Accurate prediction of core-level spectra of radicals at density functional theory cost via square gradient minimization and recoupling of mixed configurations.

State-specific orbital optimized approaches are more effective at predicting core-level spectra than traditional linear-response protocols, but their utility had been restricted on account of the risk of `variational collapse' down to the ground state. We employ the recently developed square gradient minimization (SGM, J. Chem. Theory Comput. 16, 1699-1710, 2020) algorithm to reliably avoid variational collapse and study the effectiveness of orbital optimized density functional theory (DFT) at predicting second period element 1s core-level spectra of open-shell systems. The SCAN, PBE0 and $\omega$B97X-V functionals are found to predict excitation energies from the core to singly occupied levels to high accuracy ($< 0.5$ eV RMS error), against available experimental data. Higher excited states are however more challenging by virtue of being intrinsically multiconfigurational. We thus present a CI inspired route to self-consistently recouple single determinant mixed configurations obtained from DFT, in order to obtain approximate doublet states. This recoupling scheme is used to predict the C K-edge spectra of the allyl radical and the N K-edge of NO$_2$ to high accuracy relative to experiment, indicating substantial promise in using this approach for computation of core-level spectra for doublet species (vs more traditional time dependent DFT or using unrecoupled mixed configurations). We also present general guidelines for computing core-excited states from orbital optimized DFT.