AEGIS couples Mars atmospheric physics to a differentiable Dinosaur core, yielding stable ten-Mars-year integrations that reproduce the seasonal CO2 cycle

AEGIS extends the Dinosaur core with planetary-unit rescaling, a coupled atmosphere–surface–frost–regolith state, CO2 phase change fed into the prognostic log surface pressure, and a per-step atmosphere–frost inventory projection, making mass exchange and surface memory part of the differentiated trajectory. Earlier Dinosaur-based models (NeuralGCM, JCM) target Earth configurations; AEGIS uses Mars radius, rotation, and thermodynamic scales at the dynamical interface and adds Mars radiation, surface, and CO2 condensation-cycle operators. The methods give the coupled-state equation, the CO2 condensation contribution to prognostic log pressure, and the per-step inventory projection; Appendix A details the radiation, surface, boundary-layer, and CO2-inventory operators.

Gradients through coupled trajectories are verified against finite differences: all 54 physical-sensitivity checks pass, and nine multi-diagnostic Jacobian derivatives agree with centered finite differences with a small maximum relative discrepancy. Established Mars GCMs are legacy Fortran finite-difference or finite-volume codes that do not expose gradients; AEGIS provides forward- and reverse-mode differentiation within the same implementation. Table 1 lists four gradient-check categories, all passing; the multi-diagnostic experiment compares AD with finite differences at 0.25, 1, and 5 sol horizons and notes that pressure and frost derivatives are equal and opposite, their sum bounded by the inventory projection.

Gradient-based calibration lowers atmospheric temperature RMSE on every data split in the twelve-window pilot: training 2.45 to 2.38 K, validation 2.49 to 2.44 K, and test 2.46 to 2.40 K, with all four test windows improving. This shows that differentiating the coupled trajectory loss yields a direction-correct optimization signal, whereas parameter sweeps and ensemble calibration are typically computationally expensive. Table 2 reports per-split RMSE for physical baseline, calibrated physics, and coupled neural heating; Appendix B documents the twelve ARCO-MACDA window splits and the 50 Adam updates.

Long integrations are stable and conservative: one Mars year completes in 86.6 minutes and ten Mars years in 11.4 hours, all saved states are finite, the maximum relative atmosphere–frost CO2 inventory drift is very small, and the seasonal cycle in year ten closely repeats year nine. The model reproduces the seasonal CO2 cycle (global-mean surface pressure 485 to 654 Pa, about a 30% annual range) and the MOLA-terrain-driven surface-pressure structure, with pressure pattern correlations of 0.975 to 0.989 across twelve season–reference comparisons. Performance and conservation figures come from single-GPU run statistics; Table 3 gives RMSE and correlations against NASA Ames, MCD 6.1, and ARCO-MACDA references; all thirteen ablation configurations complete with finite states.

The work targets the specific physical regime of the Mars atmosphere and is evaluated at coarse resolution with T21 triangular truncation, 12 sigma layers, and a 300 s timestep, suited to physical sensitivity studies, gradient-based calibration, and controlled neural-closure experiments. The authors note the framework already supports broader observational comparison, repeated neural experiments across seeds, and multi-year equilibration studies, and that altered surface albedo, imposed heating, or specified changes in volatile inventory could be introduced through the explicit planetary, radiation, and reservoir interfaces and evaluated with spatial climate diagnostics; larger changes in atmospheric composition and mass would require extending body thermodynamics, opacity data, condensable processes, and applicable dynamical assumptions together.

In the seasonal comparisons, near-surface wind speed is the field with the largest spread across reference products because the products differ in reported level and averaging, which the authors frame as model-versus-reanalysis comparison; surface-temperature pattern correlations of 0.732 to 0.821 are lower than the pressure correlations. In the neural radiation experiment, continued local fitting outperforms trajectory fine-tuning on both interpolation and higher-dust extrapolation, indicating that the trade-off between local and coupled-trajectory objectives remains an open question. The ablations reach chaotic saturation at the 44,400 s horizon, so process sensitivity must be read from the short horizon. In addition, several numeric values in the main text (such as gradient discrepancies, inventory drift, and closure residuals) appear as symbols or placeholders, so readers needing exact figures should consult the original tables and figures.