Evaluating Mars 2.0 Across Past and Present Climates: Interpolation, Extrapolation, and Reanalysis

Deep-learning emulators of planetary atmospheres are usually evaluated under conditions similar to those used for training, but their generalization to different climate regimes is largely untested. Early Mars provides a useful test case, as its obliquity varied by tens of degrees and produced distinct climate regimes that can be reproduced with a global climate model. We evaluate Mars 2.0, a HEALPix U-Net surrogate of the Generic Planetary Climate Model in an early-Mars configuration, under four test cases: in-domain, interpolation to a withheld obliquity, extrapolation beyond the training range, and transfer to present-day Mars by fine-tuning on the OpenMARS reanalysis. In-domain, autoregressive rollout error is comparable to the GCM’s own internal-variability level, while the model also generalizes well when interpolating and extrapolating across different obliquity-driven climate regimes, with surface-temperature seasonal damping emerging as the primary out-of-distribution error. Fine-tuning on reanalysis equivalent to 8% of the pretraining data produces stable rollouts for nearly six held-out Mars years and reproduces the annual surface-temperature cycle to 1% in amplitude and $0.4^\circ$ in phase.