Mars Weather Transfer Learning with Aurora and OpenMARS
Mars weather prediction is critical for future robotic and human missions, but Mars has far less observational data than Earth. Our project tests a practical question: can a weather model pretrained on Earth generalize to Mars after fine-tuning on limited Martian data? We adapted Aurora, a pretrained atmospheric foundation model, to predict global Martian weather fields from the OpenMARS reanalysis dataset. The model forecasts surface variables every 2 Martian hours, including surface pressure ps in Pa, surface temperature tsurf in K, CO2 surface ice co2ice in kg/m², and dust column opacity dustcol. It also predicts 3D atmospheric fields across 35 sigma-pressure levels: zonal wind u in m/s, meridional wind v in m/s, and atmospheric temperature temp in K. We fine-tuned on OpenMARS Mars Years 28-34 and validated on Mars Year 35. Because of hackathon time constraints, we ran only two main experiments: one initialized from Earth-weather pretraining and one trained without that pretrained initialization. We evaluated both with 20-step autoregressive rollouts, measuring normalized MSE along with physical RMSE and MAE for each predicted field. Overall, the pretrained and non-pretrained models performed similarly across most rollout metrics. The clearest benefit appeared in surface temperature prediction: the Earth-pretrained model reduced tsurf prediction error by roughly 30% compared with the non-pretrained model. This suggests that some learned atmospheric structure from Earth weather modeling can transfer to Mars, even though the planets differ substantially in pressure, composition, thermal forcing, and dust dynamics. The result is preliminary, but promising: with more training runs, stronger hyperparameter sweeps, and longer validation across multiple Mars years and dust seasons, Earth-pretrained weather foundation models may become a practical starting point for data-scarce planetary forecasting.
Sorry, didn't have time to get the video up in time. :/
Log in or sign up for Devpost to join the conversation.