Orbital Chaos Maps as a Test of Transferability in Continuum-Dynamics Foundation Models

Foundation models trained on PDE-governed physical systems have shown promise in predicting the evolution of fluid, thermal, and elastic fields. We investigate whether their learned representations transfer to a structurally different problem: mapping orbital chaos in the Sun–Jupiter system. Using the MEGNO chaos indicator, we construct two-dimensional maps of orbital stability over semi-major axis and eccentricity in the Jovian co-orbital region and its surroundings. Although these maps resemble continuous physical fields, each grid point is generated by an independent orbital integration rather than through local spatial coupling. We use these maps to evaluate Walrus, a pretrained continuum-dynamics foundation model, across two modes of variation: changes in orbital phase (phase-offset) and increasing integration time (temporal). Zero-shot prediction fails to recover the phase-dependent structure, while retaining substantial fidelity in the temporal evolution of the maps. Frozen-backbone fine-tuning does not resolve the phase-offset failure, whereas full-backbone fine-tuning yields modest, configuration-dependent improvements in temporal prediction. These results identify a specific limitation in the transferability of the pretrained model and suggest that orbital chaos maps provide a useful benchmark for evaluating physics foundation models beyond their original training domains.