We present a general approach to converge excited state solutions to any quantum chemistry orbital optimization process, without the risk of variational collapse.The resulting Square Gradient Minimization (SGM) approach only requires analytic energy/Lagrangian orbital gradients and merely costs 3 times as much as ground state orbital optimization (per iteration), when implemented via a finite difference approach.SGM is applied to both single determinant ∆SCF and spin-purified Restricted Open-Shell Kohn-Sham (ROKS) approaches to study the accuracy of orbital optimized DFT excited states. It is found that SGM can converge challenging states where the Maximum Overlap Method (MOM) or analogues either collapse to the ground state or fail to converge. We also report that ∆SCF/ROKS predict highly accurate excitation 1 arXiv:1911.04709v1 [physics.chem-ph] 12 Nov 2019 energies for doubly excited states (which are inaccessible via TDDFT). Singly excited states obtained via ROKS are also found to be quite accurate, especially for Rydberg states that frustrate (semi)local TDDFT. Our results suggest that orbital optimized excited state DFT methods can be used to push past the limitations of TDDFT to doubly excited, charge-transfer or Rydberg states, making them a useful tool for the practical quantum chemist's toolbox for studying excited states in large systems.
IntroductionAccurate quantum chemical methods for modeling electronic excited states are essential for gaining insight into the photophysics and photochemistry of molecules and materials. The most widely used technique for excited state calculations at present is time dependent density functional theory (TDDFT), 1-5 on account of its relatively low computational complexity (O(N 2−3 ), where N is the molecule size) and reasonable accuracy for many problems. 6,7 TDDFT excited states are computed via determining the linear response of a ground state DFT solution to time-dependent external electric fields, 5 permitting simultaneous modeling of multiple excited states. In principle, TDDFT is formally exact 1 when the exact exchangecorrelation (xc) ground state functional is employed, although lack of that functional, and the need for the widely used adiabatic local density approximation 3-5 (ALDA) prevents this from being the case in practice. ALDA in fact restricts utility of TDDFT to single excitations out of the reference alone, with large errors arising whenever the target excited state has significant double (or higher) excitation character. 8-11 Furthermore, TDDFT is known to systematically underestimate excitation energies for charge-transfer 5,12-14 and Rydberg 8,15 states, and yields qualitatively erroneous potential energy surfaces along single bond dissociation coordinates. 16 These effects originally stem from errors in the ground state DFT solution like delocalization error 17,18 or spin symmetry breaking, 19,20 but the linear response protocol augments these deficiencies in the reference to catastrophic levels in excited states, on account of insufficient or...