strong short-horizon dust-column forecasting during MY34, though rare-regime performance remains challenging for meridional wind, uplift, and surface temperature.The framework supports both forecasting and diagnosis of where the coupled atmospheric response remains difficult to model.Source: §8 · p. 32
  • The graph augments recent atmospheric history with relational context across space, time, and atmospheric-state similarity rather than replacing temporal information.Supplementary Venus and Titan demonstrations indicate that the graph-construction principles can be reused beyond Mars.Source: §8 · p. 32
  • Future work includes improving rare dust-storm generalisation and vertical-structure use, and comparing MaGMA with modern spatiotemporal forecasting architectures not exhaustively tested here.The listed comparison targets include graph-based global models, neural operators, and transformer-based weather models.Source: §8 · p. 32

Declaration on Generative AI

The authors report using ChatGPT for manuscript and workflow support, while retaining responsibility for the final manuscript.

Declaration on Generative AI

  • The authors used ChatGPT for language polishing and clarity improvement.They state that they reviewed all content and take full responsibility for the final manuscript.Source: Declaration on Generative AI · p. 33
  • The authors report using ChatGPT for LaTeX formatting support.They state that they reviewed all content and take full responsibility for the final manuscript.Source: Declaration on Generative AI · p. 33
  • The authors report using ChatGPT for code debugging and workflow assistance.They state that they reviewed all content and take full responsibility for the final manuscript.Source: Declaration on Generative AI · p. 33
Loading 2609.35042v1…