vdW fits | TG | DC: Did the SAPT energies include electrostatic penetration? JR: are you talking about the Pauli repulsion term Danny? DS: Include charge penetration that is exponentially decaying BS: I think in this dataset many of the entries are ions and not molecules JR: So these calculations came from opts with the standard protocol? TG: No, I’m doing a single point with B3LYP-D3BJ, and most of the structures don’t have negative frequencies TG: It escapes me what model chemistry is used to minimize the geometries
BS: You said you find bond/angle/torsions and non bonded all at the same time? Could those degrees of freedom have allowed a compensation effect? TG: My last slide will expand on that. But the bonds/angles/valence didn’t change, that happened a little with torsions but not to a large extent. JR: So you didn’t fit charges? TG: No
DC: Try condensed phase simulations as often as possible in this process, and expect things to “explode” TG: Do my errors when compared to the openff fit seem reasonable? BS: Are openff parameters such that all carbon are the same type? or is it more specialized? DM: There’s a bit of a spread, much like GAFF LW: There are about 3 or 4 types for carbon. JW: 2 types of oxygen and 1 type of nitrogen TG: Yes, so the goal in this is that the BESMART could identify additional typings that are required.
BS: You show the QM contribution to LJ, some on the order of -50 kj/mol, how is that? Shouldn’t they always be positive? DM: I wonder if the QM method chosen shows different physics? DC: Yes this is where charge penetration comes in. Subtract multipole contribution from QM to get LJ contribution. Don’t worry as much about what you’re fitting, and more on how you’re making that split. DM: So you know what he’s fitting to, so once he gets the splitting correct you can step in? DC: Yes, we just don’t have the bandwidth, and we don’t have the splitting finalized DM: we have TG for a month so let us know DC: We are trying to get a GitHub repo
JR: How are you separating charge and other contributions: TG: However SAPT does it JR: I expect a string dispersion term TG: many go out to C4, although I’m not. A paper I looked at suggested the dispersion term is more subtle and two other terms dominate. TG: There are many energy terms and I need to know what to include and what not to include DC: Grabbed an electrostatics model and subtracted that from a CCSD model. It took quite some time. We took it and used it, so there’s room for improvement. I’ll send you some papers. TG: Ok that’ll be helpful
JAC: You mused before that you’ll stop simultaneously fitting all but charge contributions, but also that torsions are changing and some are dominating, maybe consider keeping those? TG: I’ll take this as an opportunity to ask, is simultaneous fit of these ok? DM: We iterate between the two so the torsions certainly have the ability to compensate JR: Torsions add a lot of opportunity to take in the “slop”, just to play devils advocate. DM: I’m saying what we did, not that you should do it JAC: I’m concerned that you found that large torsions are present, and that keeping them will bias you target of non bonded negatively, maybe limit the dataset to smaller molecules with less torsion affects TG: Yeah, I’ll have to figure it out
BS: You might remove charges, since fitting the non bonded at that point is noise compared to the charge TG: That could be one of the grouping I have, I’ll have to find out. JR: I agree to remove charges as a first draft. Also light atoms are difficult to get right, like He and hydrogen gas. TG: Ok so focus on neutral molecules without lighter gasses
|
