Abstract Recent work has pointed out the potential existence of a tight relation between the cosmological parameter Ω_m, at fixed Ω_b, and the properties of individual galaxies in state-of-the-art cosmological hydrodynamic simulations. In this paper, we investigate whether such a relation also holds for galaxies from simulations run with a different code that makes use of a distinct subgrid physics: Astrid. We also find that in this case, neural networks are able to infer the value of Ω_m with a ∼10% precision from the properties of individual galaxies, while accounting for astrophysics uncertainties, as modeled in Cosmology and Astrophysics with MachinE Learning (CAMELS). This tight relationship is present at all considered redshifts, z ≤ 3, and the stellar mass, the stellar metallicity, and the maximum circular velocity are among the most important galaxy properties behind the relation. In order to use this method with real galaxies, one needs to quantify its robustness: the accuracy of the model when tested on galaxies generated by codes different from the one used for training. We quantify the robustness of the models by testing them on galaxies from four different codes: IllustrisTNG, SIMBA, Astrid, and Magneticum. We show that the models perform well on a large fraction of the galaxies, but fail dramatically on a small fraction of them. Removing these outliers significantly improves the accuracy of the models across simulation codes.