additional SOAP (turbo) descriptors for multi-species GAP

[richardjana]

I am currently working on a multi-species GAP for two elements with previously existing GAP potentials. As a starting point, I joined both training data sets and used all the descriptors from the original GAPs. I then extended the SOAP descriptors to both species and added an additional 2b descriptor (mixed):

distance_2b: A-A
angle_3b: A-A-A
soap_turbo: AB, A-centered
distance_2b: A-B
distance_2b: B-B
angle_3b: B-B-B
soap_turbo: AB, B-centered

I expected this to work reasonably well on the original training data, as that was still included in the new training set and the necessary descriptors were also still there. However, I found that for one of the elements (but not the other!), the new GAP performed much worse than the old (single-species) one. I then added 2 more SOAP descriptors, one for each pure element:

distance_2b: A-A
angle_3b: A-A-A
soap_turbo: AB, A-centered
**soap_turbo: A, A-centered**
distance_2b: A-B
distance_2b: B-B
angle_3b: B-B-B
soap_turbo: AB, B-centered
**soap_turbo: B, B-centered**

This fixed the issues I had before, giving RMSE on the training data similar to the original GAPs for both elements.

Has anyone experienced such behavior before? Is it the result of a bug in gap_fit, or to be expected for physical reasons?

[gabor1]

Interesting. I've never done this. Is it possible that you didn't get too many sparse points with just A or just B for the first fit? You can control how the sparse points are selected eg you could request a certain number of sparse points from the pure A and pure B part of the training set. This is done with the config_type_n_sparse keyword

[richardjana]

I used the config_type_n_sparse in both cases, with the same number of sparse configs in each case. (Of course, the mixed and pure SOAP descriptors might pick up different environments. So, in a sense, I had more sparse configurations in the case with 4 SOAP descriptors...)
I'll run a test doubling the numbers for the 2*SOAP GAP, so the sums are the same.

[richardjana]

An update on this issue: I pinned the problem down to the atom_sigma_r and atom_sigma_t parameters of the SOAP descriptors -- as I had made my initial gap_fit command by combining the single-species commands the atom sigma were not the same between the SOAPs. I assumed this was just fine.
However, I found that I can get a working potential with

distance_2b: A-A
angle_3b: A-A-A
soap_turbo: AB, A-centered
distance_2b: A-B
distance_2b: B-B
angle_3b: B-B-B
soap_turbo: AB, B-centered

when the atom sigmas are the same in both SOAP descriptors, but not when they are different (as described above). Is this a bug, or expected behavior? (At least I did not expect this.)

The problem also goes away when using 4 SOAP descriptors as above, regardless if the atom sigmas are the same or not.

[mcaroba]

I've transferred this issue to the soap_turbo repo. I'll try to take a look when I have the chance.