K-point integration and ISMEAR for ML training of metals and insulators

[akretschmer]

I am trying to train a force field for a solid/liquid interface between a metallic solid like Zn and water.

Regarding the k-point mesh and smearing settings, there are very different requirements for metallic and insulating materials. The general recommendation for for ISMEAR settings when simulating different phases together is to use the optimal ISMEAR setting for each phase consecutively, see here for example: https://www.vasp.at/forum/viewtopic.php?t=446

But this advice is for relaxations, when I train a force field of such an interface, I first train the forces for water, then for the metal, and then the combined interface. Since the force field needs to be trained on a single potential energy landscape, does this apply to the k-point mesh and smearing parameter too?

Do I need to use the same k-point density in all calculations? Should I use ISMEAR = 0 for all cells since this is the best compromise for all possible phases? Should the SIGMA parameter be kept constant?

[alexey.tal]

Dear akretschmer,

Thank you for your question.

The post you are referring to mostly concerns the density of states, which can be better represented with the tetrahedron method. For the MD calculations, the Gaussian smearing method leads to very reasonable results in both cases metals and semiconductors.

In principle, by using a different number of k-point one might have a different accuracy for the water and the solid phases. To avoid such a situation using consistent k-point meshes is recommended. Nevertheless, if the calculation is well converged with the number of k-points the difference in the k-points sampling should not affect the results.

Best wishes,
Alexey