search

ACEsuit / ACEpotentials.jl

Machine Learning Interatomic Potentials with the Atomic Cluster Expansion
MIT License
45 stars 11 forks source link

Application of Prior is Inconsistent #250

Open cortner opened 1 week ago

cortner commented 1 week ago

The way we apply the prior is different across different scenarios. This creates confusion. Fundamentally, the question seems to be whether the prior is part of the model or part of the solver. In other words should it rescale the basis in ACEpotentials, or should it rescale the lsq system inside of ACEfit? There is no rush to address this, but in a future version it would be nice to become more systematic about this.

wcwitt commented 1 week ago

In some ways, I like rescaling the basis, such that the prior is completely decoupled from the solver.

However, we have also discussed solvers that might adjust the prior (e.g. a BLR that optimizes the prior strength). So maybe the prior has to go to ACEfit?

cortner commented 1 week ago

I'm really not sure. I do think that there is a "true prior" which lives in the model. (And in a near-future rewrite I want this to simply become a rescaling of the radial basis!)

Maybe we just need to accept the fact it can be in both. And for some solvers there can be an "adjusted prior".