WindSightNet: The Inter-Annual Variability of Martian Winds Retrieved From InSight's Seismic Data With Machine Learning

Abstract

Wind measurements from landed missions on Mars are vital to characterize the near surface atmospheric behavior on Mars and improve atmospheric models. These winds are responsible for aeolian change and the mixing of dust in and out of the atmosphere, which has a significant effect on global circulation. The NASA InSight mission recorded wind data for around 750 sols. The seismometer, however, recorded data for around 1400 sols. The dominant source of energy in the seismic data is in fact due to winds. To this end, we propose a machine learning model, dubbed WindSightNet, to map the seismic data to wind speed and direction. The trained network achieves wind speed and direction measurements with errors of 0.932 m/s and 32.6°. We use WindSightNet to retrieve winds from the entire time the seismometer was recording to compare year-to-year wind variations at InSight. The continuous nature of the data set enables the extraction of periodic behavior. We observe a pattern of waves due to baroclinic activity with periods of $\sim$2–3, $\sim$4, $\sim$5–7 and $\sim$9–20 sols occurring $L_s=$180–360°. We also observe periodicity during the day due to convective cells. This is used to estimate the boundary layer height, yielding values between 2.3 and 7.7 km. A data-science based metric is proposed to provide a quantification of the year-to-year differences in the wind speeds. This highlights variations linked to dust activity as well as other transient differences. On the whole, the seismic-derived winds confirm the dominance of the global circulation leading to repeatable weather patterns.