Should Model Parameters be Unitful?

[cortner]

At the moment model parameters such as r0 or rcut are given dimension-less, implicitly assuming Å. Maybe it could be useful to make all ACE models dimension-aware, while still allowing dimensionless models as well. No rush to decide this.

[cortner]

Some notes on this

• r_cut always has length-units in AtomsBase, some neighbourlists simply won't run dimensionless; the same should apply for r0 and rin
• The loss can be made dimensionless by giving weights suitable dimensions, e.g. |E - Eref|^2 --> | (E - Eref) / sig |^2 then sig should have the same dimension as energy. This need not be transferred to the linear algebra level by u-stripping with compatible units.