A research team has developed a graph-based data engineering framework named MaGMA to advance multi-horizon Martian atmospheric forecasting. The methodology directly addresses longstanding challenges in modelling atmospheric behaviour on Mars, where complex dynamics simultaneously integrate spatial, temporal, vertical, and dust-driven physical processes.
Traditional machine learning investigations targeting Martian weather have frequently reduced complex atmospheric dynamics to local time series. This limits their ability to capture wider atmospheric dynamics.
To overcome these spatial and temporal constraints, the researchers introduced the MaGMA framework, short for Martian Graph-based Multi-horizon Atmospheric Forecasting. The methodology transforms historical OpenMARS reanalysis data fields into structured learning objects specifically engineered for planetary atmospheric prediction models across diverse horizons.
Under the MaGMA framework, local atmospheric patches are explicitly structured as graph nodes. The system establishes connectivity by linking these regional patches across neighbouring geographical territories, successive temporal steps, extended temporal dependencies, and dynamically similar atmospheric states observed across the Martian climate system.
The constructed representation further integrates recent atmospheric historical sequences, engineered physical descriptors, and comprehensive vertical atmospheric information. This supports forecasting across multiple horizons.
The scientists evaluated the performance of MaGMA across five previously unseen Martian years, testing the architecture under both regular seasonal conditions and an extreme global dust storm year. In regular years, the framework demonstrated overall coefficients of determination ranging between approximately 0.73 and 0.85 across target atmospheric variables.
For specific dust-column forecasting, the graph-based model outperformed classical and deep temporal baseline models in most evaluated year-and-horizon comparisons. During the global dust storm year, dust-column prediction remained robust at shorter forecast horizons, achieving coefficients of determination exceeding 0.8 across the first two horizons.
Despite these strengths, broader multivariate forecasting performance experienced measurable declines during the global dust storm year, demonstrating that rare extreme regimes continue to challenge machine learning generalisation. The authors concluded that graph-based engineering provides reusable and diagnostically valuable representations, highlighting improved learning under extreme regimes and vertical structure integration as future priorities.
The overall coefficients of determination indicate that the framework captures a substantial proportion of the variation in target atmospheric variables during regular years.
Subject of Research: Graph-based machine learning for Martian atmospheric forecasting
Article Title: Graph-based learning for multi-horizon Martian atmospheric forecasting
Article References: Myler, G., Holmes, J., Patel, M., & Bennaceur, A. (2026). Graph-based learning for multi-horizon Martian atmospheric forecasting. Journal of Big Data. https://doi.org/10.1186/s40537-026-01582-7
Image Credits: AI Generated
DOI: 10.1186/s40537-026-01582-7
Keywords: Mars, Martian atmosphere, atmospheric forecasting, weather forecasting, graph-based learning, machine learning, MaGMA, OpenMARS, reanalysis data, dust storms, planetary science, spatial dynamics, temporal forecasting, data engineering, atmospheric modelling, dust column
Cite Scienmag News
Denise Maddox. (November 20, 2026). Graph-Based Model Outperforms Baselines in Most Martian Dust-Column Forecast Comparisons. Scienmag. https://scienmag.com/graph-based-model-outperforms-baselines-in-most-martian-dust-column-forecast-comparisons/
Denise Maddox. "Graph-Based Model Outperforms Baselines in Most Martian Dust-Column Forecast Comparisons." Scienmag, 20 November 2026, https://scienmag.com/graph-based-model-outperforms-baselines-in-most-martian-dust-column-forecast-comparisons/. Accessed 20 November 2026.
Denise Maddox. "Graph-Based Model Outperforms Baselines in Most Martian Dust-Column Forecast Comparisons." Scienmag. November 20, 2026. https://scienmag.com/graph-based-model-outperforms-baselines-in-most-martian-dust-column-forecast-comparisons/








