I'm generating a BP86/def2-TZVPD COSMO .cosmo (FINE cavity, for a later COSMO-RS / BP_TZVPD_FINE screen) for hen egg-white lysozyme — 1963 atoms, net charge +11, 7620 electrons, ~46,340 basis functions, on a GFN-FF-preoptimized geometry. TURBOMOLE 7.8, ridft, RI-J, m4 grid, COSMO (ε=∞).

The SCF diverges hard: cycle 1 from a good guess is physical (~ −52,426 Eh), then cycle 2 the density-change norm jumps to ~10¹⁰ and the energy runs to ~+1.7×10⁵ Eh; the Fermi level fails to bracket. This looks like near-linear-dependence of the diffuse (def2-TZVPD) set on the dense folded protein (near-singular AO overlap).

Tried (all diverge unless noted):

• Fermi smearing, constant 300 K and annealed 2000→300 K.

• Bootstrap: BP86/def2-TZVP (no diffuse, ~36,863 BF) converges cleanly (−52,426 Eh); projecting those MOs to def2-TZVPD via define's use → diverges on cycle 2 with the same 10¹⁰ norm — so it's not the start vectors.

• Heavy damping (\\\$scfdamp start=5.0 step=0.05 min=0.5) + large level shift (\\\$scforbitalshift automatic=1.0, applied shift ~12 Eh) from the projected guess → still diverges at cycle 2.

• Mixed basis (def2-TZVP on H,C / def2-TZVPD on N,O,S, ~39,776 BF) converges — but that's no longer plain def2-TZVPD.

• Geometry is clash-free (min non-bonded distance > 0.9 Å).

Questions:

Is there a supported way in 7.8 to project out / remove near-linearly-dependent basis functions (canonical orthogonalization with an overlap-eigenvalue threshold) inside a COSMO SCF in ridft/dscf? I only find lsdiag in riper.

Does riper support COSMO for finite molecules (so riper+lsdiag could yield a .cosmo)? The COSMO manual chapter lists dscf/ridft/grad/rdgrad/escf/ricc2 but not riper.

Is there a recommended route to a BP-TZVPD-FINE .cosmo for protein-scale solutes (>1500 atoms), or is def2-TZVP the practical ceiling?

Thanks!