Basis sparsification

[cortner]

If we want to allow solvers that sparsity, then we must also then be able to reduce the basis size accordingly. This is currently getting implemented in ACE.jl, but we may wish to backport it to ACE1.jl.

[cortner]

@casv2 @davkovacs - thoughts? Useful? Important enough to give it a milestone?

[casv2]

I think this would be very useful indeed, ideally we'd get a lot of freedom of adding/removing basis functions when fitting. It would allow use to really explore the entire ACE basis and guide us towards tailor-made bases for every fit. We've seen that often the fit revolves around just a fraction of the entire ACE basis so I think there's a lot of interesting stuff to be done here.

[davkovacs]

As for me, I have entirely abandoned sparsification, but would be happy to re-explore if it actually removed the basis functions. Previously I saw that I could get the same accuracy with 80% less basis functions, but it didn’t translate into speed or memory saving and was adding an extra parameter, so I stopped using it.