I’ve been experimenting with ROKS (including the variants that use the squared-gradient minimization and state-targeted energy projection algorithms). For some of my compounds, all of these methods converge to a solution that has a significant overlap with S0 (~0.17). Using the original implementation of ROKS, this persists even when I use a pretty big ROKS_LEVEL_SHIFT.

The one thing that appears to effectively eliminate the overlap with S0 (reducing it to ~0.009) is performing a ROKS calculation and then reading those orbitals in as guesses for a second ROKS calculation.

Is this a meaningful way to improve ROKS calculations?

Thank you in advance for any guidance!

-Andrew