Participants
CS: We should use ELF10 charges while fitting the WBO bond orders, these are the most conformationally independent and the current wbo calculation method in openff-tk uses only one single conformer and AM1BCC, they’re not ELF10. Openeye backend used to have it back in the summer of 2019 that I used in fragmenter, but somehow not openly available now and I raised a ticket.
DLM: Tag me, Simon and Jeff on that, I will check with openeye.
CS: JM, did you notice any difference in trends while doing the non-ELF10 calculated wbos?
JM: Nope, I din’t see any noticeable change in the trends with the curated substituted phenyl dataset using the current openff-tk calculated wbos.
CS: Okay, still I prefer ELF10 based wbos in the paper, Openeye is going to send me a script which I will pass on and we may have to redo some of the plots.
DLM & JM: Sounds good. This should help in wrapping up the paper, let's talk with JC and decide on things left to do to get the preprint out.
SB: We discovered a bug with the sigma values in toolkit verisons 0.8.1 and 0.8.2, so going back to 0.8.0 resolves the energy blowing up issue.
PB: Yes, I ran those again and there is no energy blowup. I still get some negative energy values, is that consistent or need to change the optimizer?
DLM & SB: You can get negative energy values, that's not wrong.
JM: I am working on the k,TB vs WBO plots, adding visualization of molecules and doing the analysis for a general torsion term that includes the in-ring torsions too. And the undergrads working with me are doing a comparison of WBO calculated with Openeye and ambertools.
DLM: Sounds good.
PB: What's the difference between those two calculations of wbo with openeye and ambertools?
DLM: Application of symmetrization of charges is different with these backends. For ELF10 we feed in a bunch of conformers and Openeye picks up electrostatically least clashing which gives us a better value when compared to Ambertools. In some cases Openeye insisits on one direction of proton, etc., so to deal with these kind of confomer issues, and symmsterization which results in better charge assignment. Also, we don't want to end up with a FF that works with only one backend, this comparison helps in getting a clear picture.
SB: It would be great to add more tests like checking how wbo varies for a single conformer, diversifying and averaging conformers, etc.
JM: Sure, will pass on to the undergrad.
SB: PB, what's the status of your work?
PB: When we last talked we were comparing the forcefields that originated from 1.3.0, that TG helped us fit, one with interpolated parameters, another with the general torsion parameters in place of each of those interpolated parameters. We hit a roadblock with the energy blow-up issues and started doing our experimental FB fits with 1.2.0 as our starting point. It looks like there is no need to do that now, I will go back to doing benchmarking with 0.8.0 toolkit version.
SB: DLM, are we going to do a paper for 1.2.0? I am also in favor of doing the science as a standalone publication that gives more rationale about the choices we make in making these FFs.
DLM: Right now the focus is on getting out Sage release out, we will eventually do that given how much time people have in writing and implementing stuff.
JM: A followup on this, every version of release gets a publication, or do we have to bundle?
DLM: Dataset coverge you did can be a separate paper or the main focus of 1.2 release paper, since primarily 1.2.0 is dataset re-design. We can discuss that with Hyesu.
SB: TG, are you doing any science with the new infrastrucutre on Vsites?
TG: It's kind of on the roadmap, right now working on chemper, automation of smarts patterns.
DLM: We were talking about splitting of LJ terms
TG: The code I work on is splitting terms which involves fitting LJ parameters...
Add Comment