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

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

Problem Statement
Recent advances in machine learning have established Earth-based weather foundation models, such as GraphCast, as state-of-the-art tools for terrestrial forecasting, offering significant computational advantages over traditional Numerical Weather Prediction (NWP). However, the applicability of these models to non-Earth planetary environments remains unexplored. The Martian atmosphere presents a distinct challenge due to its thin, CO₂-dominated composition, surface pressures two orders of magnitude lower than Earth's, and unique dynamics driven by strong diurnal variability, seasonal mass exchange, and dust aerosol heating. Traditional Mars Global Climate Models (MGCMs) are physically comprehensive but computationally intensive, limiting their utility for rapid forecasting or ensemble-based uncertainty quantification. The central question addressed is whether Earth-trained atmospheric foundation models can be transferred to planetary environments with fundamentally different physics and composition, and if so, what limitations exist in zero-shot application and how they can be overcome through fine-tuning.

Methodology
The study adapts GraphCast, a graph neural network (GNN) weather forecasting model originally trained on Earth's ERA5 reanalysis data, to simulate Martian atmospheric dynamics. The methodology proceeds in two phases:

1. Data Preparation and Preprocessing:

   • Source Data: The Mars Climate Database (MCD) provides global atmospheric fields, including 2-meter air temperature, surface temperature, and atmospheric variables at 13 vertical pressure levels (50–1000 Pa), along with top-of-atmosphere (TOA) solar radiation forcing.
   • Rescaling: To maintain numerical consistency with the pretrained GraphCast weights, Mars temperature fields were rescaled to match the dynamic range of the ERA5 dataset using a linear transformation.
   • Variable Handling: The model inputs included 2-meter temperature, 10-meter wind components, and 3D wind/temperature fields at 13 levels. Crucially, humidity was held constant at a global mean value to isolate temperature variability and reduce dimensionality, and the land-sea mask was set to unity (treating Mars as entirely land).
   • Temporal Sampling: Data was sampled at 00, 06, 12, and 18 hours to capture diurnal cycles, with the seasonal cycle represented by solar longitude ($L_s$).

2. Experimental Design:

   • Zero-Shot Evaluation: GraphCast was applied autoregressively to Martian states without retraining to assess its ability to generalize Earth-learned dynamics to Mars.
   • Fine-Tuning Strategy: A targeted fine-tuning approach was employed using the AdamW optimizer (learning rate $1 \times 10^{-6}$) and Mean Squared Error (MSE) loss. The model was unfrozen to allow adaptation of the encoder, processor, and decoder. Training utilized sequential triplets of time steps (e.g., $T_{00}, T_{06} \rightarrow T_{12}$) to learn diurnal heating and cooling driven by solar forcing.
   • Sensitivity Analysis: Experiments evaluated performance across varying training durations (up to ~1000 epochs) and sample sizes (subsets of 30, 100, and 360 days).

Key Contributions

• Demonstration of Transferability: The study provides the first evidence that an Earth-trained foundation model can be adapted to simulate Martian atmospheric dynamics, moving beyond the "from-scratch" training approaches previously attempted for planetary AI models.
• Identification of Zero-Shot Failure Modes: The research characterizes the specific failure of Earth-trained models in zero-shot scenarios, noting a rapid decay toward climatological means and a complete inability to reproduce diurnal variability.
• Efficient Adaptation Protocol: The paper establishes a fine-tuning protocol that requires minimal data (as few as 30 days of training samples) and computational resources (hours on a single GPU) to achieve rapid learning of planetary thermal dynamics.
• Open-Source Model: The resulting fine-tuned model, termed "MarsCast," along with its architecture and checkpoints, has been made available via GitHub and Hugging Face.

Results

• Zero-Shot Performance: Without fine-tuning, GraphCast produced discernible spatial patterns but failed to capture diurnal cycles. Forecasts rapidly collapsed toward climatological equilibrium, indicating that Earth-specific radiative balance and thermodynamic assumptions do not generalize to Mars.
• Rapid Learning via Fine-Tuning:
  • Diurnal Cycle: Within just 10 training epochs, the fine-tuned model began to capture the Martian diurnal heating and cooling cycle.
  • Accuracy: After ~1000 epochs, the model reproduced diurnal temperature variability consistent with solar forcing, with errors reduced to less than ±2 K in most locations.
  • Data Efficiency: The model demonstrated high data efficiency, learning planetary dynamics effectively from small subsets of the MCD.
• Vertical and Seasonal Structure:
  • Vertical Profiles: MarsCast successfully reproduced vertical temperature profiles at 13 levels, showing strong agreement with MCD in the middle atmosphere (250–800 Pa). Discrepancies were slightly higher in the lower and upper atmosphere but preserved major gradients.
  • Seasonality: The model maintained forecast skill across four representative seasons ($L_s = 60^\circ, 150^\circ, 240^\circ, 330^\circ$), accurately capturing seasonal dependence.
• Wind Prediction: When wind variables were included in the final fine-tuning phase, the model successfully predicted both meridional and zonal winds, with wind speed differences generally within ±4 m/s for the first 48 hours.

Significance and Claims
The paper claims that the primary significance of this work lies in demonstrating the feasibility of adapting Earth weather foundation models for planetary science, offering a pathway toward rapid, AI-augmented weather prediction. The authors highlight three specific advantages of this approach:

1. Speed: Forecast generation is orders of magnitude faster than running full physics MGCMs.
2. Computational Efficiency: Generating a 10-day, 1° resolution forecast requires approximately 2 minutes on a single GPU, compared to 30 minutes on a single CPU for a coarser 5° resolution physics-based forecast. This efficiency enables the practical use of ensemble forecasting.
3. Resolution Enhancement: Despite being initialized with MCD data downscaled to 1° via bilinear interpolation, the model's outputs visually capture finer resolution details, suggesting that foundation models can effectively downscale coarser inputs.

The study concludes that while the current implementation simplifies humidity and dust dynamics, the success of MarsCast validates the concept of using a "base" foundation model for planetary adaptation. This approach provides a cost-effective method to generate new planetary models, supporting future mission operations, dust storm risk mitigation, and human exploration planning.