Torsion barrier issue/ Next valence parameter fitting plan
|
@Hyesu Jang
| (Add link to page here) HJ – Upon refitting, the t70 parameter seems to have improved, but there are still some outliers. DM – It looks like previous parameter values may have been more correct CB – In FF lore, CH and NH hydrogens always got different types. But SMIRNOFF format lets us do wildcards to avoid this. Old t70 was overly specific for hydrogen. But the new t70 may be too general. If we were going to mimic the lore, we would have a t70, which is very general, followed by a new t70a with hydrogen in the 1-position. HJ – Sure. I’ll check that. HJ – What I noticed is that, at 0 degrees, we largely improved performance on t70.
Biaryl torsion profiles A few hypotheses – Link to page here DM – Does FB allow us to down-weight high energy points? HJ – Yes, I’m pretty sure that’s doable. DM – We should run the Rowley set through QCArchive CB – If, in our new optimization, Parsley 1.2.0 is worse with biaryls, then it probably got better with something else. This might mean that we can EITHER do well on our training set, OR on these biaryls. HJ – I don’t know of the specific molecules off the top of my head. (Link to slide here) CB – To determine whether we need to split the parameter types, we might do something like what Simon did. That is, at the global minimum… (something I didn’t understand) DM – Do you have the ability to extract the direction of the derivative, HJ? CB – LPW implied a while back that this is possible. It should indicate when a parameter should be “split”. JW – Barrier heights always increase. Maybe regularization was constraining them artificially low. Also more congested molecules in the 1.2.0 training set HJ – Both 1.0.0 and 1.2.0 started from the same initial guesses (so JW’s first point is not what’s happening) HJ – So I’ll check the first derivative of each barrier and see if that indicates we need to split torsion groups. CB – Yes, Simon plotted distribution of first derivative…. CB – I wonder if this is similar to what we saw with amides, where there was a single parameter initially, but we saw a need to split the parameter later. In the training set I see lots of chemistry that we may be able to assign different torsions to over a biaryl. So we may be able tos tart adding more, less general torsions. CB – In 2-pyrazol-1-ylpyrimidine, the interaction between the virtual sites on the furyl nitrogens with the hydrogen on the pyridyl will be highly repulsive. Right now, the only place we can account for that repulsion is the torsion. But we know that this will mangle the general parameter. DM – Virtualsites will not be out in Sage. Once the infrastructure is ready, we’ll need a few months to fit them, and then we’ll need to wait until we achieve greater interoperability with other packages. In the meantime, we’ll release virtualsites in a non-mainstream FF line. JW – VSites likely in 3 months.
|