add support for standard machine-learning-type descriptors?

[jkshenton]

It might be useful to have a SOAP gene. i.e. have soprano calculate (possibly via another package such as https://github.com/SINGROUP/dscribe) standard machine learning descriptors such as SOAP.

[stur86]

That looks a bit similar to what BondOrder is doing (https://github.com/CCP-NC/soprano/blob/master/soprano/properties/order/order.py and there's also a gene version). I agree these sort of descriptors could be useful; given they're already implemented in this library, we could add it as a dependency (even just an optional one) and then have some kind of general interface to it.

About the SOAP descriptor in particular, it seems to me like the general idea is that for each atom it produces one vector with one parameter per species and angular momentum channel representing its neighbourhood. Thinking about turning these into Soprano genes, you could linearise the entire vector (so for example a (32, 422) descriptor array would become a (13504,) one), but the problem I see with that is that it wouldn't necessarily work to compare structures that are almost identical if they happen to have atomic positions swapped. It would work if we either standardised the order of atoms somehow or did something else like reducing the descriptors only to species averages. Or we could possibly combine this with spglib, find the inequivalent sites in the crystal and only compare atoms in those, with some order again standardised on the basis of the symmetry (anything goes and we can figure it out later, as long as it's a consistent convention).

[jkshenton]

An optional dependency could be the way to go, though actually I just tried to pip install dscribe alongside soprano and things didn't go very well (the numba required by dscribe wanted a slightly older version of numpy).

Anyway regarding SOAP, I can see how the re-ordering etc could get very messy! Luckily we're not the first to face such questions. I came across this from the Ceriotti group: https://doi.org/10.1039/C6CP00415F see section 2.2 for various approaches including their "regularized-entropy match" (REMatch) kernel (eqn 12).

[stur86]

Ah, good, I'll look into it! Right now I haven't made the phylogenetic stuff much of a priority since I'm not aware of anyone that's using it at the moment, but that could change if we did have some expressions of interest (and of course revamping or adding some functionality could generate interest too, so there's that).

The installation failing is a problem, at least a good reason to have it be optional. But if it's hung on an old Numpy version that doesn't bode well. Is it possible that we can just do SOAP with the machinery we already have in Soprano anyway? As I mentioned, the BondOrder functionality is very similar, using spherical harmonics too. Spherical harmonics are defined in Scipy, and Clebsch-Gordan and Wigner 3j coefficients are in Soprano's utils module. Or there could be other libraries supporting it.

[dch0ph]

We're experimenting the Ceriotti group's ShiftML2 which uses SOAP descriptors for machine learned chemical shifts. We'll have a dig to see how "accessible" this functionality is.