search

ACEsuit / ACE.jl

Parameterisation of Equivariant Properties of Particle Systems
66 stars 15 forks source link

Constant term in ACE basis #80

Closed zhanglw0521 closed 2 years ago

zhanglw0521 commented 2 years ago

Adding a constant term ($c\in\mathbb{C}$ for invariant basis and $$cI(2L+1)$$ for SphericalMatrix(L,L)) to the corresponding ACE basis would not destroy the symmetry of the basis and may help improve the fitting result as well as stability (Maybe?).

We could add it directly to the basis while another thought could be that we add a mean field for ACELinearModels, which contains the mean value of data. This means that we didn't fit the coefficient for the constant term but give an expected value to it.

cortner commented 2 years ago

Note that in principle there is the OneBody potential in JuLIP, which is just a separate model that can then be combined with an ACE model. This could be done here as well.

But it is a fair proposition to keep it inside ACE and the option of having it fixed or variable. The reason it would be natural is that the constant term is just the null-correlation. But I would like to think about it.

cortner commented 2 years ago

This has been merged and is tagged as v0.12.24