Converge EPS_SCHWARZ and CUTOFF_RADIUS for hybrid functionals

[Max H]

Dear all,

I plan on doing some single point calculations with the PBE0 functional and a ccGRB basis and would appreciate advice on the selection of appropriate parameters, because I have not found information of sufficient detail.

I read in the recent review (JPhysChemB, "The CP2K Program Package Made Simple") and in various other sources that the EPS_SCHWARZ and CUTOFF_RADIUS parameters are crucial for accuracy and computational efficiency.

What is the typical approach to converge these values? Are single-point calculations with one SCF cycle sufficient as soon as energy convergence is observed (comparable to the convergence protocols for CUTOFF and REL_CUTOFF)?

When studying metallic systems, which values of CUTOFF_RADIUS can be expected?
Supposing that the CUTOFF_RADIUS did not converge at half the smallest cell dimension, is the best approach then to set MULTIPLE_UNIT_CELL 2 2 1 (xy periodic system) in the &CELL block or do special restrictions arise?

I appreciate your kind advice.

Best,
Max

[Frederick Stein]

Dear Max,

Both parameters are determined differently.

CUTOFF_RADIUS needs to be less than half of the distance of an atom to its nearest image. For sufficiently large cells, you may even reduce it to values at the order of 5-7 A (please check convergence of your results in that case). If this parameter is not yet converged, you employ the MULTIPLE_UNIT_CELL as you suggested. Convergence is expected if CUTOFF_RADIUS can be set to values of at least 5 A.

EPS_SCHWARZ is determined by checking convergence of total energies. A good starting point is 1E-6 but larger values may be necessary, especially if you set "SCREEN_ON_INITIAL_P T" or with ADMM (ADMM is recommended).

Best,

Frederick

[Max H]

Dear Frederick,

thank you for your kind explanation.
I have been doing a few tests with CUTOFF_RADIUS 8.2 (slightly smaller than half the smallest cell dimension) and various values of EPS_SCHWARZ (from 1E-6 to 1E-8).
Unfortunately, the SCF does not converge at all. The first iteration yields a negative energy, but the energies become positive after the second.
I am restarting the SCF from a previously converged calculation at the PBE level of theory (convergence with fewer than 50 iterations) and employ the "MIN_PAIR_LIST_RADIUS -1" setting in the QS block of the input (attached).

Is it possible to conclude from the SCF behaviour that calculations on multiple unit cells with a yet larger CUTOFF_RADIUS are required?

Kind regards,
Max

[Frederick Stein]

Dear Max,

a larger cell should not be necessary. I have just seen that you set SCF_GUESS ATOMIC instead of RESTART. Thus, you start from an atomic guess instead of your restart file. I also would not use XC_GRID and I would not touch the *_CUTOFF parameters in the XC section.

Best,

Frederick

[Max H]

Dear Frederick,

I apologize for the confusion. (I tested both options, restarting from the PBE converged guess and the ATOMIC guess, although both fail.)

I removed the *_CUTOFF and XC_GRID options in recent tests, but this does not improve the outcome. I have attached a recent input file and parts of the output file, which include information on the SCF. I would appreciate it if you could take a look at it and perhaps point out the error for me.

A test with EPS_SCHWARZ 1E-9 instead of 1E-8 yields somewhat more consistently negative energies in the SCF. Could it be necessary to test yet smaller EPS_SCHWARZ values?

All the best,
Max

[Frederick Stein]

Dear Max,

Please double-check whether the geometry and the multiplicity are correct. Apart from that

- I do not know how the numbers looked like before but you can only test if the energies stabilize if you decrease EPS_SCHWARZ.

- You may try to increase the electronic temperature, 5 K is pretty low. Commonly, you use room temperature or the actual temperature of your system. To have some guidance, you print all MOs using the DFT/PRINT/MO section and check the occupancies. If you do not find partial occupation, turn smearing off. Then, you may even try the OT method instead.

-You may try to use PERIODIC XYZ instead of XY which is much better tested and has much less bugs.

-Otherwise, you may play around with the convergence parameters. It may help to reduce ALPHA.

Unrelated to that, your performance is pretty poor. Can you increase the memory further? If no and possible, you may use more OpenMP threads in favor of MPI ranks while increasing MAX_MEMORY to reduce the memory overhead of MPI.

Best,

Frederick