Do Better Scores Mean Better Physics? Physics-Grounded Explanations for Sim2Real Neural Operators

Machine-learning surrogates accelerate physical simulation, but lower prediction error does not necessarily indicate more faithful physical behavior. We study this problem for flow around a NACA4418 airfoil using paired computational-fluid-dynamics simulations and experimental particle-image-velocimetry measurements in a simulation-to-experiment setting. We introduce a physics-grounded perturbation test that removes velocity fluctuations from selected regions of the observed flow history while preserving the local mean flow. Across four neural operators, removing fluctuations from the most energetic 10% of the observed region changes future predictions more than equal-area random removal. We further find that a CNO can have lower field error yet substantially higher TKE error than the reference on both analysis subsets. These results show that scalar accuracy metrics alone cannot determine physical fidelity.