NCIPLOT issue with all-electron density cubes

[SimonAlbrechts]

Hello,
I am encountering a problem when using NCIPLOT on all-electron density cubes generated with CP2K using the GAPW method. The <root>-dens.cube files contain negative regions aligned along the axes of the PBC cell, as illustrated in the attached figure for a He dimer. I tried increasing the grid density but that didn't help. I also experience the same issue for other tested systems.

After reading the following remark in the NCIPLOT documentation:

“If the reference field that provides the density for NCIPLOT is a grid field, then it is usually much faster if core-augmentation (ZPSP field option) is not used… However, not using the core augmentation can sometimes result in noisy Hessian components, which may result in alternate blue/red domains.”

I tested valence density cubes generated with the GPW module of CP2K. When these valence cubes are used together with ZPSP, the <root>-dens.cube files behave as expected. Without ZPSP, similar artefacts appear. When I manually construct a sign(λ₂)·ρ cube in Critic2 from the all-electron density, the resulting cube no longer shows these artefacts.

Is there a way to avoid this issue while still using NCIPLOT on all-electron density cubes or maybe a possibility to fix this bug?

Thank you!

[aoterodelaroza]

Sorry for the wait. This is very like an interpolation problem caused by doing a Fourier transform on all-electron densities. In the output, does it say:

  Values of NSTEP derived from reference field grid dimensions.
  Using fast Fourier transform for derivative calculation.

or:

  Using scalar field differentiation for derivative calculation.

? If this is the case, I think what makes sense is to use SMOOTHRHO instead, and it is probably a good idea to implement a keyword to switch between Fourier and grid interpolation. I'll try to do it today.

[aoterodelaroza]

I have implemented a keyword that controls whether Fourier interpolation or direct differentiation from the grid interpolant is used. Please, see the manual and also the modified tips section. In your case, try SMOOTHRHO when you LOAD the field and then INTERPOLATION GRID in the NCIPLOT environment.

[SimonAlbrechts]

Thank you for the implementation! Using INTERPOLATION GRID, the -dens.cube files now look correct.

I also tested the calculation both with and without SMOOTHRHO, and the difference in computation time was very large (about 9 hours versus ~100 seconds cpu time). Does this seem realistic to you?

How important do you think SMOOTHRHO is when generating NCI plots with INTERPOLATION GRID? Could it perhaps be avoided to reduce computational cost?

[aoterodelaroza]

Unfortunately, SMOOTHRHO is (much) slower than the other interpolation methods. But it's the only way to get "correct" derivatives from a density grid. You could try with the default (trispline) interpolation method but the results may be wonky. I suggest you try and see what happens. Another option is reducing the number of points and/or the region in which you are making the plots.