Mars GCM climate model¶
The primary Mars model resolves atmospheric dynamics and column physics on a sphere with terrain-following sigma layers. It evolves circulation and spatial temperature/pressure fields alongside surface temperature, CO₂ frost and optional multilayer soil. The torch global-mean ODE model is deprecated.
Dynamics and state¶
Dinosaur advances hydrostatic primitive equations with spectral vorticity, divergence, temperature variation and logarithmic surface pressure. Mars supplies its body constants and MOLA lower boundary. The grid is spherical harmonics horizontally and sigma coordinates vertically; surface reservoirs are not advected as atmospheric tracers.
Energy exchange¶
Solar geometry follows a Keplerian eccentric orbit and either moving diurnal illumination or a daily mean. The radiation scheme returns directional interface fluxes whose convergence drives atmospheric heating and surface energy input. Resolved CO₂ radiation uses bundled Ames correlated-k tables or an explicitly selected compact-band alternative; simple CLI configurations use a grey balance. Dust optical depth is prescribed and can vary seasonally.
Surface sensible heat is transferred to the lowest layer with its physical pressure-dependent heat capacity. Surface loss and atmospheric gain cancel in the exchange budget. Optional conduction transfers heat through the regolith; PBL diffusion and smooth dry convective exchange redistribute column energy.
CO₂ reservoirs¶
Condensation removes atmospheric mass and releases surface latent heat; sublimation reverses the transfer. Frost is stored in pressure-equivalent units, and its local saturation temperature depends on pressure. Energy-limited phase exchange is tied to the surface-energy residual.
Advancing logarithmic pressure introduces finite-step inventory error. The complete-step wrapper repairs negative frost and rescales atmospheric pressure so global Gaussian-quadrature atmosphere-plus-frost matches the pre-step inventory minus prescribed escape. This is a global budget correction, not proof of exact local transport or full-system energy conservation.
Integration and learning¶
The GCM uses IMEX-RK-SIL3: implicit gravity-wave dynamics and explicit physics. It is separate from the deprecated model's RK4/fast modes. JAX scan rollouts support gradients, batching and optional rematerialization; tests compare selected gradients with finite differences. Physical calibration controls gas/dust thermal opacity and surface exchange. Learned radiation replaces fluxes, while bounded neural temperature heating adds a source. Temperature-only postprocessing instead fits frozen forecast outputs.
What results mean¶
Short maps are transient diagnostics. A duration above 668 sols does not establish equilibrium; phase-matched atmospheric, surface and deep-soil repeatability must be checked. Existing reference comparisons differ in sampling and wind height and are not independent observational validation. Water/cloud physics, interactive dust and composition-dependent high-pressure validation remain absent.