A new model learns the Martian atmosphere from observations

MarsNet uses observations to forecast temperatures and surface pressures on Mars.

Researchers built MarsNet to learn the Martian atmosphere using only observations. This model alternates a two-hour forecast with an update that assimilates observations.

The model uses a fixed physical structure because the Martian record is too sparse for an unrestricted network. The state is in physical units. The forecast uses physical tendencies of fixed form and learned strength. The update is a bounded Kalman update.

MarsNet was trained on Mars Years (MY) 24--31. It was evaluated on independent observations. It matched or improved on three reanalyses in the lowest three scale heights. The near-surface root-mean-square error (RMSE) was 4.2 K4.2\text{ K} against 4.4–5.4 K4.4\text{--}5.4\text{ K}.

Surface pressures were the most accurate with an RMSE of 10 Pa10\text{ Pa} against Curiosity in MY32. This was compared to 20–46 Pa20\text{--}46\text{ Pa}. The model is less accurate aloft. Training mainly improves surface pressures but comes at some cost to dust.

The model cycles stably through nine Mars years. Its speed makes experiments on the observing record practical. It offers an independent reference for general circulation models (GCMs) and reanalyses.

Key facts

  • MarsNet was trained on Mars Years (MY) 24--31.
  • The model achieved a near-surface root-mean-square error (RMSE) of 4.2 K4.2\text{ K}.
  • Surface pressure error was 10 Pa10\text{ Pa} against Curiosity in MY32.

All sections

Edition