A new preprint from the Impact AI team at NASA’s Marshall Space Flight Center in Huntsville, Alabama, and IBM Research presents early results in the process to build a foundation model for the Martian atmosphere, combining data-driven learning with novel algorithmic approaches for sparse planetary data.

The arXiv preprint paper submitted on May 16, 2026 lays out the design and algorithmic roadmap for a Mars atmospheric Foundation Model. “Towards a Foundation Model for the Martian Atmosphere” describes how a single model with physics-informed pretraining and data assimilation can represent, forecast, and detect distinct phenomena within the Martian atmosphere.

Building such a model is not straightforward. General circulation models can simulate Martian atmospheric behavior, but they grow computationally expensive at resolutions needed to capture mesoscale features. The observational record compounds the difficulty as these observations are sparse, short, and fragmented across instrument generations.

To work within those constraints, the team characterized available data sources, including OpenMARS reanalysis, orbital retrievals, and Mars Global Climate Model (GCM) outputs. They then evaluated three architectures: a Mars-adapted GraphCast (graph neural network), a Mars-adapted Prithvi Weather and Climate (vision transformer), and a Spectral-LS transformer trained on OpenMARS surface variables. They also explored pretraining based on partial differential equations as a strategy to improve model skill in data-limited settings.

The Spectral-LS transformer outperforms persistence for wind and surface pressure forecasts over 24-hour horizons, an encouraging early result. Dust column opacity remains challenging across all three architectures and is an open problem the team continues to tackle. Beyond forecasting, the paper surveys AI-based data assimilation approaches as a path toward ingesting sparse orbital observations directly into the model state, a critical capability given that observational records of Mars are fragmented.