Technical Summary: Towards a Foundation Model for the Martian Atmosphere

1. Problem Statement and Motivation

The Martian atmosphere exhibits complex dynamical phenomena, ranging from planet-encircling dust storms to mesoscale orographic clouds and nocturnal low-level jets. While General Circulation Models (GCMs) like the NASA Ames Mars Global Climate Model (MGCM) and MarsWRF can simulate these processes, they are computationally prohibitive at the resolutions required to resolve mesoscale features (e.g., a single sol at 54 km resolution in MarsWRF requires ~34 wall-hours). Furthermore, existing reanalysis products, such as OpenMARS, are limited by coarse spatial resolution (5°×5°), short temporal records (approx. 8 Mars years), and fragmentation across different instrument generations.

Unlike Earth, where data-driven foundation models have recently matched or exceeded operational numerical weather prediction (NWP) using dense reanalysis datasets like ERA5, Mars presents unique challenges:

• Data Scarcity: The observational record is sparse, heterogeneous, and instrument-dependent.
• Physical Differences: The dominance of the CO2 mass cycle (seasonal freezing/sublimation), the radiative dominance of dust, and the absence of liquid water mean Earth-based loss functions and inductive biases do not transfer directly.
• Observational Gaps: No single instrument provides continuous global coverage, and direct wind retrievals are largely absent from current orbital datasets.

The paper proposes the development of a Mars Atmospheric Foundation Model (MAFM): a single, multi-scale model designed to represent, forecast, and detect principal Martian atmospheric phenomena in a data- and compute-efficient manner.

2. Methodology and Design Landscape

2.1 Data Perspective

The authors characterize the data landscape to define the scope and constraints of the MAFM:

• Reanalysis Data: OpenMARS (versions 2–5) and EMARS provide physically consistent, global gridded states. OpenMARS v5 (MY 34–35) offers hourly cadence and resolved water vapor, while earlier versions lack water vapor. Resolution is 5°×5° horizontally with 35–70 vertical sigma levels.
• Atmospheric Retrievals: Five instrument suites (TES, MCS, ACS, NOMAD, EMIRS) provide higher-resolution data (5–300 km) but are spatially/temporally irregular and lack direct wind profiles.
• GCMs as Data Sources: MGCM and MarsWRF serve dual roles: generating additional training data with richer physics (e.g., explicit water-ice microphysics, CO2 frost) and providing baselines for validation. However, their computational cost limits the generation of multi-year high-resolution archives.
• Scope Tradeoffs: The model must navigate the asymmetry between coarse reanalysis and fine retrievals, the absence of water vapor in pre-v5 reanalysis, and the lack of vertical velocity and electrification data.

2.2 Downstream Applications

The paper identifies a wide range of candidate applications, categorized by atmospheric phenomenon:

• Dust Storms: Detection, tracking, and life-cycle prediction of regional to global storms.
• Synoptic Systems: Frontal systems and Double Annular Cyclones (DACs).
• Mesoscale/Diurnal Features: Low-level jets (LLJs), orographic water-ice clouds, the Arsia Mons Elongated Cloud (AMEC), and the Aphelion Cloud Belt (ACB).
• Global Cycles: CO2 and H2O frost cycles and global surface-pressure/mass-budget forecasting.
• Cross-Cutting Tasks: Spatial downscaling (5° to 1°/2°) and reanalysis-observation fusion.

2.3 AI Perspective and Architecture

The paper surveys recent AI developments relevant to atmospheric physics, including:

• Data-Driven Forecasting: Autoregressive models (e.g., GraphCast, Prithvi-WxC) that predict the next state from previous states.
• Data Assimilation: Variational, regression-based, and diffusion-based approaches (e.g., DiffDA, Score-Based DA) to integrate sparse observations.
• Data Efficiency: Techniques for limited data settings, including pre-training on generic PDEs, transfer learning, and architectural inductive biases (e.g., conservation laws, locality).

Three candidate architecture families were implemented and evaluated on OpenMARS data:

1. Mars SpectFormer: A hybrid spectral-attention transformer combining global frequency-domain processing with local spatial attention.
2. Mars-Adapted GraphCast: A graph neural network (GNN) representing the atmosphere on an icosahedral multi-mesh.
3. Mars-Adapted Prithvi-WxC: A Vision Transformer (ViT) based model adapted for surface variables.

Additionally, the study explored PDE Pretraining, extending the Poseidon PDE foundation model (pretrained on Navier-Stokes/Euler equations) to 3D to accelerate training on OpenMARS data.

3. Key Results

3.1 Predictive Performance on OpenMARS

The authors evaluated the three architectures on OpenMARS data (MY 24–35) for short-to-medium-range forecasting (up to 24–72 hours):

• Mars SpectFormer: Demonstrated substantial skill over a persistence baseline for dynamical variables (zonal/meridional wind, surface pressure) across a 24-hour horizon. However, it struggled with temperature (crossing persistence error at ~20 hours) and dust column opacity (worse than persistence at all lead times), likely due to the intermittent nature of dust events and systematic thermal bias accumulation.
• Mars-Adapted GraphCast: Showed significant systematic biases. It failed to reproduce mesoscale structures and exhibited rapid divergence in temperature and wind components, exceeding persistence error from the earliest lead times. Surface pressure errors showed continuous drift.
• Mars-Adapted Prithvi-WxC: Maintained reasonable skill for zonal wind and moderate skill for meridional wind, preserving large-scale circulation patterns. However, it exhibited systematic biases in surface pressure (accumulating drift) and dust column opacity. It partially captured diurnal thermal forcing but struggled with dust prediction.

3.2 PDE Pretraining Efficacy

Experiments using a 3D-extended Poseidon model pretrained on PDE solutions showed that pretraining significantly outperformed random initialization. The pretrained model converged faster and avoided the overfitting observed in the randomly initialized model. Crucially, the performance gap widened when initial conditions were sparsified, suggesting PDE pretraining is highly effective for data-limited settings.

3.3 Explainability

The paper outlines a framework for Mechanistic Interpretability (MI) for MAFMs, proposing three strands:

1. Causal Component Analysis: Identifying internal units (neurons, attention heads) responsible for specific behaviors.
2. Representation-Level Analysis: Using sparse autoencoders to uncover latent features corresponding to physical variables.
3. Steering and Control: Intervening on internal representations to correct biases or enforce physical consistency without full retraining.

4. Significance and Claims

The paper does not claim to have solved Martian weather forecasting but rather elucidates the design landscape for a Mars Atmospheric Foundation Model. Its primary contributions are:

1. Data-Task Mapping: A systematic linkage of downstream applications (e.g., AMEC, DACs, LLJs) to specific data sources (OpenMARS versions, specific instruments), highlighting critical constraints such as the lack of water vapor in pre-v5 reanalysis and the absence of wind retrievals.
2. Feasibility Assessment: Preliminary results indicate that while data-driven models can outperform persistence for dynamical variables, they currently struggle with variables requiring high-fidelity microphysics (dust, temperature) and long-term stability (pressure drift).
3. Architectural Direction: The comparison of SpectFormer, GraphCast, and Prithvi-WxC suggests that hybrid architectures or those with strong inductive biases may be necessary. The success of PDE pretraining highlights a viable path for overcoming data scarcity.
4. Strategic Roadmap: The paper argues that a single foundation model is preferable to multiple specialized models given the small size of the Mars science community and the coupled nature of CO2, dust, and water cycles. It emphasizes that multi-instrument fusion and cross-Mars-year generalization are mandatory, not optional, for a successful MAFM.

The authors conclude that while short-term rollouts are an achievable target, long-term stability and the accurate representation of "black swan" events (global dust storms) and microscale phenomena remain significant challenges requiring further architectural innovation and data fusion strategies.