MarsNet reconstructs Martian weather from observations—with physical guardrails

Why try another approach?

Existing Martian reanalyses combine spacecraft measurements with general circulation models: simulations that calculate how the atmosphere evolves. Where measurements are missing, the reconstruction depends heavily on the simulation. Different reanalyses can therefore disagree substantially without an obvious way to identify the correct answer.

Sebastian Mengwall and J. Michael Battalio developed MarsNet to provide a reference built differently. No general circulation model output or gridded atmospheric reanalysis supplies its training targets, initial state or normalization statistics. Existing reanalyses are used only for evaluation.

“Observations alone” does not mean “without assumptions.” MarsNet contains substantial physical structure, and the instrument retrievals—the atmospheric quantities inferred from measurements—can themselves incorporate model assumptions.

How MarsNet works

Every two hours of Mars time, MarsNet forecasts the atmosphere, predicts what instruments should observe, and then updates its estimate using the measurements available in that window. Forecasting and updating are trained jointly, with observations scored before they are assimilated.

The model represents temperature, surface pressure, winds, dust, water ice and other quantities in physical units. Its structure combines recurring seasonal and daily patterns fitted to observations, statistical models of persistent large-scale departures, and finer spatial detail.

A graph neural network helps set the strengths of named processes, including transport, dust lifting, sedimentation and heating. Physical signs and numerical bounds restrict those adjustments. A bounded Kalman-style update balances uncertainty in the forecast against measurement error, limiting how strongly observations can change the estimate.

This design addresses a central problem: Mars has too few observations to constrain an unrestricted neural network. Without fresh measurements, departures gradually fade toward the model’s fitted climatology, or typical seasonal and daily conditions.

What the tests found

MarsNet was trained on Mars Years 24–31 and validated on Mars Year 32. The researchers reserved entire seasons of radio-occultation temperatures, inferred from the bending of radio signals through the atmosphere, from training, assimilation and fitted tables.

Against held-out Mars Global Surveyor measurements near the surface, MarsNet’s root-mean-square temperature error was 4.2 kelvin, compared with 4.4–5.4 kelvin for MACDA, OpenMARS and EMARS. This error measure gives greater weight to larger mistakes. In the lowest three atmospheric scale heights—layers measured by how rapidly pressure decreases—MarsNet generally outperformed MACDA and OpenMARS and was comparable to EMARS. Although three seasons were withheld, the shared comparison covered two because some reanalyses lacked the third.

Higher up, the advantage disappeared. At four to five scale heights, MarsNet’s error was 6.9 kelvin, versus 5.5–5.9 kelvin for the reanalyses. A separate Mars Express occultation set also did not show uniform superiority.

Surface pressure was a clearer strength. Against withheld Curiosity measurements in Mars Year 32, MarsNet’s error was 9.7 pascals, compared with 19.9–46.3 pascals. However, MarsNet had learned from pressure observations in other periods, while the comparison reanalyses did not assimilate surface pressure. The advantage therefore reflects different observational inputs as well as modeling choices.

Forecasting through data gaps

The team also launched forecasts from 218 starting points, withholding every subsequent observation. Against reserved Mars Climate Sounder temperatures, MarsNet remained more skillful than a separately fitted climatology and a fading-memory persistence forecast through ten sols, or Martian days.

These comparisons used a probability-based score that assesses both predictions and their uncertainty, not just average temperature errors. Predicted temperature uncertainty was reasonably close to observed errors, though somewhat too narrow. A separate calibration analysis excluded exceptionally cold observations below the model’s carbon-dioxide frost threshold.

Crucially, temperature forecasts before and after neural-network training differed little. Much of the skill apparently came from the fixed structure and observation-fitted tables. Training mainly improved surface pressure, while degrading dust performance.

Limits and significance

Dust and water-ice predictions often failed to beat simple reference forecasts, and the authors claim no wind skill. Temperature uncertainty was poorly calibrated during years dominated by the older Thermal Emission Spectrometer record. Stability checks succeeded across nine separately initialized Mars-year runs; that is not proof of accurate atmospheric behavior everywhere.

The Mars Year 32 validation data also helped select the released model, so those results may be optimistic. Later test years were not examined. The strongest fully independent temperature checks cover earlier years and limited locations.

The November 2026 preprint presents provisional findings. Code, weights and evaluation data were promised for release in coming months, rather than documented as already available. MarsNet’s potential lies in offering a fast, traceable alternative reference and enabling experiments that test how much particular instruments contribute to reconstructing Martian weather.