Testing noncollinear spin-flip, collinear spin-flip, and conventional time-dependent density functional theory for predicting electronic excitation energies of closed-shell atoms

Conventional time-dependent density functional theory (TDDFT) is based on a closed-shell Kohn-Sham (KS) singlet ground state with the adiabatic approximation, using either linear response (KS-LR) or the Tamm-Dancoff approximation (KS-TDA); these methods can only directly predict singly excited states. This deficiency can be overcome by using a triplet state as the reference in the KS-TDA approximation and "exciting" the singlet by a spin flip (SF) from the triplet; this is the method suggested by Krylov and co-workers, and we abbreviate this procedure as SF-KS-TDA. SF-KS-TDA can be applied either with the original collinear kernel of Krylov and co-workers or with a noncollinear kernel, as suggested by Wang and Ziegler. The SF-KS-TDA method does bring some new practical difficulties into play, but it can at least formally model doubly excited states and states with double-excitation character, so it might be more useful than conventional TDDFT (both KS-LR and KS-TDA) for photochemistry if these additional difficulties can be surmounted and if it is accurate with existing approximate exchange-correlation functionals. In the present work, we carried out calculations specifically designed to understand better the accuracy and limitations of the conventional TDDFT and SF-KS-TDA methods; we did this by studying closed-shell atoms and closed-shell monatomic cations because they provide a simple but challenging testing ground for what we might expect in studying the photochemistry of molecules with closed-shell ground states. To test their accuracy, we applied conventional KS-LR and KS-TDA and 18 versions of SF-KS-TDA (nine collinear and nine noncollinear) to the same set of vertical excitation energies (including both Rydberg and valence excitations) of Be, B⁺, Ne, Na⁺, Mg, and Al⁺. We did this for 10 exchange-correlation functionals of various types, both local and nonlocal. We found that the GVWN5 and M06 functionals with nonlocal kernels in spin-flip calculations can both have accuracy competitive to CASPT2 calculations. When the results were averaged over all 36 test energy differences, seven (GVWN5, M06, B3PW91, LRC-ωPBE, LRC-ωPBEh, PBE, and M06-2X) of the 10 studied density functionals had smaller mean unsigned errors for noncollinear calculations than the mean unsigned error of the best functional (M06-2X) for either conventional KS-TDA or KS-LR.