MarsNet learns the Martian atmosphere from observations alone: 4.2 K near-surface temperature error and the best surface pressures of four products
Synopsis
The Martian atmosphere has been described by reanalyses that assimilate spacecraft retrievals into general circulation models, so where observations are sparse their fields largely reflect the model. MarsNet trains a two-hour forecast and a per-window observation update jointly inside a fixed physical structure, using no reanalysis or model field as input or target. On MGS radio-occultation temperatures assimilated by no dataset, it matches or improves on the three reanalyses in the lowest three scale heights, with a near-surface RMSE of 4.2 K against 4.4–5.4 K, but is less accurate aloft.
Interpretation
An estimate of the Martian atmosphere can be learned from observations alone, without a reanalysis field as a training target. The three existing reanalyses assimilate spacecraft retrievals into their own general circulation models, so in sparse regions and for rarely observed quantities the fields come from the model's dynamical core, dust and cloud parameterizations, and boundary layer. On MGS radio-occultation temperatures assimilated by no dataset, MarsNet matches or improves on MACDA and OpenMARS in the lowest three scale heights and is level with EMARS; its near-surface RMSE is 4.2 K against 4.4–5.4 K for the reanalyses.
Forecast and observation update are learned jointly inside one fixed physical structure, with the network adjusting named physical terms and update weights. The Martian record is too sparse to constrain an unrestricted network, so learning is confined within the structure classical data assimilation already provides. The state is in physical units, the forecast uses physical tendencies of fixed form and learned strength, and the update is a bounded Kalman update; the model trains on MY24–31 and is scored on MY32 and held-out occultation seasons.
Surface pressures are the most accurate of the four products, and the model cycles stably through nine Mars years. The reanalyses do not assimilate surface pressure, so this comparison concerns a quantity none of them used. Against Curiosity's pressures in MY32, MarsNet's RMSE is 10 Pa against 20–46 Pa for the three reanalyses; cycled through each of MY24–32 from its own start, the state stays within its bounds in every year.
Perspective
The comparison is made only on observations no product assimilated: the held MGS occultation seasons, the MEX occultations, and Curiosity's MY32 surface pressures; the fully independent temperatures cover MY24–28, and the MEX set is small.
The MY32 scored sets contain no occultation temperatures, so temperature skill is measured against held-out MCS profiles that share retrieval errors with the profiles assimilated around them; the MY32 holdout also served to stop training and choose the released run, so its skill may be slightly optimistic. In MY24–27, when only TES is assimilated, the stated uncertainty is not calibrated; the model claims no skill for winds, and it holds too little dust in the MY28 storm.