"Custom" use of SLATM and mbtypes

[puckvg]

Hello,

I would like to use your nice all-python SLATM function. I notice that the typical use case in the examples is to have a full dataset pre-defined, which sets the mbtypes as combinations of elements in each individual molecule. What about if I instead wanted to specify only the nuclear charges I expect to see up-front, e.g. unique_elements=[1,6] , such that I can generate mbtypes once at the beginning and then construct new SLATM representations "on the fly" for new molecules with elements falling into those pre-defined?
I tried just giving one example molecule with the above unique elements, but the three-body terms are empty, presumably since there are only two atoms in my pseudo-molecule. What is the best way to set up mbtypes to cover all use cases? Set up a pseudo-molecule with at least three of each element I want to consider?

Thanks in advance!

[YAY-C]

Hi @puckvg !
Sorry for the late reply!

I've been having the same issue to run SLATM computations in parallel !
As far as I can tell, the best option is to generate and save the mbtypes for a given dataset and then load 'on-the-fly' whenever you want to generate a representation for a molecule contained in the dataset.
I have a little snippet to do this if you want.
(but this is still dangerous if you don't know beforehand the complete set of molecules that can exist in the dataset)

A more robust way of doing it, I think, would be to do what you suggested. But you don't really need to create a molecule just a set of atomic_charges, with each atom-type in triplicate, the consequence though is that you might have big bunches of features equal to zero and a longer feature vector.
(but this I haven't tested yet)