At the surface velocity inversions, MURaM-ARs dataset. Patches of 50 × 50 pixels² extracted from 30-timestep-averaged transverse velocity field maps which were computed at optical depth τ ≈ 1 by (a) the MURaM-ARs simulation (Rempel & Cheung, 2014), (b) the DeepVel neural network that generated Figure 1a, and (c) a version of the DeepVel neural network which was trained using the MURaM-ARs simulation resampled at the SDO/HMI spatial resolution (Δx ≈ 368 km). The vertical component of the velocity computed by the MURaM-ARs simulation is displayed as colored background. (Right column) Scatterplots comparing amplitudes ||v_D,t|| to ||v_ref,t||. The black line represents the expected solution (i.e., a coefficient of determination R² = 1).