MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres
Earth-trained GraphCast was adapted to Mars and learned useful Martian temperature forecasts after fine-tuning.
The paper tests zero-shot and fine-tuned GraphCast on Mars Climate Database fields for temperature and winds. Zero-shot runs captured current conditions but drifted toward climatological means and missed the diurnal cycle. Fine-tuning with Martian variables and solar radiation forcing let the model learn thermal variability within about 10 epochs. Forecasts up to 10 days reproduced seasonal and vertical temperature structure, with performance depending on sample size and seasonal initialization. ArXiv · AI/CL/LG's note
The paper tests zero-shot and fine-tuned GraphCast on Mars Climate Database fields for temperature and winds. Zero-shot runs captured current conditions but drifted toward climatological means and missed the diurnal cycle. Fine-tuning with Martian variables and solar radiation forcing let the model learn thermal variability within about 10 epochs. Forecasts up to 10 days reproduced seasonal and vertical temperature structure, with performance depending on sample size and seasonal initialization. ArXiv · AI/CL/LG's note
score 4