Questions on feasibility double hybrids >500 atom system

[bmk]

Dear CP2K users/developers,

I am training machine learning potentials on hybrid single points, but I would like to try something as described in here and perform DH single points in CP2K. In order to do that I would (probably) need access to a GPU cluster, but before I write a proposal I would like to know if this is a good idea: 

1. Is it possible to perform revDSD-PBEP86-D3(BJ) single points in CP2K? If not, are there other DHF you recommend for organic reactions?
2. Are DH single points possible/feasible for my systems? See below for some details on my systems
3. With my systems, would it feasible to obtain the nuclear gradients?

Some details on my system:

1. periodic cubic box with side lengths of 22.5 A
2. 545 atoms: 215 H, 11 O, 215 C, 6 F, 2 S and 200 Cl

[Frederick Stein]

Dear bmk,

1+3. Calculations using double hybrid functionals are generally possible. CP2K supports double hybrid functionals based on MP2, SOS-MP2 and RPA including analytical nuclear gradients and stress tensors. I have already done that in CP2K pretty extensively, see my list of publications: https://scholar.google.com/citations?view_op=list_works&hl=de&user=Pn4YqtsAAAAJ for some ideas. You may also check the supporting information with the respective input files for ideas on how to set up these calculations. Meanwhile, CP2K supports the D4 correction as well.

2. A system of your size was already feasible 10 years ago using CP2K. Mauro Del Ben (see https://doi.org/10.1021/ct4002202) ran calculations using up to 512 water molecules. I have only tried 128 water molecules so far (with RPA gradients on the now decommissioned Piz Daint).

Best,

Frederick

[bmk]

Thank you very much!

Op donderdag 2 oktober 2025 om 12:17:32 UTC+2 schreef Frederick Stein: