r2SCAN CUTOFF convergence

[Max H]

Dear all,

I am trying to implement the r2SCAN functional for my calculations and struggle to converge the values for the CUTOFF and REL_CUTOFF. As some have observed earlier, it is difficult to get the accuracy below 10^-3 Hartree and I encounter the same challenge for my system (Cu, O, C and H).
I tried the following steps to resolve this issue, albeit without success:

1. Raise CUTOFF systematically up to 3000 at REL_CUTOFF 100 (same at 200).
2. Apply smoothing in the XC_GRID section, e.g. XC_DERIV SPLINE3_SMOOTH and XC_SMOOTH_RHO NN10 (and NN50).
3. Use the GPAW method instead of GPW (as some noted that CUTOFF convergence is easier).

I have been using the DZVP-MOLOPT-SCAN-GTH basis sets and the GTH-SCAN pseudopotentials for the tests with GPW and the 6-31G* basis set for GPAW.
The input files are attached (although I left out the exact structures, as the problem should be quite general).

I would appreciate your suggestions to solve this issue or point it out if I made an obvious mistake.

All the best,
Max

[Jürg Hutter]

Hi

maybe you can try to see if the energy difference between two structures converges?

BTW you can also use GAPW with your original GPW setting (basis & PP).

best regards
JH

________________________________________
From: 'Max H' via cp2k <cp...@googlegroups.com>
Sent: Friday, January 30, 2026 12:59 PM
To: cp2k
Subject: [CP2K:22071] r2SCAN CUTOFF convergence

Dear all,

I am trying to implement the r2SCAN functional for my calculations and struggle to converge the values for the CUTOFF and REL_CUTOFF. As some have observed earlier, it is difficult to get the accuracy below 10^-3 Hartree and I encounter the same challenge for my system (Cu, O, C and H).
I tried the following steps to resolve this issue, albeit without success:

1. Raise CUTOFF systematically up to 3000 at REL_CUTOFF 100 (same at 200).
2. Apply smoothing in the XC_GRID section, e.g. XC_DERIV SPLINE3_SMOOTH and XC_SMOOTH_RHO NN10 (and NN50).
3. Use the GPAW method instead of GPW (as some noted that CUTOFF convergence is easier).

I have been using the DZVP-MOLOPT-SCAN-GTH basis sets and the GTH-SCAN pseudopotentials for the tests with GPW and the 6-31G* basis set for GPAW.
The input files are attached (although I left out the exact structures, as the problem should be quite general).

I would appreciate your suggestions to solve this issue or point it out if I made an obvious mistake.

All the best,
Max

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com<mailto:cp2k+uns...@googlegroups.com>.
To view this discussion visit https://groups.google.com/d/msgid/cp2k/683dde59-9bbd-4de0-8064-79c0501b8167n%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/683dde59-9bbd-4de0-8064-79c0501b8167n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[Max H]

Dear Prof. Hutter,

thank you for your kind suggestion (and thank you for pointing out the option for the GAPW settings).
I tested it on two isomers of the adsorption structure on the surface, but it did not resolve the issue, unfortunately.

I guess that I could also do 'ENERGY_FORCE' calculations with a converged SCF and see if the forces converge, but I moved on to using the PBE functional now.

During my tests I came to wonder about a set of two possible remedies.

1. The "XC_GRID" option offers the Keyword 'USE_FINER_GRID True", but it seems to be unavailable for meta-GGA functionals such as r2SCAN (Please correct me if I am wrong). The printed XC energies for atomic density guess show at least fluctuations of up to 10^-3 Ha in the range of the tested CUTOFF values (200 to 2000) for my structures, whereas the other printed energies essentially converged  (Overlap energy of the core charge distribution, Self energy of the core charge distribution, Core Hamiltonian energy and  Hartree energy).
2. Could it be related to the pseuopotentials? Maybe the regularization of r2SCAN introduces some issues for the SCAN-optimized PP?

Kind regards,

MH