Perezhogin et al. propose in this preprint , a physics-informed neural network to improve the generalization of data-driven mesoscale eddy parameterizations in ocean models. By applying local input-output scaling based on dimensional analysis, their method adapts to different grid resolutions and depths. This approach enhances energy representation and affects biases in both idealized and global ocean simulations. The scaling framework is broadly applicable and robust across configurations. Results show competitive performance compared to traditional parameterizations.