there seems to be something off with the emb_lin_neg option of the MEAM pair potential definitions. I actually "fixed" this a while back in my local version, but re-checking before creating a PR, I noticed this may have been intentional.
In summary, according to documentation:
emb_lin_neg = integer value to select embedding function for negative densities
0 = F(rho) = 0
1 = F(rho) = -asub*esub*rho (linear in rho, matches DYNAMO)
default = 0
So with emb_lin_neg=0, the embedding function is clamped to 0.
However, in code it looks like this:
So actually, with emb_lin_neg=0, the embedding function for negative rho/rho0 is NaN (because log(-val)).
This sort-of makes sense if one wanted to carry the NaN in case of negative densities - but it's not what was documented.
Wagner's Tech Report on the implementation details does not mention this at all, and Lee's KISSMD code does what our documentation says.
Does anyone know if this really was a bug? If it was, I already have the improved version ready and will include it in the next PR.