Energy transfer during molecular collisions at a metal surface represents a sensitive probe of the molecule–surface interaction potential. Here, via molecular dynamics calculations on several first-principles neural network potentials, we find that the vibrational energy transfer dynamics of highly vibrationally excited NO and CO molecules scattered from Au(111) are strongly correlated with their respective potential energy landscapes in the vicinity of the dissociation barrier. Our results not only reproduce the observed significantly less vibrational relaxation of CO (vi = 17) than NO (vi = 16) scattered from Au(111) and attribute it to the different dissociation barriers in the two systems, but also show a dramatic change of dynamics due to a minor adjustment of the energy landscape near the barrier in the CO case. We find that the BEEF-vdW density functional based potential largely overestimates the vibrational relaxation of CO (vi = 17) even in the absence of any nonadiabatic energy loss, despite its good description for CO adsorption on Au(111). We also discuss the possibility of integrating the validated adiabatic potential with first-principles determined diabatic states towards a more complete description of the non-adiabatic energy transfer in these benchmark systems.