AEGIS: a differentiable climate model of Mars, designed to incorporate neural networks

To replace Martian models written in old Fortran, researchers are proposing AEGIS, which is capable of simulating ten Martian years and providing gradients for calibration and training.

Predicting the Martian weather is not an academic exercise: the entry, descent and landing of missions depend on it. The task is arduous, because observations are rare, the CO₂ atmosphere is very tenuous and the dust cycle, which absorbs and re-emits radiation, causes huge temperature differences between day and night, with no terrestrial equivalent to validate the models.

Leaving Fortran

Reference general circulation models (LMD PCM, NASA Ames model, PlanetWRF) are detailed and proven, but written in Fortran with classical solvers that do not compute gradients, which prevents them from being automatically calibrated or grafted with machine learning. Sameera Kashyap, Razvan Marinescu and colleagues present AEGIS, a modular, differentiable model that combines Mars-specific physics with the dynamic Dinosaur core and provides locations for "neural closures," i.e., learned components.

The model holds stable simulations over ten Martian years, reproduces the seasonal cycle of CO₂ while conserving its total quantity, finds the structure of surface temperatures and a pressure that follows the relief measured by the MOLA laser altimeter. The gradients obtained agree with those calculated by finite differences, which opens the way to physical calibration and training of neural networks directly in the model.