DFT keeps breaking spin symmetry on a µ-oxo Fe(III) dimer, and the electron correlation looks too static for a clean KS picture, so I’m testing a DLPNO-CCSD(T) route instead. In ORCA 5.0.4 with def2-TZVPP/def2/J+def2/MP2fit, TightPNO and RIJK on 64 cores/256 GB RAM, my quintet–triplet gap shifts about 6–8 kcal/mol depending on PNO cutoffs and the HF orbital generator — what PNO thresholds and auxiliary choices have given you stable, defensible spin-splittings in similar open-shell clusters?
@OP Switching to ROHF + DLPNO-CCSD(T1) with VeryTightPNO shrank gap drift to about 2 kcal; pricier but consistent.
And i had the same thing on a µ‑oxo Fe(III) dimer; the gap settled once I forced identical pair lists/PNO domains across states — write them on one state and read them for the other (ORCA’s WritePairs/ReadPairs) while keeping VeryTightPNO. It costs a bit and I’d still spot‑check with “NoFrozenCore,” but it cut the drift to about 1 kcal for me. @grayson_lee92’s ROHF+(T1) advice plus locked PNOs keeps the method from changing the rules mid‑game.
@OP I had the same 6–8 kcal swing in ORCA 5 with def2‑TZVPP and it drove me nuts until I stopped freezing Fe 3s/3p — set “NoFrozenCore” — and forced Pipek–Mezey localization in both states; after that, TightPNO + RIJK settled the quintet–triplet gap to about 1–2 kcal regardless of the HF generator. It costs more and you might need to nudge TCutPNO tighter, but it was far more reproducible on my µ‑oxo dimer. Ref: https://orcaforum.kofo.mpg.de.