How accurate are TD-DFT excited-state geometries compared to DFT ground-state geometries?

In this work, we take a different angle to the benchmarking of time-dependent density functional theory (TD-DFT) for the calculation of excited-state geometries by extensively assessing how accurate such geometries are compared to ground-state geometries calculated with ordinary DFT. To this end, we consider 20 medium-sized aromatic organic compounds whose lowest singlet excited states are ideally suited for TD-DFT modeling and are very well described by the approximate coupled-cluster singles and doubles (CC2) method, and then use this method and six different density functionals (BP86, B3LYP, PBE0, M06-2X, CAM-B3LYP, and omega B97XD) to optimize the corresponding ground- and excited-state geometries. The results show that although each hybrid functional reproduces the CC2 excited-state bond lengths very satisfactorily, achieving an overall root mean square error of 0.011 angstrom for all 336 bonds in the 20 molecules, these errors are distinctly larger than those of only 0.004-0.006 angstrom with which the hybrid functionals reproduce the CC2 ground-state bond lengths. Furthermore, for each functional employed, the variation in the error relative to CC2 between different molecules is found to be much larger (by at least a factor of 3) for the excited-state geometries than for the ground-state geometries, despite the fact that the molecules/states under investigation have rather uniform chemical and spectroscopic character. Overall, the study finds that even in favorable circumstances, TD-DFT excited-state geometries appear intrinsically and comparatively less accurate than DFT ground-state ones.