Execluding some species from training

[m-a-saleh]

Dear developers,

I have a Si slab with varying thickness t_si and whose top and bottom surfaces are passivated by "H" atoms. I need to train an SGP model on AIMD data for it. Training the model on the AIMD data with the two species (H & Si) gives a positive likelihood. However, I think that excluding the "H" atoms from the training will lead to a more accurate trained model as I am interested only in the forces on Si atoms.

To exclude the "H" atoms, I have tried the following, which all resulted in a negative likelihood:
-Remove lines starting with "H" from the AIMD.xyz file for all frames
-Assign zero forces to "H" atoms in the AIMD.xyz file
-Use small cutoffs in cutoff_matrix for "H" (e.g, ctuoff_matrix = [[1.0, 1.0], [1.0, 5.0]])

Any idea on how to exclude those surface "H" atoms?

Thanks

[cjowen1]

Hi m-a-saleh,

Thanks for your question. I would be cautious in using this approach, since the presence of H atoms also influence the forces on the Si atoms at the interface. Moreover, H atoms reduce the total energy of the system by filling the 'dangling-bonds' of Si that are formed from cleaving the bulk crystal, so learning the energy by just removing H in this way might be problematic. Would it be too expensive to generate some pure Si data? Moreover, you can train on both H and Si, but use the resultant model to model pure Si slabs and bulk (as long as you have also seen pure Si slabs and bulk in the training).

Best,
Cameron